Update release notes and doxygen for 1.1.0 release

This manifested itself by a call to an overloaded function in a launch count 
expression that wasn't being resolved.

.gitignore

@@ -1,2 +1,8 @@
 *.pyc
 *~
+depend
+ispc
+ispc_test
+objs
+docs/doxygen
+docs/ispc.html

LICENSE.txt

@@ -114,3 +114,30 @@ CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH THE
 SOFTWARE.
+
+---------------------------------------------------------------------------
+
+ispc's code to convert to and from half-precision floats is based on James
+Tursa's code, which is covered by the following license:
+
+Redistribution and use in source and binary forms, with or without 
+modification, are permitted provided that the following conditions are 
+met:
+
+   * Redistributions of source code must retain the above copyright 
+     notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above copyright 
+     notice, this list of conditions and the following disclaimer in 
+     the documentation and/or other materials provided with the distribution
+      
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE 
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 
+POSSIBILITY OF SUCH DAMAGE.

Makefile

@@ -2,26 +2,42 @@
 # ispc Makefile
 #
 
-ARCH = $(shell uname)
+ARCH_OS = $(shell uname)
+ARCH_TYPE = $(shell arch)
 
 CLANG=clang
-LLVM_LIBS=$(shell llvm-config --ldflags --libs) -lpthread -ldl
+CLANG_LIBS = -lclangFrontend -lclangDriver \
+             -lclangSerialization -lclangParse -lclangSema \
+             -lclangAnalysis -lclangAST -lclangLex -lclangBasic
+
+ISPC_LIBS=$(CLANG_LIBS) \
+	$(shell llvm-config --ldflags --libs) \
+	-lpthread -ldl
+ISPC_TEST_LIBS=$(shell llvm-config --ldflags --libs) \
+	-lpthread -ldl
+
 LLVM_CXXFLAGS=$(shell llvm-config --cppflags)
-LLVM_VERSION_DEF=-DLLVM_$(shell llvm-config --version | sed s/\\./_/)
+LLVM_VERSION=$(shell llvm-config --version | sed s/\\./_/)
+LLVM_VERSION_DEF=-DLLVM_$(LLVM_VERSION)
 
 BUILD_DATE=$(shell date +%Y%m%d)
-BUILD_VERSION=$(shell git log | head -1)
+BUILD_VERSION=$(shell git log --abbrev-commit --abbrev=16 | head -1)
 
 CXX=g++
 CPP=cpp
-CXXFLAGS=-g3 $(LLVM_CXXFLAGS) -I. -Iobjs/ -Wall $(LLVM_VERSION_DEF) \
+OPT=-g3
+CXXFLAGS=$(OPT) $(LLVM_CXXFLAGS) -I. -Iobjs/ -Wall $(LLVM_VERSION_DEF) \
 	-DBUILD_DATE="\"$(BUILD_DATE)\"" -DBUILD_VERSION="\"$(BUILD_VERSION)\""
 
 LDFLAGS=
-ifeq ($(ARCH),Linux)
+ifeq ($(ARCH_OS),Linux)
   # try to link everything statically under Linux (including libstdc++) so
   # that the binaries we generate will be portable across distributions...
-  LDFLAGS=-static -L/usr/lib/gcc/x86_64-linux-gnu/4.4
+  ifeq ($(ARCH_TYPE),x86_64)
+    LDFLAGS=-static -L/usr/lib/gcc/x86_64-linux-gnu/4.4
+  else
+    LDFLAGS=-L/usr/lib/gcc/i686-redhat-linux/4.6.0
+  endif
 endif
 
 LEX=flex
@@ -29,22 +45,24 @@ YACC=bison -d -v -t
 
 ###########################################################################
 
-CXX_SRC=builtins.cpp ctx.cpp decl.cpp expr.cpp ispc.cpp \
+CXX_SRC=ast.cpp builtins.cpp ctx.cpp decl.cpp expr.cpp func.cpp ispc.cpp \
 	llvmutil.cpp main.cpp module.cpp opt.cpp stmt.cpp sym.cpp type.cpp \
 	util.cpp
-HEADERS=builtins.h ctx.h decl.h expr.h ispc.h llvmutil.h module.h \
+HEADERS=ast.h builtins.h ctx.h decl.h expr.h func.h ispc.h llvmutil.h module.h \
 	opt.h stmt.h sym.h type.h util.h
-STDLIB_SRC=stdlib-avx.ll stdlib-sse2.ll stdlib-sse4.ll stdlib-sse4x2.ll
+BUILTINS_SRC=builtins-avx.ll builtins-avx-x2.ll builtins-sse2.ll builtins-sse2-x2.ll \
+	builtins-sse4.ll builtins-sse4-x2.ll builtins-dispatch.ll
 BISON_SRC=parse.yy
 FLEX_SRC=lex.ll
 
-OBJS=$(addprefix objs/, $(CXX_SRC:.cpp=.o) $(STDLIB_SRC:.ll=.o) stdlib-c.o stdlib_ispc.o \
-	$(BISON_SRC:.yy=.o) $(FLEX_SRC:.ll=.o))
+OBJS=$(addprefix objs/, $(CXX_SRC:.cpp=.o) $(BUILTINS_SRC:.ll=.o) \
+	builtins-c-32.o builtins-c-64.o stdlib_ispc.o $(BISON_SRC:.yy=.o) \
+	$(FLEX_SRC:.ll=.o))
 
 default: ispc ispc_test
 
-.PHONY: dirs clean depend doxygen
-.PRECIOUS: objs/stdlib-%.cpp
+.PHONY: dirs clean depend doxygen print_llvm_src
+.PRECIOUS: objs/builtins-%.cpp
 
 depend: $(CXX_SRC) $(HEADERS)
 	@echo Updating dependencies
@@ -56,6 +74,9 @@ dirs:
 	@echo Creating objs/ directory
 	@/bin/mkdir -p objs
 
+print_llvm_src:
+	@echo Using LLVM `llvm-config --version` from `llvm-config --libdir`
+
 clean:
 	/bin/rm -rf objs ispc ispc_test
 
@@ -63,13 +84,13 @@ doxygen:
 	/bin/rm -rf docs/doxygen
 	doxygen doxygen.cfg
 
-ispc: dirs $(OBJS)
+ispc: print_llvm_src dirs $(OBJS)
 	@echo Creating ispc executable
-	@$(CXX) $(LDFLAGS) -o $@ $(OBJS) $(LLVM_LIBS)
+	@$(CXX) $(LDFLAGS) -o $@ $(OBJS) $(ISPC_LIBS)
 
 ispc_test: dirs ispc_test.cpp
 	@echo Creating ispc_test executable
-	@$(CXX) $(LDFLAGS) $(CXXFLAGS) -o $@ ispc_test.cpp $(LLVM_LIBS)
+	@$(CXX) $(LDFLAGS) $(CXXFLAGS) -o $@ ispc_test.cpp $(ISPC_TEST_LIBS)
 
 objs/%.o: %.cpp
 	@echo Compiling $<
@@ -83,36 +104,49 @@ objs/parse.o: objs/parse.cc $(HEADERS)
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
 
-objs/lex.cpp: lex.ll
+objs/lex.cpp: lex.ll 
 	@echo Running flex on $<
 	@$(LEX) -o $@ $<
 
-objs/lex.o: objs/lex.cpp $(HEADERS)
+objs/lex.o: objs/lex.cpp $(HEADERS) objs/parse.cc
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
 
-$(STDLIB_SRC): stdlib.m4
+objs/builtins-%.cpp: builtins-%.ll
+	@echo Creating C++ source from builtin definitions file $<
+	@m4 -DLLVM_VERSION=$(LLVM_VERSION) builtins.m4 $< | ./bitcode2cpp.py $< > $@
 
-objs/stdlib-%.cpp: stdlib-%.ll
-	@echo Creating C++ source from stdlib file $<
-	@m4 stdlib.m4 $< | ./bitcode2cpp.py $< > $@
-
-objs/stdlib-%.o: objs/stdlib-%.cpp
+objs/builtins-%.o: objs/builtins-%.cpp
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
 
-objs/stdlib-c.cpp: stdlib-c.c
-	@echo Creating C++ source from stdlib file $<
-	@$(CLANG) -I /opt/l1om/usr/include/ -emit-llvm -c $< -o - | llvm-dis - | ./bitcode2cpp.py $< > $@
+objs/builtins-c-32.cpp: builtins-c.c
+	@echo Creating C++ source from builtins definition file $<
+	@$(CLANG) -m32 -emit-llvm -c $< -o - | llvm-dis - | ./bitcode2cpp.py builtins-c-32.c > $@
 
-objs/stdlib-c.o: objs/stdlib-c.cpp
+objs/builtins-c-32.o: objs/builtins-c-32.cpp
+	@echo Compiling $<
+	@$(CXX) $(CXXFLAGS) -o $@ -c $<
+
+objs/builtins-c-64.cpp: builtins-c.c
+	@echo Creating C++ source from builtins definition file $<
+	@$(CLANG) -m64 -emit-llvm -c $< -o - | llvm-dis - | ./bitcode2cpp.py builtins-c-64.c > $@
+
+objs/builtins-c-64.o: objs/builtins-c-64.cpp
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
 
 objs/stdlib_ispc.cpp: stdlib.ispc
 	@echo Creating C++ source from $<
-	@$(CPP) -DISPC=1 -DPI=3.1415936535 $< | ./stdlib2cpp.py > $@
+	@$(CLANG) -E -x c -DISPC=1 -DPI=3.1415926536 $< -o - | ./stdlib2cpp.py > $@
 
 objs/stdlib_ispc.o: objs/stdlib_ispc.cpp
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
+
+objs/builtins-sse2.cpp: builtins.m4 builtins-sse2-common.ll builtins-sse2.ll
+objs/builtins-sse2-x2.cpp: builtins.m4 builtins-sse2-common.ll builtins-sse2-x2.ll
+objs/builtins-sse4.cpp: builtins.m4 builtins-sse4-common.ll builtins-sse4.ll
+objs/builtins-sse4-x2.cpp: builtins.m4 builtins-sse4-common.ll builtins-sse4-x2.ll
+objs/builtins-avx.cpp: builtins.m4 builtins-avx-common.ll builtins-avx.ll
+objs/builtins-avx-x2.cpp: builtins.m4 builtins-avx-common.ll builtins-avx-x2.ll

README.txt

@@ -15,8 +15,8 @@ code.
 
 ispc is an open source compiler under the BSD license; see the file
 LICENSE.txt.  ispc supports Windows, Mac, and Linux, with both x86 and
-x86-64 targets. It currently supports the SSE2 and SSE4 instruction sets,
-though support for AVX should be available soon.
+x86-64 targets.  It currently supports the SSE2, SSE4, and AVX instruction
+sets.
 
 For more information and examples, as well as a wiki and the bug database,
 see the ispc distribution site, http://ispc.github.com.

ast.cpp

@@ -1,5 +1,5 @@
 /*
-  Copyright (c) 2010-2011, Intel Corporation
+  Copyright (c) 2011, Intel Corporation
   All rights reserved.
 
   Redistribution and use in source and binary forms, with or without
@@ -31,60 +31,35 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
 */
 
-/* A simple task system for ispc programs based on Apple's Grand Central
-   Dispatch. */
+/** @file ast.cpp
+    @brief 
+*/
 
-#include <dispatch/dispatch.h>
+#include "ast.h"
+#include "func.h"
+#include "sym.h"
 
-static dispatch_queue_t gcdQueue;
-static dispatch_group_t gcdGroup;
+///////////////////////////////////////////////////////////////////////////
+// ASTNode
 
-// ispc expects these functions to have C linkage / not be mangled
-extern "C" {
-    void ISPCLaunch(void *f, void *data);
-    void ISPCSync();
+ASTNode::~ASTNode() {
 }
 
-struct TaskInfo {
-    void *func;
-    void *data;
-};
+
+///////////////////////////////////////////////////////////////////////////
+// AST
+
+void
+AST::AddFunction(Symbol *sym, const std::vector<Symbol *> &args, Stmt *code) {
+    if (sym == NULL)
+        return;
+    functions.push_back(new Function(sym, args, code));
+}
 
 
 void
-TasksInit() {
-    gcdQueue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
-    gcdGroup = dispatch_group_create();
+AST::GenerateIR() {
+    for (unsigned int i = 0; i < functions.size(); ++i)
+        functions[i]->GenerateIR();
 }
 
-
-static void
-lRunTask(void *ti) {
-    typedef void (*TaskFuncType)(void *, int, int);
-    TaskInfo *taskInfo = (TaskInfo *)ti;
-
-    TaskFuncType func = (TaskFuncType)(taskInfo->func);
-
-    // FIXME: these are bogus values; may cause bugs in code that depends
-    // on them having unique values in different threads.
-    int threadIndex = 0;
-    int threadCount = 1;
-    // Actually run the task
-    func(taskInfo->data, threadIndex, threadCount);
-
-    // FIXME: taskInfo leaks...
-}
-
-
-void ISPCLaunch(void *func, void *data) {
-    TaskInfo *ti = new TaskInfo;
-    ti->func = func;
-    ti->data = data;
-    dispatch_group_async_f(gcdGroup, gcdQueue, ti, lRunTask);
-}
-
-
-void ISPCSync() {
-    // Wait for all of the tasks in the group to complete before returning
-    dispatch_group_wait(gcdGroup, DISPATCH_TIME_FOREVER);
-}

ast.h

@@ -0,0 +1,94 @@
+/*
+  Copyright (c) 2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+/** @file ast.h
+    @brief 
+*/
+
+#ifndef ISPC_AST_H
+#define ISPC_AST_H 1
+
+#include "ispc.h"
+#include <vector>
+
+/** @brief Abstract base class for nodes in the abstract syntax tree (AST).
+
+    This class defines a basic interface that all abstract syntax tree
+    (AST) nodes must implement.  The base classes for both expressions
+    (Expr) and statements (Stmt) inherit from this class.
+*/
+class ASTNode {
+public:
+    ASTNode(SourcePos p) : pos(p) { }
+    virtual ~ASTNode();
+
+    /** The Optimize() method should perform any appropriate early-stage
+        optimizations on the node (e.g. constant folding).  The caller
+        should use the returned ASTNode * in place of the original node.
+        This method may return NULL if an error is encountered during
+        optimization. */
+    virtual ASTNode *Optimize() = 0;
+
+    /** Type checking should be performed by the node when this method is
+        called.  In the event of an error, a NULL value may be returned.
+        As with ASTNode::Optimize(), the caller should store the returned
+        pointer in place of the original ASTNode *. */
+    virtual ASTNode *TypeCheck() = 0;
+
+    virtual int EstimateCost() const = 0;
+
+    /** All AST nodes must track the file position where they are
+        defined. */
+    SourcePos pos;
+};
+
+
+/** Simple representation of the abstract syntax trees for all of the
+    functions declared in a compilation unit.
+ */
+class AST {
+public:
+    /** Add the AST for a function described by the given declaration
+        information and source code. */
+    void AddFunction(Symbol *sym, const std::vector<Symbol *> &args, 
+                     Stmt *code);
+
+    /** Generate LLVM IR for all of the functions into the current
+        module. */
+    void GenerateIR();
+
+private:
+    std::vector<Function *> functions;
+};
+
+#endif // ISPC_AST_H

bitcode2cpp.py

@@ -4,30 +4,36 @@ import sys
 import string
 import re
 import subprocess
+import platform
+import os
 
 length=0
 
 src=str(sys.argv[1])
 
-target = re.sub(".*stdlib-", "", src)
+target = re.sub(".*builtins-", "", src)
 target = re.sub("\.ll$", "", target)
 target = re.sub("\.c$", "", target)
 target = re.sub("-", "_", target)
 
+llvm_as="llvm-as"
+if platform.system() == 'Windows' or string.find(platform.system(), "CYGWIN_NT") != -1:
+    llvm_as = os.getenv("LLVM_INSTALL_DIR").replace("\\", "/") + "/bin/" + llvm_as
+
 try:
-    as_out=subprocess.Popen([ "llvm-as", "-", "-o", "-"], stdout=subprocess.PIPE)
+    as_out=subprocess.Popen([llvm_as, "-", "-o", "-"], stdout=subprocess.PIPE)
 except IOError:
     print >> sys.stderr, "Couldn't open " + src
     sys.exit(1)
 
-print "unsigned char stdlib_bitcode_" + target + "[] = {"
+print "unsigned char builtins_bitcode_" + target + "[] = {"
 for line in as_out.stdout.readlines():
     length = length + len(line)
     for c in line:
         print ord(c)
         print ", "
 print " 0 };\n\n"
-print "int stdlib_bitcode_" + target + "_length = " + str(length) + ";\n"
+print "int builtins_bitcode_" + target + "_length = " + str(length) + ";\n"
 
 as_out.wait()
 

buildall.bat

@@ -0,0 +1,16 @@
+@echo off
+
+REM If LLVM_INSTALL_DIR isn't set globally in your environment,
+REM it can be set here_
+set LLVM_INSTALL_DIR=c:\users\mmp\llvm-dev
+
+REM Both the LLVM binaries and python need to be in the path
+set path=%LLVM_INSTALL_DIR%\bin;%PATH%;c:\cygwin\bin
+
+msbuild ispc.vcxproj /V:m /p:Platform=Win32 /p:Configuration=Release
+msbuild ispc_test.vcxproj /V:m /p:Platform=Win32 /p:Configuration=Release
+
+msbuild examples\examples.sln /V:m /p:Platform=x64 /p:Configuration=Release /t:rebuild
+msbuild examples\examples.sln /V:m /p:Platform=x64 /p:Configuration=Debug /t:rebuild
+msbuild examples\examples.sln /V:m /p:Platform=Win32 /p:Configuration=Release /t:rebuild
+msbuild examples\examples.sln /V:m /p:Platform=Win32 /p:Configuration=Debug /t:rebuild

builtins-avx-common.ll

@@ -0,0 +1,274 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; AVX target implementation.
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
+
+define float @__rcp_uniform_float(float) nounwind readonly alwaysinline {
+;    uniform float iv = extract(__rcp_u(v), 0);
+;    return iv * (2. - v * iv);
+  %vecval = insertelement <4 x float> undef, float %0, i32 0
+  %call = call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %vecval)
+  %scall = extractelement <4 x float> %call, i32 0
+
+  ; do one N-R iteration
+  %v_iv = fmul float %0, %scall
+  %two_minus = fsub float 2., %v_iv  
+  %iv_mul = fmul float %scall, %two_minus
+  ret float %iv_mul
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding floats
+
+declare <4 x float> @llvm.x86.sse41.round.ss(<4 x float>, <4 x float>, i32) nounwind readnone
+
+define float @__round_uniform_float(float) nounwind readonly alwaysinline {
+  ; roundss, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
+  ; the roundss intrinsic is a total mess--docs say:
+  ;
+  ;  __m128 _mm_round_ss (__m128 a, __m128 b, const int c)
+  ;       
+  ;  b is a 128-bit parameter. The lowest 32 bits are the result of the rounding function
+  ;  on b0. The higher order 96 bits are copied directly from input parameter a. The
+  ;  return value is described by the following equations:
+  ;
+  ;  r0 = RND(b0)
+  ;  r1 = a1
+  ;  r2 = a2
+  ;  r3 = a3
+  ;
+  ;  It doesn't matter what we pass as a, since we only need the r0 value
+  ;  here.  So we pass the same register for both.
+  %xi = insertelement <4 x float> undef, float %0, i32 0
+  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 8)
+  %rs = extractelement <4 x float> %xr, i32 0
+  ret float %rs
+}
+
+define float @__floor_uniform_float(float) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <4 x float> undef, float %0, i32 0
+  ; roundps, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 9)
+  %rs = extractelement <4 x float> %xr, i32 0
+  ret float %rs
+}
+
+define float @__ceil_uniform_float(float) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <4 x float> undef, float %0, i32 0
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 10)
+  %rs = extractelement <4 x float> %xr, i32 0
+  ret float %rs
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare <2 x double> @llvm.x86.sse41.round.sd(<2 x double>, <2 x double>, i32) nounwind readnone
+
+define double @__round_uniform_double(double) nounwind readonly alwaysinline {
+  %xi = insertelement <2 x double> undef, double %0, i32 0
+  %xr = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %xi, <2 x double> %xi, i32 8)
+  %rs = extractelement <2 x double> %xr, i32 0
+  ret double %rs
+}
+
+define double @__floor_uniform_double(double) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <2 x double> undef, double %0, i32 0
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  %xr = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %xi, <2 x double> %xi, i32 9)
+  %rs = extractelement <2 x double> %xr, i32 0
+  ret double %rs
+}
+
+define double @__ceil_uniform_double(double) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <2 x double> undef, double %0, i32 0
+  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  %xr = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %xi, <2 x double> %xi, i32 10)
+  %rs = extractelement <2 x double> %xr, i32 0
+  ret double %rs
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rsqrt
+
+declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
+
+define float @__rsqrt_uniform_float(float) nounwind readonly alwaysinline {
+  ;  uniform float is = extract(__rsqrt_u(v), 0);
+  %v = insertelement <4 x float> undef, float %0, i32 0
+  %vis = call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %v)
+  %is = extractelement <4 x float> %vis, i32 0
+
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul float %0, %is
+  %v_is_is = fmul float %v_is, %is
+  %three_sub = fsub float 3., %v_is_is
+  %is_mul = fmul float %is, %three_sub
+  %half_scale = fmul float 0.5, %is_mul
+  ret float %half_scale
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; sqrt
+
+declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone
+
+define float @__sqrt_uniform_float(float) nounwind readonly alwaysinline {
+  sse_unary_scalar(ret, 4, float, @llvm.x86.sse.sqrt.ss, %0)
+  ret float %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; fastmath
+
+declare void @llvm.x86.sse.stmxcsr(i8 *) nounwind
+declare void @llvm.x86.sse.ldmxcsr(i8 *) nounwind
+
+define void @__fastmath() nounwind alwaysinline {
+  %ptr = alloca i32
+  %ptr8 = bitcast i32 * %ptr to i8 *
+  call void @llvm.x86.sse.stmxcsr(i8 * %ptr8)
+  %oldval = load i32 *%ptr
+
+  ; turn on DAZ (64)/FTZ (32768) -> 32832
+  %update = or i32 %oldval, 32832
+  store i32 %update, i32 *%ptr
+  call void @llvm.x86.sse.ldmxcsr(i8 * %ptr8)
+  ret void
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <4 x float> @llvm.x86.sse.max.ss(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @llvm.x86.sse.min.ss(<4 x float>, <4 x float>) nounwind readnone
+
+define float @__max_uniform_float(float, float) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.max.ss, %0, %1)
+  ret float %ret
+}
+
+define float @__min_uniform_float(float, float) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.min.ss, %0, %1)
+  ret float %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; int min/max
+
+declare <4 x i32> @llvm.x86.sse41.pminsd(<4 x i32>, <4 x i32>) nounwind readnone
+declare <4 x i32> @llvm.x86.sse41.pmaxsd(<4 x i32>, <4 x i32>) nounwind readnone
+
+define i32 @__min_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pminsd, %0, %1)
+  ret i32 %ret
+}
+
+define i32 @__max_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
+  ret i32 %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unsigned int min/max
+
+declare <4 x i32> @llvm.x86.sse41.pminud(<4 x i32>, <4 x i32>) nounwind readnone
+declare <4 x i32> @llvm.x86.sse41.pmaxud(<4 x i32>, <4 x i32>) nounwind readnone
+
+define i32 @__min_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pminud, %0, %1)
+  ret i32 %ret
+}
+
+define i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pmaxud, %0, %1)
+  ret i32 %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops
+
+declare i32 @llvm.ctpop.i32(i32) nounwind readnone
+
+define i32 @__popcnt_int32(i32) nounwind readonly alwaysinline {
+  %call = call i32 @llvm.ctpop.i32(i32 %0)
+  ret i32 %call
+}
+
+declare i64 @llvm.ctpop.i64(i64) nounwind readnone
+
+define i64 @__popcnt_int64(i64) nounwind readonly alwaysinline {
+  %call = call i64 @llvm.ctpop.i64(i64 %0)
+  ret i64 %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <2 x double> @llvm.x86.sse.sqrt.sd(<2 x double>) nounwind readnone
+
+define double @__sqrt_uniform_double(double) nounwind alwaysinline {
+  sse_unary_scalar(ret, 2, double, @llvm.x86.sse.sqrt.sd, %0)
+  ret double %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <2 x double> @llvm.x86.sse2.max.sd(<2 x double>, <2 x double>) nounwind readnone
+declare <2 x double> @llvm.x86.sse2.min.sd(<2 x double>, <2 x double>) nounwind readnone
+
+define double @__min_uniform_double(double, double) nounwind readnone alwaysinline {
+  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.min.sd, %0, %1)
+  ret double %ret
+}
+
+define double @__max_uniform_double(double, double) nounwind readnone alwaysinline {
+  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.max.sd, %0, %1)
+  ret double %ret
+}

builtins-avx-x2.ll

@@ -0,0 +1,658 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Basic 16-wide definitions
+
+stdlib_core(16)
+packed_load_and_store(16)
+scans(16)
+int64minmax(16)
+
+include(`builtins-avx-common.ll')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <8 x float> @llvm.x86.avx.rcp.ps.256(<8 x float>) nounwind readnone
+
+define <16 x float> @__rcp_varying_float(<16 x float>) nounwind readonly alwaysinline {
+  ;  float iv = __rcp_v(v);
+  ;  return iv * (2. - v * iv);
+
+  unary8to16(call, float, @llvm.x86.avx.rcp.ps.256, %0)
+  ; do one N-R iteration
+  %v_iv = fmul <16 x float> %0, %call
+  %two_minus = fsub <16 x float> <float 2., float 2., float 2., float 2.,
+                                  float 2., float 2., float 2., float 2.,
+                                  float 2., float 2., float 2., float 2.,
+                                  float 2., float 2., float 2., float 2.>, %v_iv  
+  %iv_mul = fmul <16 x float> %call, %two_minus
+  ret <16 x float> %iv_mul
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding floats
+
+declare <8 x float> @llvm.x86.avx.round.ps.256(<8 x float>, i32) nounwind readnone
+
+define <16 x float> @__round_varying_float(<16 x float>) nounwind readonly alwaysinline {
+  ; roundps, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
+  round8to16(%0, 8)
+}
+
+define <16 x float> @__floor_varying_float(<16 x float>) nounwind readonly alwaysinline {
+  ; roundps, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  round8to16(%0, 9)
+}
+
+define <16 x float> @__ceil_varying_float(<16 x float>) nounwind readonly alwaysinline {
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  round8to16(%0, 10)
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare <4 x double> @llvm.x86.avx.round.pd.256(<4 x double>, i32) nounwind readnone
+
+define <16 x double> @__round_varying_double(<16 x double>) nounwind readonly alwaysinline {
+  round4to16double(%0, 8)
+}
+
+define <16 x double> @__floor_varying_double(<16 x double>) nounwind readonly alwaysinline {
+  round4to16double(%0, 9)
+}
+
+define <16 x double> @__ceil_varying_double(<16 x double>) nounwind readonly alwaysinline {
+  round4to16double(%0, 10)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rsqrt
+
+declare <8 x float> @llvm.x86.avx.rsqrt.ps.256(<8 x float>) nounwind readnone
+
+define <16 x float> @__rsqrt_varying_float(<16 x float> %v) nounwind readonly alwaysinline {
+  ;  float is = __rsqrt_v(v);
+  unary8to16(is, float, @llvm.x86.avx.rsqrt.ps.256, %v)
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul <16 x float> %v, %is
+  %v_is_is = fmul <16 x float> %v_is, %is
+  %three_sub = fsub <16 x float> <float 3., float 3., float 3., float 3.,
+                                  float 3., float 3., float 3., float 3.,
+                                  float 3., float 3., float 3., float 3.,
+                                  float 3., float 3., float 3., float 3.>, %v_is_is
+  %is_mul = fmul <16 x float> %is, %three_sub
+  %half_scale = fmul <16 x float> <float 0.5, float 0.5, float 0.5, float 0.5,
+                                   float 0.5, float 0.5, float 0.5, float 0.5,
+                                   float 0.5, float 0.5, float 0.5, float 0.5,
+                                   float 0.5, float 0.5, float 0.5, float 0.5>, %is_mul
+  ret <16 x float> %half_scale
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; sqrt
+
+declare <8 x float> @llvm.x86.avx.sqrt.ps.256(<8 x float>) nounwind readnone
+
+define <16 x float> @__sqrt_varying_float(<16 x float>) nounwind readonly alwaysinline {
+  unary8to16(call, float, @llvm.x86.avx.sqrt.ps.256, %0)
+  ret <16 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; svml
+
+; FIXME: need either to wire these up to the 8-wide SVML entrypoints,
+; or, use the macro to call the 4-wide ones 4x with our 16-wide
+; vectors...
+
+declare <16 x float> @__svml_sin(<16 x float>)
+declare <16 x float> @__svml_cos(<16 x float>)
+declare void @__svml_sincos(<16 x float>, <16 x float> *, <16 x float> *)
+declare <16 x float> @__svml_tan(<16 x float>)
+declare <16 x float> @__svml_atan(<16 x float>)
+declare <16 x float> @__svml_atan2(<16 x float>, <16 x float>)
+declare <16 x float> @__svml_exp(<16 x float>)
+declare <16 x float> @__svml_log(<16 x float>)
+declare <16 x float> @__svml_pow(<16 x float>, <16 x float>)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <8 x float> @llvm.x86.avx.max.ps.256(<8 x float>, <8 x float>) nounwind readnone
+declare <8 x float> @llvm.x86.avx.min.ps.256(<8 x float>, <8 x float>) nounwind readnone
+
+define <16 x float> @__max_varying_float(<16 x float>,
+                                                  <16 x float>) nounwind readonly alwaysinline {
+  binary8to16(call, float, @llvm.x86.avx.max.ps.256, %0, %1)
+  ret <16 x float> %call
+}
+
+define <16 x float> @__min_varying_float(<16 x float>,
+                                                  <16 x float>) nounwind readonly alwaysinline {
+  binary8to16(call, float, @llvm.x86.avx.min.ps.256, %0, %1)
+  ret <16 x float> %call
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; int min/max
+
+define <16 x i32> @__min_varying_int32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
+  binary4to16(ret, i32, @llvm.x86.sse41.pminsd, %0, %1)
+  ret <16 x i32> %ret
+}
+
+define <16 x i32> @__max_varying_int32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
+  binary4to16(ret, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
+  ret <16 x i32> %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unsigned int min/max
+
+define <16 x i32> @__min_varying_uint32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
+  binary4to16(ret, i32, @llvm.x86.sse41.pminud, %0, %1)
+  ret <16 x i32> %ret
+}
+
+define <16 x i32> @__max_varying_uint32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
+  binary4to16(ret, i32, @llvm.x86.sse41.pmaxud, %0, %1)
+  ret <16 x i32> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops
+
+declare i32 @llvm.x86.avx.movmsk.ps.256(<8 x float>) nounwind readnone
+
+define i32 @__movmsk(<16 x i32>) nounwind readnone alwaysinline {
+  %floatmask = bitcast <16 x i32> %0 to <16 x float>
+  %mask0 = shufflevector <16 x float> %floatmask, <16 x float> undef,
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %v0 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> %mask0) nounwind readnone
+  %mask1 = shufflevector <16 x float> %floatmask, <16 x float> undef,
+          <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %v1 = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> %mask1) nounwind readnone
+
+  %v1shift = shl i32 %v1, 8
+  %v = or i32 %v1shift, %v0
+  ret i32 %v
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal float ops
+
+declare <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float>, <8 x float>) nounwind readnone
+
+define float @__reduce_add_float(<16 x float>) nounwind readonly alwaysinline {
+  %va = shufflevector <16 x float> %0, <16 x float> undef,
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %vb = shufflevector <16 x float> %0, <16 x float> undef,
+          <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %v1 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %va, <8 x float> %vb)
+  %v2 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %v1, <8 x float> %v1)
+  %v3 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %v2, <8 x float> %v2)
+  %scalar1 = extractelement <8 x float> %v3, i32 0
+  %scalar2 = extractelement <8 x float> %v3, i32 4
+  %sum = fadd float %scalar1, %scalar2
+  ret float %sum
+}
+
+
+define float @__reduce_min_float(<16 x float>) nounwind readnone alwaysinline {
+  reduce16(float, @__min_varying_float, @__min_uniform_float)
+}
+
+
+define float @__reduce_max_float(<16 x float>) nounwind readnone alwaysinline {
+  reduce16(float, @__max_varying_float, @__max_uniform_float)
+}
+
+reduce_equal(16)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal int32 ops
+
+define <16 x i32> @__add_varying_int32(<16 x i32>,
+                                                <16 x i32>) nounwind readnone alwaysinline {
+  %s = add <16 x i32> %0, %1
+  ret <16 x i32> %s
+}
+
+define i32 @__add_uniform_int32(i32, i32) nounwind readnone alwaysinline {
+  %s = add i32 %0, %1
+  ret i32 %s
+}
+
+define i32 @__reduce_add_int32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__add_varying_int32, @__add_uniform_int32)
+}
+
+
+define i32 @__reduce_min_int32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__min_varying_int32, @__min_uniform_int32)
+}
+
+
+define i32 @__reduce_max_int32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__max_varying_int32, @__max_uniform_int32)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;; horizontal uint32 ops
+
+define i32 @__reduce_add_uint32(<16 x i32> %v) nounwind readnone alwaysinline {
+  %r = call i32 @__reduce_add_int32(<16 x i32> %v)
+  ret i32 %r
+}
+
+define i32 @__reduce_min_uint32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__min_varying_uint32, @__min_uniform_uint32)
+}
+
+
+define i32 @__reduce_max_uint32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__max_varying_uint32, @__max_uniform_uint32)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal double ops
+
+declare <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double>, <4 x double>) nounwind readnone
+
+define double @__reduce_add_double(<16 x double>) nounwind readonly alwaysinline {
+  %va = shufflevector <16 x double> %0, <16 x double> undef,
+         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %vb = shufflevector <16 x double> %0, <16 x double> undef,
+         <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %vc = shufflevector <16 x double> %0, <16 x double> undef,
+         <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  %vd = shufflevector <16 x double> %0, <16 x double> undef,
+         <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+  %vab = fadd <4 x double> %va, %vb
+  %vcd = fadd <4 x double> %vc, %vd
+
+  %sum0 = call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %vab, <4 x double> %vcd)
+  %sum1 = call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %sum0, <4 x double> %sum0)
+  %final0 = extractelement <4 x double> %sum1, i32 0
+  %final1 = extractelement <4 x double> %sum1, i32 2
+  %sum = fadd double %final0, %final1
+  ret double %sum
+}
+
+define double @__reduce_min_double(<16 x double>) nounwind readnone alwaysinline {
+  reduce16(double, @__min_varying_double, @__min_uniform_double)
+}
+
+
+define double @__reduce_max_double(<16 x double>) nounwind readnone alwaysinline {
+  reduce16(double, @__max_varying_double, @__max_uniform_double)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal int64 ops
+
+define <16 x i64> @__add_varying_int64(<16 x i64>,
+                                                <16 x i64>) nounwind readnone alwaysinline {
+  %s = add <16 x i64> %0, %1
+  ret <16 x i64> %s
+}
+
+define i64 @__add_uniform_int64(i64, i64) nounwind readnone alwaysinline {
+  %s = add i64 %0, %1
+  ret i64 %s
+}
+
+define i64 @__reduce_add_int64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__add_varying_int64, @__add_uniform_int64)
+}
+
+
+define i64 @__reduce_min_int64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+
+define i64 @__reduce_max_int64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;; horizontal uint64 ops
+
+define i64 @__reduce_add_uint64(<16 x i64> %v) nounwind readnone alwaysinline {
+  %r = call i64 @__reduce_add_int64(<16 x i64> %v)
+  ret i64 %r
+}
+
+define i64 @__reduce_min_uint64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+
+define i64 @__reduce_max_uint64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__max_varying_uint64, @__max_uniform_uint64)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unaligned loads/loads+broadcasts
+
+load_and_broadcast(16, i8, 8)
+load_and_broadcast(16, i16, 16)
+load_and_broadcast(16, i32, 32)
+load_and_broadcast(16, i64, 64)
+
+; no masked load instruction for i8 and i16 types??
+load_masked(16, i8,  8,  1)
+load_masked(16, i16, 16, 2)
+
+declare <8 x float> @llvm.x86.avx.maskload.ps.256(i8 *, <8 x float> %mask)
+declare <4 x double> @llvm.x86.avx.maskload.pd.256(i8 *, <4 x double> %mask)
+ 
+define <16 x i32> @__load_masked_32(i8 *, <16 x i32> %mask) nounwind alwaysinline {
+  %floatmask = bitcast <16 x i32> %mask to <16 x float>
+  %mask0 = shufflevector <16 x float> %floatmask, <16 x float> undef,
+     <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %val0 = call <8 x float> @llvm.x86.avx.maskload.ps.256(i8 * %0, <8 x float> %mask0)
+  %mask1 = shufflevector <16 x float> %floatmask, <16 x float> undef,
+     <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %ptr1 = getelementptr i8 * %0, i32 32   ;; 8x4 bytes = 32
+  %val1 = call <8 x float> @llvm.x86.avx.maskload.ps.256(i8 * %ptr1, <8 x float> %mask1)
+
+  %retval = shufflevector <8 x float> %val0, <8 x float> %val1,
+     <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7,
+                 i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %reti32 = bitcast <16 x float> %retval to <16 x i32>
+  ret <16 x i32> %reti32
+}
+
+
+define <16 x i64> @__load_masked_64(i8 *, <16 x i32> %mask) nounwind alwaysinline {
+  ; double up masks, bitcast to doubles
+  %mask0 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  %mask1 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+  %mask2 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 8, i32 8, i32 9, i32 9, i32 10, i32 10, i32 11, i32 11>
+  %mask3 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 12, i32 12, i32 13, i32 13, i32 14, i32 14, i32 15, i32 15>
+  %mask0d = bitcast <8 x i32> %mask0 to <4 x double>
+  %mask1d = bitcast <8 x i32> %mask1 to <4 x double>
+  %mask2d = bitcast <8 x i32> %mask2 to <4 x double>
+  %mask3d = bitcast <8 x i32> %mask3 to <4 x double>
+
+  %val0d = call <4 x double> @llvm.x86.avx.maskload.pd.256(i8 * %0, <4 x double> %mask0d)
+  %ptr1 = getelementptr i8 * %0, i32 32
+  %val1d = call <4 x double> @llvm.x86.avx.maskload.pd.256(i8 * %ptr1, <4 x double> %mask1d)
+  %ptr2 = getelementptr i8 * %0, i32 64
+  %val2d = call <4 x double> @llvm.x86.avx.maskload.pd.256(i8 * %ptr2, <4 x double> %mask2d)
+  %ptr3 = getelementptr i8 * %0, i32 96
+  %val3d = call <4 x double> @llvm.x86.avx.maskload.pd.256(i8 * %ptr3, <4 x double> %mask3d)
+
+  %val01 = shufflevector <4 x double> %val0d, <4 x double> %val1d,
+      <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %val23 = shufflevector <4 x double> %val2d, <4 x double> %val3d,
+      <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %val0123 = shufflevector <8 x double> %val01, <8 x double> %val23,
+      <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7,
+                  i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %val = bitcast <16 x double> %val0123 to <16 x i64>
+  ret <16 x i64> %val
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+; FIXME: there is no AVX instruction for these, but we could be clever
+; by packing the bits down and setting the last 3/4 or half, respectively,
+; of the mask to zero...  Not sure if this would be a win in the end
+gen_masked_store(16, i8, 8)
+gen_masked_store(16, i16, 16)
+
+; note that mask is the 2nd parameter, not the 3rd one!!
+declare void @llvm.x86.avx.maskstore.ps.256(i8 *, <8 x float>, <8 x float>)
+declare void @llvm.x86.avx.maskstore.pd.256(i8 *, <4 x double>, <4 x double>)
+
+define void @__masked_store_32(<16 x i32>* nocapture, <16 x i32>, 
+                               <16 x i32>) nounwind alwaysinline {
+  %ptr = bitcast <16 x i32> * %0 to i8 *
+  %val = bitcast <16 x i32> %1 to <16 x float>
+  %mask = bitcast <16 x i32> %2 to <16 x float>
+
+  %val0 = shufflevector <16 x float> %val, <16 x float> undef,
+        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %val1 = shufflevector <16 x float> %val, <16 x float> undef,
+        <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+
+  %mask0 = shufflevector <16 x float> %mask, <16 x float> undef,
+        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %mask1 = shufflevector <16 x float> %mask, <16 x float> undef,
+        <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+
+  call void @llvm.x86.avx.maskstore.ps.256(i8 * %ptr, <8 x float> %mask0, <8 x float> %val0)
+  %ptr1 = getelementptr i8 * %ptr, i32 32
+  call void @llvm.x86.avx.maskstore.ps.256(i8 * %ptr1, <8 x float> %mask1, <8 x float> %val1)
+
+  ret void
+}
+
+define void @__masked_store_64(<16 x i64>* nocapture, <16 x i64>,
+                               <16 x i32> %mask) nounwind alwaysinline {
+  %ptr = bitcast <16 x i64> * %0 to i8 *
+  %val = bitcast <16 x i64> %1 to <16 x double>
+
+  ; double up masks, bitcast to doubles
+  %mask0 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  %mask1 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+  %mask2 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 8, i32 8, i32 9, i32 9, i32 10, i32 10, i32 11, i32 11>
+  %mask3 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 12, i32 12, i32 13, i32 13, i32 14, i32 14, i32 15, i32 15>
+  %mask0d = bitcast <8 x i32> %mask0 to <4 x double>
+  %mask1d = bitcast <8 x i32> %mask1 to <4 x double>
+  %mask2d = bitcast <8 x i32> %mask2 to <4 x double>
+  %mask3d = bitcast <8 x i32> %mask3 to <4 x double>
+
+  %val0 = shufflevector <16 x double> %val, <16 x double> undef,
+     <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %val1 = shufflevector <16 x double> %val, <16 x double> undef,
+     <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %val2 = shufflevector <16 x double> %val, <16 x double> undef,
+     <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  %val3 = shufflevector <16 x double> %val, <16 x double> undef,
+     <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+
+  call void @llvm.x86.avx.maskstore.pd.256(i8 * %ptr, <4 x double> %mask0d, <4 x double> %val0)
+  %ptr1 = getelementptr i8 * %ptr, i32 32
+  call void @llvm.x86.avx.maskstore.pd.256(i8 * %ptr1, <4 x double> %mask1d, <4 x double> %val1)
+  %ptr2 = getelementptr i8 * %ptr, i32 64
+  call void @llvm.x86.avx.maskstore.pd.256(i8 * %ptr2, <4 x double> %mask2d, <4 x double> %val2)
+  %ptr3 = getelementptr i8 * %ptr, i32 96
+  call void @llvm.x86.avx.maskstore.pd.256(i8 * %ptr3, <4 x double> %mask3d, <4 x double> %val3)
+
+  ret void
+}
+
+
+masked_store_blend_8_16_by_16()
+
+declare <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float>, <8 x float>,
+                                                <8 x float>) nounwind readnone
+
+define void @__masked_store_blend_32(<16 x i32>* nocapture, <16 x i32>, 
+                                     <16 x i32>) nounwind alwaysinline {
+  %maskAsFloat = bitcast <16 x i32> %2 to <16 x float>
+  %oldValue = load <16 x i32>* %0, align 4
+  %oldAsFloat = bitcast <16 x i32> %oldValue to <16 x float>
+  %newAsFloat = bitcast <16 x i32> %1 to <16 x float>
+ 
+  %old0 = shufflevector <16 x float> %oldAsFloat, <16 x float> undef,
+        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %old1 = shufflevector <16 x float> %oldAsFloat, <16 x float> undef,
+        <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %new0 = shufflevector <16 x float> %newAsFloat, <16 x float> undef,
+        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %new1 = shufflevector <16 x float> %newAsFloat, <16 x float> undef,
+        <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %mask0 = shufflevector <16 x float> %maskAsFloat, <16 x float> undef,
+        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %mask1 = shufflevector <16 x float> %maskAsFloat, <16 x float> undef,
+        <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+
+  %blend0 = call <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float> %old0,
+                                                         <8 x float> %new0,
+                                                         <8 x float> %mask0)
+  %blend1 = call <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float> %old1,
+                                                         <8 x float> %new1,
+                                                         <8 x float> %mask1)
+  %blend = shufflevector <8 x float> %blend0, <8 x float> %blend1,
+    <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7,
+                i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %blendAsInt = bitcast <16 x float> %blend to <16 x i32>
+  store <16 x i32> %blendAsInt, <16 x i32>* %0, align 4
+  ret void
+}
+
+
+declare <4 x double> @llvm.x86.avx.blendv.pd.256(<4 x double>, <4 x double>,
+                                                 <4 x double>) nounwind readnone
+
+define void @__masked_store_blend_64(<16 x i64>* nocapture %ptr, <16 x i64> %newi64, 
+                                     <16 x i32> %mask) nounwind alwaysinline {
+  %oldValue = load <16 x i64>* %ptr, align 8
+  %old = bitcast <16 x i64> %oldValue to <16 x double>
+  %old0d = shufflevector <16 x double> %old, <16 x double> undef,
+     <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %old1d = shufflevector <16 x double> %old, <16 x double> undef,
+     <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %old2d = shufflevector <16 x double> %old, <16 x double> undef,
+     <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  %old3d = shufflevector <16 x double> %old, <16 x double> undef,
+     <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+
+  %new = bitcast <16 x i64> %newi64 to <16 x double>
+  %new0d = shufflevector <16 x double> %new, <16 x double> undef,
+     <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %new1d = shufflevector <16 x double> %new, <16 x double> undef,
+     <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %new2d = shufflevector <16 x double> %new, <16 x double> undef,
+     <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  %new3d = shufflevector <16 x double> %new, <16 x double> undef,
+     <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+
+  %mask0 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  %mask1 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+  %mask2 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 8, i32 8, i32 9, i32 9, i32 10, i32 10, i32 11, i32 11>
+  %mask3 = shufflevector <16 x i32> %mask, <16 x i32> undef,
+     <8 x i32> <i32 12, i32 12, i32 13, i32 13, i32 14, i32 14, i32 15, i32 15>
+  %mask0d = bitcast <8 x i32> %mask0 to <4 x double>
+  %mask1d = bitcast <8 x i32> %mask1 to <4 x double>
+  %mask2d = bitcast <8 x i32> %mask2 to <4 x double>
+  %mask3d = bitcast <8 x i32> %mask3 to <4 x double>
+
+  %result0d = call <4 x double> @llvm.x86.avx.blendv.pd.256(<4 x double> %old0d,
+                                 <4 x double> %new0d, <4 x double> %mask0d)
+  %result1d = call <4 x double> @llvm.x86.avx.blendv.pd.256(<4 x double> %old1d,
+                                 <4 x double> %new1d, <4 x double> %mask1d)
+  %result2d = call <4 x double> @llvm.x86.avx.blendv.pd.256(<4 x double> %old2d,
+                                 <4 x double> %new2d, <4 x double> %mask2d)
+  %result3d = call <4 x double> @llvm.x86.avx.blendv.pd.256(<4 x double> %old3d,
+                                 <4 x double> %new3d, <4 x double> %mask3d)
+
+  %result01 = shufflevector <4 x double> %result0d, <4 x double> %result1d,
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %result23 = shufflevector <4 x double> %result2d, <4 x double> %result3d,
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+
+  %result = shufflevector <8 x double> %result01, <8 x double> %result23,
+           <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7,
+                       i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %result64 = bitcast <16 x double> %result to <16 x i64>
+  store <16 x i64> %result64, <16 x i64> * %ptr
+  ret void
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; gather/scatter
+
+gen_gather(16, i8)
+gen_gather(16, i16)
+gen_gather(16, i32)
+gen_gather(16, i64)
+
+gen_scatter(16, i8)
+gen_scatter(16, i16)
+gen_scatter(16, i32)
+gen_scatter(16, i64)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <4 x double> @llvm.x86.avx.sqrt.pd.256(<4 x double>) nounwind readnone
+
+define <16 x double> @__sqrt_varying_double(<16 x double>) nounwind alwaysinline {
+  unary4to16(ret, double, @llvm.x86.avx.sqrt.pd.256, %0)
+  ret <16 x double> %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <4 x double> @llvm.x86.avx.max.pd.256(<4 x double>, <4 x double>) nounwind readnone
+declare <4 x double> @llvm.x86.avx.min.pd.256(<4 x double>, <4 x double>) nounwind readnone
+
+define <16 x double> @__min_varying_double(<16 x double>, <16 x double>) nounwind readnone alwaysinline {
+  binary4to16(ret, double, @llvm.x86.avx.min.pd.256, %0, %1)
+  ret <16 x double> %ret
+}
+
+define <16 x double> @__max_varying_double(<16 x double>, <16 x double>) nounwind readnone alwaysinline {
+  binary4to16(ret, double, @llvm.x86.avx.max.pd.256, %0, %1)
+  ret <16 x double> %ret
+}

builtins-avx.ll

@@ -0,0 +1,557 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Basic 8-wide definitions
+
+stdlib_core(8)
+packed_load_and_store(8)
+scans(8)
+int64minmax(8)
+
+include(`builtins-avx-common.ll')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <8 x float> @llvm.x86.avx.rcp.ps.256(<8 x float>) nounwind readnone
+
+define <8 x float> @__rcp_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ;  float iv = __rcp_v(v);
+  ;  return iv * (2. - v * iv);
+
+  %call = call <8 x float> @llvm.x86.avx.rcp.ps.256(<8 x float> %0)
+  ; do one N-R iteration
+  %v_iv = fmul <8 x float> %0, %call
+  %two_minus = fsub <8 x float> <float 2., float 2., float 2., float 2.,
+                                 float 2., float 2., float 2., float 2.>, %v_iv  
+  %iv_mul = fmul <8 x float> %call, %two_minus
+  ret <8 x float> %iv_mul
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding floats
+
+declare <8 x float> @llvm.x86.avx.round.ps.256(<8 x float>, i32) nounwind readnone
+
+define <8 x float> @__round_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ; roundps, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
+  %call = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %0, i32 8)
+  ret <8 x float> %call
+}
+
+define <8 x float> @__floor_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ; roundps, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  %call = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %0, i32 9)
+  ret <8 x float> %call
+}
+
+define <8 x float> @__ceil_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  %call = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %0, i32 10)
+  ret <8 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare <4 x double> @llvm.x86.avx.round.pd.256(<4 x double>, i32) nounwind readnone
+
+define <8 x double> @__round_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  round4to8double(%0, 8)
+}
+
+define <8 x double> @__floor_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1000 = 9
+  round4to8double(%0, 9)
+}
+
+
+define <8 x double> @__ceil_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1000 = 10
+  round4to8double(%0, 10)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rsqrt
+
+declare <8 x float> @llvm.x86.avx.rsqrt.ps.256(<8 x float>) nounwind readnone
+
+define <8 x float> @__rsqrt_varying_float(<8 x float> %v) nounwind readonly alwaysinline {
+  ;  float is = __rsqrt_v(v);
+  %is = call <8 x float> @llvm.x86.avx.rsqrt.ps.256(<8 x float> %v)
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul <8 x float> %v, %is
+  %v_is_is = fmul <8 x float> %v_is, %is
+  %three_sub = fsub <8 x float> <float 3., float 3., float 3., float 3.,
+                                 float 3., float 3., float 3., float 3.>, %v_is_is
+  %is_mul = fmul <8 x float> %is, %three_sub
+  %half_scale = fmul <8 x float> <float 0.5, float 0.5, float 0.5, float 0.5,
+                                  float 0.5, float 0.5, float 0.5, float 0.5>, %is_mul
+  ret <8 x float> %half_scale
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; sqrt
+
+declare <8 x float> @llvm.x86.avx.sqrt.ps.256(<8 x float>) nounwind readnone
+
+define <8 x float> @__sqrt_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  %call = call <8 x float> @llvm.x86.avx.sqrt.ps.256(<8 x float> %0)
+  ret <8 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; svml
+
+; FIXME: need either to wire these up to the 8-wide SVML entrypoints,
+; or, use the macro to call the 4-wide ones twice with our 8-wide
+; vectors...
+
+declare <8 x float> @__svml_sin(<8 x float>)
+declare <8 x float> @__svml_cos(<8 x float>)
+declare void @__svml_sincos(<8 x float>, <8 x float> *, <8 x float> *)
+declare <8 x float> @__svml_tan(<8 x float>)
+declare <8 x float> @__svml_atan(<8 x float>)
+declare <8 x float> @__svml_atan2(<8 x float>, <8 x float>)
+declare <8 x float> @__svml_exp(<8 x float>)
+declare <8 x float> @__svml_log(<8 x float>)
+declare <8 x float> @__svml_pow(<8 x float>, <8 x float>)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <8 x float> @llvm.x86.avx.max.ps.256(<8 x float>, <8 x float>) nounwind readnone
+declare <8 x float> @llvm.x86.avx.min.ps.256(<8 x float>, <8 x float>) nounwind readnone
+
+define <8 x float> @__max_varying_float(<8 x float>,
+                                                 <8 x float>) nounwind readonly alwaysinline {
+  %call = call <8 x float> @llvm.x86.avx.max.ps.256(<8 x float> %0, <8 x float> %1)
+  ret <8 x float> %call
+}
+
+define <8 x float> @__min_varying_float(<8 x float>,
+                                                 <8 x float>) nounwind readonly alwaysinline {
+  %call = call <8 x float> @llvm.x86.avx.min.ps.256(<8 x float> %0, <8 x float> %1)
+  ret <8 x float> %call
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; int min/max
+
+define <8 x i32> @__min_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  binary4to8(ret, i32, @llvm.x86.sse41.pminsd, %0, %1)
+  ret <8 x i32> %ret
+}
+
+define <8 x i32> @__max_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  binary4to8(ret, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
+  ret <8 x i32> %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unsigned int min/max
+
+define <8 x i32> @__min_varying_uint32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  binary4to8(ret, i32, @llvm.x86.sse41.pminud, %0, %1)
+  ret <8 x i32> %ret
+}
+
+define <8 x i32> @__max_varying_uint32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  binary4to8(ret, i32, @llvm.x86.sse41.pmaxud, %0, %1)
+  ret <8 x i32> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops
+
+declare i32 @llvm.x86.avx.movmsk.ps.256(<8 x float>) nounwind readnone
+
+define i32 @__movmsk(<8 x i32>) nounwind readnone alwaysinline {
+  %floatmask = bitcast <8 x i32> %0 to <8 x float>
+  %v = call i32 @llvm.x86.avx.movmsk.ps.256(<8 x float> %floatmask) nounwind readnone
+  ret i32 %v
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal float ops
+
+declare <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float>, <8 x float>) nounwind readnone
+
+define float @__reduce_add_float(<8 x float>) nounwind readonly alwaysinline {
+  %v1 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %0, <8 x float> %0)
+  %v2 = call <8 x float> @llvm.x86.avx.hadd.ps.256(<8 x float> %v1, <8 x float> %v1)
+  %scalar1 = extractelement <8 x float> %v2, i32 0
+  %scalar2 = extractelement <8 x float> %v2, i32 4
+  %sum = fadd float %scalar1, %scalar2
+  ret float %sum
+}
+
+
+define float @__reduce_min_float(<8 x float>) nounwind readnone alwaysinline {
+  reduce8(float, @__min_varying_float, @__min_uniform_float)
+}
+
+
+define float @__reduce_max_float(<8 x float>) nounwind readnone alwaysinline {
+  reduce8(float, @__max_varying_float, @__max_uniform_float)
+}
+
+reduce_equal(8)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal int32 ops
+
+define <8 x i32> @__add_varying_int32(<8 x i32>,
+                                               <8 x i32>) nounwind readnone alwaysinline {
+  %s = add <8 x i32> %0, %1
+  ret <8 x i32> %s
+}
+
+define i32 @__add_uniform_int32(i32, i32) nounwind readnone alwaysinline {
+  %s = add i32 %0, %1
+  ret i32 %s
+}
+
+define i32 @__reduce_add_int32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__add_varying_int32, @__add_uniform_int32)
+}
+
+
+define i32 @__reduce_min_int32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__min_varying_int32, @__min_uniform_int32)
+}
+
+
+define i32 @__reduce_max_int32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__max_varying_int32, @__max_uniform_int32)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;; horizontal uint32 ops
+
+define i32 @__reduce_add_uint32(<8 x i32> %v) nounwind readnone alwaysinline {
+  %r = call i32 @__reduce_add_int32(<8 x i32> %v)
+  ret i32 %r
+}
+
+define i32 @__reduce_min_uint32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__min_varying_uint32, @__min_uniform_uint32)
+}
+
+
+define i32 @__reduce_max_uint32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__max_varying_uint32, @__max_uniform_uint32)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal double ops
+
+declare <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double>, <4 x double>) nounwind readnone
+
+define double @__reduce_add_double(<8 x double>) nounwind readonly alwaysinline {
+  %v0 = shufflevector <8 x double> %0, <8 x double> undef,
+                      <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %v1 = shufflevector <8 x double> %0, <8 x double> undef,
+                      <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %sum0 = call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %v0, <4 x double> %v1)
+  %sum1 = call <4 x double> @llvm.x86.avx.hadd.pd.256(<4 x double> %sum0, <4 x double> %sum0)
+  %final0 = extractelement <4 x double> %sum1, i32 0
+  %final1 = extractelement <4 x double> %sum1, i32 2
+  %sum = fadd double %final0, %final1
+
+  ret double %sum
+}
+
+define double @__reduce_min_double(<8 x double>) nounwind readnone alwaysinline {
+  reduce8(double, @__min_varying_double, @__min_uniform_double)
+}
+
+
+define double @__reduce_max_double(<8 x double>) nounwind readnone alwaysinline {
+  reduce8(double, @__max_varying_double, @__max_uniform_double)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal int64 ops
+
+define <8 x i64> @__add_varying_int64(<8 x i64>,
+                                               <8 x i64>) nounwind readnone alwaysinline {
+  %s = add <8 x i64> %0, %1
+  ret <8 x i64> %s
+}
+
+define i64 @__add_uniform_int64(i64, i64) nounwind readnone alwaysinline {
+  %s = add i64 %0, %1
+  ret i64 %s
+}
+
+define i64 @__reduce_add_int64(<8 x i64>) nounwind readnone alwaysinline {
+  reduce8(i64, @__add_varying_int64, @__add_uniform_int64)
+}
+
+
+define i64 @__reduce_min_int64(<8 x i64>) nounwind readnone alwaysinline {
+  reduce8(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+
+define i64 @__reduce_max_int64(<8 x i64>) nounwind readnone alwaysinline {
+  reduce8(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;; horizontal uint64 ops
+
+define i64 @__reduce_add_uint64(<8 x i64> %v) nounwind readnone alwaysinline {
+  %r = call i64 @__reduce_add_int64(<8 x i64> %v)
+  ret i64 %r
+}
+
+define i64 @__reduce_min_uint64(<8 x i64>) nounwind readnone alwaysinline {
+  reduce8(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+
+define i64 @__reduce_max_uint64(<8 x i64>) nounwind readnone alwaysinline {
+  reduce8(i64, @__max_varying_uint64, @__max_uniform_uint64)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unaligned loads/loads+broadcasts
+
+load_and_broadcast(8, i8, 8)
+load_and_broadcast(8, i16, 16)
+load_and_broadcast(8, i32, 32)
+load_and_broadcast(8, i64, 64)
+
+; no masked load instruction for i8 and i16 types??
+load_masked(8, i8,  8,  1)
+load_masked(8, i16, 16, 2)
+
+declare <8 x float> @llvm.x86.avx.maskload.ps.256(i8 *, <8 x float> %mask)
+declare <4 x double> @llvm.x86.avx.maskload.pd.256(i8 *, <4 x double> %mask)
+ 
+define <8 x i32> @__load_masked_32(i8 *, <8 x i32> %mask) nounwind alwaysinline {
+  %floatmask = bitcast <8 x i32> %mask to <8 x float>
+  %floatval = call <8 x float> @llvm.x86.avx.maskload.ps.256(i8 * %0, <8 x float> %floatmask)
+  %retval = bitcast <8 x float> %floatval to <8 x i32>
+  ret <8 x i32> %retval
+}
+
+
+define <8 x i64> @__load_masked_64(i8 *, <8 x i32> %mask) nounwind alwaysinline {
+  ; double up masks, bitcast to doubles
+  %mask0 = shufflevector <8 x i32> %mask, <8 x i32> undef,
+     <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  %mask1 = shufflevector <8 x i32> %mask, <8 x i32> undef,
+     <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+  %mask0d = bitcast <8 x i32> %mask0 to <4 x double>
+  %mask1d = bitcast <8 x i32> %mask1 to <4 x double>
+
+  %val0d = call <4 x double> @llvm.x86.avx.maskload.pd.256(i8 * %0, <4 x double> %mask0d)
+  %ptr1 = getelementptr i8 * %0, i32 32
+  %val1d = call <4 x double> @llvm.x86.avx.maskload.pd.256(i8 * %ptr1, <4 x double> %mask1d)
+
+  %vald = shufflevector <4 x double> %val0d, <4 x double> %val1d,
+      <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %val = bitcast <8 x double> %vald to <8 x i64>
+  ret <8 x i64> %val
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+; FIXME: there is no AVX instruction for these, but we could be clever
+; by packing the bits down and setting the last 3/4 or half, respectively,
+; of the mask to zero...  Not sure if this would be a win in the end
+gen_masked_store(8, i8, 8)
+gen_masked_store(8, i16, 16)
+
+; note that mask is the 2nd parameter, not the 3rd one!!
+declare void @llvm.x86.avx.maskstore.ps.256(i8 *, <8 x float>, <8 x float>)
+declare void @llvm.x86.avx.maskstore.pd.256(i8 *, <4 x double>, <4 x double>)
+
+define void @__masked_store_32(<8 x i32>* nocapture, <8 x i32>, 
+                               <8 x i32>) nounwind alwaysinline {
+  %ptr = bitcast <8 x i32> * %0 to i8 *
+  %val = bitcast <8 x i32> %1 to <8 x float>
+  %mask = bitcast <8 x i32> %2 to <8 x float>
+  call void @llvm.x86.avx.maskstore.ps.256(i8 * %ptr, <8 x float> %mask, <8 x float> %val)
+  ret void
+}
+
+define void @__masked_store_64(<8 x i64>* nocapture, <8 x i64>,
+                               <8 x i32> %mask) nounwind alwaysinline {
+  %ptr = bitcast <8 x i64> * %0 to i8 *
+  %val = bitcast <8 x i64> %1 to <8 x double>
+
+  %mask0 = shufflevector <8 x i32> %mask, <8 x i32> undef,
+     <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  %mask1 = shufflevector <8 x i32> %mask, <8 x i32> undef,
+     <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+
+  %mask0d = bitcast <8 x i32> %mask0 to <4 x double>
+  %mask1d = bitcast <8 x i32> %mask1 to <4 x double>
+
+  %val0 = shufflevector <8 x double> %val, <8 x double> undef,
+     <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %val1 = shufflevector <8 x double> %val, <8 x double> undef,
+     <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+
+  call void @llvm.x86.avx.maskstore.pd.256(i8 * %ptr, <4 x double> %mask0d, <4 x double> %val0)
+  %ptr1 = getelementptr i8 * %ptr, i32 32
+  call void @llvm.x86.avx.maskstore.pd.256(i8 * %ptr1, <4 x double> %mask1d, <4 x double> %val1)
+  ret void
+}
+
+
+
+masked_store_blend_8_16_by_8()
+
+declare <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float>, <8 x float>,
+                                                <8 x float>) nounwind readnone
+
+define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>, 
+                                     <8 x i32>) nounwind alwaysinline {
+  %mask_as_float = bitcast <8 x i32> %2 to <8 x float>
+  %oldValue = load <8 x i32>* %0, align 4
+  %oldAsFloat = bitcast <8 x i32> %oldValue to <8 x float>
+  %newAsFloat = bitcast <8 x i32> %1 to <8 x float>
+  %blend = call <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float> %oldAsFloat,
+                                                        <8 x float> %newAsFloat,
+                                                        <8 x float> %mask_as_float)
+  %blendAsInt = bitcast <8 x float> %blend to <8 x i32>
+  store <8 x i32> %blendAsInt, <8 x i32>* %0, align 4
+  ret void
+}
+
+
+define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new, 
+                                     <8 x i32> %i32mask) nounwind alwaysinline {
+  %oldValue = load <8 x i64>* %ptr, align 8
+  %mask = bitcast <8 x i32> %i32mask to <8 x float>
+
+  ; Do 4x64-bit blends by doing two <8 x i32> blends, where the <8 x i32> values
+  ; are actually bitcast <4 x i64> values
+  ;
+  ; set up the first four 64-bit values
+  %old01  = shufflevector <8 x i64> %oldValue, <8 x i64> undef,
+                          <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %old01f = bitcast <4 x i64> %old01 to <8 x float>
+  %new01  = shufflevector <8 x i64> %new, <8 x i64> undef,
+                          <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %new01f = bitcast <4 x i64> %new01 to <8 x float>
+  ; compute mask--note that the indices are all doubled-up
+  %mask01 = shufflevector <8 x float> %mask, <8 x float> undef,
+                          <8 x i32> <i32 0, i32 0, i32 1, i32 1,
+                                     i32 2, i32 2, i32 3, i32 3>
+  ; and blend them
+  %result01f = call <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float> %old01f,
+                                                            <8 x float> %new01f,
+                                                            <8 x float> %mask01)
+  %result01 = bitcast <8 x float> %result01f to <4 x i64>
+
+  ; and again
+  %old23  = shufflevector <8 x i64> %oldValue, <8 x i64> undef,
+                          <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %old23f = bitcast <4 x i64> %old23 to <8 x float>
+  %new23  = shufflevector <8 x i64> %new, <8 x i64> undef,
+                          <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %new23f = bitcast <4 x i64> %new23 to <8 x float>
+  ; compute mask--note that the values are doubled-up...
+  %mask23 = shufflevector <8 x float> %mask, <8 x float> undef,
+                          <8 x i32> <i32 4, i32 4, i32 5, i32 5,
+                                     i32 6, i32 6, i32 7, i32 7>
+  ; and blend them
+  %result23f = call <8 x float> @llvm.x86.avx.blendv.ps.256(<8 x float> %old23f,
+                                                            <8 x float> %new23f,
+                                                            <8 x float> %mask23)
+  %result23 = bitcast <8 x float> %result23f to <4 x i64>
+
+  ; reconstruct the final <8 x i64> vector
+  %final = shufflevector <4 x i64> %result01, <4 x i64> %result23,
+                         <8 x i32> <i32 0, i32 1, i32 2, i32 3,
+                                    i32 4, i32 5, i32 6, i32 7>
+  store <8 x i64> %final, <8 x i64> * %ptr, align 8
+  ret void
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; gather/scatter
+
+gen_gather(8, i8)
+gen_gather(8, i16)
+gen_gather(8, i32)
+gen_gather(8, i64)
+
+gen_scatter(8, i8)
+gen_scatter(8, i16)
+gen_scatter(8, i32)
+gen_scatter(8, i64)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <4 x double> @llvm.x86.avx.sqrt.pd.256(<4 x double>) nounwind readnone
+
+define <8 x double> @__sqrt_varying_double(<8 x double>) nounwind alwaysinline {
+  unary4to8(ret, double, @llvm.x86.avx.sqrt.pd.256, %0)
+  ret <8 x double> %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <4 x double> @llvm.x86.avx.max.pd.256(<4 x double>, <4 x double>) nounwind readnone
+declare <4 x double> @llvm.x86.avx.min.pd.256(<4 x double>, <4 x double>) nounwind readnone
+
+define <8 x double> @__min_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
+  binary4to8(ret, double, @llvm.x86.avx.min.pd.256, %0, %1)
+  ret <8 x double> %ret
+}
+
+define <8 x double> @__max_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
+  binary4to8(ret, double, @llvm.x86.avx.max.pd.256, %0, %1)
+  ret <8 x double> %ret
+}
+

builtins-c.c

@@ -31,7 +31,7 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
 */
 
-/** @file stdlib-c.c
+/** @file builtins-c.c
     @brief Standard library function implementations written in C.
 
     This file provides C implementations of various functions that can be
@@ -51,8 +51,13 @@
   */
 
 
+#ifndef _MSC_VER
+#include <unistd.h>
+#endif // !_MSC_VER
+
 #include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <stdarg.h>
 
 typedef int Bool;
@@ -128,6 +133,8 @@ void __do_print(const char *format, const char *types, int width, int mask,
                 case 'V': PRINT_VECTOR("%llu", unsigned long long);
                 case 'd': PRINT_SCALAR("%f", double);
                 case 'D': PRINT_VECTOR("%f", double);
+                case 'p': PRINT_SCALAR("%p", void *);
+                case 'P': PRINT_VECTOR("%p", void *);
                 default:
                     printf("UNKNOWN TYPE ");
                     putchar(*types);
@@ -139,3 +146,28 @@ void __do_print(const char *format, const char *types, int width, int mask,
     }
     fflush(stdout);
 }
+
+
+int __num_cores() {
+#ifdef _MSC_VER
+    // This is quite a hack.  Including all of windows.h to get this definition
+    // pulls in a bunch of stuff that leads to undefined symbols at link time.
+    // So we don't #include <windows.h> but instead have the equivalent declarations
+    // here.  Presumably this struct declaration won't be changing in the future
+    // anyway...
+    struct SYSTEM_INFO {
+        int pad0[2];
+        void *pad1[2];
+        int *pad2;
+        int dwNumberOfProcessors;
+        int pad3[3];
+    };
+
+    struct SYSTEM_INFO sysInfo;
+    extern void __stdcall GetSystemInfo(struct SYSTEM_INFO *);
+    GetSystemInfo(&sysInfo);
+    return sysInfo.dwNumberOfProcessors;
+#else
+    return sysconf(_SC_NPROCESSORS_ONLN);
+#endif // !_MSC_VER
+}

builtins-dispatch.ll

@@ -0,0 +1,123 @@
+;;  Copyright (c) 2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;; This file defines various functions that are used when generating the
+;; the "dispatch" object/assembly file that has entrypoints for each
+;; exported function in a module that dispatch to the best available
+;; variant of that function that will run on the system's CPU.
+
+;; Stores the best target ISA that the system on which we're actually
+;; running supports.  -1 represents "uninitialized", otherwise this value
+;; should correspond to one of the enumerant values of Target::ISA from
+;; ispc.h.
+
+@__system_best_isa = internal global i32 -1
+
+declare void @abort() noreturn
+
+;; The below is the result of running "clang -O2 -emit-llvm -c -o -" on the
+;; following code...  Specifically, __get_system_isa should return a value
+;; corresponding to one of the Target::ISA enumerant values that gives the
+;; most capable ISA that the curremt system can run.
+;;
+;; #ifdef _MSC_VER
+;; extern void __stdcall __cpuid(int info[4], int infoType);
+;; #else
+;; static void __cpuid(int info[4], int infoType) {
+;;     __asm__ __volatile__ ("cpuid"
+;;                           : "=a" (info[0]), "=b" (info[1]), "=c" (info[2]), "=d" (info[3])
+;;                           : "0" (infoType));
+;; }
+;; #endif
+;; 
+;; int32_t __get_system_isa() {
+;;     int info[4];
+;;     __cpuid(info, 1);
+;;     /* NOTE: the values returned below must be the same as the
+;;        corresponding enumerant values in Target::ISA. */
+;;     if ((info[2] & (1 << 28)) != 0)
+;;         return 2; // AVX
+;;     else if ((info[2] & (1 << 19)) != 0)
+;;         return 1; // SSE4
+;;     else if ((info[3] & (1 << 26)) != 0)
+;;         return 0; // SSE2
+;;     else
+;;         abort();
+;; }
+
+%0 = type { i32, i32, i32, i32 }
+
+define i32 @__get_system_isa() nounwind ssp {
+  %1 = tail call %0 asm sideeffect "cpuid", "={ax},={bx},={cx},={dx},0,~{dirflag},~{fpsr},~{flags}"(i32 1) nounwind
+  %2 = extractvalue %0 %1, 2
+  %3 = extractvalue %0 %1, 3
+  %4 = and i32 %2, 268435456
+  %5 = icmp eq i32 %4, 0
+  br i1 %5, label %6, label %13
+
+; <label>:6                                       ; preds = %0
+  %7 = and i32 %2, 524288
+  %8 = icmp eq i32 %7, 0
+  br i1 %8, label %9, label %13
+
+; <label>:9                                       ; preds = %6
+  %10 = and i32 %3, 67108864
+  %11 = icmp eq i32 %10, 0
+  br i1 %11, label %12, label %13
+
+; <label>:12                                      ; preds = %9
+  tail call void @abort() noreturn nounwind
+  unreachable
+
+; <label>:13                                      ; preds = %9, %6, %0
+  %.0 = phi i32 [ 2, %0 ], [ 1, %6 ], [ 0, %9 ]
+  ret i32 %.0
+}
+
+
+;; This function is called by each of the dispatch functions we generate;
+;; it sets @__system_best_isa if it is unset.
+
+define void @__set_system_isa() {
+entry:
+  %bi = load i32* @__system_best_isa
+  %unset = icmp eq i32 %bi, -1
+  br i1 %unset, label %set_system_isa, label %done
+
+set_system_isa:
+  %bival = call i32 @__get_system_isa()
+  store i32 %bival, i32* @__system_best_isa
+  ret void
+
+done:
+  ret void
+}
+

builtins-sse2-common.ll

@@ -0,0 +1,266 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
+
+define float @__rcp_uniform_float(float) nounwind readonly alwaysinline {
+  ; do the rcpss call
+  %vecval = insertelement <4 x float> undef, float %0, i32 0
+  %call = call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %vecval)
+  %scall = extractelement <4 x float> %call, i32 0
+
+  ; do one N-R iteration to improve precision, as above
+  %v_iv = fmul float %0, %scall
+  %two_minus = fsub float 2., %v_iv  
+  %iv_mul = fmul float %scall, %two_minus
+  ret float %iv_mul
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; rsqrt
+
+declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
+
+define float @__rsqrt_uniform_float(float) nounwind readonly alwaysinline {
+  ;  uniform float is = extract(__rsqrt_u(v), 0);
+  %v = insertelement <4 x float> undef, float %0, i32 0
+  %vis = call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %v)
+  %is = extractelement <4 x float> %vis, i32 0
+
+  ; Newton-Raphson iteration to improve precision
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul float %0, %is
+  %v_is_is = fmul float %v_is, %is
+  %three_sub = fsub float 3., %v_is_is
+  %is_mul = fmul float %is, %three_sub
+  %half_scale = fmul float 0.5, %is_mul
+  ret float %half_scale
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; sqrt
+
+declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone
+
+
+define float @__sqrt_uniform_float(float) nounwind readonly alwaysinline {
+  sse_unary_scalar(ret, 4, float, @llvm.x86.sse.sqrt.ss, %0)
+  ret float %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; fast math mode
+
+declare void @llvm.x86.sse.stmxcsr(i8 *) nounwind
+declare void @llvm.x86.sse.ldmxcsr(i8 *) nounwind
+
+define void @__fastmath() nounwind alwaysinline {
+  %ptr = alloca i32
+  %ptr8 = bitcast i32 * %ptr to i8 *
+  call void @llvm.x86.sse.stmxcsr(i8 * %ptr8)
+  %oldval = load i32 *%ptr
+
+  ; turn on DAZ (64)/FTZ (32768) -> 32832
+  %update = or i32 %oldval, 32832
+  store i32 %update, i32 *%ptr
+  call void @llvm.x86.sse.ldmxcsr(i8 * %ptr8)
+  ret void
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <4 x float> @llvm.x86.sse.max.ss(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @llvm.x86.sse.min.ss(<4 x float>, <4 x float>) nounwind readnone
+
+define float @__max_uniform_float(float, float) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.max.ss, %0, %1)
+  ret float %ret
+}
+
+
+define float @__min_uniform_float(float, float) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.min.ss, %0, %1)
+  ret float %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double>) nounwind readnone
+
+define double @__sqrt_uniform_double(double) nounwind alwaysinline {
+  sse_unary_scalar(ret, 2, double, @llvm.x86.sse2.sqrt.sd, %0)
+  ret double %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <2 x double> @llvm.x86.sse2.max.sd(<2 x double>, <2 x double>) nounwind readnone
+declare <2 x double> @llvm.x86.sse2.min.sd(<2 x double>, <2 x double>) nounwind readnone
+
+define double @__min_uniform_double(double, double) nounwind readnone {
+  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.min.sd, %0, %1)
+  ret double %ret
+}
+
+define double @__max_uniform_double(double, double) nounwind readnone {
+  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.max.sd, %0, %1)
+  ret double %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding
+;;
+;; There are not any rounding instructions in SSE2, so we have to emulate
+;; the functionality with multiple instructions...
+
+; The code for __round_* is the result of compiling the following source
+; code.
+;
+; export float Round(float x) {
+;    unsigned int sign = signbits(x);
+;    unsigned int ix = intbits(x);
+;    ix ^= sign;
+;    x = floatbits(ix);
+;    x += 0x1.0p23f;
+;    x -= 0x1.0p23f;
+;    ix = intbits(x);
+;    ix ^= sign;
+;    x = floatbits(ix);
+;    return x;
+;}
+
+define float @__round_uniform_float(float) nounwind readonly alwaysinline {
+  %float_to_int_bitcast.i.i.i.i = bitcast float %0 to i32
+  %bitop.i.i = and i32 %float_to_int_bitcast.i.i.i.i, -2147483648
+  %bitop.i = xor i32 %bitop.i.i, %float_to_int_bitcast.i.i.i.i
+  %int_to_float_bitcast.i.i40.i = bitcast i32 %bitop.i to float
+  %binop.i = fadd float %int_to_float_bitcast.i.i40.i, 8.388608e+06
+  %binop21.i = fadd float %binop.i, -8.388608e+06
+  %float_to_int_bitcast.i.i.i = bitcast float %binop21.i to i32
+  %bitop31.i = xor i32 %float_to_int_bitcast.i.i.i, %bitop.i.i
+  %int_to_float_bitcast.i.i.i = bitcast i32 %bitop31.i to float
+  ret float %int_to_float_bitcast.i.i.i
+}
+
+;; Similarly, for implementations of the __floor* functions below, we have the
+;; bitcode from compiling the following source code...
+
+;export float Floor(float x) {
+;    float y = Round(x);
+;    unsigned int cmp = y > x ? 0xffffffff : 0;
+;    float delta = -1.f;
+;    unsigned int idelta = intbits(delta);
+;    idelta &= cmp;
+;    delta = floatbits(idelta);
+;    return y + delta;
+;}
+
+define float @__floor_uniform_float(float) nounwind readonly alwaysinline {
+  %calltmp.i = tail call float @__round_uniform_float(float %0) nounwind
+  %bincmp.i = fcmp ogt float %calltmp.i, %0
+  %selectexpr.i = sext i1 %bincmp.i to i32
+  %bitop.i = and i32 %selectexpr.i, -1082130432
+  %int_to_float_bitcast.i.i.i = bitcast i32 %bitop.i to float
+  %binop.i = fadd float %calltmp.i, %int_to_float_bitcast.i.i.i
+  ret float %binop.i
+}
+
+;; And here is the code we compiled to get the __ceil* functions below
+;
+;export uniform float Ceil(uniform float x) {
+;    uniform float y = Round(x);
+;    uniform int yltx = y < x ? 0xffffffff : 0;
+;    uniform float delta = 1.f;
+;    uniform int idelta = intbits(delta);
+;    idelta &= yltx;
+;    delta = floatbits(idelta);
+;    return y + delta;
+;}
+
+define float @__ceil_uniform_float(float) nounwind readonly alwaysinline {
+  %calltmp.i = tail call float @__round_uniform_float(float %0) nounwind
+  %bincmp.i = fcmp olt float %calltmp.i, %0
+  %selectexpr.i = sext i1 %bincmp.i to i32
+  %bitop.i = and i32 %selectexpr.i, 1065353216
+  %int_to_float_bitcast.i.i.i = bitcast i32 %bitop.i to float
+  %binop.i = fadd float %calltmp.i, %int_to_float_bitcast.i.i.i
+  ret float %binop.i
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare double @round(double)
+declare double @floor(double)
+declare double @ceil(double)
+
+define double @__round_uniform_double(double) nounwind readonly alwaysinline {
+  %r = call double @round(double %0)
+  ret double %r
+}
+
+define double @__floor_uniform_double(double) nounwind readonly alwaysinline {
+  %r = call double @floor(double %0)
+  ret double %r
+}
+
+define double @__ceil_uniform_double(double) nounwind readonly alwaysinline {
+  %r = call double @ceil(double %0)
+  ret double %r
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops / reductions
+
+declare i32 @llvm.ctpop.i32(i32)
+declare i64 @llvm.ctpop.i64(i64)
+
+define i32 @__popcnt_int32(i32) nounwind readonly alwaysinline {
+  %val = call i32 @llvm.ctpop.i32(i32 %0)
+  ret i32 %val
+}
+
+define i64 @__popcnt_int64(i64) nounwind readnone alwaysinline {
+  %val = call i64 @llvm.ctpop.i64(i64 %0)
+  ret i64 %val
+}
+
+

builtins-sse2-x2.ll

@@ -0,0 +1,631 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+
+;; This file defines the target for "double-pumped" SSE2, i.e. running
+;; with 8-wide vectors
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; standard 8-wide definitions from m4 macros
+
+stdlib_core(8)
+packed_load_and_store(8)
+scans(8)
+int64minmax(8)
+
+include(`builtins-sse2-common.ll')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ps(<4 x float>) nounwind readnone
+
+define <8 x float> @__rcp_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ;  float iv = __rcp_v(v);
+  ;  return iv * (2. - v * iv);
+
+  unary4to8(call, float, @llvm.x86.sse.rcp.ps, %0)
+  ; do one N-R iteration
+  %v_iv = fmul <8 x float> %0, %call
+  %two_minus = fsub <8 x float> <float 2., float 2., float 2., float 2.,
+                                 float 2., float 2., float 2., float 2.>, %v_iv  
+  %iv_mul = fmul <8 x float> %call, %two_minus
+  ret <8 x float> %iv_mul
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rsqrt
+
+declare <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float>) nounwind readnone
+
+define <8 x float> @__rsqrt_varying_float(<8 x float> %v) nounwind readonly alwaysinline {
+  ;  float is = __rsqrt_v(v);
+  unary4to8(is, float, @llvm.x86.sse.rsqrt.ps, %v)
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul <8 x float> %v, %is
+  %v_is_is = fmul <8 x float> %v_is, %is
+  %three_sub = fsub <8 x float> <float 3., float 3., float 3., float 3.,
+                                 float 3., float 3., float 3., float 3.>, %v_is_is
+  %is_mul = fmul <8 x float> %is, %three_sub
+  %half_scale = fmul <8 x float> <float 0.5, float 0.5, float 0.5, float 0.5,
+                                  float 0.5, float 0.5, float 0.5, float 0.5>, %is_mul
+  ret <8 x float> %half_scale
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; sqrt
+
+declare <4 x float> @llvm.x86.sse.sqrt.ps(<4 x float>) nounwind readnone
+
+define <8 x float> @__sqrt_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  unary4to8(call, float, @llvm.x86.sse.sqrt.ps, %0)
+  ret <8 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; svml stuff
+
+declare <4 x float> @__svml_sinf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_cosf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_sincosf4(<4 x float> *, <4 x float>) nounwind readnone
+declare <4 x float> @__svml_tanf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_atanf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_atan2f4(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @__svml_expf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_logf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_powf4(<4 x float>, <4 x float>) nounwind readnone
+
+
+define <8 x float> @__svml_sin(<8 x float>) nounwind readnone alwaysinline {
+  unary4to8(ret, float, @__svml_sinf4, %0)
+  ret <8 x float> %ret
+}
+
+define <8 x float> @__svml_cos(<8 x float>) nounwind readnone alwaysinline {
+  unary4to8(ret, float, @__svml_cosf4, %0)
+  ret <8 x float> %ret
+}
+
+define void @__svml_sincos(<8 x float>, <8 x float> *,
+                                    <8 x float> *) nounwind readnone alwaysinline {
+  ; call svml_sincosf4 two times with the two 4-wide sub-vectors
+  %a = shufflevector <8 x float> %0, <8 x float> undef,
+         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %b = shufflevector <8 x float> %0, <8 x float> undef,
+         <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+
+  %cospa = alloca <4 x float>
+  %sa = call <4 x float> @__svml_sincosf4(<4 x float> * %cospa, <4 x float> %a)
+
+  %cospb = alloca <4 x float>
+  %sb = call <4 x float> @__svml_sincosf4(<4 x float> * %cospb, <4 x float> %b)
+
+  %sin = shufflevector <4 x float> %sa, <4 x float> %sb,
+         <8 x i32> <i32 0, i32 1, i32 2, i32 3,
+                    i32 4, i32 5, i32 6, i32 7>
+  store <8 x float> %sin, <8 x float> * %1
+
+  %cosa = load <4 x float> * %cospa
+  %cosb = load <4 x float> * %cospb
+  %cos = shufflevector <4 x float> %cosa, <4 x float> %cosb,
+         <8 x i32> <i32 0, i32 1, i32 2, i32 3,
+                    i32 4, i32 5, i32 6, i32 7>
+  store <8 x float> %cos, <8 x float> * %2
+
+  ret void
+}
+
+define <8 x float> @__svml_tan(<8 x float>) nounwind readnone alwaysinline {
+  unary4to8(ret, float, @__svml_tanf4, %0)
+  ret <8 x float> %ret
+}
+
+define <8 x float> @__svml_atan(<8 x float>) nounwind readnone alwaysinline {
+  unary4to8(ret, float, @__svml_atanf4, %0)
+  ret <8 x float> %ret
+}
+
+define <8 x float> @__svml_atan2(<8 x float>,
+                                          <8 x float>) nounwind readnone alwaysinline {
+  binary4to8(ret, float, @__svml_atan2f4, %0, %1)
+  ret <8 x float> %ret
+}
+
+define <8 x float> @__svml_exp(<8 x float>) nounwind readnone alwaysinline {
+  unary4to8(ret, float, @__svml_expf4, %0)
+  ret <8 x float> %ret
+}
+
+define <8 x float> @__svml_log(<8 x float>) nounwind readnone alwaysinline {
+  unary4to8(ret, float, @__svml_logf4, %0)
+  ret <8 x float> %ret
+}
+
+define <8 x float> @__svml_pow(<8 x float>,
+                                        <8 x float>) nounwind readnone alwaysinline {
+  binary4to8(ret, float, @__svml_powf4, %0, %1)
+  ret <8 x float> %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <4 x float> @llvm.x86.sse.max.ps(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @llvm.x86.sse.min.ps(<4 x float>, <4 x float>) nounwind readnone
+
+define <8 x float> @__max_varying_float(<8 x float>, <8 x float>) nounwind readonly alwaysinline {
+  binary4to8(call, float, @llvm.x86.sse.max.ps, %0, %1)
+  ret <8 x float> %call
+}
+
+define <8 x float> @__min_varying_float(<8 x float>, <8 x float>) nounwind readonly alwaysinline {
+  binary4to8(call, float, @llvm.x86.sse.min.ps, %0, %1)
+  ret <8 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; min/max
+
+; There is no blend instruction with SSE2, so we simulate it with bit
+; operations on i32s.  For these two vselect functions, for each
+; vector element, if the mask is on, we return the corresponding value
+; from %1, and otherwise return the value from %0.
+
+define <8 x i32> @__vselect_i32(<8 x i32>, <8 x i32> ,
+                                         <8 x i32> %mask) nounwind readnone alwaysinline {
+  %notmask = xor <8 x i32> %mask, <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1>
+  %cleared_old = and <8 x i32> %0, %notmask
+  %masked_new = and <8 x i32> %1, %mask
+  %new = or <8 x i32> %cleared_old, %masked_new
+  ret <8 x i32> %new
+}
+
+define <8 x float> @__vselect_float(<8 x float>, <8 x float>,
+                                             <8 x i32> %mask) nounwind readnone alwaysinline {
+  %v0 = bitcast <8 x float> %0 to <8 x i32>
+  %v1 = bitcast <8 x float> %1 to <8 x i32>
+  %r = call <8 x i32> @__vselect_i32(<8 x i32> %v0, <8 x i32> %v1, <8 x i32> %mask)
+  %rf = bitcast <8 x i32> %r to <8 x float>
+  ret <8 x float> %rf
+}
+
+
+; To do vector integer min and max, we do the vector compare and then sign
+; extend the i1 vector result to an i32 mask.  The __vselect does the
+; rest...
+
+define <8 x i32> @__min_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  %c = icmp slt <8 x i32> %0, %1
+  %mask = sext <8 x i1> %c to <8 x i32>
+  %v = call <8 x i32> @__vselect_i32(<8 x i32> %1, <8 x i32> %0, <8 x i32> %mask)
+  ret <8 x i32> %v
+}
+
+define i32 @__min_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp slt i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+define <8 x i32> @__max_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  %c = icmp sgt <8 x i32> %0, %1
+  %mask = sext <8 x i1> %c to <8 x i32>
+  %v = call <8 x i32> @__vselect_i32(<8 x i32> %1, <8 x i32> %0, <8 x i32> %mask)
+  ret <8 x i32> %v
+}
+
+define i32 @__max_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp sgt i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+; The functions for unsigned ints are similar, just with unsigned
+; comparison functions...
+
+define <8 x i32> @__min_varying_uint32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  %c = icmp ult <8 x i32> %0, %1
+  %mask = sext <8 x i1> %c to <8 x i32>
+  %v = call <8 x i32> @__vselect_i32(<8 x i32> %1, <8 x i32> %0, <8 x i32> %mask)
+  ret <8 x i32> %v
+}
+
+define i32 @__min_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp ult i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+define <8 x i32> @__max_varying_uint32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+  %c = icmp ugt <8 x i32> %0, %1
+  %mask = sext <8 x i1> %c to <8 x i32>
+  %v = call <8 x i32> @__vselect_i32(<8 x i32> %1, <8 x i32> %0, <8 x i32> %mask)
+  ret <8 x i32> %v
+}
+
+define i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp ugt i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops / reductions
+
+declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) nounwind readnone
+
+define i32 @__movmsk(<8 x i32>) nounwind readnone alwaysinline {
+  ; first do two 4-wide movmsk calls
+  %floatmask = bitcast <8 x i32> %0 to <8 x float>
+  %m0 = shufflevector <8 x float> %floatmask, <8 x float> undef,
+          <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %v0 = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %m0) nounwind readnone
+  %m1 = shufflevector <8 x float> %floatmask, <8 x float> undef,
+          <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %v1 = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %m1) nounwind readnone
+
+  ; and shift the first one over by 4 before ORing it with the value 
+  ; of the second one
+  %v1s = shl i32 %v1, 4
+  %v = or i32 %v0, %v1s
+  ret i32 %v
+}
+
+define <4 x float> @__vec4_add_float(<4 x float> %v0,
+                                            <4 x float> %v1) nounwind readnone alwaysinline {
+  %v = fadd <4 x float> %v0, %v1
+  ret <4 x float> %v
+}
+
+define float @__add_float(float, float) nounwind readnone alwaysinline {
+  %v = fadd float %0, %1
+  ret float %v
+}
+
+define float @__reduce_add_float(<8 x float>) nounwind readnone alwaysinline {
+  reduce8by4(float, @__vec4_add_float, @__add_float)
+}
+
+define float @__reduce_min_float(<8 x float>) nounwind readnone alwaysinline {
+  reduce8(float, @__min_varying_float, @__min_uniform_float)
+}
+
+define float @__reduce_max_float(<8 x float>) nounwind readnone alwaysinline {
+  reduce8(float, @__max_varying_float, @__max_uniform_float)
+}
+
+; helper function for reduce_add_int32
+define <4 x i32> @__vec4_add_int32(<4 x i32> %v0,
+                                            <4 x i32> %v1) nounwind readnone alwaysinline {
+  %v = add <4 x i32> %v0, %v1
+  ret <4 x i32> %v
+}
+
+; helper function for reduce_add_int32
+define i32 @__add_int32(i32, i32) nounwind readnone alwaysinline {
+  %v = add i32 %0, %1
+  ret i32 %v
+}
+
+define i32 @__reduce_add_int32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8by4(i32, @__vec4_add_int32, @__add_int32)
+}
+
+define i32 @__reduce_min_int32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__min_varying_int32, @__min_uniform_int32)
+}
+
+define i32 @__reduce_max_int32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__max_varying_int32, @__max_uniform_int32)
+}
+
+define i32 @__reduce_add_uint32(<8 x i32> %v) nounwind readnone alwaysinline {
+  %r = call i32 @__reduce_add_int32(<8 x i32> %v)
+  ret i32 %r
+}
+
+define i32 @__reduce_min_uint32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__min_varying_uint32, @__min_uniform_uint32)
+}
+
+define i32 @__reduce_max_uint32(<8 x i32>) nounwind readnone alwaysinline {
+  reduce8(i32, @__max_varying_uint32, @__max_uniform_uint32)
+}
+
+define <4 x double> @__add_varying_double(<4 x double>,
+                                     <4 x double>) nounwind readnone alwaysinline {
+  %r = fadd <4 x double> %0, %1
+  ret <4 x double> %r
+}
+
+define double @__add_uniform_double(double, double) nounwind readnone alwaysinline {
+  %r = fadd double %0, %1
+  ret double %r
+}
+
+define double @__reduce_add_double(<8 x double>) nounwind readnone {
+  reduce8by4(double, @__add_varying_double, @__add_uniform_double)
+}
+
+define double @__reduce_min_double(<8 x double>) nounwind readnone {
+  reduce8(double, @__min_varying_double, @__min_uniform_double)
+}
+
+define double @__reduce_max_double(<8 x double>) nounwind readnone {
+  reduce8(double, @__max_varying_double, @__max_uniform_double)
+}
+
+define <4 x i64> @__add_varying_int64(<4 x i64>,
+                                               <4 x i64>) nounwind readnone alwaysinline {
+  %r = add <4 x i64> %0, %1
+  ret <4 x i64> %r
+}
+
+define i64 @__add_uniform_int64(i64, i64) nounwind readnone alwaysinline {
+  %r = add i64 %0, %1
+  ret i64 %r
+}
+
+define i64 @__reduce_add_int64(<8 x i64>) nounwind readnone {
+  reduce8by4(i64, @__add_varying_int64, @__add_uniform_int64)
+}
+
+define i64 @__reduce_min_int64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+define i64 @__reduce_max_int64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+define i64 @__reduce_min_uint64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+define i64 @__reduce_max_uint64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__max_varying_uint64, @__max_uniform_uint64)
+}
+
+reduce_equal(8)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unaligned loads/loads+broadcasts
+
+load_and_broadcast(8, i8, 8)
+load_and_broadcast(8, i16, 16)
+load_and_broadcast(8, i32, 32)
+load_and_broadcast(8, i64, 64)
+
+load_masked(8, i8,  8,  1)
+load_masked(8, i16, 16, 2)
+load_masked(8, i32, 32, 4)
+load_masked(8, i64, 64, 8)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; gather/scatter
+
+gen_gather(8, i8)
+gen_gather(8, i16)
+gen_gather(8, i32)
+gen_gather(8, i64)
+
+gen_scatter(8, i8)
+gen_scatter(8, i16)
+gen_scatter(8, i32)
+gen_scatter(8, i64)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float rounding
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding
+;;
+;; There are not any rounding instructions in SSE2, so we have to emulate
+;; the functionality with multiple instructions...
+
+; The code for __round_* is the result of compiling the following source
+; code.
+;
+; export float Round(float x) {
+;    unsigned int sign = signbits(x);
+;    unsigned int ix = intbits(x);
+;    ix ^= sign;
+;    x = floatbits(ix);
+;    x += 0x1.0p23f;
+;    x -= 0x1.0p23f;
+;    ix = intbits(x);
+;    ix ^= sign;
+;    x = floatbits(ix);
+;    return x;
+;}
+
+define <8 x float> @__round_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  %float_to_int_bitcast.i.i.i.i = bitcast <8 x float> %0 to <8 x i32>
+  %bitop.i.i = and <8 x i32> %float_to_int_bitcast.i.i.i.i, <i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648>
+  %bitop.i = xor <8 x i32> %float_to_int_bitcast.i.i.i.i, %bitop.i.i
+  %int_to_float_bitcast.i.i40.i = bitcast <8 x i32> %bitop.i to <8 x float>
+  %binop.i = fadd <8 x float> %int_to_float_bitcast.i.i40.i, <float 8.388608e+06, float 8.388608e+06, float 8.388608e+06, float 8.388608e+06, float 8.388608e+06, float 8.388608e+06, float 8.388608e+06, float 8.388608e+06>
+  %binop21.i = fadd <8 x float> %binop.i, <float -8.388608e+06, float -8.388608e+06, float -8.388608e+06, float -8.388608e+06, float -8.388608e+06, float -8.388608e+06, float -8.388608e+06, float -8.388608e+06>
+  %float_to_int_bitcast.i.i.i = bitcast <8 x float> %binop21.i to <8 x i32>
+  %bitop31.i = xor <8 x i32> %float_to_int_bitcast.i.i.i, %bitop.i.i
+  %int_to_float_bitcast.i.i.i = bitcast <8 x i32> %bitop31.i to <8 x float>
+  ret <8 x float> %int_to_float_bitcast.i.i.i
+}
+
+;; Similarly, for implementations of the __floor* functions below, we have the
+;; bitcode from compiling the following source code...
+
+;export float Floor(float x) {
+;    float y = Round(x);
+;    unsigned int cmp = y > x ? 0xffffffff : 0;
+;    float delta = -1.f;
+;    unsigned int idelta = intbits(delta);
+;    idelta &= cmp;
+;    delta = floatbits(idelta);
+;    return y + delta;
+;}
+
+define <8 x float> @__floor_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  %calltmp.i = tail call <8 x float> @__round_varying_float(<8 x float> %0) nounwind
+  %bincmp.i = fcmp ogt <8 x float> %calltmp.i, %0
+  %val_to_boolvec32.i = sext <8 x i1> %bincmp.i to <8 x i32>
+  %bitop.i = and <8 x i32> %val_to_boolvec32.i, <i32 -1082130432, i32 -1082130432, i32 -1082130432, i32 -1082130432, i32 -1082130432, i32 -1082130432, i32 -1082130432, i32 -1082130432>
+  %int_to_float_bitcast.i.i.i = bitcast <8 x i32> %bitop.i to <8 x float>
+  %binop.i = fadd <8 x float> %calltmp.i, %int_to_float_bitcast.i.i.i
+  ret <8 x float> %binop.i
+}
+
+;; And here is the code we compiled to get the __ceil* functions below
+;
+;export uniform float Ceil(uniform float x) {
+;    uniform float y = Round(x);
+;    uniform int yltx = y < x ? 0xffffffff : 0;
+;    uniform float delta = 1.f;
+;    uniform int idelta = intbits(delta);
+;    idelta &= yltx;
+;    delta = floatbits(idelta);
+;    return y + delta;
+;}
+
+define <8 x float> @__ceil_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  %calltmp.i = tail call <8 x float> @__round_varying_float(<8 x float> %0) nounwind
+  %bincmp.i = fcmp olt <8 x float> %calltmp.i, %0
+  %val_to_boolvec32.i = sext <8 x i1> %bincmp.i to <8 x i32>
+  %bitop.i = and <8 x i32> %val_to_boolvec32.i, <i32 1065353216, i32 1065353216, i32 1065353216, i32 1065353216, i32 1065353216, i32 1065353216, i32 1065353216, i32 1065353216>
+  %int_to_float_bitcast.i.i.i = bitcast <8 x i32> %bitop.i to <8 x float>
+  %binop.i = fadd <8 x float> %calltmp.i, %int_to_float_bitcast.i.i.i
+  ret <8 x float> %binop.i
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+define <8 x double> @__round_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  unary1to8(double, @round)
+}
+
+define <8 x double> @__floor_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  unary1to8(double, @floor)
+}
+
+define <8 x double> @__ceil_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  unary1to8(double, @ceil)
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+gen_masked_store(8, i8, 8)
+gen_masked_store(8, i16, 16)
+gen_masked_store(8, i32, 32)
+gen_masked_store(8, i64, 64)
+
+masked_store_blend_8_16_by_8()
+
+define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>, 
+                                     <8 x i32> %mask) nounwind alwaysinline {
+  %val = load <8 x i32> * %0, align 4
+  %newval = call <8 x i32> @__vselect_i32(<8 x i32> %val, <8 x i32> %1, <8 x i32> %mask) 
+  store <8 x i32> %newval, <8 x i32> * %0, align 4
+  ret void
+}
+
+define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
+                                     <8 x i32> %mask) nounwind alwaysinline {
+  %oldValue = load <8 x i64>* %ptr, align 8
+
+  ; Do 8x64-bit blends by doing two <8 x i32> blends, where the <8 x i32> values
+  ; are actually bitcast <2 x i64> values
+  ;
+  ; set up the first two 64-bit values
+  %old0123  = shufflevector <8 x i64> %oldValue, <8 x i64> undef,
+                            <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %old0123f = bitcast <4 x i64> %old0123 to <8 x float>
+  %new0123  = shufflevector <8 x i64> %new, <8 x i64> undef,
+                            <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %new0123f = bitcast <4 x i64> %new0123 to <8 x float>
+  ; compute mask--note that the indices are doubled-up
+  %mask0123 = shufflevector <8 x i32> %mask, <8 x i32> undef,
+              <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  ; and blend the first 4 values
+  %result0123f = call <8 x float> @__vselect_float(<8 x float> %old0123f, <8 x float> %new0123f,
+                                                   <8 x i32> %mask0123)
+  %result0123 = bitcast <8 x float> %result0123f to <4 x i64>
+
+  ; and again
+  %old4567  = shufflevector <8 x i64> %oldValue, <8 x i64> undef,
+                            <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %old4567f = bitcast <4 x i64> %old4567 to <8 x float>
+  %new4567  = shufflevector <8 x i64> %new, <8 x i64> undef,
+                            <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %new4567f = bitcast <4 x i64> %new4567 to <8 x float>
+  ; compute mask--note that the values are doubled-up
+  %mask4567 = shufflevector <8 x i32> %mask, <8 x i32> undef,
+              <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+  ; and blend the two of the values
+  %result4567f = call <8 x float> @__vselect_float(<8 x float> %old4567f, <8 x float> %new4567f,
+                                                   <8 x i32> %mask4567)
+  %result4567 = bitcast <8 x float> %result4567f to <4 x i64>
+
+  ; reconstruct the final <8 x i64> vector
+  %final = shufflevector <4 x i64> %result0123, <4 x i64> %result4567,
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  store <8 x i64> %final, <8 x i64> * %ptr, align 8
+  ret void
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <2 x double> @llvm.x86.sse2.sqrt.pd(<2 x double>) nounwind readnone
+
+define <8 x double> @__sqrt_varying_double(<8 x double>) nounwind alwaysinline {
+  unary2to8(ret, double, @llvm.x86.sse2.sqrt.pd, %0)
+  ret <8 x double> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision float min/max
+
+declare <2 x double> @llvm.x86.sse2.max.pd(<2 x double>, <2 x double>) nounwind readnone
+declare <2 x double> @llvm.x86.sse2.min.pd(<2 x double>, <2 x double>) nounwind readnone
+
+define <8 x double> @__min_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
+  binary2to8(ret, double, @llvm.x86.sse2.min.pd, %0, %1)
+  ret <8 x double> %ret
+}
+
+define <8 x double> @__max_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
+  binary2to8(ret, double, @llvm.x86.sse2.max.pd, %0, %1)
+  ret <8 x double> %ret
+}

builtins-sse2.ll

@@ -0,0 +1,573 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; Define the standard library builtins for the SSE2 target
+
+; Define some basics for a 4-wide target
+stdlib_core(4)
+packed_load_and_store(4)
+scans(4)
+int64minmax(4)
+
+include(`builtins-sse2-common.ll')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding
+;;
+;; There are not any rounding instructions in SSE2, so we have to emulate
+;; the functionality with multiple instructions...
+
+; The code for __round_* is the result of compiling the following source
+; code.
+;
+; export float Round(float x) {
+;    unsigned int sign = signbits(x);
+;    unsigned int ix = intbits(x);
+;    ix ^= sign;
+;    x = floatbits(ix);
+;    x += 0x1.0p23f;
+;    x -= 0x1.0p23f;
+;    ix = intbits(x);
+;    ix ^= sign;
+;    x = floatbits(ix);
+;    return x;
+;}
+
+define <4 x float> @__round_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %float_to_int_bitcast.i.i.i.i = bitcast <4 x float> %0 to <4 x i32>
+  %bitop.i.i = and <4 x i32> %float_to_int_bitcast.i.i.i.i, <i32 -2147483648, i32 -2147483648, i32 -2147483648, i32 -2147483648>
+  %bitop.i = xor <4 x i32> %float_to_int_bitcast.i.i.i.i, %bitop.i.i
+  %int_to_float_bitcast.i.i40.i = bitcast <4 x i32> %bitop.i to <4 x float>
+  %binop.i = fadd <4 x float> %int_to_float_bitcast.i.i40.i, <float 8.388608e+06, float 8.388608e+06, float 8.388608e+06, float 8.388608e+06>
+  %binop21.i = fadd <4 x float> %binop.i, <float -8.388608e+06, float -8.388608e+06, float -8.388608e+06, float -8.388608e+06>
+  %float_to_int_bitcast.i.i.i = bitcast <4 x float> %binop21.i to <4 x i32>
+  %bitop31.i = xor <4 x i32> %float_to_int_bitcast.i.i.i, %bitop.i.i
+  %int_to_float_bitcast.i.i.i = bitcast <4 x i32> %bitop31.i to <4 x float>
+  ret <4 x float> %int_to_float_bitcast.i.i.i
+}
+
+;; Similarly, for implementations of the __floor* functions below, we have the
+;; bitcode from compiling the following source code...
+
+;export float Floor(float x) {
+;    float y = Round(x);
+;    unsigned int cmp = y > x ? 0xffffffff : 0;
+;    float delta = -1.f;
+;    unsigned int idelta = intbits(delta);
+;    idelta &= cmp;
+;    delta = floatbits(idelta);
+;    return y + delta;
+;}
+
+define <4 x float> @__floor_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %calltmp.i = tail call <4 x float> @__round_varying_float(<4 x float> %0) nounwind
+  %bincmp.i = fcmp ogt <4 x float> %calltmp.i, %0
+  %val_to_boolvec32.i = sext <4 x i1> %bincmp.i to <4 x i32>
+  %bitop.i = and <4 x i32> %val_to_boolvec32.i, <i32 -1082130432, i32 -1082130432, i32 -1082130432, i32 -1082130432>
+  %int_to_float_bitcast.i.i.i = bitcast <4 x i32> %bitop.i to <4 x float>
+  %binop.i = fadd <4 x float> %calltmp.i, %int_to_float_bitcast.i.i.i
+  ret <4 x float> %binop.i
+}
+
+;; And here is the code we compiled to get the __ceil* functions below
+;
+;export uniform float Ceil(uniform float x) {
+;    uniform float y = Round(x);
+;    uniform int yltx = y < x ? 0xffffffff : 0;
+;    uniform float delta = 1.f;
+;    uniform int idelta = intbits(delta);
+;    idelta &= yltx;
+;    delta = floatbits(idelta);
+;    return y + delta;
+;}
+
+define <4 x float> @__ceil_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %calltmp.i = tail call <4 x float> @__round_varying_float(<4 x float> %0) nounwind
+  %bincmp.i = fcmp olt <4 x float> %calltmp.i, %0
+  %val_to_boolvec32.i = sext <4 x i1> %bincmp.i to <4 x i32>
+  %bitop.i = and <4 x i32> %val_to_boolvec32.i, <i32 1065353216, i32 1065353216, i32 1065353216, i32 1065353216>
+  %int_to_float_bitcast.i.i.i = bitcast <4 x i32> %bitop.i to <4 x float>
+  %binop.i = fadd <4 x float> %calltmp.i, %int_to_float_bitcast.i.i.i
+  ret <4 x float> %binop.i
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+define <4 x double> @__round_varying_double(<4 x double>) nounwind readonly alwaysinline {
+  unary1to4(double, @round)
+}
+
+define <4 x double> @__floor_varying_double(<4 x double>) nounwind readonly alwaysinline {
+  unary1to4(double, @floor)
+}
+
+define <4 x double> @__ceil_varying_double(<4 x double>) nounwind readonly alwaysinline {
+  unary1to4(double, @ceil)
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; min/max
+
+; There is no blend instruction with SSE2, so we simulate it with bit
+; operations on i32s.  For these two vselect functions, for each
+; vector element, if the mask is on, we return the corresponding value
+; from %1, and otherwise return the value from %0.
+
+define <4 x i32> @__vselect_i32(<4 x i32>, <4 x i32> ,
+                                         <4 x i32> %mask) nounwind readnone alwaysinline {
+  %notmask = xor <4 x i32> %mask, <i32 -1, i32 -1, i32 -1, i32 -1>
+  %cleared_old = and <4 x i32> %0, %notmask
+  %masked_new = and <4 x i32> %1, %mask
+  %new = or <4 x i32> %cleared_old, %masked_new
+  ret <4 x i32> %new
+}
+
+define <4 x float> @__vselect_float(<4 x float>, <4 x float>,
+                                             <4 x i32> %mask) nounwind readnone alwaysinline {
+  %v0 = bitcast <4 x float> %0 to <4 x i32>
+  %v1 = bitcast <4 x float> %1 to <4 x i32>
+  %r = call <4 x i32> @__vselect_i32(<4 x i32> %v0, <4 x i32> %v1, <4 x i32> %mask)
+  %rf = bitcast <4 x i32> %r to <4 x float>
+  ret <4 x float> %rf
+}
+
+
+; To do vector integer min and max, we do the vector compare and then sign
+; extend the i1 vector result to an i32 mask.  The __vselect does the
+; rest...
+
+define <4 x i32> @__min_varying_int32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %c = icmp slt <4 x i32> %0, %1
+  %mask = sext <4 x i1> %c to <4 x i32>
+  %v = call <4 x i32> @__vselect_i32(<4 x i32> %1, <4 x i32> %0, <4 x i32> %mask)
+  ret <4 x i32> %v
+}
+
+define i32 @__min_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp slt i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+define <4 x i32> @__max_varying_int32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %c = icmp sgt <4 x i32> %0, %1
+  %mask = sext <4 x i1> %c to <4 x i32>
+  %v = call <4 x i32> @__vselect_i32(<4 x i32> %1, <4 x i32> %0, <4 x i32> %mask)
+  ret <4 x i32> %v
+}
+
+define i32 @__max_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp sgt i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+; The functions for unsigned ints are similar, just with unsigned
+; comparison functions...
+
+define <4 x i32> @__min_varying_uint32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %c = icmp ult <4 x i32> %0, %1
+  %mask = sext <4 x i1> %c to <4 x i32>
+  %v = call <4 x i32> @__vselect_i32(<4 x i32> %1, <4 x i32> %0, <4 x i32> %mask)
+  ret <4 x i32> %v
+}
+
+define i32 @__min_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp ult i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+define <4 x i32> @__max_varying_uint32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %c = icmp ugt <4 x i32> %0, %1
+  %mask = sext <4 x i1> %c to <4 x i32>
+  %v = call <4 x i32> @__vselect_i32(<4 x i32> %1, <4 x i32> %0, <4 x i32> %mask)
+  ret <4 x i32> %v
+}
+
+define i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  %c = icmp ugt i32 %0, %1
+  %r = select i1 %c, i32 %0, i32 %1
+  ret i32 %r
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops / reductions
+
+declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) nounwind readnone
+
+define i32 @__movmsk(<4 x i32>) nounwind readnone alwaysinline {
+  %floatmask = bitcast <4 x i32> %0 to <4 x float>
+  %v = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %floatmask) nounwind readnone
+  ret i32 %v
+}
+
+define float @__reduce_add_float(<4 x float> %v) nounwind readonly alwaysinline {
+  %v1 = shufflevector <4 x float> %v, <4 x float> undef,
+                      <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %m1 = fadd <4 x float> %v1, %v
+  %m1a = extractelement <4 x float> %m1, i32 0
+  %m1b = extractelement <4 x float> %m1, i32 1
+  %sum = fadd float %m1a, %m1b
+  ret float %sum
+}
+
+define float @__reduce_min_float(<4 x float>) nounwind readnone {
+  reduce4(float, @__min_varying_float, @__min_uniform_float)
+}
+
+define float @__reduce_max_float(<4 x float>) nounwind readnone {
+  reduce4(float, @__max_varying_float, @__max_uniform_float)
+}
+
+define i32 @__reduce_add_int32(<4 x i32> %v) nounwind readnone {
+  %v1 = shufflevector <4 x i32> %v, <4 x i32> undef,
+                      <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %m1 = add <4 x i32> %v1, %v
+  %m1a = extractelement <4 x i32> %m1, i32 0
+  %m1b = extractelement <4 x i32> %m1, i32 1
+  %sum = add i32 %m1a, %m1b
+  ret i32 %sum
+}
+
+define i32 @__reduce_min_int32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__min_varying_int32, @__min_uniform_int32)
+}
+
+define i32 @__reduce_max_int32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__max_varying_int32, @__max_uniform_int32)
+}
+
+define i32 @__reduce_add_uint32(<4 x i32> %v) nounwind readnone {
+  %r = call i32 @__reduce_add_int32(<4 x i32> %v)
+  ret i32 %r
+}
+
+define i32 @__reduce_min_uint32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__min_varying_uint32, @__min_uniform_uint32)
+}
+
+define i32 @__reduce_max_uint32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__max_varying_uint32, @__max_uniform_uint32)
+ }
+
+
+define double @__reduce_add_double(<4 x double>) nounwind readnone {
+  %v0 = shufflevector <4 x double> %0, <4 x double> undef,
+                      <2 x i32> <i32 0, i32 1>
+  %v1 = shufflevector <4 x double> %0, <4 x double> undef,
+                      <2 x i32> <i32 2, i32 3>
+  %sum = fadd <2 x double> %v0, %v1
+  %e0 = extractelement <2 x double> %sum, i32 0
+  %e1 = extractelement <2 x double> %sum, i32 1
+  %m = fadd double %e0, %e1
+  ret double %m
+}
+
+define double @__reduce_min_double(<4 x double>) nounwind readnone {
+  reduce4(double, @__min_varying_double, @__min_uniform_double)
+}
+
+define double @__reduce_max_double(<4 x double>) nounwind readnone {
+  reduce4(double, @__max_varying_double, @__max_uniform_double)
+}
+
+define i64 @__reduce_add_int64(<4 x i64>) nounwind readnone {
+  %v0 = shufflevector <4 x i64> %0, <4 x i64> undef,
+                      <2 x i32> <i32 0, i32 1>
+  %v1 = shufflevector <4 x i64> %0, <4 x i64> undef,
+                      <2 x i32> <i32 2, i32 3>
+  %sum = add <2 x i64> %v0, %v1
+  %e0 = extractelement <2 x i64> %sum, i32 0
+  %e1 = extractelement <2 x i64> %sum, i32 1
+  %m = add i64 %e0, %e1
+  ret i64 %m
+}
+
+define i64 @__reduce_min_int64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+define i64 @__reduce_max_int64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+define i64 @__reduce_min_uint64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+define i64 @__reduce_max_uint64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__max_varying_uint64, @__max_uniform_uint64)
+}
+
+reduce_equal(4)
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+define void @__masked_store_blend_32(<4 x i32>* nocapture, <4 x i32>, 
+                                     <4 x i32> %mask) nounwind alwaysinline {
+  %val = load <4 x i32> * %0, align 4
+  %newval = call <4 x i32> @__vselect_i32(<4 x i32> %val, <4 x i32> %1, <4 x i32> %mask) 
+  store <4 x i32> %newval, <4 x i32> * %0, align 4
+  ret void
+}
+
+define void @__masked_store_blend_64(<4 x i64>* nocapture %ptr, <4 x i64> %new,
+                                     <4 x i32> %mask) nounwind alwaysinline {
+  %oldValue = load <4 x i64>* %ptr, align 8
+
+  ; Do 4x64-bit blends by doing two <4 x i32> blends, where the <4 x i32> values
+  ; are actually bitcast <2 x i64> values
+  ;
+  ; set up the first two 64-bit values
+  %old01  = shufflevector <4 x i64> %oldValue, <4 x i64> undef,
+                          <2 x i32> <i32 0, i32 1>
+  %old01f = bitcast <2 x i64> %old01 to <4 x float>
+  %new01  = shufflevector <4 x i64> %new, <4 x i64> undef,
+                          <2 x i32> <i32 0, i32 1>
+  %new01f = bitcast <2 x i64> %new01 to <4 x float>
+  ; compute mask--note that the indices 0 and 1 are doubled-up
+  %mask01 = shufflevector <4 x i32> %mask, <4 x i32> undef,
+                          <4 x i32> <i32 0, i32 0, i32 1, i32 1>
+  ; and blend the two of the values
+  %result01f = call <4 x float> @__vselect_float(<4 x float> %old01f, <4 x float> %new01f, <4 x i32> %mask01)
+  %result01 = bitcast <4 x float> %result01f to <2 x i64>
+
+  ; and again
+  %old23  = shufflevector <4 x i64> %oldValue, <4 x i64> undef,
+                          <2 x i32> <i32 2, i32 3>
+  %old23f = bitcast <2 x i64> %old23 to <4 x float>
+  %new23  = shufflevector <4 x i64> %new, <4 x i64> undef,
+                          <2 x i32> <i32 2, i32 3>
+  %new23f = bitcast <2 x i64> %new23 to <4 x float>
+  ; compute mask--note that the values 2 and 3 are doubled-up
+  %mask23 = shufflevector <4 x i32> %mask, <4 x i32> undef,
+                          <4 x i32> <i32 2, i32 2, i32 3, i32 3>
+  ; and blend the two of the values
+  %result23f = call <4 x float> @__vselect_float(<4 x float> %old23f, <4 x float> %new23f, <4 x i32> %mask23)
+  %result23 = bitcast <4 x float> %result23f to <2 x i64>
+
+  ; reconstruct the final <4 x i64> vector
+  %final = shufflevector <2 x i64> %result01, <2 x i64> %result23,
+                         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  store <4 x i64> %final, <4 x i64> * %ptr, align 8
+  ret void
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ps(<4 x float>) nounwind readnone
+
+define <4 x float> @__rcp_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %0)
+  ; do one N-R iteration to improve precision
+  ;  float iv = __rcp_v(v);
+  ;  return iv * (2. - v * iv);
+  %v_iv = fmul <4 x float> %0, %call
+  %two_minus = fsub <4 x float> <float 2., float 2., float 2., float 2.>, %v_iv  
+  %iv_mul = fmul <4 x float> %call, %two_minus
+  ret <4 x float> %iv_mul
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; rsqrt
+
+declare <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float>) nounwind readnone
+
+define <4 x float> @__rsqrt_varying_float(<4 x float> %v) nounwind readonly alwaysinline {
+  ;  float is = __rsqrt_v(v);
+  %is = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %v)
+  ; Newton-Raphson iteration to improve precision
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul <4 x float> %v, %is
+  %v_is_is = fmul <4 x float> %v_is, %is
+  %three_sub = fsub <4 x float> <float 3., float 3., float 3., float 3.>, %v_is_is
+  %is_mul = fmul <4 x float> %is, %three_sub
+  %half_scale = fmul <4 x float> <float 0.5, float 0.5, float 0.5, float 0.5>, %is_mul
+  ret <4 x float> %half_scale
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; sqrt
+
+declare <4 x float> @llvm.x86.sse.sqrt.ps(<4 x float>) nounwind readnone
+
+define <4 x float> @__sqrt_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.sqrt.ps(<4 x float> %0)
+  ret <4 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; svml stuff
+
+declare <4 x float> @__svml_sinf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_cosf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_sincosf4(<4 x float> *, <4 x float>) nounwind readnone
+declare <4 x float> @__svml_tanf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_atanf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_atan2f4(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @__svml_expf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_logf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_powf4(<4 x float>, <4 x float>) nounwind readnone
+
+
+define <4 x float> @__svml_sin(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_sinf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_cos(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_cosf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define void @__svml_sincos(<4 x float>, <4 x float> *, <4 x float> *) nounwind readnone alwaysinline {
+  %s = call <4 x float> @__svml_sincosf4(<4 x float> * %2, <4 x float> %0)
+  store <4 x float> %s, <4 x float> * %1
+  ret void
+}
+
+define <4 x float> @__svml_tan(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_tanf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_atan(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_atanf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_atan2(<4 x float>, <4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_atan2f4(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_exp(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_expf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_log(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_logf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_pow(<4 x float>, <4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_powf4(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <4 x float> @llvm.x86.sse.max.ps(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @llvm.x86.sse.min.ps(<4 x float>, <4 x float>) nounwind readnone
+
+define <4 x float> @__max_varying_float(<4 x float>, <4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.max.ps(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %call
+}
+
+define <4 x float> @__min_varying_float(<4 x float>, <4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.min.ps(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <2 x double> @llvm.x86.sse2.sqrt.pd(<2 x double>) nounwind readnone
+
+define <4 x double> @__sqrt_varying_double(<4 x double>) nounwind alwaysinline {
+  unary2to4(ret, double, @llvm.x86.sse2.sqrt.pd, %0)
+  ret <4 x double> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <2 x double> @llvm.x86.sse2.max.pd(<2 x double>, <2 x double>) nounwind readnone
+declare <2 x double> @llvm.x86.sse2.min.pd(<2 x double>, <2 x double>) nounwind readnone
+
+define <4 x double> @__min_varying_double(<4 x double>, <4 x double>) nounwind readnone {
+  binary2to4(ret, double, @llvm.x86.sse2.min.pd, %0, %1)
+  ret <4 x double> %ret
+}
+
+define <4 x double> @__max_varying_double(<4 x double>, <4 x double>) nounwind readnone {
+  binary2to4(ret, double, @llvm.x86.sse2.max.pd, %0, %1)
+  ret <4 x double> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+masked_store_blend_8_16_by_4()
+
+gen_masked_store(4, i8, 8)
+gen_masked_store(4, i16, 16)
+gen_masked_store(4, i32, 32)
+gen_masked_store(4, i64, 64)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unaligned loads/loads+broadcasts
+
+load_and_broadcast(4, i8, 8)
+load_and_broadcast(4, i16, 16)
+load_and_broadcast(4, i32, 32)
+load_and_broadcast(4, i64, 64)
+
+load_masked(4, i8,  8,  1)
+load_masked(4, i16, 16, 2)
+load_masked(4, i32, 32, 4)
+load_masked(4, i64, 64, 8)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; gather/scatter
+
+; define these with the macros from stdlib.m4
+
+gen_gather(4, i8)
+gen_gather(4, i16)
+gen_gather(4, i32)
+gen_gather(4, i64)
+
+gen_scatter(4, i8)
+gen_scatter(4, i16)
+gen_scatter(4, i32)
+gen_scatter(4, i64)

builtins-sse4-common.ll

@@ -0,0 +1,271 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding floats
+
+declare <4 x float> @llvm.x86.sse41.round.ss(<4 x float>, <4 x float>, i32) nounwind readnone
+
+define float @__round_uniform_float(float) nounwind readonly alwaysinline {
+  ; roundss, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
+  ; the roundss intrinsic is a total mess--docs say:
+  ;
+  ;  __m128 _mm_round_ss (__m128 a, __m128 b, const int c)
+  ;       
+  ;  b is a 128-bit parameter. The lowest 32 bits are the result of the rounding function
+  ;  on b0. The higher order 96 bits are copied directly from input parameter a. The
+  ;  return value is described by the following equations:
+  ;
+  ;  r0 = RND(b0)
+  ;  r1 = a1
+  ;  r2 = a2
+  ;  r3 = a3
+  ;
+  ;  It doesn't matter what we pass as a, since we only need the r0 value
+  ;  here.  So we pass the same register for both.  Further, only the 0th
+  ;  element of the b parameter matters
+  %xi = insertelement <4 x float> undef, float %0, i32 0
+  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 8)
+  %rs = extractelement <4 x float> %xr, i32 0
+  ret float %rs
+}
+
+define float @__floor_uniform_float(float) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <4 x float> undef, float %0, i32 0
+  ; roundps, round down 0b01 | don't signal precision exceptions 0b1010 = 9
+  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 9)
+  %rs = extractelement <4 x float> %xr, i32 0
+  ret float %rs
+}
+
+define float @__ceil_uniform_float(float) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <4 x float> undef, float %0, i32 0
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 10)
+  %rs = extractelement <4 x float> %xr, i32 0
+  ret float %rs
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare <2 x double> @llvm.x86.sse41.round.sd(<2 x double>, <2 x double>, i32) nounwind readnone
+
+define double @__round_uniform_double(double) nounwind readonly alwaysinline {
+  %xi = insertelement <2 x double> undef, double %0, i32 0
+  %xr = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %xi, <2 x double> %xi, i32 8)
+  %rs = extractelement <2 x double> %xr, i32 0
+  ret double %rs
+}
+
+define double @__floor_uniform_double(double) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <2 x double> undef, double %0, i32 0
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  %xr = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %xi, <2 x double> %xi, i32 9)
+  %rs = extractelement <2 x double> %xr, i32 0
+  ret double %rs
+}
+
+define double @__ceil_uniform_double(double) nounwind readonly alwaysinline {
+  ; see above for round_ss instrinsic discussion...
+  %xi = insertelement <2 x double> undef, double %0, i32 0
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  %xr = call <2 x double> @llvm.x86.sse41.round.sd(<2 x double> %xi, <2 x double> %xi, i32 10)
+  %rs = extractelement <2 x double> %xr, i32 0
+  ret double %rs
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
+
+define float @__rcp_uniform_float(float) nounwind readonly alwaysinline {
+  ; do the rcpss call
+  %vecval = insertelement <4 x float> undef, float %0, i32 0
+  %call = call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %vecval)
+  %scall = extractelement <4 x float> %call, i32 0
+
+  ; do one N-R iteration to improve precision, as above
+  %v_iv = fmul float %0, %scall
+  %two_minus = fsub float 2., %v_iv  
+  %iv_mul = fmul float %scall, %two_minus
+  ret float %iv_mul
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; rsqrt
+
+declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
+
+define float @__rsqrt_uniform_float(float) nounwind readonly alwaysinline {
+  ;  uniform float is = extract(__rsqrt_u(v), 0);
+  %v = insertelement <4 x float> undef, float %0, i32 0
+  %vis = call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %v)
+  %is = extractelement <4 x float> %vis, i32 0
+
+  ; Newton-Raphson iteration to improve precision
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul float %0, %is
+  %v_is_is = fmul float %v_is, %is
+  %three_sub = fsub float 3., %v_is_is
+  %is_mul = fmul float %is, %three_sub
+  %half_scale = fmul float 0.5, %is_mul
+  ret float %half_scale
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; sqrt
+
+declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone
+
+define float @__sqrt_uniform_float(float) nounwind readonly alwaysinline {
+  sse_unary_scalar(ret, 4, float, @llvm.x86.sse.sqrt.ss, %0)
+  ret float %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; fast math mode
+
+declare void @llvm.x86.sse.stmxcsr(i8 *) nounwind
+declare void @llvm.x86.sse.ldmxcsr(i8 *) nounwind
+
+define void @__fastmath() nounwind alwaysinline {
+  %ptr = alloca i32
+  %ptr8 = bitcast i32 * %ptr to i8 *
+  call void @llvm.x86.sse.stmxcsr(i8 * %ptr8)
+  %oldval = load i32 *%ptr
+
+  ; turn on DAZ (64)/FTZ (32768) -> 32832
+  %update = or i32 %oldval, 32832
+  store i32 %update, i32 *%ptr
+  call void @llvm.x86.sse.ldmxcsr(i8 * %ptr8)
+  ret void
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <4 x float> @llvm.x86.sse.max.ss(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @llvm.x86.sse.min.ss(<4 x float>, <4 x float>) nounwind readnone
+
+define float @__max_uniform_float(float, float) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.max.ss, %0, %1)
+  ret float %ret
+}
+
+define float @__min_uniform_float(float, float) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.min.ss, %0, %1)
+  ret float %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double>) nounwind readnone
+
+define double @__sqrt_uniform_double(double) nounwind alwaysinline {
+  sse_unary_scalar(ret, 2, double, @llvm.x86.sse2.sqrt.sd, %0)
+  ret double %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <2 x double> @llvm.x86.sse2.max.sd(<2 x double>, <2 x double>) nounwind readnone
+declare <2 x double> @llvm.x86.sse2.min.sd(<2 x double>, <2 x double>) nounwind readnone
+
+define double @__min_uniform_double(double, double) nounwind readnone {
+  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.min.sd, %0, %1)
+  ret double %ret
+}
+
+
+define double @__max_uniform_double(double, double) nounwind readnone {
+  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.max.sd, %0, %1)
+  ret double %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; int32 min/max
+
+declare <4 x i32> @llvm.x86.sse41.pminsd(<4 x i32>, <4 x i32>) nounwind readnone
+declare <4 x i32> @llvm.x86.sse41.pmaxsd(<4 x i32>, <4 x i32>) nounwind readnone
+
+define i32 @__min_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pminsd, %0, %1)
+  ret i32 %ret
+}
+
+define i32 @__max_uniform_int32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
+  ret i32 %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; unsigned int min/max
+
+declare <4 x i32> @llvm.x86.sse41.pminud(<4 x i32>, <4 x i32>) nounwind readnone
+declare <4 x i32> @llvm.x86.sse41.pmaxud(<4 x i32>, <4 x i32>) nounwind readnone
+
+define i32 @__min_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pminud, %0, %1)
+  ret i32 %ret
+}
+
+define i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
+  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pmaxud, %0, %1)
+  ret i32 %ret
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops / reductions
+
+declare i32 @llvm.ctpop.i32(i32) nounwind readnone
+
+define i32 @__popcnt_int32(i32) nounwind readonly alwaysinline {
+  %call = call i32 @llvm.ctpop.i32(i32 %0)
+  ret i32 %call
+}
+
+declare i64 @llvm.ctpop.i64(i64) nounwind readnone
+
+define i64 @__popcnt_int64(i64) nounwind readonly alwaysinline {
+  %call = call i64 @llvm.ctpop.i64(i64 %0)
+  ret i64 %call
+}

builtins-sse4-x2.ll

@@ -38,15 +38,17 @@
 
 stdlib_core(8)
 packed_load_and_store(8)
-int8_16(8)
+scans(8)
+int64minmax(8)
+
+include(`builtins-sse4-common.ll')
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; rcp
 
 declare <4 x float> @llvm.x86.sse.rcp.ps(<4 x float>) nounwind readnone
-declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
 
-define internal <8 x float> @__rcp_varying_float(<8 x float>) nounwind readonly alwaysinline {
+define <8 x float> @__rcp_varying_float(<8 x float>) nounwind readonly alwaysinline {
   ;  float iv = __rcp_v(v);
   ;  return iv * (2. - v * iv);
 
@@ -59,27 +61,12 @@ define internal <8 x float> @__rcp_varying_float(<8 x float>) nounwind readonly
   ret <8 x float> %iv_mul
 }
 
-define internal float @__rcp_uniform_float(float) nounwind readonly alwaysinline {
-;    uniform float iv = extract(__rcp_u(v), 0);
-;    return iv * (2. - v * iv);
-  %vecval = insertelement <4 x float> undef, float %0, i32 0
-  %call = call <4 x float> @llvm.x86.sse.rcp.ss(<4 x float> %vecval)
-  %scall = extractelement <4 x float> %call, i32 0
-
-  ; do one N-R iteration
-  %v_iv = fmul float %0, %scall
-  %two_minus = fsub float 2., %v_iv  
-  %iv_mul = fmul float %scall, %two_minus
-  ret float %iv_mul
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; rsqrt
 
 declare <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float>) nounwind readnone
-declare <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float>) nounwind readnone
 
-define internal <8 x float> @__rsqrt_varying_float(<8 x float> %v) nounwind readonly alwaysinline {
+define <8 x float> @__rsqrt_varying_float(<8 x float> %v) nounwind readonly alwaysinline {
   ;  float is = __rsqrt_v(v);
   unary4to8(is, float, @llvm.x86.sse.rsqrt.ps, %v)
   ;  return 0.5 * is * (3. - (v * is) * is);
@@ -93,56 +80,16 @@ define internal <8 x float> @__rsqrt_varying_float(<8 x float> %v) nounwind read
   ret <8 x float> %half_scale
 }
 
-define internal float @__rsqrt_uniform_float(float) nounwind readonly alwaysinline {
-  ;  uniform float is = extract(__rsqrt_u(v), 0);
-  %v = insertelement <4 x float> undef, float %0, i32 0
-  %vis = call <4 x float> @llvm.x86.sse.rsqrt.ss(<4 x float> %v)
-  %is = extractelement <4 x float> %vis, i32 0
-
-  ;  return 0.5 * is * (3. - (v * is) * is);
-  %v_is = fmul float %0, %is
-  %v_is_is = fmul float %v_is, %is
-  %three_sub = fsub float 3., %v_is_is
-  %is_mul = fmul float %is, %three_sub
-  %half_scale = fmul float 0.5, %is_mul
-  ret float %half_scale
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; sqrt
 
 declare <4 x float> @llvm.x86.sse.sqrt.ps(<4 x float>) nounwind readnone
-declare <4 x float> @llvm.x86.sse.sqrt.ss(<4 x float>) nounwind readnone
 
-define internal <8 x float> @__sqrt_varying_float(<8 x float>) nounwind readonly alwaysinline {
+define <8 x float> @__sqrt_varying_float(<8 x float>) nounwind readonly alwaysinline {
   unary4to8(call, float, @llvm.x86.sse.sqrt.ps, %0)
   ret <8 x float> %call
 }
 
-define internal float @__sqrt_uniform_float(float) nounwind readonly alwaysinline {
-  sse_unary_scalar(ret, 4, float, @llvm.x86.sse.sqrt.ss, %0)
-  ret float %ret
-}
-
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; fast math
-
-declare void @llvm.x86.sse.stmxcsr(i32 *) nounwind
-declare void @llvm.x86.sse.ldmxcsr(i32 *) nounwind
-
-define internal void @__fastmath() nounwind alwaysinline {
-  %ptr = alloca i32
-  call void @llvm.x86.sse.stmxcsr(i32 * %ptr)
-  %oldval = load i32 *%ptr
-
-  ; turn on DAZ (64)/FTZ (32768) -> 32832
-  %update = or i32 %oldval, 32832
-  store i32 %update, i32 *%ptr
-  call void @llvm.x86.sse.ldmxcsr(i32 * %ptr)
-  ret void
-}
-
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; svml stuff
 
@@ -157,17 +104,17 @@ declare <4 x float> @__svml_logf4(<4 x float>) nounwind readnone
 declare <4 x float> @__svml_powf4(<4 x float>, <4 x float>) nounwind readnone
 
 
-define internal <8 x float> @__svml_sin(<8 x float>) nounwind readnone alwaysinline {
+define <8 x float> @__svml_sin(<8 x float>) nounwind readnone alwaysinline {
   unary4to8(ret, float, @__svml_sinf4, %0)
   ret <8 x float> %ret
 }
 
-define internal <8 x float> @__svml_cos(<8 x float>) nounwind readnone alwaysinline {
+define <8 x float> @__svml_cos(<8 x float>) nounwind readnone alwaysinline {
   unary4to8(ret, float, @__svml_cosf4, %0)
   ret <8 x float> %ret
 }
 
-define internal void @__svml_sincos(<8 x float>, <8 x float> *,
+define void @__svml_sincos(<8 x float>, <8 x float> *,
                                     <8 x float> *) nounwind readnone alwaysinline {
   ; call svml_sincosf4 two times with the two 4-wide sub-vectors
   %a = shufflevector <8 x float> %0, <8 x float> undef,
@@ -196,33 +143,33 @@ define internal void @__svml_sincos(<8 x float>, <8 x float> *,
   ret void
 }
 
-define internal <8 x float> @__svml_tan(<8 x float>) nounwind readnone alwaysinline {
+define <8 x float> @__svml_tan(<8 x float>) nounwind readnone alwaysinline {
   unary4to8(ret, float, @__svml_tanf4, %0)
   ret <8 x float> %ret
 }
 
-define internal <8 x float> @__svml_atan(<8 x float>) nounwind readnone alwaysinline {
+define <8 x float> @__svml_atan(<8 x float>) nounwind readnone alwaysinline {
   unary4to8(ret, float, @__svml_atanf4, %0)
   ret <8 x float> %ret
 }
 
-define internal <8 x float> @__svml_atan2(<8 x float>,
+define <8 x float> @__svml_atan2(<8 x float>,
                                           <8 x float>) nounwind readnone alwaysinline {
   binary4to8(ret, float, @__svml_atan2f4, %0, %1)
   ret <8 x float> %ret
 }
 
-define internal <8 x float> @__svml_exp(<8 x float>) nounwind readnone alwaysinline {
+define <8 x float> @__svml_exp(<8 x float>) nounwind readnone alwaysinline {
   unary4to8(ret, float, @__svml_expf4, %0)
   ret <8 x float> %ret
 }
 
-define internal <8 x float> @__svml_log(<8 x float>) nounwind readnone alwaysinline {
+define <8 x float> @__svml_log(<8 x float>) nounwind readnone alwaysinline {
   unary4to8(ret, float, @__svml_logf4, %0)
   ret <8 x float> %ret
 }
 
-define internal <8 x float> @__svml_pow(<8 x float>,
+define <8 x float> @__svml_pow(<8 x float>,
                                         <8 x float>) nounwind readnone alwaysinline {
   binary4to8(ret, float, @__svml_powf4, %0, %1)
   ret <8 x float> %ret
@@ -233,91 +180,52 @@ define internal <8 x float> @__svml_pow(<8 x float>,
 ;; float min/max
 
 declare <4 x float> @llvm.x86.sse.max.ps(<4 x float>, <4 x float>) nounwind readnone
-declare <4 x float> @llvm.x86.sse.max.ss(<4 x float>, <4 x float>) nounwind readnone
 declare <4 x float> @llvm.x86.sse.min.ps(<4 x float>, <4 x float>) nounwind readnone
-declare <4 x float> @llvm.x86.sse.min.ss(<4 x float>, <4 x float>) nounwind readnone
 
-define internal <8 x float> @__max_varying_float(<8 x float>, <8 x float>) nounwind readonly alwaysinline {
+define <8 x float> @__max_varying_float(<8 x float>, <8 x float>) nounwind readonly alwaysinline {
   binary4to8(call, float, @llvm.x86.sse.max.ps, %0, %1)
   ret <8 x float> %call
 }
 
-define internal float @__max_uniform_float(float, float) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.max.ss, %0, %1)
-  ret float %ret
-}
-
-define internal <8 x float> @__min_varying_float(<8 x float>, <8 x float>) nounwind readonly alwaysinline {
+define <8 x float> @__min_varying_float(<8 x float>, <8 x float>) nounwind readonly alwaysinline {
   binary4to8(call, float, @llvm.x86.sse.min.ps, %0, %1)
   ret <8 x float> %call
 }
 
-define internal float @__min_uniform_float(float, float) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 4, float, @llvm.x86.sse.min.ss, %0, %1)
-  ret float %ret
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; int min/max
+;; int32 min/max
 
-declare <4 x i32> @llvm.x86.sse41.pminsd(<4 x i32>, <4 x i32>) nounwind readnone
-declare <4 x i32> @llvm.x86.sse41.pmaxsd(<4 x i32>, <4 x i32>) nounwind readnone
-
-define internal <8 x i32> @__min_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+define <8 x i32> @__min_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
   binary4to8(call, i32, @llvm.x86.sse41.pminsd, %0, %1)
   ret <8 x i32> %call
 }
 
-define internal i32 @__min_uniform_int32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pminsd, %0, %1)
-  ret i32 %ret
-}
-
-define internal <8 x i32> @__max_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
+define <8 x i32> @__max_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
   binary4to8(call, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
   ret <8 x i32> %call
 }
 
-define internal i32 @__max_uniform_int32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
-  ret i32 %ret
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; unsigned int min/max
 
-declare <4 x i32> @llvm.x86.sse41.pminud(<4 x i32>, <4 x i32>) nounwind readnone
-declare <4 x i32> @llvm.x86.sse41.pmaxud(<4 x i32>, <4 x i32>) nounwind readnone
-
-define internal <8 x i32> @__min_varying_uint32(<8 x i32>,
+define <8 x i32> @__min_varying_uint32(<8 x i32>,
                                                 <8 x i32>) nounwind readonly alwaysinline {
   binary4to8(call, i32, @llvm.x86.sse41.pminud, %0, %1)
   ret <8 x i32> %call
 }
 
-define internal i32 @__min_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pminud, %0, %1)
-  ret i32 %ret
-}
-
-define internal <8 x i32> @__max_varying_uint32(<8 x i32>,
+define <8 x i32> @__max_varying_uint32(<8 x i32>,
                                                 <8 x i32>) nounwind readonly alwaysinline {
   binary4to8(call, i32, @llvm.x86.sse41.pmaxud, %0, %1)
   ret <8 x i32> %call
 }
 
-define internal i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 4, i32, @llvm.x86.sse41.pmaxud, %0, %1)
-  ret i32 %ret
-}
-
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; horizontal ops / reductions
 
 declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) nounwind readnone
 
-define internal i32 @__movmsk(<8 x i32>) nounwind readnone alwaysinline {
+define i32 @__movmsk(<8 x i32>) nounwind readnone alwaysinline {
   ; first do two 4-wide movmsk calls
   %floatmask = bitcast <8 x i32> %0 to <8 x float>
   %m0 = shufflevector <8 x float> %floatmask, <8 x float> undef,
@@ -334,211 +242,179 @@ define internal i32 @__movmsk(<8 x i32>) nounwind readnone alwaysinline {
   ret i32 %v
 }
 
-define internal float @__reduce_min_float(<8 x float>) nounwind readnone alwaysinline {
+define float @__reduce_min_float(<8 x float>) nounwind readnone alwaysinline {
   reduce8by4(float, @llvm.x86.sse.min.ps, @__min_uniform_float)
 }
 
-define internal float @__reduce_max_float(<8 x float>) nounwind readnone alwaysinline {
+define float @__reduce_max_float(<8 x float>) nounwind readnone alwaysinline {
   reduce8by4(float, @llvm.x86.sse.max.ps, @__max_uniform_float)
 }
 
 ; helper function for reduce_add_int32
-define internal <4 x i32> @__vec4_add_int32(<4 x i32> %v0,
+define <4 x i32> @__vec4_add_int32(<4 x i32> %v0,
                                             <4 x i32> %v1) nounwind readnone alwaysinline {
   %v = add <4 x i32> %v0, %v1
   ret <4 x i32> %v
 }
 
 ; helper function for reduce_add_int32
-define internal i32 @__add_int32(i32, i32) nounwind readnone alwaysinline {
+define i32 @__add_int32(i32, i32) nounwind readnone alwaysinline {
   %v = add i32 %0, %1
   ret i32 %v
 }
 
-define internal i32 @__reduce_add_int32(<8 x i32>) nounwind readnone alwaysinline {
+define i32 @__reduce_add_int32(<8 x i32>) nounwind readnone alwaysinline {
   reduce8by4(i32, @__vec4_add_int32, @__add_int32)
 }
 
-define internal i32 @__reduce_min_int32(<8 x i32>) nounwind readnone alwaysinline {
+define i32 @__reduce_min_int32(<8 x i32>) nounwind readnone alwaysinline {
   reduce8by4(i32, @llvm.x86.sse41.pminsd, @__min_uniform_int32)
 }
 
-define internal i32 @__reduce_max_int32(<8 x i32>) nounwind readnone alwaysinline {
+define i32 @__reduce_max_int32(<8 x i32>) nounwind readnone alwaysinline {
   reduce8by4(i32, @llvm.x86.sse41.pmaxsd, @__max_uniform_int32)
 }
 
-define internal i32 @__reduce_add_uint32(<8 x i32> %v) nounwind readnone alwaysinline {
+define i32 @__reduce_add_uint32(<8 x i32> %v) nounwind readnone alwaysinline {
   %r = call i32 @__reduce_add_int32(<8 x i32> %v)
   ret i32 %r
 }
 
-define internal i32 @__reduce_min_uint32(<8 x i32>) nounwind readnone alwaysinline {
+define i32 @__reduce_min_uint32(<8 x i32>) nounwind readnone alwaysinline {
   reduce8by4(i32, @llvm.x86.sse41.pminud, @__min_uniform_uint32)
 }
 
-define internal i32 @__reduce_max_uint32(<8 x i32>) nounwind readnone alwaysinline {
+define i32 @__reduce_max_uint32(<8 x i32>) nounwind readnone alwaysinline {
   reduce8by4(i32, @llvm.x86.sse41.pmaxud, @__max_uniform_uint32)
 }
 
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; masked store
-
-define void @__masked_store_32(<8 x i32>* nocapture, <8 x i32>,
-                               <8 x i32>) nounwind alwaysinline {
-  per_lane(8, <8 x i32> %2, `
-      ; compute address for this one
-      %ptr_ID = getelementptr <8 x i32> * %0, i32 0, i32 LANE
-      %storeval_ID = extractelement <8 x i32> %1, i32 LANE
-      store i32 %storeval_ID, i32 * %ptr_ID')
-  ret void
+define <4 x double> @__add_varying_double(<4 x double>,
+                                     <4 x double>) nounwind readnone alwaysinline {
+  %r = fadd <4 x double> %0, %1
+  ret <4 x double> %r
 }
 
-
-define void @__masked_store_64(<8 x i64>* nocapture, <8 x i64>,
-                               <8 x i32>) nounwind alwaysinline {
-  per_lane(8, <8 x i32> %2, `
-      ; compute address for this one
-      %ptr_ID = getelementptr <8 x i64> * %0, i32 0, i32 LANE
-      %storeval_ID = extractelement <8 x i64> %1, i32 LANE
-      store i64 %storeval_ID, i64 * %ptr_ID')
-  ret void
+define double @__add_uniform_double(double, double) nounwind readnone alwaysinline {
+  %r = fadd double %0, %1
+  ret double %r
 }
 
+define double @__reduce_add_double(<8 x double>) nounwind readnone {
+  reduce8by4(double, @__add_varying_double, @__add_uniform_double)
+}
+
+define double @__reduce_min_double(<8 x double>) nounwind readnone {
+  reduce8(double, @__min_varying_double, @__min_uniform_double)
+}
+
+define double @__reduce_max_double(<8 x double>) nounwind readnone {
+  reduce8(double, @__max_varying_double, @__max_uniform_double)
+}
+
+define <4 x i64> @__add_varying_int64(<4 x i64>,
+                                               <4 x i64>) nounwind readnone alwaysinline {
+  %r = add <4 x i64> %0, %1
+  ret <4 x i64> %r
+}
+
+define i64 @__add_uniform_int64(i64, i64) nounwind readnone alwaysinline {
+  %r = add i64 %0, %1
+  ret i64 %r
+}
+
+define i64 @__reduce_add_int64(<8 x i64>) nounwind readnone {
+  reduce8by4(i64, @__add_varying_int64, @__add_uniform_int64)
+}
+
+define i64 @__reduce_min_int64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+define i64 @__reduce_max_int64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+define i64 @__reduce_min_uint64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+define i64 @__reduce_max_uint64(<8 x i64>) nounwind readnone {
+  reduce8(i64, @__max_varying_uint64, @__max_uniform_uint64)
+}
+
+reduce_equal(8)
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; unaligned loads/loads+broadcasts
 
-; FIXME: I think this and the next one need to verify that the mask isn't
-; all off before doing the load!!!  (See e.g. stdlib-sse.ll)
+load_and_broadcast(8, i8, 8)
+load_and_broadcast(8, i16, 16)
+load_and_broadcast(8, i32, 32)
+load_and_broadcast(8, i64, 64)
 
-define <8 x i32> @__load_and_broadcast_32(i8 *, <8 x i32> %mask) nounwind alwaysinline {
-  %ptr = bitcast i8 * %0 to i32 *
-  %val = load i32 * %ptr
-
-  %ret0 = insertelement <8 x i32> undef, i32 %val, i32 0
-  %ret1 = insertelement <8 x i32> %ret0, i32 %val, i32 1
-  %ret2 = insertelement <8 x i32> %ret1, i32 %val, i32 2
-  %ret3 = insertelement <8 x i32> %ret2, i32 %val, i32 3
-  %ret4 = insertelement <8 x i32> %ret3, i32 %val, i32 4
-  %ret5 = insertelement <8 x i32> %ret4, i32 %val, i32 5
-  %ret6 = insertelement <8 x i32> %ret5, i32 %val, i32 6
-  %ret7 = insertelement <8 x i32> %ret6, i32 %val, i32 7
-  ret <8 x i32> %ret7
-}
-
-
-define <8 x i64> @__load_and_broadcast_64(i8 *, <8 x i32> %mask) nounwind alwaysinline {
-  %ptr = bitcast i8 * %0 to i64 *
-  %val = load i64 * %ptr
-
-  %ret0 = insertelement <8 x i64> undef, i64 %val, i32 0
-  %ret1 = insertelement <8 x i64> %ret0, i64 %val, i32 1
-  %ret2 = insertelement <8 x i64> %ret1, i64 %val, i32 2
-  %ret3 = insertelement <8 x i64> %ret2, i64 %val, i32 3
-  %ret4 = insertelement <8 x i64> %ret3, i64 %val, i32 4
-  %ret5 = insertelement <8 x i64> %ret4, i64 %val, i32 5
-  %ret6 = insertelement <8 x i64> %ret5, i64 %val, i32 6
-  %ret7 = insertelement <8 x i64> %ret6, i64 %val, i32 7
-  ret <8 x i64> %ret7
-}
-
-
-define <8 x i32> @__load_masked_32(i8 *, <8 x i32> %mask) nounwind alwaysinline {
-  %ptr = bitcast i8 * %0 to <8 x i32> *
-  %val = load <8 x i32> * %ptr, align 4
-  ret <8 x i32> %val
-}
-
-
-define <8 x i64> @__load_masked_64(i8 *, <8 x i32> %mask) nounwind alwaysinline {
-  %ptr = bitcast i8 * %0 to <8 x i64> *
-  %val = load <8 x i64> * %ptr, align 8
-  ret <8 x i64> %val
-}
+load_masked(8, i8,  8,  1)
+load_masked(8, i16, 16, 2)
+load_masked(8, i32, 32, 4)
+load_masked(8, i64, 64, 8)
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; gather/scatter
 
+gen_gather(8, i8)
+gen_gather(8, i16)
 gen_gather(8, i32)
 gen_gather(8, i64)
+
+gen_scatter(8, i8)
+gen_scatter(8, i16)
 gen_scatter(8, i32)
 gen_scatter(8, i64)
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; math
+;; float rounding
 
 declare <4 x float> @llvm.x86.sse41.round.ps(<4 x float>, i32) nounwind readnone
-declare <4 x float> @llvm.x86.sse41.round.ss(<4 x float>, <4 x float>, i32) nounwind readnone
 
-define internal <8 x float> @__round_varying_float(<8 x float>) nounwind readonly alwaysinline {
+define <8 x float> @__round_varying_float(<8 x float>) nounwind readonly alwaysinline {
   ; roundps, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
   round4to8(%0, 8)
 }
 
-define internal float @__round_uniform_float(float) nounwind readonly alwaysinline {
-  ; roundss, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
-  ; the roundss intrinsic is a total mess--docs say:
-  ;
-  ;  __m128 _mm_round_ss (__m128 a, __m128 b, const int c)
-  ;       
-  ;  b is a 128-bit parameter. The lowest 32 bits are the result of the rounding function
-  ;  on b0. The higher order 96 bits are copied directly from input parameter a. The
-  ;  return value is described by the following equations:
-  ;
-  ;  r0 = RND(b0)
-  ;  r1 = a1
-  ;  r2 = a2
-  ;  r3 = a3
-  ;
-  ;  It doesn't matter what we pass as a, since we only need the r0 value
-  ;  here.  So we pass the same register for both.  
-  %xi = insertelement <4 x float> undef, float %0, i32 0
-  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 8)
-  %rs = extractelement <4 x float> %xr, i32 0
-  ret float %rs
-}
-
-define internal <8 x float> @__floor_varying_float(<8 x float>) nounwind readonly alwaysinline {
-  ; roundps, round down 0b01 | don't signal precision exceptions 0b1000 = 9
+define <8 x float> @__floor_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ; roundps, round down 0b01 | don't signal precision exceptions 0b1001 = 9
   round4to8(%0, 9)
 }
 
-define internal float @__floor_uniform_float(float) nounwind readonly alwaysinline {
-  ; see above for round_ss instrinsic discussion...
-  %xi = insertelement <4 x float> undef, float %0, i32 0
-  ; roundps, round down 0b01 | don't signal precision exceptions 0b1000 = 9
-  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 9)
-  %rs = extractelement <4 x float> %xr, i32 0
-  ret float %rs
-}
-
-define internal <8 x float> @__ceil_varying_float(<8 x float>) nounwind readonly alwaysinline {
-  ; roundps, round up 0b10 | don't signal precision exceptions 0b1000 = 10
+define <8 x float> @__ceil_varying_float(<8 x float>) nounwind readonly alwaysinline {
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
   round4to8(%0, 10)
 }
 
-define internal float @__ceil_uniform_float(float) nounwind readonly alwaysinline {
-  ; see above for round_ss instrinsic discussion...
-  %xi = insertelement <4 x float> undef, float %0, i32 0
-  ; roundps, round up 0b10 | don't signal precision exceptions 0b1000 = 10
-  %xr = call <4 x float> @llvm.x86.sse41.round.ss(<4 x float> %xi, <4 x float> %xi, i32 10)
-  %rs = extractelement <4 x float> %xr, i32 0
-  ret float %rs
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare <2 x double> @llvm.x86.sse41.round.pd(<2 x double>, i32) nounwind readnone
+
+define <8 x double> @__round_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  round2to8double(%0, 8)
+}
+
+define <8 x double> @__floor_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  round2to8double(%0, 9)
+}
+
+define <8 x double> @__ceil_varying_double(<8 x double>) nounwind readonly alwaysinline {
+  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  round2to8double(%0, 10)
 }
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; horizontal ops / reductions
 
-declare i32 @llvm.ctpop.i32(i32) nounwind readnone
-
-define internal i32 @__popcnt(i32) nounwind readonly alwaysinline {
-  %call = call i32 @llvm.ctpop.i32(i32 %0)
-  ret i32 %call
-}
-
 declare <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float>, <4 x float>) nounwind readnone
 
-define internal float @__reduce_add_float(<8 x float>) nounwind readonly alwaysinline {
+define float @__reduce_add_float(<8 x float>) nounwind readonly alwaysinline {
   %a = shufflevector <8 x float> %0, <8 x float> undef,
          <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %b = shufflevector <8 x float> %0, <8 x float> undef,
@@ -555,6 +431,13 @@ define internal float @__reduce_add_float(<8 x float>) nounwind readonly alwaysi
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; masked store
 
+gen_masked_store(8, i8, 8)
+gen_masked_store(8, i16, 16)
+gen_masked_store(8, i32, 32)
+gen_masked_store(8, i64, 64)
+
+masked_store_blend_8_16_by_8()
+
 declare <4 x float> @llvm.x86.sse41.blendvps(<4 x float>, <4 x float>,
                                              <4 x float>) nounwind readnone
 
@@ -566,7 +449,7 @@ define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>,
                 <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %mask_b = shufflevector <8 x float> %mask_as_float, <8 x float> undef,
                 <4 x i32> <i32 4, i32 5, i32 6, i32 7>
-  %oldValue = load <8 x i32>* %0
+  %oldValue = load <8 x i32>* %0, align 4
   %oldAsFloat = bitcast <8 x i32> %oldValue to <8 x float>
   %newAsFloat = bitcast <8 x i32> %1 to <8 x float>
   %old_a = shufflevector <8 x float> %oldAsFloat, <8 x float> undef,
@@ -584,7 +467,7 @@ define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>,
   %blend = shufflevector <4 x float> %blend_a, <4 x float> %blend_b,
                <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
   %blendAsInt = bitcast <8 x float> %blend to <8 x i32>
-  store <8 x i32> %blendAsInt, <8 x i32>* %0
+  store <8 x i32> %blendAsInt, <8 x i32>* %0, align 4
   ret void
 }
 
@@ -595,7 +478,7 @@ define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
 
   %mask_as_float = bitcast <8 x i32> %mask to <8 x float>
 
-  %old = load <8 x i64>* %ptr
+  %old = load <8 x i64>* %ptr, align 8
 
   ; set up the first two 64-bit values
   %old01 = shufflevector <8 x i64> %old, <8 x i64> undef, <2 x i32> <i32 0, i32 1>
@@ -651,7 +534,7 @@ define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
        <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %final = shufflevector <4 x i64> %final0123, <4 x i64> %final4567,
        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
-  store <8 x i64> %final, <8 x i64> * %ptr
+  store <8 x i64> %final, <8 x i64> * %ptr, align 8
   ret void
 }
 
@@ -660,44 +543,24 @@ define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
 ;; double precision sqrt
 
 declare <2 x double> @llvm.x86.sse2.sqrt.pd(<2 x double>) nounwind readnone
-declare <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double>) nounwind readnone
 
-define internal <8 x double> @__sqrt_varying_double(<8 x double>) nounwind alwaysinline {
+define <8 x double> @__sqrt_varying_double(<8 x double>) nounwind alwaysinline {
   unary2to8(ret, double, @llvm.x86.sse2.sqrt.pd, %0)
   ret <8 x double> %ret
 }
 
-
-define internal double @__sqrt_uniform_double(double) nounwind alwaysinline {
-  sse_unary_scalar(ret, 2, double, @llvm.x86.sse2.sqrt.pd, %0)
-  ret double %ret
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; double precision float min/max
 
 declare <2 x double> @llvm.x86.sse2.max.pd(<2 x double>, <2 x double>) nounwind readnone
-declare <2 x double> @llvm.x86.sse2.max.sd(<2 x double>, <2 x double>) nounwind readnone
 declare <2 x double> @llvm.x86.sse2.min.pd(<2 x double>, <2 x double>) nounwind readnone
-declare <2 x double> @llvm.x86.sse2.min.sd(<2 x double>, <2 x double>) nounwind readnone
 
-define internal <8 x double> @__min_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
+define <8 x double> @__min_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
   binary2to8(ret, double, @llvm.x86.sse2.min.pd, %0, %1)
   ret <8 x double> %ret
 }
 
-define internal double @__min_uniform_double(double, double) nounwind readnone alwaysinline {
-  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.min.pd, %0, %1)
-  ret double %ret
-}
-
-define internal <8 x double> @__max_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
+define <8 x double> @__max_varying_double(<8 x double>, <8 x double>) nounwind readnone alwaysinline {
   binary2to8(ret, double, @llvm.x86.sse2.max.pd, %0, %1)
   ret <8 x double> %ret
 }
-
-define internal double @__max_uniform_double(double, double) nounwind readnone alwaysinline {
-  sse_binary_scalar(ret, 2, double, @llvm.x86.sse2.max.pd, %0, %1)
-  ret double %ret
-
-}

builtins-sse4.ll

@@ -0,0 +1,472 @@
+;;  Copyright (c) 2010-2011, Intel Corporation
+;;  All rights reserved.
+;;
+;;  Redistribution and use in source and binary forms, with or without
+;;  modification, are permitted provided that the following conditions are
+;;  met:
+;;
+;;    * Redistributions of source code must retain the above copyright
+;;      notice, this list of conditions and the following disclaimer.
+;;
+;;    * Redistributions in binary form must reproduce the above copyright
+;;      notice, this list of conditions and the following disclaimer in the
+;;      documentation and/or other materials provided with the distribution.
+;;
+;;    * Neither the name of Intel Corporation nor the names of its
+;;      contributors may be used to endorse or promote products derived from
+;;      this software without specific prior written permission.
+;;
+;;
+;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+; Define common 4-wide stuff
+stdlib_core(4)
+packed_load_and_store(4)
+scans(4)
+int64minmax(4)
+
+include(`builtins-sse4-common.ll')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ps(<4 x float>) nounwind readnone
+
+define <4 x float> @__rcp_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.rcp.ps(<4 x float> %0)
+  ; do one N-R iteration to improve precision
+  ;  float iv = __rcp_v(v);
+  ;  return iv * (2. - v * iv);
+  %v_iv = fmul <4 x float> %0, %call
+  %two_minus = fsub <4 x float> <float 2., float 2., float 2., float 2.>, %v_iv  
+  %iv_mul = fmul <4 x float> %call, %two_minus
+  ret <4 x float> %iv_mul
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; rsqrt
+
+declare <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float>) nounwind readnone
+
+define <4 x float> @__rsqrt_varying_float(<4 x float> %v) nounwind readonly alwaysinline {
+  ;  float is = __rsqrt_v(v);
+  %is = call <4 x float> @llvm.x86.sse.rsqrt.ps(<4 x float> %v)
+  ; Newton-Raphson iteration to improve precision
+  ;  return 0.5 * is * (3. - (v * is) * is);
+  %v_is = fmul <4 x float> %v, %is
+  %v_is_is = fmul <4 x float> %v_is, %is
+  %three_sub = fsub <4 x float> <float 3., float 3., float 3., float 3.>, %v_is_is
+  %is_mul = fmul <4 x float> %is, %three_sub
+  %half_scale = fmul <4 x float> <float 0.5, float 0.5, float 0.5, float 0.5>, %is_mul
+  ret <4 x float> %half_scale
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; sqrt
+
+declare <4 x float> @llvm.x86.sse.sqrt.ps(<4 x float>) nounwind readnone
+
+define <4 x float> @__sqrt_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.sqrt.ps(<4 x float> %0)
+  ret <4 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision sqrt
+
+declare <2 x double> @llvm.x86.sse2.sqrt.pd(<2 x double>) nounwind readnone
+
+define <4 x double> @__sqrt_varying_double(<4 x double>) nounwind alwaysinline {
+  unary2to4(ret, double, @llvm.x86.sse2.sqrt.pd, %0)
+  ret <4 x double> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding floats
+
+declare <4 x float> @llvm.x86.sse41.round.ps(<4 x float>, i32) nounwind readnone
+
+define <4 x float> @__round_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  ; roundps, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
+  %call = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %0, i32 8)
+  ret <4 x float> %call
+}
+
+define <4 x float> @__floor_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  ; roundps, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  %call = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %0, i32 9)
+  ret <4 x float> %call
+}
+
+define <4 x float> @__ceil_varying_float(<4 x float>) nounwind readonly alwaysinline {
+  ; roundps, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  %call = call <4 x float> @llvm.x86.sse41.round.ps(<4 x float> %0, i32 10)
+  ret <4 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding doubles
+
+declare <2 x double> @llvm.x86.sse41.round.pd(<2 x double>, i32) nounwind readnone
+
+define <4 x double> @__round_varying_double(<4 x double>) nounwind readonly alwaysinline {
+  round2to4double(%0, 8)
+}
+
+define <4 x double> @__floor_varying_double(<4 x double>) nounwind readonly alwaysinline {
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1001 = 9
+  round2to4double(%0, 9)
+}
+
+define <4 x double> @__ceil_varying_double(<4 x double>) nounwind readonly alwaysinline {
+  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1010 = 10
+  round2to4double(%0, 10)
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; float min/max
+
+declare <4 x float> @llvm.x86.sse.max.ps(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @llvm.x86.sse.min.ps(<4 x float>, <4 x float>) nounwind readnone
+
+define <4 x float> @__max_varying_float(<4 x float>, <4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.max.ps(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %call
+}
+
+define <4 x float> @__min_varying_float(<4 x float>, <4 x float>) nounwind readonly alwaysinline {
+  %call = call <4 x float> @llvm.x86.sse.min.ps(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; int32 min/max
+
+define <4 x i32> @__min_varying_int32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %call = call <4 x i32> @llvm.x86.sse41.pminsd(<4 x i32> %0, <4 x i32> %1)
+  ret <4 x i32> %call
+}
+
+define <4 x i32> @__max_varying_int32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %call = call <4 x i32> @llvm.x86.sse41.pmaxsd(<4 x i32> %0, <4 x i32> %1)
+  ret <4 x i32> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; unsigned int min/max
+
+define <4 x i32> @__min_varying_uint32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %call = call <4 x i32> @llvm.x86.sse41.pminud(<4 x i32> %0, <4 x i32> %1)
+  ret <4 x i32> %call
+}
+
+define <4 x i32> @__max_varying_uint32(<4 x i32>, <4 x i32>) nounwind readonly alwaysinline {
+  %call = call <4 x i32> @llvm.x86.sse41.pmaxud(<4 x i32> %0, <4 x i32> %1)
+  ret <4 x i32> %call
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; double precision min/max
+
+declare <2 x double> @llvm.x86.sse2.max.pd(<2 x double>, <2 x double>) nounwind readnone
+declare <2 x double> @llvm.x86.sse2.min.pd(<2 x double>, <2 x double>) nounwind readnone
+
+define <4 x double> @__min_varying_double(<4 x double>, <4 x double>) nounwind readnone {
+  binary2to4(ret, double, @llvm.x86.sse2.min.pd, %0, %1)
+  ret <4 x double> %ret
+}
+
+define <4 x double> @__max_varying_double(<4 x double>, <4 x double>) nounwind readnone {
+  binary2to4(ret, double, @llvm.x86.sse2.max.pd, %0, %1)
+  ret <4 x double> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; svml stuff
+
+declare <4 x float> @__svml_sinf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_cosf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_sincosf4(<4 x float> *, <4 x float>) nounwind readnone
+declare <4 x float> @__svml_tanf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_atanf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_atan2f4(<4 x float>, <4 x float>) nounwind readnone
+declare <4 x float> @__svml_expf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_logf4(<4 x float>) nounwind readnone
+declare <4 x float> @__svml_powf4(<4 x float>, <4 x float>) nounwind readnone
+
+
+define <4 x float> @__svml_sin(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_sinf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_cos(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_cosf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define void @__svml_sincos(<4 x float>, <4 x float> *, <4 x float> *) nounwind readnone alwaysinline {
+  %s = call <4 x float> @__svml_sincosf4(<4 x float> * %2, <4 x float> %0)
+  store <4 x float> %s, <4 x float> * %1
+  ret void
+}
+
+define <4 x float> @__svml_tan(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_tanf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_atan(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_atanf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_atan2(<4 x float>, <4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_atan2f4(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_exp(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_expf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_log(<4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_logf4(<4 x float> %0)
+  ret <4 x float> %ret
+}
+
+define <4 x float> @__svml_pow(<4 x float>, <4 x float>) nounwind readnone alwaysinline {
+  %ret = call <4 x float> @__svml_powf4(<4 x float> %0, <4 x float> %1)
+  ret <4 x float> %ret
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+; horizontal ops / reductions
+
+declare i32 @llvm.x86.sse.movmsk.ps(<4 x float>) nounwind readnone
+
+define i32 @__movmsk(<4 x i32>) nounwind readnone alwaysinline {
+  %floatmask = bitcast <4 x i32> %0 to <4 x float>
+  %v = call i32 @llvm.x86.sse.movmsk.ps(<4 x float> %floatmask) nounwind readnone
+  ret i32 %v
+}
+
+declare <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float>, <4 x float>) nounwind readnone
+
+define float @__reduce_add_float(<4 x float>) nounwind readonly alwaysinline {
+  %v1 = call <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float> %0, <4 x float> %0)
+  %v2 = call <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float> %v1, <4 x float> %v1)
+  %scalar = extractelement <4 x float> %v2, i32 0
+  ret float %scalar
+}
+
+define float @__reduce_min_float(<4 x float>) nounwind readnone {
+  reduce4(float, @__min_varying_float, @__min_uniform_float)
+}
+
+define float @__reduce_max_float(<4 x float>) nounwind readnone {
+  reduce4(float, @__max_varying_float, @__max_uniform_float)
+}
+
+define i32 @__reduce_add_int32(<4 x i32> %v) nounwind readnone {
+  %v1 = shufflevector <4 x i32> %v, <4 x i32> undef,
+                      <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %m1 = add <4 x i32> %v1, %v
+  %m1a = extractelement <4 x i32> %m1, i32 0
+  %m1b = extractelement <4 x i32> %m1, i32 1
+  %sum = add i32 %m1a, %m1b
+  ret i32 %sum
+}
+
+define i32 @__reduce_min_int32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__min_varying_int32, @__min_uniform_int32)
+}
+
+define i32 @__reduce_max_int32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__max_varying_int32, @__max_uniform_int32)
+}
+
+define i32 @__reduce_add_uint32(<4 x i32> %v) nounwind readnone {
+  %r = call i32 @__reduce_add_int32(<4 x i32> %v)
+  ret i32 %r
+}
+
+define i32 @__reduce_min_uint32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__min_varying_uint32, @__min_uniform_uint32)
+}
+
+define i32 @__reduce_max_uint32(<4 x i32>) nounwind readnone {
+  reduce4(i32, @__max_varying_uint32, @__max_uniform_uint32)
+ }
+
+
+define double @__reduce_add_double(<4 x double>) nounwind readnone {
+  %v0 = shufflevector <4 x double> %0, <4 x double> undef,
+                      <2 x i32> <i32 0, i32 1>
+  %v1 = shufflevector <4 x double> %0, <4 x double> undef,
+                      <2 x i32> <i32 2, i32 3>
+  %sum = fadd <2 x double> %v0, %v1
+  %e0 = extractelement <2 x double> %sum, i32 0
+  %e1 = extractelement <2 x double> %sum, i32 1
+  %m = fadd double %e0, %e1
+  ret double %m
+}
+
+define double @__reduce_min_double(<4 x double>) nounwind readnone {
+  reduce4(double, @__min_varying_double, @__min_uniform_double)
+}
+
+define double @__reduce_max_double(<4 x double>) nounwind readnone {
+  reduce4(double, @__max_varying_double, @__max_uniform_double)
+}
+
+define i64 @__reduce_add_int64(<4 x i64>) nounwind readnone {
+  %v0 = shufflevector <4 x i64> %0, <4 x i64> undef,
+                      <2 x i32> <i32 0, i32 1>
+  %v1 = shufflevector <4 x i64> %0, <4 x i64> undef,
+                      <2 x i32> <i32 2, i32 3>
+  %sum = add <2 x i64> %v0, %v1
+  %e0 = extractelement <2 x i64> %sum, i32 0
+  %e1 = extractelement <2 x i64> %sum, i32 1
+  %m = add i64 %e0, %e1
+  ret i64 %m
+}
+
+define i64 @__reduce_min_int64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+define i64 @__reduce_max_int64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+define i64 @__reduce_min_uint64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+define i64 @__reduce_max_uint64(<4 x i64>) nounwind readnone {
+  reduce4(i64, @__max_varying_uint64, @__max_uniform_uint64)
+}
+
+reduce_equal(4)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+declare <4 x float> @llvm.x86.sse41.blendvps(<4 x float>, <4 x float>,
+                                             <4 x float>) nounwind readnone
+
+
+define void @__masked_store_blend_32(<4 x i32>* nocapture, <4 x i32>, 
+                                     <4 x i32> %mask) nounwind alwaysinline {
+  %mask_as_float = bitcast <4 x i32> %mask to <4 x float>
+  %oldValue = load <4 x i32>* %0, align 4
+  %oldAsFloat = bitcast <4 x i32> %oldValue to <4 x float>
+  %newAsFloat = bitcast <4 x i32> %1 to <4 x float>
+  %blend = call <4 x float> @llvm.x86.sse41.blendvps(<4 x float> %oldAsFloat,
+                                                     <4 x float> %newAsFloat,
+                                                     <4 x float> %mask_as_float)
+  %blendAsInt = bitcast <4 x float> %blend to <4 x i32>
+  store <4 x i32> %blendAsInt, <4 x i32>* %0, align 4
+  ret void
+}
+
+
+define void @__masked_store_blend_64(<4 x i64>* nocapture %ptr, <4 x i64> %new,
+                                     <4 x i32> %i32mask) nounwind alwaysinline {
+  %oldValue = load <4 x i64>* %ptr, align 8
+  %mask = bitcast <4 x i32> %i32mask to <4 x float>
+
+  ; Do 4x64-bit blends by doing two <4 x i32> blends, where the <4 x i32> values
+  ; are actually bitcast <2 x i64> values
+  ;
+  ; set up the first two 64-bit values
+  %old01  = shufflevector <4 x i64> %oldValue, <4 x i64> undef,
+                          <2 x i32> <i32 0, i32 1>
+  %old01f = bitcast <2 x i64> %old01 to <4 x float>
+  %new01  = shufflevector <4 x i64> %new, <4 x i64> undef,
+                          <2 x i32> <i32 0, i32 1>
+  %new01f = bitcast <2 x i64> %new01 to <4 x float>
+  ; compute mask--note that the indices 0 and 1 are doubled-up
+  %mask01 = shufflevector <4 x float> %mask, <4 x float> undef,
+                          <4 x i32> <i32 0, i32 0, i32 1, i32 1>
+  ; and blend the two of the values
+  %result01f = call <4 x float> @llvm.x86.sse41.blendvps(<4 x float> %old01f,
+                                                         <4 x float> %new01f,
+                                                         <4 x float> %mask01)
+  %result01 = bitcast <4 x float> %result01f to <2 x i64>
+
+  ; and again
+  %old23  = shufflevector <4 x i64> %oldValue, <4 x i64> undef,
+                          <2 x i32> <i32 2, i32 3>
+  %old23f = bitcast <2 x i64> %old23 to <4 x float>
+  %new23  = shufflevector <4 x i64> %new, <4 x i64> undef,
+                          <2 x i32> <i32 2, i32 3>
+  %new23f = bitcast <2 x i64> %new23 to <4 x float>
+  ; compute mask--note that the values 2 and 3 are doubled-up
+  %mask23 = shufflevector <4 x float> %mask, <4 x float> undef,
+                          <4 x i32> <i32 2, i32 2, i32 3, i32 3>
+  ; and blend the two of the values
+  %result23f = call <4 x float> @llvm.x86.sse41.blendvps(<4 x float> %old23f,
+                                                         <4 x float> %new23f,
+                                                         <4 x float> %mask23)
+  %result23 = bitcast <4 x float> %result23f to <2 x i64>
+
+  ; reconstruct the final <4 x i64> vector
+  %final = shufflevector <2 x i64> %result01, <2 x i64> %result23,
+                         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  store <4 x i64> %final, <4 x i64> * %ptr, align 8
+  ret void
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; masked store
+
+masked_store_blend_8_16_by_4()
+
+gen_masked_store(4, i8, 8)
+gen_masked_store(4, i16, 16)
+gen_masked_store(4, i32, 32)
+gen_masked_store(4, i64, 64)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; unaligned loads/loads+broadcasts
+
+load_and_broadcast(4, i8, 8)
+load_and_broadcast(4, i16, 16)
+load_and_broadcast(4, i32, 32)
+load_and_broadcast(4, i64, 64)
+
+load_masked(4, i8,  8,  1)
+load_masked(4, i16, 16, 2)
+load_masked(4, i32, 32, 4)
+load_masked(4, i64, 64, 8)
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; gather/scatter
+
+; define these with the macros from stdlib.m4
+
+gen_gather(4, i8)
+gen_gather(4, i16)
+gen_gather(4, i32)
+gen_gather(4, i64)
+
+gen_scatter(4, i8)
+gen_scatter(4, i16)
+gen_scatter(4, i32)
+gen_scatter(4, i64)

builtins.h

@@ -55,4 +55,7 @@
 void DefineStdlib(SymbolTable *symbolTable, llvm::LLVMContext *ctx, llvm::Module *module,
                   bool includeStdlib);
 
+void AddBitcodeToModule(const unsigned char *bitcode, int length,
+                        llvm::Module *module, SymbolTable *symbolTable = NULL);
+
 #endif // ISPC_STDLIB_H

contrib/ispc.vim

@@ -0,0 +1,32 @@
+" Vim syntax file
+" Language:	ISPC
+" Maintainer:	Andreas Wendleder <andreas.wendleder@gmail.com>
+" Last Change:	2011 Aug 3
+
+" Quit when a syntax file was already loaded
+if exists("b:current_syntax")
+  finish
+endif
+
+" Read the C syntax to start with
+runtime! syntax/c.vim
+unlet b:current_syntax
+
+" New keywords
+syn keyword	ispcStatement	cbreak ccontinue creturn launch print reference soa sync task
+syn keyword	ispcConditional	cif
+syn keyword	ispcRepeat	cdo cfor cwhile
+syn keyword	ispcBuiltin	programCount programIndex	
+syn keyword	ispcType	export int8 int16 int32 int64
+
+" Default highlighting
+command -nargs=+ HiLink hi def link <args>
+HiLink ispcStatement	Statement
+HiLink ispcConditional	Conditional
+HiLink ispcRepeat	Repeat
+HiLink ispcBuiltin	Statement
+HiLink ispcType		Type
+delcommand HiLink
+
+let b:current_syntax = "ispc"
+

ctx.h

@@ -41,9 +41,7 @@
 #include "ispc.h"
 #include <llvm/InstrTypes.h>
 #include <llvm/Instructions.h>
-#ifndef LLVM_2_8
 #include <llvm/Analysis/DIBuilder.h>
-#endif
 #include <llvm/Analysis/DebugInfo.h>
 
 struct CFInfo;
@@ -59,17 +57,22 @@ struct CFInfo;
 class FunctionEmitContext {
 public:
     /** Create a new FunctionEmitContext.
-        @param returnType   The return type of the function
-        @param function     LLVM function in the current module that corresponds
-                            to the function
+        @param function     The Function object representing the function
         @param funSym       Symbol that corresponds to the function
+        @param llvmFunction LLVM function in the current module that corresponds
+                            to the function
         @param firstStmtPos Source file position of the first statement in the
                             function
      */
-    FunctionEmitContext(const Type *returnType, llvm::Function *function, Symbol *funSym,
+    FunctionEmitContext(Function *function, Symbol *funSym, 
+                        llvm::Function *llvmFunction,
                         SourcePos firstStmtPos);
     ~FunctionEmitContext();
 
+    /** Returns the Function * corresponding to the function that we're
+        currently generating code for. */
+    const Function *GetFunction() const;
+
     /** @name Current basic block management
         @{
      */
@@ -83,20 +86,33 @@ public:
     /** @name Mask management
         @{
      */
-    /** Returns the current mask value */ 
-    llvm::Value *GetMask();
+    /** Returns the mask value at entry to the current function. */ 
+    llvm::Value *GetFunctionMask();
+
+    /** Returns the mask value corresponding to "varying" control flow
+        within the current function.  (i.e. this doesn't include the effect
+        of the mask at function entry. */
+    llvm::Value *GetInternalMask();
+
+    /** Returns the complete current mask value--i.e. the logical AND of
+        the function entry mask and the internal mask. */ 
+    llvm::Value *GetFullMask();
+
+    /** Provides the alloca'd pointer to memory to store the full function
+        mask.  This is only used to wire up the __mask builtin variable. */
+    void SetMaskPointer(llvm::Value *p);
 
     /** Provides the value of the mask at function entry */
-    void SetEntryMask(llvm::Value *val);
+    void SetFunctionMask(llvm::Value *val);
 
-    /** Sets the mask to a new value */
-    void SetMask(llvm::Value *val);
+    /** Sets the internal mask to a new value */
+    void SetInternalMask(llvm::Value *val);
 
-    /** Sets the mask to (oldMask & val) */
-    void MaskAnd(llvm::Value *oldMask, llvm::Value *val);
+    /** Sets the internal mask to (oldMask & val) */
+    void SetInternalMaskAnd(llvm::Value *oldMask, llvm::Value *val);
 
-    /** Sets the mask to (oldMask & ~val) */
-    void MaskAndNot(llvm::Value *oldMask, llvm::Value *test);
+    /** Sets the internal mask to (oldMask & ~val) */
+    void SetInternalMaskAndNot(llvm::Value *oldMask, llvm::Value *test);
 
     /** Emits a branch instruction to the basic block btrue if any of the
         lanes of current mask are on and bfalse if none are on. */
@@ -115,9 +131,8 @@ public:
         @{
     */
     /** Notifies the FunctionEmitContext that we're starting emission of an
-        'if' statement with a uniform test.  The value of the mask going
-        into the 'if' statement is provided in the oldMask parameter. */
-    void StartUniformIf(llvm::Value *oldMask);
+        'if' statement with a uniform test.  */
+    void StartUniformIf();
 
     /** Notifies the FunctionEmitContext that we're starting emission of an
         'if' statement with a varying test.  The value of the mask going
@@ -132,10 +147,9 @@ public:
         for a loop.  Basic blocks are provides for where 'break' and
         'continue' statements should jump to (if all running lanes want to
         break or continue), uniformControlFlow indicates whether the loop
-        condition is 'uniform', and oldMask provides the current mask going
-        into the loop. */
+        condition is 'uniform'. */
     void StartLoop(llvm::BasicBlock *breakTarget, llvm::BasicBlock *continueTarget, 
-                   bool uniformControlFlow, llvm::Value *oldMask);
+                   bool uniformControlFlow);
 
     /** Informs FunctionEmitContext of the value of the mask at the start
         of a loop body. */
@@ -145,6 +159,13 @@ public:
         finished. */
     void EndLoop();
 
+    /** Indicates that code generation for a 'foreach' or 'foreach_tiled'
+        loop is about to start.  The provided basic block pointer indicates
+        where control flow should go if a 'continue' statement is executed
+        in the loop. */
+    void StartForeach(llvm::BasicBlock *continueTarget);
+    void EndForeach();
+
     /** Emit code for a 'break' statement in a loop.  If doCoherenceCheck
         is true, then if we're in a 'varying' loop, code will be emitted to
         see if all of the lanes want to break, in which case a jump to the
@@ -169,6 +190,8 @@ public:
         flow */
     int VaryingCFDepth() const;
 
+    bool InForeachLoop() const;
+
     /** Called to generate code for 'return' statement; value is the
         expression in the return statement (if non-NULL), and
         doCoherenceCheck indicates whether instructions should be generated
@@ -210,16 +233,6 @@ public:
         i32. */
     llvm::Value *I1VecToBoolVec(llvm::Value *b);
 
-    /** Emit code to call the user-supplied ISPCMalloc function to
-        allocate space for an object of thee given type.  Returns the
-        pointer value returned by the ISPCMalloc call. */
-    llvm::Value *EmitMalloc(const llvm::Type *ty);
-
-    /** Emit code to call the user-supplied ISPCFree function, passing it
-        the given pointer to storage previously allocated by an
-        EmitMalloc() call. */
-    void EmitFree(llvm::Value *ptr);
-
     /** If the user has asked to compile the program with instrumentation,
         this inserts a callback to the user-supplied instrumentation
         function at the current point in the code. */
@@ -303,43 +316,60 @@ public:
                          llvm::CmpInst::Predicate pred,
                          llvm::Value *v0, llvm::Value *v1, const char *name = NULL);
 
-    llvm::Value *BitCastInst(llvm::Value *value, const llvm::Type *type,
+    /** Given a scalar value, return a vector of the same type (or an
+        array, for pointer types). */
+    llvm::Value *SmearUniform(llvm::Value *value, const char *name = NULL);
+
+    llvm::Value *BitCastInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type,
                              const char *name = NULL);
-    llvm::Instruction *PtrToIntInst(llvm::Value *value, const llvm::Type *type,
-                                    const char *name = NULL);
-    llvm::Instruction *IntToPtrInst(llvm::Value *value, const llvm::Type *type,
-                                    const char *name = NULL);
-    llvm::Instruction *TruncInst(llvm::Value *value, const llvm::Type *type,
+    llvm::Value *PtrToIntInst(llvm::Value *value, const char *name = NULL);
+    llvm::Value *PtrToIntInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type,
+                              const char *name = NULL);
+    llvm::Value *IntToPtrInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type,
+                              const char *name = NULL);
+
+    llvm::Instruction *TruncInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type,
                                  const char *name = NULL);
     llvm::Instruction *CastInst(llvm::Instruction::CastOps op, llvm::Value *value,
-                                const llvm::Type *type, const char *name = NULL);
-    llvm::Instruction *FPCastInst(llvm::Value *value, const llvm::Type *type, 
+                                LLVM_TYPE_CONST llvm::Type *type, const char *name = NULL);
+    llvm::Instruction *FPCastInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type, 
                                   const char *name = NULL);
-    llvm::Instruction *SExtInst(llvm::Value *value, const llvm::Type *type, 
+    llvm::Instruction *SExtInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type, 
                                 const char *name = NULL);
-    llvm::Instruction *ZExtInst(llvm::Value *value, const llvm::Type *type, 
+    llvm::Instruction *ZExtInst(llvm::Value *value, LLVM_TYPE_CONST llvm::Type *type, 
                                 const char *name = NULL);
 
-    /** This GEP method is a generalization of the standard one in LLVM; it
-        supports both uniform and varying basePtr values (an array of
-        pointers) as well as uniform and varying index values (arrays of
-        indices). */
+    /** These GEP methods are generalizations of the standard ones in LLVM;
+        they support both uniform and varying basePtr values as well as
+        uniform and varying index values (arrays of indices).  Varying base
+        pointers are expected to come in as vectors of i32/i64 (depending
+        on the target), since LLVM doesn't currently support vectors of
+        pointers.  The underlying type of the base pointer must be provided
+        via the ptrType parameter */
+    llvm::Value *GetElementPtrInst(llvm::Value *basePtr, llvm::Value *index,
+                                   const Type *ptrType, const char *name = NULL);
     llvm::Value *GetElementPtrInst(llvm::Value *basePtr, llvm::Value *index0,
-                                   llvm::Value *index1, const char *name = NULL);
-
-    /** This is a convenience method to generate a GEP instruction with
-        indices with values with known constant values as the ispc program
-        is being compiled. */
-    llvm::Value *GetElementPtrInst(llvm::Value *basePtr, int v0, int v1,
+                                   llvm::Value *index1, const Type *ptrType,
                                    const char *name = NULL);
 
-    /** Load from the memory location(s) given by lvalue.  The lvalue may
-        be varying, in which case this corresponds to a gather from the
-        multiple memory locations given by the array of pointer values
-        given by the lvalue.  If the lvalue is not varying, then the type
-        parameter may be NULL. */
-    llvm::Value *LoadInst(llvm::Value *lvalue, const Type *type,
-                          const char *name = NULL);
+    /** This method returns a new pointer that represents offsetting the
+        given base pointer to point at the given element number of the
+        structure type that the base pointer points to.  (The provided
+        pointer must be a pointer to a structure type.  The ptrType gives
+        the type of the pointer, though it may be NULL if the base pointer
+        is uniform. */
+    llvm::Value *AddElementOffset(llvm::Value *basePtr, int elementNum,
+                                  const Type *ptrType, const char *name = NULL);
+
+    /** Load from the memory location(s) given by lvalue, using the given
+        mask.  The lvalue may be varying, in which case this corresponds to
+        a gather from the multiple memory locations given by the array of
+        pointer values given by the lvalue.  If the lvalue is not varying,
+        then both the mask pointer and the type pointer may be NULL. */
+    llvm::Value *LoadInst(llvm::Value *ptr, llvm::Value *mask,
+                          const Type *ptrType, const char *name = NULL);
+
+    llvm::Value *LoadInst(llvm::Value *ptr, const char *name = NULL);
 
     /** Emits an alloca instruction to allocate stack storage for the given
         type.  If a non-zero alignment is specified, the object is also
@@ -347,21 +377,20 @@ public:
         instruction is added at the start of the function in the entry
         basic block; if it should be added to the current basic block, then
         the atEntryBlock parameter should be false. */ 
-    llvm::Value *AllocaInst(const llvm::Type *llvmType, const char *name = NULL,
-                            int align = 0, bool atEntryBlock = true);
+    llvm::Value *AllocaInst(LLVM_TYPE_CONST llvm::Type *llvmType, 
+                            const char *name = NULL, int align = 0, 
+                            bool atEntryBlock = true);
 
     /** Standard store instruction; for this variant, the lvalue must be a
         single pointer, not a varying lvalue. */
-    void StoreInst(llvm::Value *rvalue, llvm::Value *lvalue, 
-                   const char *name = NULL);
+    void StoreInst(llvm::Value *value, llvm::Value *ptr);
 
     /** In this variant of StoreInst(), the lvalue may be varying.  If so,
         this corresponds to a scatter.  Whether the lvalue is uniform of
         varying, the given storeMask is used to mask the stores so that
         they only execute for the active program instances. */
-    void StoreInst(llvm::Value *rvalue, llvm::Value *lvalue,
-                   llvm::Value *storeMask, const Type *rvalueType,
-                   const char *name = NULL);
+    void StoreInst(llvm::Value *value, llvm::Value *ptr,
+                   llvm::Value *storeMask, const Type *ptrType);
 
     void BranchInst(llvm::BasicBlock *block);
     void BranchInst(llvm::BasicBlock *trueBlock, llvm::BasicBlock *falseBlock,
@@ -378,32 +407,45 @@ public:
     llvm::Value *InsertInst(llvm::Value *v, llvm::Value *eltVal, int elt, 
                             const char *name = NULL);
 
-    llvm::PHINode *PhiNode(const llvm::Type *type, int count, const char *name = NULL);
+    llvm::PHINode *PhiNode(LLVM_TYPE_CONST llvm::Type *type, int count, 
+                           const char *name = NULL);
     llvm::Instruction *SelectInst(llvm::Value *test, llvm::Value *val0,
                                   llvm::Value *val1, const char *name = NULL);
 
-    llvm::Instruction *CallInst(llvm::Function *func, 
-                                const std::vector<llvm::Value *> &args,
-                                const char *name = NULL);
+    /** Emits IR to do a function call with the given arguments.  If the
+        function type is a varying function pointer type, its full type
+        must be provided in funcType.  funcType can be NULL if func is a
+        uniform function pointer. */
+    llvm::Value *CallInst(llvm::Value *func, const FunctionType *funcType,
+                          const std::vector<llvm::Value *> &args,
+                          const char *name = NULL);
+
     /** This is a convenience method that issues a call instruction to a
         function that takes just a single argument. */
-    llvm::Instruction *CallInst(llvm::Function *func, llvm::Value *arg,
-                                const char *name = NULL);
+    llvm::Value *CallInst(llvm::Value *func, const FunctionType *funcType,
+                          llvm::Value *arg, const char *name = NULL);
 
     /** This is a convenience method that issues a call instruction to a
         function that takes two arguments. */
-    llvm::Instruction *CallInst(llvm::Function *func, llvm::Value *arg0,
-                                llvm::Value *arg1, const char *name = NULL);
+    llvm::Value *CallInst(llvm::Value *func, const FunctionType *funcType,
+                          llvm::Value *arg0, llvm::Value *arg1,
+                          const char *name = NULL);
 
     /** Launch an asynchronous task to run the given function, passing it
         he given argument values. */
-    llvm::Instruction *LaunchInst(llvm::Function *callee, 
-                                  std::vector<llvm::Value *> &argVals);
+    llvm::Value *LaunchInst(llvm::Value *callee, 
+                            std::vector<llvm::Value *> &argVals,
+                            llvm::Value *launchCount);
+
+    void SyncInst();
 
     llvm::Instruction *ReturnInst();
     /** @} */
 
 private:
+    /** Pointer to the Function for which we're currently generating code. */
+    Function *function;
+
     /** The basic block into which we add any alloca instructions that need
         to go at the very start of the function. */
     llvm::BasicBlock *allocaBlock;
@@ -413,8 +455,16 @@ private:
     llvm::BasicBlock *bblock;
 
     /** Pointer to stack-allocated memory that stores the current value of
-        the program mask. */
-    llvm::Value *maskPtr;
+        the full program mask. */
+    llvm::Value *fullMaskPointer;
+
+    /** Pointer to stack-allocated memory that stores the current value of
+        the program mask representing varying control flow within the
+        function. */
+    llvm::Value *internalMaskPointer;
+
+    /** Value of the program mask when the function starts execution.  */
+    llvm::Value *functionMaskValue;
 
     /** Current source file position; if debugging information is being
         generated, this position is used to set file/line information for
@@ -425,12 +475,6 @@ private:
         for error messages and debugging symbols. */
     SourcePos funcStartPos;
 
-    /** Type of result that the current function returns. */
-    const Type *returnType;
-
-    /** Value of the program mask when the function starts execution.  */
-    llvm::Value *entryMask;
-
     /** If currently in a loop body, the value of the mask at the start of
         the loop. */
     llvm::Value *loopMask;
@@ -488,20 +532,29 @@ private:
     /** True if a 'launch' statement has been encountered in the function. */
     bool launchedTasks;
 
+    /** This is a pointer to a void * that is passed to the ISPCLaunch(),
+        ISPCAlloc(), and ISPCSync() routines as a handle to the group ot
+        tasks launched from the current function. */
+    llvm::Value *launchGroupHandlePtr;
+
     llvm::Value *pointerVectorToVoidPointers(llvm::Value *value);
-    static void addGSMetadata(llvm::Instruction *inst, SourcePos pos);
+    static void addGSMetadata(llvm::Value *inst, SourcePos pos);
     bool ifsInLoopAllUniform() const;
     void jumpIfAllLoopLanesAreDone(llvm::BasicBlock *target);
     llvm::Value *emitGatherCallback(llvm::Value *lvalue, llvm::Value *retPtr);
 
+    llvm::Value *applyVaryingGEP(llvm::Value *basePtr, llvm::Value *index, 
+                                 const Type *ptrType);
+
     void restoreMaskGivenReturns(llvm::Value *oldMask);
 
-    void scatter(llvm::Value *rvalue, llvm::Value *lvalue, 
-                 llvm::Value *maskPtr, const Type *rvalueType);
-    llvm::Value *gather(llvm::Value *lvalue, const Type *type,
+    void scatter(llvm::Value *value, llvm::Value *ptr, const Type *ptrType, 
+                 llvm::Value *mask);
+    void maskedStore(llvm::Value *value, llvm::Value *ptr, const Type *ptrType,
+                     llvm::Value *mask);
+    llvm::Value *gather(llvm::Value *ptr, const Type *ptrType, llvm::Value *mask,
                         const char *name);
-    void maskedStore(llvm::Value *rvalue, llvm::Value *lvalue,
-                     const Type *rvalueType, llvm::Value *maskPtr);
+    llvm::Value *addVaryingOffsetsIfNeeded(llvm::Value *ptr, const Type *ptrType);
 };
 
 #endif // ISPC_CTX_H

decl.cpp

@@ -38,10 +38,59 @@
 
 #include "decl.h"
 #include "util.h"
+#include "module.h"
 #include "sym.h"
 #include "type.h"
+#include "stmt.h"
 #include "expr.h"
 #include <stdio.h>
+#include <set>
+
+/** Given a Type and a set of type qualifiers, apply the type qualifiers to
+    the type, returning the type that is the result. 
+*/
+static const Type *
+lApplyTypeQualifiers(int typeQualifiers, const Type *type, SourcePos pos) {
+    if (type == NULL)
+        return NULL;
+
+    if ((typeQualifiers & TYPEQUAL_UNSIGNED) != 0) {
+        if ((typeQualifiers & TYPEQUAL_SIGNED) != 0)
+            Error(pos, "Illegal to apply both \"signed\" and \"unsigned\" "
+                  "qualifiers.");
+
+        const Type *unsignedType = type->GetAsUnsignedType();
+        if (unsignedType != NULL)
+            type = unsignedType;
+        else
+            Error(pos, "\"unsigned\" qualifier is illegal with \"%s\" type.",
+              type->GetString().c_str());
+
+    }
+
+    if ((typeQualifiers & TYPEQUAL_SIGNED) != 0 && type->IsIntType() == false)
+        Error(pos, "\"signed\" qualifier is illegal with non-integer type "
+              "\"%s\".", type->GetString().c_str());
+
+    if ((typeQualifiers & TYPEQUAL_CONST) != 0)
+        type = type->GetAsConstType();
+
+    if ((typeQualifiers & TYPEQUAL_UNIFORM) != 0)
+        type = type->GetAsUniformType();
+    else if ((typeQualifiers & TYPEQUAL_VARYING) != 0)
+        type = type->GetAsVaryingType();
+    else {
+        // otherwise, structs are uniform by default and everything
+        // else is varying by default
+        if (dynamic_cast<const StructType *>(type->GetBaseType()) != NULL)
+            type = type->GetAsUniformType();
+        else
+            type = type->GetAsVaryingType();
+    }
+
+    return type;
+}
+
 
 ///////////////////////////////////////////////////////////////////////////
 // DeclSpecs
@@ -49,29 +98,57 @@
 DeclSpecs::DeclSpecs(const Type *t, StorageClass sc, int tq) {
     baseType = t;
     storageClass = sc;
-    typeQualifier = tq;
+    typeQualifiers = tq;
     soaWidth = 0;
     vectorSize = 0;
 }
 
 
+const Type *
+DeclSpecs::GetBaseType(SourcePos pos) const {
+    const Type *bt = baseType;
+    if (vectorSize > 0) {
+        const AtomicType *atomicType = dynamic_cast<const AtomicType *>(bt);
+        if (atomicType == NULL) {
+            Error(pos, "Only atomic types (int, float, ...) are legal for vector "
+                  "types.");
+            return NULL;
+        }
+        bt = new VectorType(atomicType, vectorSize);
+    }
+
+    return lApplyTypeQualifiers(typeQualifiers, bt, pos);
+}
+
+
+static const char *
+lGetStorageClassName(StorageClass storageClass) {
+    switch (storageClass) {
+    case SC_NONE:     return "";
+    case SC_EXTERN:   return "extern";
+    case SC_EXTERN_C: return "extern \"C\"";
+    case SC_EXPORT:   return "export";
+    case SC_STATIC:   return "static";
+    case SC_TYPEDEF:  return "typedef";
+    default:          FATAL("Unhandled storage class in lGetStorageClassName");
+                      return "";
+    }
+}
+
+
 void
 DeclSpecs::Print() const {
-    if (storageClass == SC_EXTERN)   printf("extern ");
-    if (storageClass == SC_EXTERN_C) printf("extern \"C\" ");
-    if (storageClass == SC_EXPORT)   printf("export ");
-    if (storageClass == SC_STATIC)   printf("static ");
-    if (storageClass == SC_TYPEDEF)  printf("typedef ");
+    printf("%s ", lGetStorageClassName(storageClass));
 
     if (soaWidth > 0) printf("soa<%d> ", soaWidth);
 
-    if (typeQualifier & TYPEQUAL_INLINE)    printf("inline ");
-    if (typeQualifier & TYPEQUAL_CONST)     printf("const ");
-    if (typeQualifier & TYPEQUAL_UNIFORM)   printf("uniform ");
-    if (typeQualifier & TYPEQUAL_VARYING)   printf("varying ");
-    if (typeQualifier & TYPEQUAL_TASK)      printf("task ");
-    if (typeQualifier & TYPEQUAL_REFERENCE) printf("reference ");
-    if (typeQualifier & TYPEQUAL_UNSIGNED)  printf("unsigned ");
+    if (typeQualifiers & TYPEQUAL_INLINE)    printf("inline ");
+    if (typeQualifiers & TYPEQUAL_CONST)     printf("const ");
+    if (typeQualifiers & TYPEQUAL_UNIFORM)   printf("uniform ");
+    if (typeQualifiers & TYPEQUAL_VARYING)   printf("varying ");
+    if (typeQualifiers & TYPEQUAL_TASK)      printf("task ");
+    if (typeQualifiers & TYPEQUAL_SIGNED)    printf("signed ");
+    if (typeQualifiers & TYPEQUAL_UNSIGNED)  printf("unsigned ");
 
     printf("%s", baseType->GetString().c_str());
 
@@ -82,34 +159,46 @@ DeclSpecs::Print() const {
 ///////////////////////////////////////////////////////////////////////////
 // Declarator
 
-Declarator::Declarator(Symbol *s, SourcePos p) 
-  : pos(p) { 
-    sym = s;
-    functionArgs = NULL;
-    isFunction = false;
+Declarator::Declarator(DeclaratorKind dk, SourcePos p) 
+    : pos(p), kind(dk) { 
+    child = NULL;
+    typeQualifiers = 0;
+    arraySize = -1;
+    sym = NULL;
     initExpr = NULL;
 }
 
 
 void
-Declarator::AddArrayDimension(int size) {
-    assert(size > 0 || size == -1); // -1 -> unsized
-    arraySize.push_back(size);
+Declarator::InitFromDeclSpecs(DeclSpecs *ds) {
+    const Type *t = GetType(ds);
+    Symbol *sym = GetSymbol();
+    if (sym != NULL) {
+        sym->type = t;
+        sym->storageClass = ds->storageClass;
+    }
 }
 
 
-void
-Declarator::InitFromDeclSpecs(DeclSpecs *ds) {
-    sym->type = GetType(ds);
-
-    if (ds->storageClass == SC_STATIC)
-        sym->isStatic = true;
+Symbol *
+Declarator::GetSymbol() const {
+    // The symbol lives at the last child in the chain, so walk down there
+    // and return the one there.
+    const Declarator *d = this;
+    while (d->child != NULL)
+        d = d->child;
+    return d->sym;
 }
 
 
 void
 Declarator::Print() const {
-    printf("%s", sym->name.c_str());
+    Symbol *sym = GetSymbol();
+    if (sym != NULL)
+        printf("%s", sym->name.c_str());
+    else
+        printf("(null symbol)");
+
     if (initExpr != NULL) {
         printf(" = (");
         initExpr->Print();
@@ -119,188 +208,305 @@ Declarator::Print() const {
 }
 
 
-static const Type *
-lGetType(const Declarator *decl, DeclSpecs *ds, 
-         std::vector<int>::const_iterator arrayIter) {
-    if (arrayIter == decl->arraySize.end()) {
-        // If we don't have an array (or have processed all of the array
-        // dimensions in previous recursive calls), we can go ahead and
-        // figure out the final non-array type we have here.
-        const Type *type = ds->baseType;
-        if (type == NULL) {
-            Error(decl->pos, "Type not provided in variable declaration for variable \"%s\".",
-                  decl->sym->name.c_str());
-            return NULL;
-        }
+Symbol *
+Declarator::GetFunctionInfo(DeclSpecs *ds, std::vector<Symbol *> *funArgs) {
+    const FunctionType *type = 
+        dynamic_cast<const FunctionType *>(GetType(ds));
+    if (type == NULL)
+        return NULL;
 
-        // Account for 'unsigned' and 'const' qualifiers in the type
-        if ((ds->typeQualifier & TYPEQUAL_UNSIGNED) != 0) {
-            const Type *unsignedType = type->GetAsUnsignedType();
-            if (unsignedType != NULL)
-                type = unsignedType;
-            else
-                Error(decl->pos, "\"unsigned\" qualifier is illegal with \"%s\" type.",
-                      type->GetString().c_str());
-        }
-        if ((ds->typeQualifier & TYPEQUAL_CONST) != 0)
-            type = type->GetAsConstType();
+    Symbol *declSym = GetSymbol();
+    assert(declSym != NULL);
 
-        if (ds->vectorSize > 0) {
-            const AtomicType *atomicType = dynamic_cast<const AtomicType *>(type);
-            if (atomicType == NULL) {
-                Error(decl->pos, "Only atomic types (int, float, ...) are legal for vector "
-                      "types.");
-                return NULL;
-            }
-            type = new VectorType(atomicType, ds->vectorSize);
-        }
+    // Get the symbol for the function from the symbol table.  (It should
+    // already have been added to the symbol table by AddGlobal() by the
+    // time we get here.)
+    Symbol *funSym = m->symbolTable->LookupFunction(declSym->name.c_str(), type);
+    if (funSym != NULL)
+        // May be NULL due to error earlier in compilation
+        funSym->pos = pos;
 
-        // if uniform/varying is specified explicitly, then go with that
-        if ((ds->typeQualifier & TYPEQUAL_UNIFORM) != 0)
-            return type->GetAsUniformType();
-        else if ((ds->typeQualifier & TYPEQUAL_VARYING) != 0)
-            return type->GetAsVaryingType();
-        else {
-            // otherwise, structs are uniform by default and everything
-            // else is varying by default
-            if (dynamic_cast<const StructType *>(type) != NULL)
-                return type->GetAsUniformType();
-            else
-                return type->GetAsVaryingType();
-        }
+    // Walk down to the declarator for the function.  (We have to get past
+    // the stuff that specifies the function's return type before we get to
+    // the function's declarator.)
+    Declarator *d = this;
+    while (d != NULL && d->kind != DK_FUNCTION)
+        d = d->child;
+    assert(d != NULL);
+
+    for (unsigned int i = 0; i < d->functionParams.size(); ++i) {
+        Declaration *pdecl = d->functionParams[i];
+        assert(pdecl->declarators.size() == 1);
+        funArgs->push_back(pdecl->declarators[0]->GetSymbol());
     }
-    else {
-        // Peel off one dimension of the array
-        int arraySize = *arrayIter;
-        ++arrayIter;
 
-        // Get the type, not including the arraySize dimension peeled off
-        // above.
-        const Type *childType = lGetType(decl, ds, arrayIter);
-
-        int soaWidth = ds->soaWidth;
-        if (soaWidth == 0)
-            // If there's no "soa<n>" stuff going on, just return a regular
-            // array with the appropriate size 
-            return new ArrayType(childType, arraySize == -1 ? 0 : arraySize);
-       else {
-            // Make sure we actually have an array of structs ..
-            const StructType *childStructType = 
-                dynamic_cast<const StructType *>(childType);
-            if (childStructType == NULL) {
-                Error(decl->pos, "Illegal to provide soa<%d> qualifier with non-struct "
-                      "type \"%s\".", soaWidth, childType->GetString().c_str());
-                return new ArrayType(childType, arraySize == -1 ? 0 : arraySize);
-            }
-            else if ((soaWidth & (soaWidth - 1)) != 0) {
-                Error(decl->pos, "soa<%d> width illegal.  Value must be power of two.",
-                      soaWidth);
-                return NULL;
-            }
-            else if (arraySize != -1 && (arraySize % soaWidth) != 0) {
-                Error(decl->pos, "soa<%d> width must evenly divide array size %d.",
-                      soaWidth, arraySize);
-                return NULL;
-            }
-            return new SOAArrayType(childStructType, arraySize == -1 ? 0 : arraySize,
-                                    soaWidth);
-        }
-    }
+    return funSym;
 }
 
 
 const Type *
-Declarator::GetType(DeclSpecs *ds) const {
-    bool hasUniformQual = ((ds->typeQualifier & TYPEQUAL_UNIFORM) != 0);
-    bool hasVaryingQual = ((ds->typeQualifier & TYPEQUAL_VARYING) != 0);
-    bool isTask =         ((ds->typeQualifier & TYPEQUAL_TASK) != 0);
-    bool isReference =    ((ds->typeQualifier & TYPEQUAL_REFERENCE) != 0);
+Declarator::GetType(const Type *base, DeclSpecs *ds) const {
+    bool hasUniformQual = ((typeQualifiers & TYPEQUAL_UNIFORM) != 0);
+    bool hasVaryingQual = ((typeQualifiers & TYPEQUAL_VARYING) != 0);
+    bool isTask =         ((typeQualifiers & TYPEQUAL_TASK) != 0);
+    bool isConst =        ((typeQualifiers & TYPEQUAL_CONST) != 0);
 
     if (hasUniformQual && hasVaryingQual) {
         Error(pos, "Can't provide both \"uniform\" and \"varying\" qualifiers.");
         return NULL;
     }
+    if (kind != DK_FUNCTION && isTask)
+        Error(pos, "\"task\" qualifier illegal in variable declaration.");
 
-    if (isFunction) {
+    const Type *type = base;
+    switch (kind) {
+    case DK_BASE:
+        // All of the type qualifiers should be in the DeclSpecs for the
+        // base declarator
+        assert(typeQualifiers == 0);
+        assert(child == NULL);
+        return type;
+
+    case DK_POINTER:
+        type = new PointerType(type, hasUniformQual, isConst);
+        if (child != NULL)
+            return child->GetType(type, ds);
+        else
+            return type;
+        break;
+
+    case DK_REFERENCE:
+        if (hasUniformQual)
+            Error(pos, "\"uniform\" qualifier is illegal to apply to references.");
+        if (hasVaryingQual)
+            Error(pos, "\"varying\" qualifier is illegal to apply to references.");
+        if (isConst)
+            Error(pos, "\"const\" qualifier is to illegal apply to references.");
+
+        // The parser should disallow this already, but double check.
+        if (dynamic_cast<const ReferenceType *>(type) != NULL) {
+            Error(pos, "References to references are illegal.");
+            return NULL;
+        }
+
+        type = new ReferenceType(type);
+        if (child != NULL)
+            return child->GetType(type, ds);
+        else
+            return type;
+        break;
+
+    case DK_ARRAY:
+        type = new ArrayType(type, arraySize);
+        if (child)
+            return child->GetType(type, ds);
+        else
+            return type;
+        break;
+
+    case DK_FUNCTION: {
         std::vector<const Type *> args;
         std::vector<std::string> argNames;
-        if (functionArgs) {
-            // Loop over the function arguments and get names and types for
-            // each one in the args and argNames arrays
-            for (unsigned int i = 0; i < functionArgs->size(); ++i) {
-                Declaration *d = (*functionArgs)[i];
-                Symbol *sym;
-                if (d->declarators.size() == 0) {
-                    // function declaration like foo(float), w/o a name for
-                    // the parameter
-                    char buf[32];
-                    sprintf(buf, "__anon_parameter_%d", i);
-                    sym = new Symbol(buf, pos);
-                    Declarator *declarator = new Declarator(sym, sym->pos);
-                    sym->type = declarator->GetType(ds);
-                    d->declarators.push_back(declarator);
-                }
-                else {
-                    assert(d->declarators.size() == 1);
-                    sym = d->declarators[0]->sym;
-                }
+        std::vector<ConstExpr *> argDefaults;
+        std::vector<SourcePos> argPos;
 
-                // Arrays are passed by reference, so convert array
-                // parameters to be references here.
-                if (dynamic_cast<const ArrayType *>(sym->type) != NULL)
-                    sym->type = new ReferenceType(sym->type, sym->type->IsConstType());
+        // Loop over the function arguments and store the names, types,
+        // default values (if any), and source file positions each one in
+        // the corresponding vector.
+        for (unsigned int i = 0; i < functionParams.size(); ++i) {
+            Declaration *d = functionParams[i];
 
-                args.push_back(sym->type);
-                argNames.push_back(sym->name);
+            char buf[32];
+            Symbol *sym;
+            if (d->declarators.size() == 0) {
+                // function declaration like foo(float), w/o a name for
+                // the parameter
+                sprintf(buf, "__anon_parameter_%d", i);
+                sym = new Symbol(buf, pos);
+                sym->type = d->declSpecs->GetBaseType(pos);
             }
+            else {
+                sym = d->declarators[0]->GetSymbol();
+                if (sym == NULL) {
+                    // Handle more complex anonymous declarations like
+                    // float (float **).
+                    sprintf(buf, "__anon_parameter_%d", i);
+                    sym = new Symbol(buf, d->declarators[0]->pos);
+                    sym->type = d->declarators[0]->GetType(d->declSpecs);
+                }
+            }
+
+            if (d->declSpecs->storageClass != SC_NONE)
+                Error(sym->pos, "Storage class \"%s\" is illegal in "
+                      "function parameter declaration for parameter \"%s\".", 
+                      lGetStorageClassName(d->declSpecs->storageClass),
+                      sym->name.c_str());
+
+            const ArrayType *at = dynamic_cast<const ArrayType *>(sym->type);
+            if (at != NULL) {
+                // As in C, arrays are passed to functions as pointers to
+                // their element type.  We'll just immediately make this
+                // change now.  (One shortcoming of losing the fact that
+                // the it was originally an array is that any warnings or
+                // errors later issued that print the function type will
+                // report this differently than it was originally declared
+                // in the function, but it's not clear that this is a
+                // significant problem.)
+                sym->type = PointerType::GetUniform(at->GetElementType());
+
+                // Make sure there are no unsized arrays (other than the
+                // first dimension) in function parameter lists.
+                at = dynamic_cast<const ArrayType *>(at->GetElementType());
+                while (at != NULL) {
+                    if (at->GetElementCount() == 0)
+                        Error(sym->pos, "Arrays with unsized dimensions in "
+                              "dimensions after the first one are illegal in "
+                              "function parameter lists.");
+                    at = dynamic_cast<const ArrayType *>(at->GetElementType());
+                }
+            }
+
+            args.push_back(sym->type);
+            argNames.push_back(sym->name);
+            argPos.push_back(sym->pos);
+
+            ConstExpr *init = NULL;
+            if (d->declarators.size()) {
+                // Try to find an initializer expression; if there is one,
+                // it lives down to the base declarator.
+                Declarator *decl = d->declarators[0];
+                while (decl->child != NULL) {
+                    assert(decl->initExpr == NULL);
+                    decl = decl->child;
+                }
+
+                if (decl->initExpr != NULL &&
+                    (decl->initExpr = decl->initExpr->TypeCheck()) != NULL &&
+                    (decl->initExpr = decl->initExpr->Optimize()) != NULL &&
+                    (init = dynamic_cast<ConstExpr *>(decl->initExpr)) == NULL) {
+                    Error(decl->initExpr->pos, "Default value for parameter "
+                          "\"%s\" must be a compile-time constant.", 
+                          sym->name.c_str());
+                }
+            }
+            argDefaults.push_back(init);
         }
 
-        if (ds->baseType == NULL) {
-            Warning(pos, "No return type provided in declaration of function \"%s\". "
-                    "Treating as \"void\".", sym->name.c_str());
-            ds->baseType = AtomicType::Void;
-        }
-
-        if (isReference) {
-            Error(pos, "Function return types can't be reference types.");
+        const Type *returnType = type;
+        if (returnType == NULL) {
+            Error(pos, "No return type provided in function declaration.");
             return NULL;
         }
 
-        const Type *returnType = lGetType(this, ds, arraySize.begin());
-        if (returnType == NULL)
+        bool isExported = ds && (ds->storageClass == SC_EXPORT);
+        bool isExternC =  ds && (ds->storageClass == SC_EXTERN_C);
+        bool isTask =     ds && ((ds->typeQualifiers & TYPEQUAL_TASK) != 0);
+
+        if (isExported && isTask) {
+            Error(pos, "Function can't have both \"task\" and \"export\" "
+                  "qualifiers");
+            return NULL;
+        }
+        if (isExternC && isTask) {
+            Error(pos, "Function can't have both \"extern \"C\"\" and \"task\" "
+                  "qualifiers");
+            return NULL;
+        }
+        if (isExternC && isExported) {
+            Error(pos, "Function can't have both \"extern \"C\"\" and \"export\" "
+                  "qualifiers");
             return NULL;
-
-        bool isExported = (ds->storageClass == SC_EXPORT);
-        bool isExternC =  (ds->storageClass == SC_EXTERN_C);
-        return new FunctionType(returnType, args, pos, &argNames, isTask, 
-                                isExported, isExternC);
-    }
-    else {
-        if (isTask)
-            Error(pos, "\"task\" qualifier illegal in variable declaration \"%s\".",
-                  sym->name.c_str());
-
-        const Type *type = lGetType(this, ds, arraySize.begin());
-
-        if (type != NULL && isReference) {
-            bool hasConstQual = ((ds->typeQualifier & TYPEQUAL_CONST) != 0);
-            type = new ReferenceType(type, hasConstQual);
         }
 
-        return type;
+        Type *functionType = 
+            new FunctionType(returnType, args, pos, argNames, argDefaults,
+                             argPos, isTask, isExported, isExternC);
+        return child->GetType(functionType, ds);
     }
+    default:
+        FATAL("Unexpected decl kind");
+        return NULL;
+    }
+
+#if 0
+            // Make sure we actually have an array of structs ..
+            const StructType *childStructType = 
+                dynamic_cast<const StructType *>(childType);
+            if (childStructType == NULL) {
+                Error(pos, "Illegal to provide soa<%d> qualifier with non-struct "
+                      "type \"%s\".", soaWidth, childType->GetString().c_str());
+                return new ArrayType(childType, arraySize == -1 ? 0 : arraySize);
+            }
+            else if ((soaWidth & (soaWidth - 1)) != 0) {
+                Error(pos, "soa<%d> width illegal.  Value must be power of two.",
+                      soaWidth);
+                return NULL;
+            }
+            else if (arraySize != -1 && (arraySize % soaWidth) != 0) {
+                Error(pos, "soa<%d> width must evenly divide array size %d.",
+                      soaWidth, arraySize);
+                return NULL;
+            }
+            return new SOAArrayType(childStructType, arraySize == -1 ? 0 : arraySize,
+                                    soaWidth);
+#endif
 }
 
+
+const Type *
+Declarator::GetType(DeclSpecs *ds) const {
+    const Type *baseType = ds->GetBaseType(pos);
+    const Type *type = GetType(baseType, ds);
+    return type;
+}
+
+
 ///////////////////////////////////////////////////////////////////////////
 // Declaration
 
-void
-Declaration::AddSymbols(SymbolTable *st) const {
-    assert(declSpecs->storageClass != SC_TYPEDEF);
-
+Declaration::Declaration(DeclSpecs *ds, std::vector<Declarator *> *dlist) {
+    declSpecs = ds;
+    if (dlist != NULL)
+        declarators = *dlist;
     for (unsigned int i = 0; i < declarators.size(); ++i)
-       if (declarators[i])
-           st->AddVariable(declarators[i]->sym);
+        if (declarators[i] != NULL)
+            declarators[i]->InitFromDeclSpecs(declSpecs);
+}
+
+
+Declaration::Declaration(DeclSpecs *ds, Declarator *d) {
+    declSpecs = ds;
+    if (d != NULL) {
+        d->InitFromDeclSpecs(ds);
+        declarators.push_back(d);
+    }
+}
+
+
+std::vector<VariableDeclaration>
+Declaration::GetVariableDeclarations() const {
+    assert(declSpecs->storageClass != SC_TYPEDEF);
+    std::vector<VariableDeclaration> vars;
+
+    for (unsigned int i = 0; i < declarators.size(); ++i) {
+        if (declarators[i] == NULL)
+            continue;
+        Declarator *decl = declarators[i];
+        if (decl == NULL)
+            // Ignore earlier errors
+            continue;
+
+        Symbol *sym = decl->GetSymbol();
+        if (dynamic_cast<const FunctionType *>(sym->type) != NULL) {
+            // function declaration
+            m->symbolTable->AddFunction(sym);
+        }
+        else {
+            m->symbolTable->AddVariable(sym);
+            vars.push_back(VariableDeclaration(sym, decl->initExpr));
+        }
+    }
+    return vars;
 }
 
 
@@ -318,31 +524,48 @@ Declaration::Print() const {
 ///////////////////////////////////////////////////////////////////////////
 
 void
-GetStructTypesAndNames(const std::vector<StructDeclaration *> &sd,
-                       std::vector<const Type *> *elementTypes,
-                       std::vector<std::string> *elementNames) {
+GetStructTypesNamesPositions(const std::vector<StructDeclaration *> &sd,
+                             std::vector<const Type *> *elementTypes,
+                             std::vector<std::string> *elementNames,
+                             std::vector<SourcePos> *elementPositions) {
+    std::set<std::string> seenNames;
     for (unsigned int i = 0; i < sd.size(); ++i) {
         const Type *type = sd[i]->type;
+        if (type == NULL)
+            continue;
+
         // FIXME: making this fake little DeclSpecs here is really
         // disgusting
         DeclSpecs ds(type);
         if (type->IsUniformType()) 
-            ds.typeQualifier |= TYPEQUAL_UNIFORM;
+            ds.typeQualifiers |= TYPEQUAL_UNIFORM;
         else
-            ds.typeQualifier |= TYPEQUAL_VARYING;
+            ds.typeQualifiers |= TYPEQUAL_VARYING;
 
         for (unsigned int j = 0; j < sd[i]->declarators->size(); ++j) {
             Declarator *d = (*sd[i]->declarators)[j];
             d->InitFromDeclSpecs(&ds);
 
-            // if it's an unsized array, make it a reference to an unsized
-            // array, so the caller can pass a pointer...
-            const ArrayType *at = dynamic_cast<const ArrayType *>(d->sym->type);
-            if (at && at->GetElementCount() == 0)
-                d->sym->type = new ReferenceType(d->sym->type, type->IsConstType());
+            Symbol *sym = d->GetSymbol();
 
-            elementTypes->push_back(d->sym->type);
-            elementNames->push_back(d->sym->name);
+            const ArrayType *arrayType = 
+                dynamic_cast<const ArrayType *>(sym->type);
+            if (arrayType != NULL && arrayType->GetElementCount() == 0) {
+                Error(d->pos, "Unsized arrays aren't allowed in struct "
+                      "definitions.");
+                elementTypes->push_back(NULL);
+            }
+            else
+                elementTypes->push_back(sym->type);
+
+            if (seenNames.find(sym->name) != seenNames.end())
+                Error(d->pos, "Struct member \"%s\" has same name as a "
+                      "previously-declared member.", sym->name.c_str());
+            else
+                seenNames.insert(sym->name);
+
+            elementNames->push_back(sym->name);
+            elementPositions->push_back(sym->pos);
         }
     }
 }

decl.h

@@ -56,6 +56,11 @@
 
 #include "ispc.h"
 
+struct VariableDeclaration;
+
+class Declaration;
+class Declarator;
+
 enum StorageClass {
     SC_NONE,
     SC_EXTERN,
@@ -74,7 +79,7 @@ enum StorageClass {
 #define TYPEQUAL_UNIFORM    (1<<1)
 #define TYPEQUAL_VARYING    (1<<2)
 #define TYPEQUAL_TASK       (1<<3)
-#define TYPEQUAL_REFERENCE  (1<<4)
+#define TYPEQUAL_SIGNED     (1<<4)
 #define TYPEQUAL_UNSIGNED   (1<<5)
 #define TYPEQUAL_INLINE     (1<<6)
 
@@ -92,15 +97,17 @@ public:
     StorageClass storageClass;
 
     /** Zero or more of the TYPEQUAL_* values, ANDed together. */
-    int typeQualifier;
+    int typeQualifiers;
 
     /** The basic type provided in the declaration; this should be an
-        AtomicType, a StructType, or a VectorType; other types (like
+        AtomicType, EnumType, StructType, or VectorType; other types (like
         ArrayTypes) will end up being created if a particular declaration
         has an array size, etc.
     */
     const Type *baseType;
 
+    const Type *GetBaseType(SourcePos pos) const;
+
     /** If this is a declaration with a vector type, this gives the vector
         width.  For non-vector types, this is zero.
      */
@@ -113,6 +120,14 @@ public:
 };
 
 
+enum DeclaratorKind {
+    DK_BASE,
+    DK_POINTER,
+    DK_REFERENCE,
+    DK_ARRAY,
+    DK_FUNCTION
+};
+
 /** @brief Representation of the declaration of a single variable.  
 
     In conjunction with an instance of the DeclSpecs, this gives us
@@ -120,13 +135,7 @@ public:
  */
 class Declarator {
 public:
-    Declarator(Symbol *s, SourcePos p);
-
-    /** As the parser peels off array dimension declarations after the
-        symbol name, it calls this method to provide them to the
-        Declarator.
-     */
-    void AddArrayDimension(int size);
+    Declarator(DeclaratorKind dk, SourcePos p);
 
     /** Once a DeclSpecs instance is available, this method completes the
         initialization of the Symbol, setting its Type accordingly.
@@ -134,21 +143,51 @@ public:
     void InitFromDeclSpecs(DeclSpecs *ds);
 
     /** Get the actual type of the combination of Declarator and the given
-        DeclSpecs */
+        DeclSpecs.  If an explicit base type is provided, the declarator is
+        applied to that type; otherwise the base type from the DeclSpecs is
+        used. */
     const Type *GetType(DeclSpecs *ds) const;
+    const Type *GetType(const Type *base, DeclSpecs *ds) const;
+
+    /** Returns the symbol corresponding to the function declared by this
+        declarator and symbols for its arguments in *args. */
+    Symbol *GetFunctionInfo(DeclSpecs *ds, std::vector<Symbol *> *args);
+
+    /** Returns the symbol associated with the declarator. */
+    Symbol *GetSymbol() const;
 
     void Print() const;
 
+    /** Position of the declarator in the source program. */
     const SourcePos pos;
+
+    /** The kind of this declarator; complex declarations are assembled as
+        a hierarchy of Declarators.  (For example, a pointer to an int
+        would have a root declarator with kind DK_POINTER and with the
+        Declarator::child member pointing to a DK_BASE declarator for the
+        int). */
+    const DeclaratorKind kind;
+
+    /** Child pointer if needed; this can only be non-NULL if the
+        declarator's kind isn't DK_BASE. */
+    Declarator *child;
+
+    /** Type qualifiers provided with the declarator. */
+    int typeQualifiers;
+
+    /** For array declarators, this gives the declared size of the array.
+        Unsized arrays have arraySize == 0. */ 
+    int arraySize;
+
+    /** Symbol associated with the declarator. */
     Symbol *sym;
-    /** If this declarator includes an array specification, the sizes of
-        the array dimensions are represented here.
-     */
-    std::vector<int> arraySize;
+
     /** Initialization expression for the variable.  May be NULL. */
     Expr *initExpr;
-    bool isFunction;
-    std::vector<Declaration *> *functionArgs;
+
+    /** For function declarations, this holds the Declaration *s for the
+        funciton's parameters. */
+    std::vector<Declaration *> functionParams;
 };
 
 
@@ -157,27 +196,18 @@ public:
  */
 class Declaration {
 public:
-    Declaration(DeclSpecs *ds, std::vector<Declarator *> *dlist = NULL) {
-        declSpecs = ds;
-        if (dlist != NULL)
-            declarators = *dlist;
-        for (unsigned int i = 0; i < declarators.size(); ++i)
-            if (declarators[i] != NULL)
-                declarators[i]->InitFromDeclSpecs(declSpecs);
-    }
-    Declaration(DeclSpecs *ds, Declarator *d) {
-        declSpecs = ds;
-        if (d) {
-            d->InitFromDeclSpecs(ds);
-            declarators.push_back(d);
-        }
-    }
+    Declaration(DeclSpecs *ds, std::vector<Declarator *> *dlist = NULL);
+    Declaration(DeclSpecs *ds, Declarator *d);
 
-    /** Adds the symbols for the variables in the declaration to the symbol
-        table. */
-    void AddSymbols(SymbolTable *st) const;
     void Print() const;
 
+    /** This method walks through all of the Declarators in a declaration
+        and returns a fully-initialized Symbol and (possibly) and
+        initialization expression for each one.  (This allows the rest of
+        the system to not have to worry about the mess of the general
+        Declarator representation.) */
+    std::vector<VariableDeclaration> GetVariableDeclarations() const;
+
     DeclSpecs *declSpecs;
     std::vector<Declarator *> declarators;
 };
@@ -196,8 +226,9 @@ struct StructDeclaration {
 
 /** Given a set of StructDeclaration instances, this returns the types of
     the elements of the corresponding struct and their names. */
-extern void GetStructTypesAndNames(const std::vector<StructDeclaration *> &sd,
-                                   std::vector<const Type *> *elementTypes,
-                                   std::vector<std::string> *elementNames);
+extern void GetStructTypesNamesPositions(const std::vector<StructDeclaration *> &sd,
+                                         std::vector<const Type *> *elementTypes,
+                                         std::vector<std::string> *elementNames,
+                                         std::vector<SourcePos> *elementPositions);
 
 #endif // ISPC_DECL_H

docs/ReleaseNotes.txt

@@ -0,0 +1,301 @@
+=== v1.1.0 === (5 December 2011)
+
+This is a major new release of the compiler, with significant additions to
+language functionality and capabilities.  It includes a number of small
+language syntax changes that will require modification of existing
+programs.  These changes should generally be straightforward and all are
+steps toward eliminating parts of ispc syntax that are incompatible with
+C/C++.  See
+http://ispc.github.com/ispc.html#updating-ispc-programs-for-changes-in-ispc-1-1
+for more information about these changes.
+
+ispc now fully supports pointers, including pointer arithmetic, implicit
+conversions of arrays to pointers, and all of the other capabilities of
+pointers in C.  See http://ispc.github.com/ispc.html#pointer-types for more
+information about pointers in ispc and
+http://ispc.github.com/ispc.html#function-pointer-types for information
+about function pointers in ispc.
+
+Reference types are now declared with C++ syntax (e.g. "const float &foo").
+
+ispc now supports 64-bit addressing.  For performance reasons, this
+capability is disabled by default (even on 64-bit targets), but can be
+enabled with a command-line flag:
+http://ispc.github.com/ispc.html#selecting-32-or-64-bit-addressing.
+
+This release features new parallel "foreach" statements, which make it
+easier in many instances to map program instances to data for data-parallel
+computation than the programIndex/programCount mechanism:
+http://ispc.github.com/ispc.html#parallel-iteration-statements-foreach-and-foreach-tiled.
+
+Finally, all of the system's documentation has been significantly revised.
+The documentation of ispc's parallel execution model has been rewritten:
+http://ispc.github.com/ispc.html#the-ispc-parallel-execution-model, and
+there is now a more specific discussion of similarities and differences
+between ispc and C/C++:
+http://ispc.github.com/ispc.html#relationship-to-the-c-programming-language.
+There is now a separate FAQ (http://ispc.github.com/faq.html), and a
+Performance Guide (http://ispc.github.com/perfguide.html).
+ 
+=== v1.0.12 === (20 October 2011)
+
+This release includes a new "double-pumped" 8-wide target for SSE2,
+"sse2-x2".  Like the sse4-x2 and avx-x2 targets, this target may deliver
+higher performance for some workloads than the regular sse2 target.  (For
+other workloads, it may be slower.)
+
+The ispc language now includes an "assert()" statement.  See
+http://ispc.github.com/ispc.html#assertions for more information.
+
+The compiler now sets a preprocessor #define based on the target ISA; for
+example, ISPC_TARGET_SSE4 is defined for the sse4 targets, and so forth.
+
+The standard library now provides high-performance routines for converting
+between some "array of structures" and "structure of arrays" formats.
+See
+http://ispc.github.com/ispc.html#converting-between-array-of-structures-and-structure-of-arrays-layout
+for more information.
+
+Inline functions now have static linkage.
+
+A number of improvements have been made to the optimization passes that
+detect when gathers and scatters can be transformed into vector stores and
+loads, respectively.  In particular, these passes now handle variables that
+are used as loop induction variables much better.
+
+=== v1.0.11 === (6 October 2011)
+
+The main new feature in this release is support for generating code for
+multiple targets (e.g., SSE2, SSE4, and AVX) and having the compiled code
+select the best variant at execution time.  For more information, see
+http://ispc.github.com/ispc.html#compiling-with-support-for-multiple-instruction-sets.
+
+All of the examples now take advantage of the support for multiple
+compilation targets; thus, if one has an AVX system, it's not necessary to
+recompile the examples to use the AVX target.
+
+Performance of the built-in task system that is used in the examples has
+been improved.
+
+Finally, the print() statement now works on OSX; it had been broken for the
+last few releases.
+
+=== v1.0.10 === (30 September 2011)
+
+This release features an extensive new example showing the application of
+ispc to a deferred shading algorithm for scenes with thousands of lights
+(examples/deferred).  This is an implementation of the algorithm that Johan
+Andersson described at SIGGRAPH 2009 and was implemented by Andrew
+Lauritzen and Jefferson Montgomery.  The basic idea is that a pre-rendered
+G-buffer is partitioned into tiles, and in each tile, the set of lights
+that contribute to the tile is computed.  Then, the pixels in the tile are
+then shaded using those light sources. (See slides 19-29 of
+http://s09.idav.ucdavis.edu/talks/04-JAndersson-ParallelFrostbite-Siggraph09.pdf
+for more details on the algorithm.)
+
+The mechanism for launching tasks from ispc code has been generalized to
+allow multiple tasks to be launched with a single launch call (see
+http://ispc.github.com/ispc.html#task-parallelism-language-syntax for more
+information.)
+
+A few new functions have been added to the standard library: num_cores()
+returns the number of cores in the system's CPU, and variants of all of the
+atomic operators that take 'uniform' values as parameters have been added.
+
+=== v1.0.9 === (26 September 2011)
+
+The binary release of v1.0.9 is the first that supports AVX code
+generation.  Two targets are provided: "avx", which runs with a
+programCount of 8, and "avx-x2" which runs 16 program instances
+simultaneously.  (This binary is also built using the in-progress LLVM 3.0
+development libraries, while previous ones have been built with the
+released 2.9 version of LLVM.)
+
+This release has no other significant changes beyond a number of small
+bugfixes (https://github.com/ispc/ispc/issues/100,
+https://github.com/ispc/ispc/issues/101, https://github.com/ispc/ispc/issues/103.)
+ 
+=== v1.0.8 === (19 September 2011)
+
+A number of improvements have been made to handling of 'if' statements in
+the language:
+  - A bug was fixed where invalid memory could be incorrectly accessed even
+    if none of the running program instances wanted to execute the
+    corresponding instructions (https://github.com/ispc/ispc/issues/74).
+  - The code generated for 'if' statements is a bit simpler and thus more
+    efficient.
+
+There is now '--pic' command-line argument that causes position-independent
+code to be generated (Linux and OSX only).
+
+A number of additional performance improvements:
+  - Loops are now unrolled by default; the --opt=disable-loop-unroll
+    command-line argument can be used to disable this behavior.
+    (https://github.com/ispc/ispc/issues/78)
+  - A few more cases where gathers/scatters could be determined at compile
+    time to actually access contiguous locations have been added.
+    (https://github.com/ispc/ispc/issues/79)
+
+Finally, warnings are now issued (if possible) when it can be determined
+at compile-time that an out-of-bounds array index is being used.
+(https://github.com/ispc/ispc/issues/98).
+
+
+=== v1.0.7 === (3 September 2011)
+
+The various atomic_*_global() standard library functions are generally
+substantially more efficient.  They all previously issued one hardware
+atomic instruction for each running program instance but now locally
+compute a reduction over the operands and issue a single hardware atomic,
+giving the same effect and results in the end (issue #57).
+
+CPU/ISA target handling has been substantially improved.  If no CPU is
+specified, the host CPU type is used, not just a default of "nehalem".  A
+number of bugs were fixed that ensure that LLVM doesn't generate SSE>2
+instructions when using the SSE2 target (fixes issue #82).
+
+Shift rights of unsigned integer types use a logical shift right
+instruction now, not an arithmetic shift right (fixed issue #88).
+
+When emitting header files, 'extern' declarations of globals used in ispc
+code are now outside of the ispc namespace.  Fixes issue #64.
+
+The stencil example has been modified to do runs with and without
+parallelism.
+
+Many other small bugfixes and improvements.
+
+=== v1.0.6 === (17 August 2011)
+
+Some additional cross-program instance operations have been added to the
+standard library.  reduce_equal() checks to see if the given value is the
+same across all running program instances, and exclusive_scan_{and,or,and}()
+computes a scan over the given value in the running program instances.
+See the documentation of these new routines for more information:
+http://ispc.github.com/ispc.html#cross-program-instance-operations.
+
+The simple task system implementations used in the examples have been
+improved.  The Windows version no nlonger has a hard limit on the number of
+tasks that can be launched, and all versions have less dynamic memory
+allocation and less locking.  More of the examples now have paths that also
+measure performance using tasks along with SPMD vectorization.
+
+Two new examples have been added: one that shows the implementation of a
+ray-marching volume rendering algorithm, and one that shows a 3D stencil
+computation, as might be done for PDE solutions.
+
+Standard library routines to issue prefetches have been added.  See the
+documentation for more details: http://ispc.github.com/ispc.html#prefetches.
+
+Fast versions of the float to half-precision float conversion routines have
+been added.  For more details, see:
+http://ispc.github.com/ispc.html#conversions-to-and-from-half-precision-floats.
+
+There is the usual set of small bug fixes.  Notably, a number of details
+related to handling 32 versus 64 bit targets have been fixed, which in turn
+has fixed a bug related to tasks having incorrect values for pointers
+passed to them.
+
+=== v1.0.5 === (1 August 2011)
+
+Multi-element vector swizzles are supported; for example, given a 3-wide
+vector "foo", then expressions like "foo.zyx" and "foo.yz" can be used to
+construct other short vectors.  See
+http://ispc.github.com/ispc.html#short-vector-types
+for more details.  (Thanks to Pete Couperus for implementing this code!).
+
+int8 and int16 datatypes are now supported.  It is still generally more
+efficient to use int32 for intermediate computations, even if the in-memory
+format is int8 or int16.
+
+There are now standard library routines to convert to and from 'half'-format
+floating-point values (half_to_float() and float_to_half()).
+
+There is a new example with an implementation of Perlin's Noise function
+(examples/noise).  It shows a speedup of approximately 4.2x versus a C
+implementation on OSX and a 2.9x speedup versus C on Windows.
+
+=== v1.0.4 === (18 July 2011)
+
+enums are now supported in ispc; see the section on enumeration types in
+the documentation (http://ispc.github.com/ispc.html#enumeration-types) for
+more informaiton.
+
+bools are converted to integers with zero extension, not sign extension as
+before (i.e. a 'true' bool converts to the value one, not 'all bits on'.)
+For cases where sign extension is still desired, there is a
+sign_extend(bool) function in the standard library.
+
+Support for 64-bit types in the standard library is much more complete than
+before.
+
+64-bit integer constants are now supported by the parser.
+
+Storage for parameters to tasks is now allocated dynamically on Windows,
+rather than on the stack; with this fix, all tests now run correctly on
+Windows.
+
+There is now support for atomic swap and compare/exchange with float and
+double types.
+
+A number of additional small bugs have been fixed and a number of cases
+where the compiler would crash given a malformed program have been fixed.
+
+=== v1.0.3 === (4 July 2011)
+
+ispc now has a bulit-in pre-processor (from LLVM's clang compiler).
+(Thanks to Pete Couperus for this patch!)  It is therefore no longer
+necessary to use cl.exe for preprocessing on Windows; the MSVC proejct
+files for the examples have been updated accordingly.
+
+There is another variant of the shuffle() function int the standard
+library: "<type> shuffle(<type> v0, <type> v1, int permute)", where the
+permutation vector indexes over the concatenation of the two vectors
+(e.g. the value 0 corresponds to the first element of v0, the value
+2*programCount-1 corresponds to the last element of v1, etc.)
+
+ispc now supports the usual range of atomic operations (add, subtract, min,
+max, and, or, and xor) as well as atomic swap and atomic compare and
+exchange.  There is also a facility for inserting memory fences.  See the
+"Atomic Operations and Memory Fences" section of the user's guide
+(http://ispc.github.com/ispc.html#atomic-operations-and-memory-fences) for
+more information.
+ 
+There are now both 'signed' and 'unsigned' variants of the standard library
+functions like packed_load_active() that take references to arrays of
+signed int32s and unsigned int32s respectively.  (The
+{load_from,store_to}_{int8,int16}() functions have similarly been augmented
+to have both 'signed' and 'unsigned' variants.)
+
+In initializer expressions with variable declarations, it is no longer
+legal to initialize arrays and structs with single scalar values that then
+initialize their members; they now must be initialized with initializer
+lists in braces (or initialized after of the initializer with a loop over
+array elements, etc.)
+
+=== v1.0.2 === (1 July 2011)
+
+Floating-point hexidecimal constants are now parsed correctly on Windows
+(fixes issue #16).
+
+SSE2 is now the default target if --cpu=atom is given in the command line
+arguments and another target isn't explicitly specified.
+
+The standard library now provides broadcast(), rotate(), and shuffle()
+routines for efficient communication between program instances.
+
+The MSVC solution files to build the examples on Windows now use
+/fpmath:fast when building.
+
+=== v1.0.1 === (24 June 2011)
+
+ispc no longer requires that pointers to memory that are passed in to ispc
+have alignment equal to the targets vector width; now alignment just has to
+be the regular element alignment (e.g. 4 bytes for floats, etc.)  This
+change also fixed a number of cases where it previously incorrectly
+generated aligned load/store instructions in cases where the address wasn't
+actually aligned (even if the base address passed into ispc code was).
+
+=== v1.0 === (21 June 2011)
+
+Initial Release

docs/build.sh

@@ -1,6 +1,12 @@
 #!/bin/bash
 
-rst2html ispc.txt > ispc.html
+for i in ispc perfguide faq; do
+    rst2html.py --template=template.txt --link-stylesheet \
+        --stylesheet-path=css/style.css $i.txt > $i.html
+done
+
+rst2html.py --template=template-perf.txt --link-stylesheet \
+        --stylesheet-path=css/style.css perf.txt > perf.html
 
 #rst2latex --section-numbering --documentclass=article --documentoptions=DIV=9,10pt,letterpaper ispc.txt > ispc.tex
 #pdflatex ispc.tex

docs/faq.txt

@@ -0,0 +1,482 @@
+=============================================================
+Intel® SPMD Program Compiler Frequently Asked Questions (FAQ)
+=============================================================
+
+This document includes a number of frequently (and not frequently) asked
+questions about ispc, the Intel® SPMD Program Compiler.  The source to this
+document is in the file ``docs/faq.txt`` in the ``ispc`` source
+distribution.
+
+* Understanding ispc's Output
+
+  + `How can I see the assembly language generated by ispc?`_
+  + `How can I have the assembly output be printed using Intel assembly syntax?`_
+  + `Why are there multiple versions of exported ispc functions in the assembly output?`_
+  + `How can I more easily see gathers and scatters in generated assembly?`_
+
+* Interoperability
+
+  + `How can I supply an initial execution mask in the call from the application?`_
+  + `How can I generate a single binary executable with support for multiple instruction sets?`_
+  + `How can I determine at run-time which vector instruction set's instructions were selected to execute?`_
+
+* Programming Techniques
+
+  + `What primitives are there for communicating between SPMD program instances?`_
+  + `How can a gang of program instances generate variable amounts of output efficiently?`_
+  + `Is it possible to use ispc for explicit vector programming?`_
+  + `How can I debug my ispc programs using Valgrind?`_
+
+Understanding ispc's Output
+===========================
+
+How can I see the assembly language generated by ispc?
+------------------------------------------------------
+
+The ``--emit-asm`` flag causes assembly output to be generated.  If the
+``-o`` command-line flag is also supplied, the assembly is stored in the
+given file, or printed to standard output if ``-`` is specified for the
+filename.  For example, given the simple ``ispc`` program:
+
+::
+
+    export uniform int foo(uniform int a, uniform int b) {
+        return a+b;
+    }
+
+If the SSE4 target is used, then the following assembly is printed:
+
+::
+
+    _foo:
+            addl    %esi, %edi
+            movl    %edi, %eax
+            ret
+
+
+How can I have the assembly output be printed using Intel assembly syntax?
+--------------------------------------------------------------------------
+
+The ``ispc`` compiler is currently only able to emit assembly with AT+T
+syntax, where the destination operand is the last operand after an
+instruction.  If you'd prefer Intel assembly output, one option is to use
+Agner Fog's ``objconv`` tool: have ``ispc`` emit a native object file and
+then use ``objconv`` to disassemble it, specifying the assembler syntax
+that you prefer.  ``objconv`` `is available for download here`_.
+
+.. _is available for download here: http://www.agner.org/optimize/#objconv
+
+Why are there multiple versions of exported ispc functions in the assembly output?
+----------------------------------------------------------------------------------
+
+Two generations of all functions qualified with ``export`` are generated:
+one of them is for being be called by other ``ispc`` functions, and the
+other is to be called by the application.  The application callable
+function has the original function's name, while the ``ispc``-callable
+function has a mangled name that encodes the types of the function's
+parameters.
+
+The crucial difference between these two functions is that the
+application-callable function doesn't take a parameter encoding the current
+execution mask, while ``ispc``-callable functions have a hidden mask
+parameter.  An implication of this difference is that the ``export``
+function starts with the execution mask "all on".  This allows a number of
+improvements in the generated code, particularly on architectures that
+don't have support for masked load and store instructions.
+
+As an example, consider this short function, which loads a vector's worth
+values from two arrays in memory, adds them, and writes the result to an
+output array.
+
+::
+
+    export void foo(uniform float a[], uniform float b[],
+                    uniform float result[]) {
+        float aa = a[programIndex], bb = b[programIndex];
+        result[programIndex] = aa+bb;
+    }
+
+Here is the assembly code for the application-callable instance of the
+function.
+
+::
+
+    _foo:
+            movups        (%rsi), %xmm1
+            movups        (%rdi), %xmm0
+            addps         %xmm1, %xmm0
+            movups        %xmm0, (%rdx)
+            ret
+
+
+And here is the assembly code for the ``ispc``-callable instance of the
+function.
+
+::
+
+    "_foo___uptr<Uf>uptr<Uf>uptr<Uf>":
+            movmskps      %xmm0, %eax
+            cmpl          $15, %eax
+            je            LBB0_3
+            testl         %eax, %eax
+            jne           LBB0_4
+            ret
+    LBB0_3:
+            movups        (%rsi), %xmm1
+            movups        (%rdi), %xmm0
+            addps         %xmm1, %xmm0
+            movups        %xmm0, (%rdx)
+            ret
+    LBB0_4:
+    ####
+    ####  Code elided; handle mixed mask case..
+    ####
+            ret
+
+There are a few things to notice in this code.  First, the current program
+mask is coming in via the ``%xmm0`` register and the initial few
+instructions in the function essentially check to see if the mask is all on
+or all off.  If the mask is all on, the code at the label LBB0_3 executes;
+it's the same as the code that was generated for ``_foo`` above.  If the
+mask is all off, then there's nothing to be done, and the function can
+return immediately.
+
+In the case of a mixed mask, a substantial amount of code is generated to
+load from and then store to only the array elements that correspond to
+program instances where the mask is on.  (This code is elided below).  This
+general pattern of having two-code paths for the "all on" and "mixed" mask
+cases is used in the code generated for almost all but the most simple
+functions (where the overhead of the test isn't worthwhile.)
+
+How can I more easily see gathers and scatters in generated assembly?
+---------------------------------------------------------------------
+
+Because CPU vector ISAs don't have native gather and scatter instructions,
+these memory operations are turned into sequences of a series of
+instructions in the code that ``ispc`` generates.  In some cases, it can be
+useful to see where gathers and scatters actually happen in code; there is
+an otherwise undocumented command-line flag that provides this information.
+
+Consider this simple program:
+
+::
+
+    void set(uniform int a[], int value, int index) {
+        a[index] = value;
+    }
+
+When compiled normally to the SSE4 target, this program generates this
+extensive code sequence, which makes it more difficult to see what the
+program is actually doing.
+
+::
+
+    "_set___uptr<Ui>ii":
+            pmulld        LCPI0_0(%rip), %xmm1
+            movmskps      %xmm2, %eax
+            testb         $1, %al
+            je            LBB0_2
+            movd          %xmm1, %ecx
+            movd          %xmm0, (%rcx,%rdi)
+    LBB0_2:
+            testb         $2, %al
+            je            LBB0_4
+            pextrd        $1, %xmm1, %ecx
+            pextrd        $1, %xmm0, (%rcx,%rdi)
+    LBB0_4:
+            testb         $4, %al
+            je            LBB0_6
+            pextrd        $2, %xmm1, %ecx
+            pextrd        $2, %xmm0, (%rcx,%rdi)
+    LBB0_6:
+            testb        $8, %al
+            je            LBB0_8
+            pextrd        $3, %xmm1, %eax
+            pextrd        $3, %xmm0, (%rax,%rdi)
+    LBB0_8:
+            ret
+
+If this program is compiled with the
+``--opt=disable-handle-pseudo-memory-ops`` command-line flag, then the
+scatter is left as an unresolved function call.  The resulting program
+won't link without unresolved symbols, but the assembly output is much
+easier to understand:
+
+::
+
+    "_set___uptr<Ui>ii":
+            movaps        %xmm0, %xmm3
+            pmulld        LCPI0_0(%rip), %xmm1
+            movdqa        %xmm1, %xmm0
+            movaps        %xmm3, %xmm1
+            jmp        ___pseudo_scatter_base_offsets32_32 ## TAILCALL
+
+
+Interoperability
+================
+
+How can I supply an initial execution mask in the call from the application?
+----------------------------------------------------------------------------
+
+Recall that when execution transitions from the application code to an
+``ispc`` function, all of the program instances are initially executing.
+In some cases, it may desired that only some of them are running, based on
+a data-dependent condition computed in the application program.  This
+situation can easily be handled via an additional parameter from the
+application.
+
+As a simple example, consider a case where the application code has an
+array of ``float`` values and we'd like the ``ispc`` code to update
+just specific values in that array, where which of those values to be
+updated has been determined by the application.  In C++ code, we might
+have:
+
+::
+
+    int count = ...;
+    float *array = new float[count];
+    bool *shouldUpdate = new bool[count];
+    // initialize array and shouldUpdate
+    ispc_func(array, shouldUpdate, count);
+
+Then, the ``ispc`` code could process this update as:
+
+::
+
+    export void ispc_func(uniform float array[], uniform bool update[],
+                          uniform int count) {
+        foreach (i = 0 ... count) {
+            cif (update[i] == true)
+                // update array[i+programIndex]...
+        }
+    }
+
+(In this case a "coherent" if statement is likely to be worthwhile if the
+``update`` array will tend to have sections that are either all-true or
+all-false.)
+
+How can I generate a single binary executable with support for multiple instruction sets?
+-----------------------------------------------------------------------------------------
+
+``ispc`` can also generate output that supports multiple target instruction
+sets, also generating code that chooses the most appropriate one at runtime
+if multiple targets are specified with the ``--target`` command-line
+argument.
+
+For example, if you run the command:
+
+::
+
+   ispc foo.ispc -o foo.o --target=sse2,sse4-x2,avx-x2
+
+Then four object files will be generated: ``foo_sse2.o``, ``foo_sse4.o``,
+``foo_avx.o``, and ``foo.o``.[#]_  Link all of these into your executable, and
+when you call a function in ``foo.ispc`` from your application code,
+``ispc`` will determine which instruction sets are supported by the CPU the
+code is running on and will call the most appropraite version of the
+function available.  
+
+.. [#] Similarly, if you choose to generate assembly langauage output or
+   LLVM bitcode output, multiple versions of those files will be created.
+
+In general, the version of the function that runs will be the one in the
+most general instruction set that is supported by the system.  If you only
+compile SSE2 and SSE4 variants and run on a system that supports AVX, for
+example, then the SSE4 variant will be executed.  If the system doesn't
+is not able to run any of the available variants of the function (for
+example, trying to run a function that only has SSE4 and AVX variants on a
+system that only supports SSE2), then the standard library ``abort()``
+function will be called.
+
+One subtlety is that all non-static global variables (if any) must have the
+same size and layout with all of the targets used.  For example, if you
+have the global variables:
+
+::
+
+   uniform int foo[2*programCount];
+   int bar;
+
+and compile to both SSE2 and AVX targets, both of these variables will have
+different sizes (the first due to program count having the value 4 for SSE2
+and 8 for AVX, and the second due to ``varying`` types having different
+numbers of elements with the two targets--essentially the same issue as the
+first.)  ``ispc`` issues an error in this case.
+
+
+How can I determine at run-time which vector instruction set's instructions were selected to execute?
+-----------------------------------------------------------------------------------------------------
+
+``ispc`` doesn't provide any API that allows querying which vector ISA's
+instructions are running when multi-target compilation was used.  However,
+this can be solved in "user space" by writing a small helper function.
+Specifically, if you implement a function like this
+
+::
+
+    export uniform int isa() {
+    #if defined(ISPC_TARGET_SSE2)
+        return 0;
+    #elif defined(ISPC_TARGET_SSE4)
+        return 1;
+    #elif defined(ISPC_TARGET_AVX)
+        return 2;
+    #else
+        return -1;
+    #endif
+    }
+
+And then call it from your application code at runtime, it will return 0,
+1, or 2, depending on which target's instructions are running.
+
+The way this works is a little surprising, but it's a useful trick.  Of
+course the preprocessor ``#if`` checks are all compile-time only
+operations.  What's actually happening is that the function is compiled
+multiple times, once for each target, with the appropriate ``ISPC_TARGET``
+preprocessor symbol set.  Then, a small dispatch function is generated for
+the application to actually call.  This dispatch function in turn calls the
+appropriate version of the function based on the CPU of the system it's
+executing on, which in turn returns the appropriate value.
+
+In a similar fashion, it's possible to find out at run-time the value of
+``programCount`` for the target that's actually being used.
+
+::
+
+    export uniform int width() { return programCount; }
+
+
+Programming Techniques
+======================
+
+What primitives are there for communicating between SPMD program instances?
+---------------------------------------------------------------------------
+
+The ``broadcast()``, ``rotate()``, and ``shuffle()`` standard library
+routines provide a variety of mechanisms for the running program instances
+to communicate values to each other during execution.  Note that there's no
+need to synchronize the program instances before communicating between
+them, due to the synchronized execution model of gangs of program instances
+in ``ispc``.
+
+How can a gang of program instances generate variable amounts of output efficiently?
+------------------------------------------------------------------------------------
+
+It's not unusual to have a gang of program instances where each program
+instance generates a variable amount of output (perhaps some generate no
+output, some generate one output value, some generate many output values
+and so forth), and where one would like to have the output densely packed
+in an output array.  The ``exclusive_scan_add()`` function from the
+standard library is quite useful in this situation.
+
+Consider the following function:
+
+::
+
+    uniform int func(uniform float outArray[], ...) {
+       int numOut = ...;  // figure out how many to be output
+       float outLocal[MAX_OUT]; // staging area
+
+       // each program instance in the gang puts its results in
+       //  outLocal[0], ..., outLocal[numOut-1]
+
+       int startOffset = exclusive_scan_add(numOut);
+       for (int i = 0; i < numOut; ++i)
+           outArray[startOffset + i] = outLocal[i];
+       return reduce_add(numOut);
+    }
+
+Here, each program instance has computed a number, ``numOut``, of values to
+output, and has stored them in the ``outLocal`` array.  Assume that four
+program instances are running and that the first one wants to output one
+value, the second two values, and the third and fourth three values each.
+In this case, ``exclusive_scan_add()`` will return the values (0, 1, 3, 6)
+to the four program instances, respectively.  
+
+The first program instance will then write its one result to
+``outArray[0]``, the second will write its two values to ``outArray[1]``
+and ``outArray[2]``, and so forth.  The ``reduce_add()`` call at the end
+returns the total number of values that all of the program instances have
+written to the array.
+
+FIXME: add discussion of foreach_active as an option here once that's in
+
+Is it possible to use ispc for explicit vector programming?
+-----------------------------------------------------------
+
+The typical model for programming in ``ispc`` is an *implicit* parallel
+model, where one writes a program that is apparently doing scalar
+computation on values and the program is then vectorized to run in parallel
+across the SIMD lanes of a processor.  However, ``ispc`` also has some
+support for explicit vector unit programming, where the vectorization is
+explicit.  Some computations may be more effectively described in the
+explicit model rather than the implicit model.
+
+This support is provided via ``uniform`` instances of short vectors
+Specifically, if this short program
+
+::
+
+    export uniform float<8> madd(uniform float<8> a, uniform float<8> b,
+                                 uniform float<8> c) {
+        return a + b * c;
+    }
+
+is compiled with the AVX target, ``ispc`` generates the following assembly:
+
+::
+
+    _madd:
+	vmulps	%ymm2, %ymm1, %ymm1
+	vaddps	%ymm0, %ymm1, %ymm0
+	ret
+
+(And similarly, if compiled with a 4-wide SSE target, two ``mulps`` and two
+``addps`` instructions are generated, and so forth.)
+
+Note that ``ispc`` doesn't currently support control-flow based on
+``uniform`` short vector types; it is thus not possible to write code like:
+
+::
+
+    export uniform int<8> count(uniform float<8> a, uniform float<8> b) {
+        uniform int<8> sum = 0;
+        while (a++ < b)
+            ++sum;
+    }
+
+
+How can I debug my ispc programs using Valgrind?
+------------------------------------------------
+
+The `valgrind`_ memory checker is an extremely useful memory checker for
+Linux and OSX; it detects a range of memory errors, including accessing
+memory after it has been freed, accessing memory beyond the end of an
+array, accessing uninitialized stack variables, and so forth.
+In general, applications that use ``ispc`` code run with ``valgrind``
+without modification and ``valgrind`` will detect the same range of memory
+errors in ``ispc`` code that it does in C/C++ code.  
+
+.. _valgrind: http://valgrind.org
+
+One issue to be aware of is that until recently, ``valgrind`` only
+supported the SSE2 vector instructions; if you are using a version of
+``valgrind`` older than the 3.7.0 release (5 November 2011), you should
+compile your ``ispc`` programs with ``--target=sse2`` before running them
+through ``valgrind``.  (Note that if no target is specified, then ``ispc``
+chooses a target based on the capabilities of the system you're running
+``ispc`` on.)  If you run an ``ispc`` program that uses instructions that
+``valgrind`` doesn't support, you'll see an error message like:
+
+::
+
+    vex amd64->IR: unhandled instruction bytes: 0xC5 0xFA 0x10 0x0 0xC5 0xFA 0x11 0x84
+    ==46059== valgrind: Unrecognised instruction at address 0x100002707.
+
+The just-released valgrind 3.7.0 adds support for the SSE4.2 instruction
+set; if you're using that version (and your system supports SSE4.2), then
+you can use ``--target=sse4`` when compiling to run with ``valgrind``.
+
+Note that ``valgrind`` does not yet support programs that use the AVX
+instruction set.
+

docs/perf.txt

@@ -0,0 +1,85 @@
+===========
+Performance
+===========
+
+The SPMD programming model that ``ispc`` makes it easy to harness the
+computational power available in SIMD vector units on modern CPUs, while
+its basis in C makes it easy for programmers to adopt and use
+productively.  This page summarizes the performance of ``ispc`` with the
+workloads in the ``examples/`` directory of the ``ispc`` distribution.
+
+These results were measured on a 4-core Apple iMac with a 4-core 3.4GHz
+Intel® Core-i7 processor using the Intel® AVX instruction set.  The basis
+for comparison is a reference C++ implementation compiled with gcc 4.2.1,
+the version distributed with OS X 10.7.2.  (The reference implementation is
+also included in the ``examples/`` directory.)
+
+.. list-table:: Performance of ``ispc`` with a variety of the workloads
+   from the ``examples/`` directory of the ``ispc`` distribution, compared
+   a reference C++ implementation compiled with gcc 4.2.1.
+
+  * - Workload
+    - ``ispc``, 1 core
+    - ``ispc``, 4 cores
+  * - `AOBench`_ (512 x 512 resolution)
+    - 3.99x
+    - 19.32x
+  * - `Binomial Options`_ (128k options)
+    - 7.94x
+    - 33.43x
+  * - `Black-Scholes Options`_ (128k options)
+    - 8.45x
+    - 32.48x
+  * - `Deferred Shading`_ (1280p)
+    - n/a
+    - 23.06x
+  * - `Mandelbrot Set`_
+    - 6.21x
+    - 19.90x
+  * - `Perlin Noise Function`_
+    - 5.37x
+    - n/a
+  * - `Ray Tracer`_ (Sponza dataset)
+    - 3.99x
+    - 19.32x
+  * - `3D Stencil`_
+    - 3.76x
+    - 13.79x
+  * - `Volume Rendering`_
+    - 3.11x
+    - 15.80x
+
+
+.. _AOBench: https://github.com/ispc/ispc/tree/master/examples/aobench
+.. _Binomial Options: https://github.com/ispc/ispc/tree/master/examples/options
+.. _Black-Scholes Options: https://github.com/ispc/ispc/tree/master/examples/options
+.. _Deferred Shading: https://github.com/ispc/ispc/tree/master/examples/deferred
+.. _Mandelbrot Set: https://github.com/ispc/ispc/tree/master/examples/mandelbrot_tasks
+.. _Ray Tracer: https://github.com/ispc/ispc/tree/master/examples/rt
+.. _Perlin Noise Function: https://github.com/ispc/ispc/tree/master/examples/noise
+.. _3D Stencil: https://github.com/ispc/ispc/tree/master/examples/stencil
+.. _Volume Rendering: https://github.com/ispc/ispc/tree/master/examples/volume_rendering
+
+
+The following table shows speedups for a number of the examples on a
+2.40GHz, 40-core Intel® Xeon E7-8870 system with the Intel® SSE4
+instruction set, running Microsoft Windows Server 2008 Enterprise.  Here,
+the serial C/C++ baseline code was compiled with MSVC 2010.
+ 
+.. list-table:: Performance of ``ispc`` with a variety of the workloads
+   from the ``examples/`` directory of the ``ispc`` distribution, on 
+   system with 40 CPU cores.
+
+  * - Workload
+    - ``ispc``, 40 cores
+  * - AOBench (2048 x 2048 resolution)
+    - 182.36x
+  * - Binomial Options (2m options)
+    - 63.85x
+  * - Black-Scholes Options (2m options)
+    - 83.97x
+  * - Ray Tracer (Sponza dataset)
+    - 195.67x
+  * - Volume Rendering
+    - 243.18x
+

docs/perfguide.txt

@@ -0,0 +1,714 @@
+==============================================
+Intel® SPMD Program Compiler Performance Guide
+==============================================
+
+The SPMD programming model provided by ``ispc`` naturally delivers
+excellent performance for many workloads thanks to efficient use of CPU
+SIMD vector hardware.  This guide provides more details about how to get
+the most out of ``ispc`` in practice.
+
+* `Key Concepts`_
+
+  + `Efficient Iteration With "foreach"`_
+  + `Improving Control Flow Coherence With "foreach_tiled"`_
+  + `Using Coherent Control Flow Constructs`_
+  + `Use "uniform" Whenever Appropriate`_
+
+* `Tips and Techniques`_
+
+  + `Understanding Gather and Scatter`_
+  + `Avoid 64-bit Addressing Calculations When Possible`_
+  + `Avoid Computation With 8 and 16-bit Integer Types`_
+  + `Implementing Reductions Efficiently`_
+  + `Using Low-level Vector Tricks`_
+  + `The "Fast math" Option`_
+  + `"inline" Aggressively`_
+  + `Avoid The System Math Library`_
+  + `Declare Variables In The Scope Where They're Used`_
+  + `Instrumenting ISPC Programs To Understand Runtime Behavior`_
+  + `Choosing A Target Vector Width`_
+
+* `Disclaimer and Legal Information`_
+
+* `Optimization Notice`_
+
+Key Concepts
+============
+
+This section describes the four most important concepts to understand and
+keep in mind when writing high-performance ``ispc`` programs.  It assumes
+good familiarity with the topics covered in the ``ispc`` `Users Guide`_.
+
+.. _Users Guide: ispc.html
+
+Efficient Iteration With "foreach"
+----------------------------------
+
+The ``foreach`` parallel iteration construct is semantically equivalent to
+a regular ``for()`` loop, though it offers meaningful performance benefits.
+(See the `documentation on "foreach" in the Users Guide`_ for a review of
+its syntax and semantics.)  As an example, consider this simple function
+that iterates over some number of elements in an array, doing computation
+on each one:
+
+.. _documentation on "foreach" in the Users Guide: ispc.html#parallel-iteration-statements-foreach-and-foreach-tiled
+
+::
+
+    export void foo(uniform int a[], uniform int count) {
+        for (int i = programIndex; i < count; i += programCount) {
+            // do some computation on a[i]
+        }
+    }
+
+Depending on the specifics of the computation being performed, the code
+generated for this function could likely be improved by modifying the code 
+so that the loop only goes as far through the data as is possible to pack
+an entire gang of program instances with computation each time thorugh the
+loop.  Doing so enables the ``ispc`` compiler to generate more efficient
+code for cases where it knows that the execution mask is "all on".  Then,
+an ``if`` statement at the end handles processing the ragged extra bits of
+data that didn't fully fill a gang.
+
+::
+
+    export void foo(uniform int a[], uniform int count) {
+        // First, just loop up to the point where all program instances
+        // in the gang will be active at the loop iteration start
+        uniform int countBase = count & ~(programCount-1);
+        for (uniform int i = 0; i < countBase; i += programCount) {
+            int index = i + programIndex;
+            // do some computation on a[index]
+        }
+        // Now handle the ragged extra bits at the end
+        if (countBase < count) {
+            int index = countBase + programIndex;
+            // do some computation on a[index]
+        }
+    }
+
+While the performance of the above code will likely be better than the
+first version of the function, the loop body code has been duplicated (or
+has been forced to move into a separate utility function).
+
+Using the ``foreach`` looping construct as below provides all of the
+performance benefits of the second version of this function, with the
+compactness of the first.
+
+::
+
+    export void foo(uniform int a[], uniform int count) {
+        foreach (i = 0 ... count) {
+            // do some computation on a[i]
+        }
+    }
+
+Improving Control Flow Coherence With "foreach_tiled"
+-----------------------------------------------------
+
+Depending on the computation being performed, ``foreach_tiled`` may give
+better performance than ``foreach``.  (See the `documentation in the Users
+Guide`_ for the syntax and semantics of ``foreach_tiled``.)  Given a
+multi-dimensional iteration like:
+
+.. _documentation in the Users Guide: ispc.html#parallel-iteration-statements-foreach-and-foreach-tiled
+
+::
+
+    foreach (i = 0 ... width, j = 0 ... height) {
+        // do computation on element (i,j)
+    }
+
+if the ``foreach`` statement is used, elements in the gang of program
+instances will be mapped to values of ``i`` and ``j`` by taking spans of
+``programCount`` elements across ``i`` with a single value of ``j``.  For
+example, the ``foreach`` statement above roughly corresponds to:
+
+::
+
+    for (uniform int j = 0; j < height; ++j)
+        for (int i = 0; i < width; i += programCount) {
+            // do computation 
+    }
+
+When a multi-dimensional domain is being iterated over, ``foreach_tiled``
+statement maps program instances to data in a way that tries to select
+square n-dimensional segments of the domain.  For example, on a compilation
+target with 8-wide gangs of program instances, it generates code that
+iterates over the domain the same way as the following code (though more
+efficiently):
+
+::
+
+    for (int j = programIndex/4; j < height; j += 2)
+        for (int i = programIndex%4; i < width; i += 4) {
+            // do computation 
+    }
+
+Thus, each gang of program instances operates on a 2x4 tile of the domain.
+With higher-dimensional iteration and different gang sizes, a similar
+mapping is performed--e.g. for 2D iteration with a 16-wide gang size, 4x4
+tiles are iterated over; for 4D iteration with a 8-gang, 1x2x2x2 tiles are
+processed, and so forth.  
+
+Performance benefit can come from using ``foreach_tiled`` in that it
+essentially optimizes for the benefit of iterating over *compact* regions
+of the domian (while ``foreach`` iterates over the domain in a way that
+generally allows linear memory access.)  There are two benefits from
+processing compact regions of the domain.  
+
+First, it's often the case that the control flow coherence of the program
+instances in the gang is improved; if data-dependent control flow decisions
+are related to the values of the data in the domain being processed, and if
+the data values have some coherence, iterating with compact regions will
+improve control flow coherence.
+
+Second, processing compact regions may mean that the data accessed by
+program instances in the gang is be more coherent, leading to performance
+benefits from better cache hit rates.
+
+As a concrete example, for the ray tracer example in the ``ispc``
+distribution (in the ``examples/rt`` directory), performance is 20% better
+when the pixels are iterated over using ``foreach_tiled`` than ``foreach``,
+because more coherent regions of the scene are accessed by the set of rays
+in the gang of program instances.
+
+
+Using Coherent Control Flow Constructs
+--------------------------------------
+
+Recall from the ``ispc`` Users Guide, in the `SPMD-on-SIMD Execution Model
+section`_ that ``if`` statements with a ``uniform`` test compile to more
+efficient code than ``if`` tests with varying tests.  The coherent ``cif``
+statement can provide many benefits of ``if`` with a uniform test in the
+case where the test is actually varying.
+
+.. _SPMD-on-SIMD Execution Model section: ispc.html#the-spmd-on-simd-execution-model
+
+In this case, the code the compiler generates for the ``if``
+test is along the lines of the following pseudo-code:
+
+::
+
+   bool expr = /* evaluate cif condition */
+   if (all(expr)) {
+       // run "true" case of if test only
+   } else if (!any(expr)) {
+       // run "false" case of if test only
+   } else {
+       // run both true and false cases, updating mask appropriately
+   }
+
+For ``if`` statements where the different running SPMD program instances
+don't have coherent values for the boolean ``if`` test, using ``cif``
+introduces some additional overhead from the ``all`` and ``any`` tests as
+well as the corresponding branches.  For cases where the program
+instances often do compute the same boolean value, this overhead is
+worthwhile.  If the control flow is in fact usually incoherent, this
+overhead only costs performance.
+
+In a similar fashion, ``ispc`` provides ``cfor``, ``cwhile``, and ``cdo``
+statements.  These statements are semantically the same as the
+corresponding non-"c"-prefixed functions.
+
+Use "uniform" Whenever Appropriate
+----------------------------------
+
+For any variable that will always have the same value across all of the
+program instances in a gang, declare the variable with the  ``unfiorm``
+qualifier.  Doing so enables the ``ispc`` compiler to emit better code in
+many different ways.
+
+As a simple example, consider a ``for`` loop that always does the same
+number of iterations:
+
+::
+
+    for (int i = 0; i < 10; ++i)
+        // do something ten times
+
+If this is written with ``i`` as a ``varying`` variable, as above, there's
+additional overhead in the code generated for the loop as the compiler
+emits instructions to handle the possibilty of not all program instances
+following the same control flow path (as might be the case if the loop
+limit, 10, was itself a ``varying`` value.)
+
+If the above loop is instead written with ``i`` ``uniform``, as:
+
+::
+
+    for (uniform int i = 0; i < 10; ++i)
+        // do something ten times
+
+Then better code can be generated (and the loop possibly unrolled).
+
+In some cases, the compiler may be able to detect simple cases like these,
+but it's always best to provide the compiler with as much help as possible
+to understand the actual form of your computation.
+
+
+Tips and Techniques
+===================
+
+This section introduces a number of additional techniques that are worth
+keeping in mind when writing ``ispc`` programs.
+
+Understanding Gather and Scatter
+--------------------------------
+
+Memory reads and writes from the program instances in a gang that access
+irregular memory locations (rather than a consecutive set of locations, or
+a single location) can be relatively inefficient.  As an example, consider
+the "simple" array indexing calculation below:
+
+::
+
+    int i = ....;
+    uniform float x[10] = { ... };
+    float f = x[i];
+
+Since the index ``i`` is a varying value, the program instances in the gang
+will in general be reading different locations in the array ``x``.  Because
+current CPUs have a "gather" instruction, the ``ispc`` compiler has to
+serialize these memory reads, performing a separate memory load for each
+running program instance, packing the result into ``f``.  (The analogous
+case happens for a write into ``x[i]``.)
+
+In many cases, gathers like these are unavoidable; the program instances
+just need to access incoherent memory locations.  However, if the array
+index ``i`` actually has the same value for all of the program instances or
+if it represents an access to a consecutive set of array locations, much
+more efficient load and store instructions can be generated instead of
+gathers and scatters, respectively.
+
+In many cases, the ``ispc`` compiler is able to deduce that the memory
+locations accessed by a varying index are either all the same or are
+uniform.  For example, given:
+
+::
+
+  uniform int x = ...;
+  int y = x;
+  return array[y];
+
+The compiler is able to determine that all of the program instances are
+loading from the same location, even though ``y`` is not a ``uniform``
+variable.  In this case, the compiler will transform this load to a regular
+vector load, rather than a general gather.
+
+Sometimes the running program instances will access a linear sequence of
+memory locations; this happens most frequently when array indexing is done
+based on the built-in ``programIndex`` variable.  In many of these cases,
+the compiler is also able to detect this case and then do a vector load.
+For example, given:
+
+::
+
+    for (int i = programIndex; i < count; i += programCount)
+      // process array[i];
+
+Regular vector loads and stores are issued for accesses to ``array[i]``.
+
+Both of these cases have been ones where the compiler is able to determine
+statically that the index has the same value at compile-time.  It's 
+often the case that this determination can't be made at compile time, but
+this is often the case at run time.  The ``reduce_equal()`` function from
+the standard library can be used in this case; it checks to see if the
+given value is the same across over all of the running program instances,
+returning true and its ``uniform`` value if so.
+
+The following function shows the use of ``reduce_equal()`` to check for an
+equal index at execution time and then either do a scalar load and
+broadcast or a general gather.
+
+::
+
+    uniform float array[..] = { ... };
+    float value;
+    int i = ...;
+    uniform int ui;
+    if (reduce_equal(i, &ui) == true)
+        value = array[ui]; // scalar load + broadcast
+    else
+        value = array[i];  // gather
+
+For a simple case like the one above, the overhead of doing the
+``reduce_equal()`` check is likely not worthwhile compared to just always
+doing a gather.  In more complex cases, where a number of accesses are done
+based on the index, it can be worth doing.  See the example
+``examples/volume_rendering`` in the ``ispc`` distribution for the use of
+this technique in an instance where it is beneficial to performance.
+
+Avoid 64-bit Addressing Calculations When Possible
+--------------------------------------------------
+
+Even when compiling to a 64-bit architecture target, ``ispc`` does many of
+the addressing calculations in 32-bit precision by default--this behavior
+can be overridden with the ``--addressing=64`` command-line argument.  This
+option should only be used if it's necessary to be able to address over 4GB
+of memory in the ``ispc`` code, as it essentially doubles the cost of
+memory addressing calculations in the generated code.
+
+Avoid Computation With 8 and 16-bit Integer Types
+-------------------------------------------------
+
+The code generated for 8 and 16-bit integer types is generally not as
+efficient as the code generated for 32-bit integer types.  It is generally
+worthwhile to use 32-bit integer types for intermediate computations, even
+if the final result will be stored in a smaller integer type.
+
+Implementing Reductions Efficiently
+-----------------------------------
+
+It's often necessary to compute a reduction over a data set--for example,
+one might want to add all of the values in an array, compute their minimum,
+etc.  ``ispc`` provides a few capabilities that make it easy to efficiently
+compute reductions like these.  However, it's important to use these
+capabilities appropriately for best results.
+
+As an example, consider the task of computing the sum of all of the values
+in an array.  In C code, we might have:
+
+::
+
+    /* C implementation of a sum reduction */
+    float sum(const float array[], int count) {
+        float sum = 0;
+        for (int i = 0; i < count; ++i)
+            sum += array[i];
+        return sum;
+    } 
+
+Exactly this computation could also be expressed as a purely uniform
+computation in ``ispc``, though without any benefit from vectorization:
+
+::
+
+    /* inefficient ispc implementation of a sum reduction */
+    uniform float sum(const uniform float array[], uniform int count) {
+        uniform float sum = 0;
+        for (uniform int i = 0; i < count; ++i)
+            sum += array[i];
+        return sum;
+    } 
+
+As a first try, one might try using the ``reduce_add()`` function from the
+``ispc`` standard library; it takes a ``varying`` value and returns the sum
+of that value across all of the active program instances.
+
+::
+
+    /* inefficient ispc implementation of a sum reduction */
+    uniform float sum(const uniform float array[], uniform int count) {
+        uniform float sum = 0;
+        foreach (i = 0 ... count)
+            sum += reduce_add(array[i+programIndex]);
+        return sum;
+    } 
+
+This implementation loads a gang's worth of values from the array, one for
+each of the program instances, and then uses ``reduce_add()`` to reduce
+across the program instances and then update the sum.  Unfortunately this
+approach loses most benefit from vectorization, as it does more work on the
+cross-program instance ``reduce_add()`` call than it saves from the vector
+load of values.
+
+The most efficient approach is to do the reduction in two phases: rather
+than using a ``uniform`` variable to store the sum, we maintain a varying
+value, such that each program instance is effectively computing a local
+partial sum on the subset of array values that it has loaded from the
+array.  When the loop over array elements concludes, a single call to
+``reduce_add()`` computes the final reduction across each of the program
+instances' elements of ``sum``.  This approach effectively compiles to a
+single vector load and a single vector add for each loop iteration's of
+values--very efficient code in the end.
+
+::
+
+    /* good ispc implementation of a sum reduction */
+    uniform float sum(const uniform float array[], uniform int count) {
+        float sum = 0;
+        foreach (i = 0 ... count)
+            sum += array[i+programIndex];
+        return reduce_add(sum);
+    } 
+
+Using Low-level Vector Tricks
+-----------------------------
+
+Many low-level Intel® SSE and AVX coding constructs can be implemented in
+``ispc`` code.  The ``ispc`` standard library functions ``intbits()`` and
+``floatbits()`` are often useful in this context.  Recall that
+``intbits()`` takes a ``float`` value and returns it as an integer where
+the bits of the integer are the same as the bit representation in memory of
+the ``float``.  (In other words, it does *not* perform an integer to
+floating-point conversion.)  ``floatbits()``, then, performs the inverse
+computation.
+
+As an example of the use of these functions, the following code efficiently
+reverses the sign of the given values.
+
+::
+
+  float flipsign(float a) {
+      unsigned int i = intbits(a);
+      i ^= 0x80000000;
+      return floatbits(i);
+  }
+
+This code compiles down to a single XOR instruction.
+
+The "Fast math" Option
+----------------------
+
+``ispc`` has a ``--opt=fast-math`` command-line flag that enables a number of
+optimizations that may be undesirable in code where numerical precision is
+critically important.  For many graphics applications, for example, the
+approximations introduced may be acceptable, however.  The following two
+optimizations are performed when ``--opt=fast-math`` is used.  By default, the
+``--opt=fast-math`` flag is off.
+
+* Expressions like ``x / y``, where ``y`` is a compile-time constant, are
+  transformed to ``x * (1./y)``, where the inverse value of ``y`` is
+  precomputed at compile time.
+
+* Expressions like ``x / y``, where ``y`` is not a compile-time constant,
+  are transformed to ``x * rcp(y)``, where ``rcp()`` maps to the
+  approximate reciprocal instruction from the ``ispc`` standard library.
+
+
+"inline" Aggressively
+---------------------
+
+Inlining functions aggressively is generally beneficial for performance
+with ``ispc``.  Definitely use the ``inline`` qualifier for any short
+functions (a few lines long), and experiment with it for longer functions.
+
+Avoid The System Math Library
+-----------------------------
+
+The default math library for transcendentals and the like that ``ispc`` has
+higher error than the system's math library, though is much more efficient
+due to being vectorized across the program instances and due to the fact
+that the functions can be inlined in the final code.  (It generally has
+errors in the range of 10ulps, while the system math library generally has
+no more than 1ulp of error for transcendentals.)
+
+If the ``--math-lib=system`` command-line option is used when compiling an
+``ispc`` program, then calls to the system math library will be generated
+instead.  This option should only be used if the higher precision is
+absolutely required as the performance impact of using it can be
+significant.
+
+Declare Variables In The Scope Where They're Used
+-------------------------------------------------
+
+Performance is slightly improved by declaring variables at the same block
+scope where they are first used.  For example, in code like the
+following, if the lifetime of ``foo`` is only within the scope of the
+``if`` clause, write the code like this:  
+
+::
+
+    float func() {
+        ....
+        if (x < y) {
+            float foo;
+            ... use foo ...
+        }
+    }
+
+Try not to write code as:
+
+::
+
+    float func() {
+        float foo;
+        ....
+        if (x < y) {
+            ... use foo ...
+        }
+    }
+
+Doing so can reduce the amount of masked store instructions that the
+compiler needs to generate.
+
+Instrumenting ISPC Programs To Understand Runtime Behavior
+----------------------------------------------------------
+
+``ispc`` has an optional instrumentation feature that can help you
+understand performance issues.  If a program is compiled using the
+``--instrument`` flag, the compiler emits calls to a function with the
+following signature at various points in the program (for
+example, at interesting points in the control flow, when scatters or
+gathers happen.)
+
+::
+
+    extern "C" {
+        void ISPCInstrument(const char *fn, const char *note, 
+                            int line, int mask);
+    }
+
+This function is passed the file name of the ``ispc`` file running, a short
+note indicating what is happening, the line number in the source file, and
+the current mask of active program instances in the gang.  You must provide an
+implementation of this function and link it in with your application.
+
+For example, when the ``ispc`` program runs, this function might be called
+as follows:
+
+::
+
+   ISPCInstrument("foo.ispc", "function entry", 55, 0xf);
+
+This call indicates that at the currently executing program has just
+entered the function defined at line 55 of the file ``foo.ispc``, with a
+mask of all lanes currently executing (assuming a four-wide gang size
+target machine).
+
+For a fuller example of the utility of this functionality, see
+``examples/aobench_instrumented`` in the ``ispc`` distribution.  Ths
+example includes an implementation of the ``ISPCInstrument()`` function
+that collects aggregate data about the program's execution behavior.
+
+When running this example, you will want to direct to the ``ao`` executable
+to generate a low resolution image, because the instrumentation adds
+substantial execution overhead.  For example:
+
+::
+
+    % ./ao 1 32 32
+
+After the ``ao`` program exits, a summary report along the following lines
+will be printed.  In the first few lines, you can see how many times a few
+functions were called, and the average percentage of SIMD lanes that were
+active upon function entry.
+
+:: 
+
+    ao.ispc(0067) - function entry: 342424 calls (0 / 0.00% all off!), 95.86% active lanes
+    ao.ispc(0067) - return: uniform control flow: 342424 calls (0 / 0.00% all off!), 95.86% active lanes
+    ao.ispc(0071) - function entry: 1122 calls (0 / 0.00% all off!), 97.33% active lanes
+    ao.ispc(0075) - return: uniform control flow: 1122 calls (0 / 0.00% all off!), 97.33% active lanes
+    ao.ispc(0079) - function entry: 10072 calls (0 / 0.00% all off!), 45.09% active lanes
+    ao.ispc(0088) - function entry: 36928 calls (0 / 0.00% all off!), 97.40% active lanes
+    ...
+
+
+Choosing A Target Vector Width
+------------------------------
+
+By default, ``ispc`` compiles to the natural vector width of the target
+instruction set.  For example, for SSE2 and SSE4, it compiles four-wide,
+and for AVX, it complies 8-wide.  For some programs, higher performance may
+be seen if the program is compiled to a doubled vector width--8-wide for
+SSE and 16-wide for AVX.  
+
+For workloads that don't require many of registers, this method can lead to
+significantly more efficient execution thanks to greater instruction level
+parallelism and amortization of various overhead over more program
+instances.  For other workloads, it may lead to a slowdown due to higher
+register pressure; trying both approaches for key kernels may be
+worthwhile.
+
+This option is only available for each of the SSE2, SSE4 and AVX targets.
+It is selected with the ``--target=sse2-x2``, ``--target=sse4-x2`` and
+``--target=avx-x2`` options, respectively.
+
+
+Disclaimer and Legal Information
+================================
+
+INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL(R) PRODUCTS.
+NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL
+PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL'S TERMS
+AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER,
+AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE
+OF INTEL PRODUCTS INCLUDING LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A
+PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT
+OR OTHER INTELLECTUAL PROPERTY RIGHT.
+
+UNLESS OTHERWISE AGREED IN WRITING BY INTEL, THE INTEL PRODUCTS ARE NOT DESIGNED
+NOR INTENDED FOR ANY APPLICATION IN WHICH THE FAILURE OF THE INTEL PRODUCT COULD
+CREATE A SITUATION WHERE PERSONAL INJURY OR DEATH MAY OCCUR.
+
+Intel may make changes to specifications and product descriptions at any time,
+without notice. Designers must not rely on the absence or characteristics of any
+features or instructions marked "reserved" or "undefined." Intel reserves these
+for future definition and shall have no responsibility whatsoever for conflicts
+or incompatibilities arising from future changes to them. The information here
+is subject to change without notice. Do not finalize a design with this
+information.
+
+The products described in this document may contain design defects or errors
+known as errata which may cause the product to deviate from published
+specifications. Current characterized errata are available on request.
+
+Contact your local Intel sales office or your distributor to obtain the latest
+specifications and before placing your product order.
+
+Copies of documents which have an order number and are referenced in this
+document, or other Intel literature, may be obtained by calling 1-800-548-4725,
+or by visiting Intel's Web Site.
+
+Intel processor numbers are not a measure of performance. Processor numbers
+differentiate features within each processor family, not across different
+processor families. See http://www.intel.com/products/processor_number for
+details.
+
+BunnyPeople, Celeron, Celeron Inside, Centrino, Centrino Atom,
+Centrino Atom Inside, Centrino Inside, Centrino logo, Core Inside, FlashFile,
+i960, InstantIP, Intel, Intel logo, Intel386, Intel486, IntelDX2, IntelDX4,
+IntelSX2, Intel Atom, Intel Atom Inside, Intel Core, Intel Inside,
+Intel Inside logo, Intel. Leap ahead., Intel. Leap ahead. logo, Intel NetBurst,
+Intel NetMerge, Intel NetStructure, Intel SingleDriver, Intel SpeedStep,
+Intel StrataFlash, Intel Viiv, Intel vPro, Intel XScale, Itanium,
+Itanium Inside, MCS, MMX, Oplus, OverDrive, PDCharm, Pentium, Pentium Inside,
+skoool, Sound Mark, The Journey Inside, Viiv Inside, vPro Inside, VTune, Xeon,
+and Xeon Inside are trademarks of Intel Corporation in the U.S. and other
+countries.
+
+* Other names and brands may be claimed as the property of others.
+
+Copyright(C) 2011, Intel Corporation. All rights reserved.
+
+
+Optimization Notice
+===================
+
+Intel compilers, associated libraries and associated development tools may
+include or utilize options that optimize for instruction sets that are
+available in both Intel and non-Intel microprocessors (for example SIMD
+instruction sets), but do not optimize equally for non-Intel
+microprocessors.  In addition, certain compiler options for Intel
+compilers, including some that are not specific to Intel
+micro-architecture, are reserved for Intel microprocessors.  For a detailed
+description of Intel compiler options, including the instruction sets and
+specific microprocessors they implicate, please refer to the "Intel
+Compiler User and Reference Guides" under "Compiler Options."  Many library
+routines that are part of Intel compiler products are more highly optimized
+for Intel microprocessors than for other microprocessors.  While the
+compilers and libraries in Intel compiler products offer optimizations for
+both Intel and Intel-compatible microprocessors, depending on the options
+you select, your code and other factors, you likely will get extra
+performance on Intel microprocessors.
+
+Intel compilers, associated libraries and associated development tools may
+or may not optimize to the same degree for non-Intel microprocessors for
+optimizations that are not unique to Intel microprocessors.  These
+optimizations include Intel® Streaming SIMD Extensions 2 (Intel® SSE2),
+Intel® Streaming SIMD Extensions 3 (Intel® SSE3), and Supplemental
+Streaming SIMD Extensions 3 (Intel SSSE3) instruction sets and other
+optimizations.  Intel does not guarantee the availability, functionality,
+or effectiveness of any optimization on microprocessors not manufactured by
+Intel.  Microprocessor-dependent optimizations in this product are intended
+for use with Intel microprocessors.
+
+While Intel believes our compilers and libraries are excellent choices to
+assist in obtaining the best performance on Intel and non-Intel
+microprocessors, Intel recommends that you evaluate other compilers and
+libraries to determine which best meet your requirements.  We hope to win
+your business by striving to offer the best performance of any compiler or
+library; please let us know if you find we do not.
+

docs/template-perf.txt

@@ -0,0 +1,65 @@
+%(head_prefix)s
+%(head)s
+<script type="text/javascript">
+
+  var _gaq = _gaq || [];
+  _gaq.push(['_setAccount', 'UA-1486404-4']);
+  _gaq.push(['_trackPageview']);
+
+  (function() {
+    var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true;
+    ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js';
+    var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s);
+  })();
+
+</script>
+%(stylesheet)s
+%(body_prefix)s
+<div id="wrap">
+  <div id="wrap2">
+    <div id="header">
+      <h1 id="logo">Intel SPMD Program Compiler</h1>
+      <div id="slogan">An open-source compiler for high-performance SIMD programming on
+      the CPU</div>
+    </div>
+    <div id="nav">
+      <div id="nbar">
+        <ul>
+          <li><a href="index.html">Overview</a></li>
+          <li><a href="features.html">Features</a></li>
+          <li><a href="downloads.html">Downloads</a></li>
+          <li><a href="documentation.html">Documentation</a></li>
+          <li id="selected"><a href="perf.html">Performance</a></li>
+        </ul>
+      </div>
+    </div>
+    <div id="content-wrap">
+      <div id="sidebar">
+          <div class="widgetspace">
+            <h1>Resources</h1>
+            <ul class="menu">
+              <li><a href="http://github.com/ispc/ispc/">ispc page on github</a></li>
+              <li><a href="http://groups.google.com/group/ispc-users/">ispc
+              users mailing list</a></li>
+              <li><a href="http://groups.google.com/group/ispc-dev/">ispc
+              developers mailing list</a></li>
+              <li><a href="http://github.com/ispc/ispc/wiki/">Wiki</a></li>
+              <li><a href="http://github.com/ispc/ispc/issues/">Bug tracking</a></li>
+              <li><a href="doxygen/index.html">Doxygen documentation of
+              <tt>ispc</tt> source code</a></li>
+            </ul>
+        </div>
+      </div>
+%(body_pre_docinfo)s
+%(docinfo)s
+<div id="content">
+%(body)s
+</div>
+    <div class="clearfix"></div>
+    <div id="footer"> &copy; 2011 <strong>Intel Corporation</strong> | Valid <a href="http://validator.w3.org/check?uri=referer">XHTML</a> | <a href="http://jigsaw.w3.org/css-validator/check/referer">CSS</a> | ClearBlue  by: <a href="http://www.themebin.com/">ThemeBin</a>
+      <!-- Please Do Not remove this link, thank u -->
+      </div>
+      </div>
+      </div>
+      </div>
+%(body_suffix)s

docs/template.txt

@@ -0,0 +1,65 @@
+%(head_prefix)s
+%(head)s
+<script type="text/javascript">
+
+  var _gaq = _gaq || [];
+  _gaq.push(['_setAccount', 'UA-1486404-4']);
+  _gaq.push(['_trackPageview']);
+
+  (function() {
+    var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true;
+    ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js';
+    var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s);
+  })();
+
+</script>
+%(stylesheet)s
+%(body_prefix)s
+<div id="wrap">
+  <div id="wrap2">
+    <div id="header">
+      <h1 id="logo">Intel SPMD Program Compiler</h1>
+      <div id="slogan">An open-source compiler for high-performance SIMD programming on
+      the CPU</div>
+    </div>
+    <div id="nav">
+      <div id="nbar">
+        <ul>
+          <li><a href="index.html">Overview</a></li>
+          <li><a href="features.html">Features</a></li>
+          <li><a href="downloads.html">Downloads</a></li>
+          <li id="selected"><a href="documentation.html">Documentation</a></li>
+          <li><a href="perf.html">Performance</a></li>
+        </ul>
+      </div>
+    </div>
+    <div id="content-wrap">
+      <div id="sidebar">
+          <div class="widgetspace">
+            <h1>Resources</h1>
+            <ul class="menu">
+              <li><a href="http://github.com/ispc/ispc/">ispc page on github</a></li>
+              <li><a href="http://groups.google.com/group/ispc-users/">ispc
+              users mailing list</a></li>
+              <li><a href="http://groups.google.com/group/ispc-dev/">ispc
+              developers mailing list</a></li>
+              <li><a href="http://github.com/ispc/ispc/wiki/">Wiki</a></li>
+              <li><a href="http://github.com/ispc/ispc/issues/">Bug tracking</a></li>
+              <li><a href="doxygen/index.html">Doxygen documentation of
+              <tt>ispc</tt> source code</a></li>
+            </ul>
+        </div>
+      </div>
+%(body_pre_docinfo)s
+%(docinfo)s
+<div id="content">
+%(body)s
+</div>
+    <div class="clearfix"></div>
+    <div id="footer"> &copy; 2011 <strong>Intel Corporation</strong> | Valid <a href="http://validator.w3.org/check?uri=referer">XHTML</a> | <a href="http://jigsaw.w3.org/css-validator/check/referer">CSS</a> | ClearBlue  by: <a href="http://www.themebin.com/">ThemeBin</a>
+      <!-- Please Do Not remove this link, thank u -->
+      </div>
+      </div>
+      </div>
+      </div>
+%(body_suffix)s

doxygen.cfg

@@ -31,7 +31,7 @@ PROJECT_NAME           = "Intel SPMD Program Compiler"
 # This could be handy for archiving the generated documentation or
 # if some version control system is used.
 
-PROJECT_NUMBER         = 1.0
+PROJECT_NUMBER         = 1.1.0
 
 # The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute)
 # base path where the generated documentation will be put.
@@ -585,7 +585,6 @@ INPUT                  = builtins.h \
                          ctx.h \
                          decl.h \
                          expr.h \
-                         gatherbuf.h \
                          ispc.h \
                          llvmutil.h \
                          module.h \
@@ -598,7 +597,6 @@ INPUT                  = builtins.h \
                          ctx.cpp \
                          decl.cpp \
                          expr.cpp \
-                         gatherbuf.cpp \
                          ispc.cpp \
                          llvmutil.cpp \
                          main.cpp \
@@ -610,7 +608,7 @@ INPUT                  = builtins.h \
                          util.cpp \
                          parse.yy \
                          lex.ll \
-                         stdlib-c.c
+                         builtins-c.c
 
 # This tag can be used to specify the character encoding of the source files
 # that doxygen parses. Internally doxygen uses the UTF-8 encoding, which is

examples/README.txt

@@ -13,6 +13,7 @@ against regular serial C++ implementations, printing out a comparison of
 the runtimes and the speedup delivered by ispc.  It may be instructive to
 do a side-by-side diff of the C++ and ispc implementations of these
 algorithms to learn more about wirting ispc code.
+
  
 AOBench
 =======
@@ -27,6 +28,7 @@ It executes the program for the given number of iterations, rendering an
 (xres x yres) image each time and measuring the computation time with both
 serial and ispc implementations.
 
+
 AOBench_Instrumented
 ====================
 
@@ -40,12 +42,47 @@ is provided in the instrument.cpp file.
 *** Note: on Linux, this example currently hits an assertion in LLVM during
 *** compilation
 
+
+Deferred
+========
+
+This example shows an extensive example of using ispc for efficient
+deferred shading of scenes with thousands of lights; it's an implementation
+of the algorithm that Johan Andersson described at SIGGRAPH 2009,
+implemented by Andrew Lauritzen and Jefferson Montgomery.  The basic idea
+is that a pre-rendered G-buffer is partitioned into tiles, and in each
+tile, the set of lights that contribute to the tile is first computed.
+Then, the pixels in the tile are then shaded using just those light
+sources. (See slides 19-29 of
+http://s09.idav.ucdavis.edu/talks/04-JAndersson-ParallelFrostbite-Siggraph09.pdf
+for more details on the algorithm.)
+
+This directory includes three implementations of the algorithm:
+
+- An ispc implementation that first does a static partitioning of the
+  screen into tiles to parallelize across the CPU cores.  Within each tile
+  ispc kernels provide highly efficient implementations of the light
+  culling and shading calculations.
+- A "best practices" serial C++ implementation.  This implementation does a
+  dynamic partitioning of the screen, refining tiles with significant Z
+  depth complexity (these tiles often have a large number of lights that
+  affect them).  Within each final tile, the pixels are shaded using
+  regular C++ code.
+- If the Cilk extensions are available in your compiler, an ispc
+  implementation that uses Cilk will also be built.
+  (See http://software.intel.com/en-us/articles/intel-cilk-plus/).  Like 
+  the "best practices" serial implementation, this version does dynamic
+  tile partitioning for better load balancing and then uses ispc for the
+  light culling and shading.
+
+
 Mandelbrot
 ==========
 
 Mandelbrot set generation.  This example is extensively documented at the
 http://ispc.github.com/example.html page.
 
+
 Mandelbrot_tasks
 ================
 
@@ -57,6 +94,14 @@ Linux, a pthreads-based task system is used (tasks_pthreads.cpp).  When
 using tasks with ispc, no task system is mandated; the user is free to plug
 in any task system they want, for ease of interoperating with existing task
 systems.
+
+
+Noise
+=====
+
+This example has an implementation of Ken Perlin's procedural "noise"
+function, as described in his 2002 "Improving Noise" SIGGRAPH paper.
+
  
 Options
 =======
@@ -64,6 +109,7 @@ Options
 This program implements both the Black-Scholes and Binomial options pricing
 models in both ispc and regular serial C++ code.
 
+
 RT
 ==
 
@@ -80,9 +126,25 @@ and triangle intersection code from pbrt; see the pbrt source code and/or
 "Physically Based Rendering" book for more about the basic algorithmic
 details.
 
+
 Simple
 ======
 
 This is a simple "hello world" type program that shows a ~10 line
 application program calling out to a ~5 line ispc program to do a simple
 computation.
+
+
+Volume
+======
+
+Ray-marching volume rendering, with single scattering lighting model.  To
+run it, specify a camera parameter file and a volume density file, e.g.:
+
+volume camera.dat density_highres.vol
+
+(See, e.g. Chapters 11 and 16 of "Physically Based Rendering" for
+information about the algorithm implemented here.)  The volume data set
+included here was generated by the example implementation of the "Wavelet
+Turbulence for Fluid Simulation" SIGGRAPH 2008 paper by Kim et
+al. (http://www.cs.cornell.edu/~tedkim/WTURB/)

examples/aobench/.gitignore

@@ -0,0 +1,2 @@
+ao
+*.ppm

examples/aobench/Makefile

@@ -1,8 +1,18 @@
 
+ARCH = $(shell uname)
+
+TASK_CXX=../tasksys.cpp
+TASK_LIB=-lpthread
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
 CXX=g++
-CXXFLAGS=-Iobjs/ -O3 -Wall
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
 ISPC=ispc
-ISPCFLAGS=-O2 --fast-math
+ISPCFLAGS=-O2 --target=sse2,sse4,avx --arch=x86-64
+
+ISPC_OBJS=objs/ao_ispc.o objs/ao_ispc_sse2.o objs/ao_ispc_sse4.o \
+	objs/ao_ispc_avx.o
+OBJS=objs/ao.o objs/ao_serial.o $(ISPC_OBJS) $(TASK_OBJ)
 
 default: ao
 
@@ -14,13 +24,16 @@ dirs:
 clean:
 	/bin/rm -rf objs *~ ao
 
-ao: dirs objs/ao.o objs/ao_serial.o objs/ao_ispc.o
-	$(CXX) $(CXXFLAGS) -o $@ objs/ao.o objs/ao_ispc.o objs/ao_serial.o -lm -lpthread
+ao: dirs $(OBJS) $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm $(TASK_LIB)
 
 objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
 objs/ao.o: objs/ao_ispc.h 
 
-objs/%_ispc.h objs/%_ispc.o: %.ispc
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
 	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/aobench/ao.cpp

@@ -100,24 +100,7 @@ savePPM(const char *fname, int w, int h)
     fprintf(fp, "255\n");
     fwrite(img, w * h * 3, 1, fp);
     fclose(fp);
-}
-
-
-// Allocate memory with 64-byte alignment.
-float *
-AllocAligned(int size) {
-#if defined(_WIN32) || defined(_WIN64)
-    return (float *)_aligned_malloc(size, 64);
-#elif defined (__APPLE__)
-    // Allocate excess memory to ensure an aligned pointer can be returned
-    void *mem = malloc(size + (64-1) + sizeof(void*));
-    char *amem = ((char*)mem) + sizeof(void*);
-    amem += 64 - (reinterpret_cast<uint64_t>(amem) & (64 - 1));
-    ((void**)amem)[-1] = mem;
-    return (float *)amem;
-#else
-    return (float *)memalign(64, size);
-#endif
+    printf("Wrote image file %s\n", fname);
 }
 
 
@@ -136,8 +119,8 @@ int main(int argc, char **argv)
     }
 
     // Allocate space for output images
-    img = (unsigned char *)AllocAligned(width * height * 3);
-    fimg = (float *)AllocAligned(sizeof(float) * width * height * 3);
+    img = new unsigned char[width * height * 3];
+    fimg = new float[width * height * 3];
 
     //
     // Run the ispc path, test_iterations times, and report the minimum
@@ -155,10 +138,30 @@ int main(int argc, char **argv)
     }
 
     // Report results and save image
-    printf("[aobench ispc]:\t\t\t[%.3f] M cycles (%d x %d image)\n", minTimeISPC, 
-           width, height);
+    printf("[aobench ispc]:\t\t\t[%.3f] M cycles (%d x %d image)\n", 
+           minTimeISPC, width, height);
     savePPM("ao-ispc.ppm", width, height); 
 
+    //
+    // Run the ispc + tasks path, test_iterations times, and report the
+    // minimum time for any of them.
+    //
+    double minTimeISPCTasks = 1e30;
+    for (unsigned int i = 0; i < test_iterations; i++) {
+        memset((void *)fimg, 0, sizeof(float) * width * height * 3);
+        assert(NSUBSAMPLES == 2);
+
+        reset_and_start_timer();
+        ao_ispc_tasks(width, height, NSUBSAMPLES, fimg);
+        double t = get_elapsed_mcycles();
+        minTimeISPCTasks = std::min(minTimeISPCTasks, t);
+    }
+
+    // Report results and save image
+    printf("[aobench ispc + tasks]:\t\t[%.3f] M cycles (%d x %d image)\n", 
+           minTimeISPCTasks, width, height);
+    savePPM("ao-ispc-tasks.ppm", width, height); 
+
     //
     // Run the serial path, again test_iteration times, and report the
     // minimum time.
@@ -175,7 +178,8 @@ int main(int argc, char **argv)
     // Report more results, save another image...
     printf("[aobench serial]:\t\t[%.3f] M cycles (%d x %d image)\n", minTimeSerial, 
            width, height);
-    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", minTimeSerial / minTimeISPC);
+    printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n", 
+           minTimeSerial / minTimeISPC, minTimeSerial / minTimeISPCTasks);
     savePPM("ao-serial.ppm", width, height); 
         
     return 0;

examples/aobench/ao.ispc

@@ -75,7 +75,7 @@ static inline vec vcross(vec v0, vec v1) {
     return ret;
 }
 
-static inline void vnormalize(reference vec v) {
+static inline void vnormalize(vec &v) {
     float len2 = dot(v, v);
     float invlen = rsqrt(len2);
     v *= invlen;
@@ -83,8 +83,7 @@ static inline void vnormalize(reference vec v) {
 
 
 static inline void
-ray_plane_intersect(reference Isect isect, reference Ray ray, 
-                    reference Plane plane) {
+ray_plane_intersect(Isect &isect, Ray &ray, Plane &plane) {
     float d = -dot(plane.p, plane.n);
     float v = dot(ray.dir, plane.n);
 
@@ -104,8 +103,7 @@ ray_plane_intersect(reference Isect isect, reference Ray ray,
 
 
 static inline void
-ray_sphere_intersect(reference Isect isect, reference Ray ray, 
-                     reference Sphere sphere) {
+ray_sphere_intersect(Isect &isect, Ray &ray, Sphere &sphere) {
     vec rs = ray.org - sphere.center;
 
     float B = dot(rs, ray.dir);
@@ -127,7 +125,7 @@ ray_sphere_intersect(reference Isect isect, reference Ray ray,
 
 
 static inline void
-orthoBasis(reference vec basis[3], vec n) {
+orthoBasis(vec basis[3], vec n) {
     basis[2] = n;
     basis[1].x = 0.0; basis[1].y = 0.0; basis[1].z = 0.0;
 
@@ -150,8 +148,8 @@ orthoBasis(reference vec basis[3], vec n) {
 
 
 static inline float
-ambient_occlusion(reference Isect isect, reference Plane plane, 
-                  reference Sphere spheres[3], reference RNGState rngstate) {
+ambient_occlusion(Isect &isect, Plane &plane, Sphere spheres[3], 
+                  RNGState &rngstate) {
     float eps = 0.0001f;
     vec p, n;
     vec basis[3];
@@ -168,8 +166,8 @@ ambient_occlusion(reference Isect isect, reference Plane plane,
             Ray ray;
             Isect occIsect;
 
-            float theta = sqrt(frandom(rngstate));
-            float phi   = 2.0f * M_PI * frandom(rngstate);
+            float theta = sqrt(frandom(&rngstate));
+            float phi   = 2.0f * M_PI * frandom(&rngstate);
             float x = cos(phi) * theta;
             float y = sin(phi) * theta;
             float z = sqrt(1.0 - theta * theta);
@@ -203,8 +201,9 @@ ambient_occlusion(reference Isect isect, reference Plane plane,
 /* Compute the image for the scanlines from [y0,y1), for an overall image
    of width w and height h.
  */
-void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h, 
-                  uniform int nsubsamples, reference uniform float image[]) {
+static void ao_scanlines(uniform int y0, uniform int y1, uniform int w, 
+                         uniform int h,  uniform int nsubsamples, 
+                         uniform float image[]) {
     static Plane plane = { { 0.0f, -0.5f, 0.0f }, { 0.f, 1.f, 0.f } };
     static Sphere spheres[3] = {
         { { -2.0f, 0.0f, -3.5f }, 0.5f },
@@ -212,7 +211,7 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
         { { 1.0f, 0.0f, -2.2f }, 0.5f } };
     RNGState rngstate;
 
-    seed_rng(rngstate, y0);
+    seed_rng(&rngstate, y0);
 
     // Compute the mapping between the 'programCount'-wide program
     // instances running in parallel and samples in the image.  
@@ -231,6 +230,9 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
     // direction we do per iteration and ny the number in y.
     uniform int nx = 1, ny = 1;
 
+    // FIXME: We actually need ny to be 1 regardless of the decomposition,
+    // since the task decomposition is one scanline high.
+
     if (programCount == 8) {
         // Do two pixels at once in the x direction
         nx = 2;
@@ -239,19 +241,21 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
             ++du;
     }
     else if (programCount == 16) {
-        // Two at once in both x and y
-        nx = ny = 2;
-        if ((programIndex >= 4 && programIndex < 8) || programIndex >= 12)
+        nx = 4;
+        ny = 1;
+        if (programIndex >= 4 && programIndex < 8)
             ++du;
-        if (programIndex >= 8)  
-            ++dv;
+        if (programIndex >= 8 && programIndex < 12)
+            du += 2;
+        if (programIndex >= 12)
+            du += 3;
     }
 
     // Now loop over all of the pixels, stepping in x and y as calculated
     // above.  (Assumes that ny divides y and nx divides x...)
     for (uniform int y = y0; y < y1; y += ny) {
         for (uniform int x = 0; x < w; x += nx)  {
-            // Figur out x,y pixel in NDC
+            // Figure out x,y pixel in NDC
             float px =  (x + du - (w / 2.0f)) / (w / 2.0f);
             float py = -(y + dv - (h / 2.0f)) / (h / 2.0f);
             float ret = 0.f;
@@ -293,7 +297,7 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
 
             // offset to the first pixel in the image
             uniform int offset = 3 * (y * w + x);
-            for (uniform int p = 0; p < programCount; p += 4, ++offset) {
+            for (uniform int p = 0; p < programCount; p += 4, offset += 3) {
                 // Get the four sample values for this pixel
                 uniform float sumret = retArray[p] + retArray[p+1] + retArray[p+2] +
                     retArray[p+3];
@@ -315,3 +319,15 @@ export void ao_ispc(uniform int w, uniform int h, uniform int nsubsamples,
                     uniform float image[]) {
     ao_scanlines(0, h, w, h, nsubsamples, image);
 }
+
+
+static void task ao_task(uniform int width, uniform int height, 
+                         uniform int nsubsamples, uniform float image[]) {
+    ao_scanlines(taskIndex, taskIndex+1, width, height, nsubsamples, image);
+}
+
+
+export void ao_ispc_tasks(uniform int w, uniform int h, uniform int nsubsamples, 
+                          uniform float image[]) {
+    launch[h] < ao_task(w, h, nsubsamples, image) >;
+}

examples/aobench/ao_serial.cpp

@@ -140,7 +140,7 @@ ray_plane_intersect(Isect &isect, Ray &ray,
     float d = -dot(plane.p, plane.n);
     float v = dot(ray.dir, plane.n);
 
-    if (fabsf(v) < 1.0e-17) 
+    if (fabsf(v) < 1.0e-17f) 
         return;
     else {
         float t = -(dot(ray.org, plane.n) + d) / v;
@@ -183,11 +183,11 @@ orthoBasis(vec basis[3], const vec &n) {
     basis[2] = n;
     basis[1].x = 0.0; basis[1].y = 0.0; basis[1].z = 0.0;
 
-    if ((n.x < 0.6) && (n.x > -0.6)) {
+    if ((n.x < 0.6f) && (n.x > -0.6f)) {
         basis[1].x = 1.0;
-    } else if ((n.y < 0.6) && (n.y > -0.6)) {
+    } else if ((n.y < 0.6f) && (n.y > -0.6f)) {
         basis[1].y = 1.0;
-    } else if ((n.z < 0.6) && (n.z > -0.6)) {
+    } else if ((n.z < 0.6f) && (n.z > -0.6f)) {
         basis[1].z = 1.0;
     } else {
         basis[1].x = 1.0;
@@ -224,7 +224,7 @@ ambient_occlusion(Isect &isect, Plane &plane,
             float phi   = 2.0f * M_PI * drand48();
             float x = cosf(phi) * theta;
             float y = sinf(phi) * theta;
-            float z = sqrtf(1.0 - theta * theta);
+            float z = sqrtf(1.0f - theta * theta);
 
             // local . global
             float rx = x * basis[0].x + y * basis[1].x + z * basis[2].x;
@@ -236,14 +236,14 @@ ambient_occlusion(Isect &isect, Plane &plane,
             ray.dir.y = ry;
             ray.dir.z = rz;
 
-            occIsect.t   = 1.0e+17;
+            occIsect.t   = 1.0e+17f;
             occIsect.hit = 0;
 
             for (int snum = 0; snum < 3; ++snum)
                 ray_sphere_intersect(occIsect, ray, spheres[snum]); 
             ray_plane_intersect (occIsect, ray, plane); 
 
-            if (occIsect.hit) occlusion += 1.0;
+            if (occIsect.hit) occlusion += 1.f;
         }
     }
 
@@ -280,10 +280,10 @@ static void ao_scanlines(int y0, int y1, int w, int h, int nsubsamples,
 
                     ray.dir.x = px;
                     ray.dir.y = py;
-                    ray.dir.z = -1.0;
+                    ray.dir.z = -1.0f;
                     vnormalize(ray.dir);
 
-                    isect.t   = 1.0e+17;
+                    isect.t   = 1.0e+17f;
                     isect.hit = 0;
 
                     for (int snum = 0; snum < 3; ++snum)

examples/aobench/aobench.vcxproj

@@ -21,22 +21,23 @@
   <ItemGroup>
     <ClCompile Include="ao.cpp" />
     <ClCompile Include="ao_serial.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="ao.ispc">
       <FileType>Document</FileType>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4,avx
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4,avx
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj</Outputs>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4,avx
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4,avx
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <PropertyGroup Label="Globals">
@@ -85,15 +86,19 @@
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ExecutablePath);$(ProjectDir)..\..</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -102,6 +107,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -115,6 +123,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -130,6 +141,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -147,6 +160,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>

examples/aobench_instrumented/.gitignore

@@ -0,0 +1,2 @@
+ao
+*.ppm

examples/aobench_instrumented/Makefile

@@ -1,8 +1,8 @@
 
-CXX=g++
+CXX=g++ -m64
 CXXFLAGS=-Iobjs/ -g3 -Wall
 ISPC=ispc
-ISPCFLAGS=-O2 --fast-math --instrument
+ISPCFLAGS=-O2 --instrument --arch=x86-64 --target=sse2
 
 default: ao
 

examples/aobench_instrumented/ao.cpp

@@ -32,7 +32,6 @@
 */
 
 #ifdef _MSC_VER
-#define _CRT_SECURE_NO_WARNINGS
 #define NOMINMAX
 #pragma warning (disable: 4244)
 #pragma warning (disable: 4305)
@@ -51,7 +50,7 @@
 #include <algorithm>
 #include <sys/types.h>
 
-#include "ao_ispc.h"
+#include "ao_instrumented_ispc.h"
 using namespace ispc;
 
 #include "instrument.h"
@@ -99,26 +98,10 @@ savePPM(const char *fname, int w, int h)
     fprintf(fp, "255\n");
     fwrite(img, w * h * 3, 1, fp);
     fclose(fp);
+    printf("Wrote image file %s\n", fname);
 }
 
 
-// Allocate memory with 64-byte alignment.
-float *
-AllocAligned(int size) {
-#if defined(_WIN32) || defined(_WIN64)
-    return (float *)_aligned_malloc(size, 64);
-#elif defined (__APPLE__)
-    // Allocate excess memory to ensure an aligned pointer can be returned
-    void *mem = malloc(size + (64-1) + sizeof(void*));
-    char *amem = ((char*)mem) + sizeof(void*);
-    amem += 64 - (reinterpret_cast<uint64_t>(amem) & (64 - 1));
-    ((void**)amem)[-1] = mem;
-    return (float *)amem;
-#else
-    return (float *)memalign(64, size);
-#endif
-}
-
 
 int main(int argc, char **argv)
 {
@@ -135,8 +118,8 @@ int main(int argc, char **argv)
     }
 
     // Allocate space for output images
-    img = (unsigned char *)AllocAligned(width * height * 3);
-    fimg = (float *)AllocAligned(sizeof(float) * width * height * 3);
+    img = new unsigned char[width * height * 3];
+    fimg = new float[width * height * 3];
 
     ao_ispc(width, height, NSUBSAMPLES, fimg);
 

examples/aobench_instrumented/ao.ispc

@@ -75,7 +75,7 @@ static inline vec vcross(vec v0, vec v1) {
     return ret;
 }
 
-static inline void vnormalize(reference vec v) {
+static inline void vnormalize(vec &v) {
     float len2 = dot(v, v);
     float invlen = rsqrt(len2);
     v *= invlen;
@@ -83,8 +83,7 @@ static inline void vnormalize(reference vec v) {
 
 
 static inline void
-ray_plane_intersect(reference Isect isect, reference Ray ray, 
-                    reference Plane plane) {
+ray_plane_intersect(Isect &isect, Ray &ray, Plane &plane) {
     float d = -dot(plane.p, plane.n);
     float v = dot(ray.dir, plane.n);
 
@@ -104,8 +103,7 @@ ray_plane_intersect(reference Isect isect, reference Ray ray,
 
 
 static inline void
-ray_sphere_intersect(reference Isect isect, reference Ray ray, 
-                     reference Sphere sphere) {
+ray_sphere_intersect(Isect &isect, Ray &ray, Sphere &sphere) {
     vec rs = ray.org - sphere.center;
 
     float B = dot(rs, ray.dir);
@@ -127,7 +125,7 @@ ray_sphere_intersect(reference Isect isect, reference Ray ray,
 
 
 static inline void
-orthoBasis(reference vec basis[3], vec n) {
+orthoBasis(vec basis[3], vec n) {
     basis[2] = n;
     basis[1].x = 0.0; basis[1].y = 0.0; basis[1].z = 0.0;
 
@@ -150,8 +148,8 @@ orthoBasis(reference vec basis[3], vec n) {
 
 
 static inline float
-ambient_occlusion(reference Isect isect, reference Plane plane, 
-                  reference Sphere spheres[3], reference RNGState rngstate) {
+ambient_occlusion(Isect &isect, Plane &plane, Sphere spheres[3], 
+                  RNGState &rngstate) {
     float eps = 0.0001f;
     vec p, n;
     vec basis[3];
@@ -168,8 +166,8 @@ ambient_occlusion(reference Isect isect, reference Plane plane,
             Ray ray;
             Isect occIsect;
 
-            float theta = sqrt(frandom(rngstate));
-            float phi   = 2.0f * M_PI * frandom(rngstate);
+            float theta = sqrt(frandom(&rngstate));
+            float phi   = 2.0f * M_PI * frandom(&rngstate);
             float x = cos(phi) * theta;
             float y = sin(phi) * theta;
             float z = sqrt(1.0 - theta * theta);
@@ -203,8 +201,9 @@ ambient_occlusion(reference Isect isect, reference Plane plane,
 /* Compute the image for the scanlines from [y0,y1), for an overall image
    of width w and height h.
  */
-void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h, 
-                  uniform int nsubsamples, reference uniform float image[]) {
+static void ao_scanlines(uniform int y0, uniform int y1, uniform int w, 
+                         uniform int h,  uniform int nsubsamples, 
+                         uniform float image[]) {
     static Plane plane = { { 0.0f, -0.5f, 0.0f }, { 0.f, 1.f, 0.f } };
     static Sphere spheres[3] = {
         { { -2.0f, 0.0f, -3.5f }, 0.5f },
@@ -212,7 +211,7 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
         { { 1.0f, 0.0f, -2.2f }, 0.5f } };
     RNGState rngstate;
 
-    seed_rng(rngstate, y0);
+    seed_rng(&rngstate, y0);
 
     // Compute the mapping between the 'programCount'-wide program
     // instances running in parallel and samples in the image.  
@@ -231,6 +230,9 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
     // direction we do per iteration and ny the number in y.
     uniform int nx = 1, ny = 1;
 
+    // FIXME: We actually need ny to be 1 regardless of the decomposition,
+    // since the task decomposition is one scanline high.
+
     if (programCount == 8) {
         // Do two pixels at once in the x direction
         nx = 2;
@@ -239,19 +241,21 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
             ++du;
     }
     else if (programCount == 16) {
-        // Two at once in both x and y
-        nx = ny = 2;
-        if ((programIndex >= 4 && programIndex < 8) || programIndex >= 12)
+        nx = 4;
+        ny = 1;
+        if (programIndex >= 4 && programIndex < 8)
             ++du;
-        if (programIndex >= 8)  
-            ++dv;
+        if (programIndex >= 8 && programIndex < 12)
+            du += 2;
+        if (programIndex >= 12)
+            du += 3;
     }
 
     // Now loop over all of the pixels, stepping in x and y as calculated
     // above.  (Assumes that ny divides y and nx divides x...)
     for (uniform int y = y0; y < y1; y += ny) {
         for (uniform int x = 0; x < w; x += nx)  {
-            // Figur out x,y pixel in NDC
+            // Figure out x,y pixel in NDC
             float px =  (x + du - (w / 2.0f)) / (w / 2.0f);
             float py = -(y + dv - (h / 2.0f)) / (h / 2.0f);
             float ret = 0.f;
@@ -293,7 +297,7 @@ void ao_scanlines(uniform int y0, uniform int y1, uniform int w, uniform int h,
 
             // offset to the first pixel in the image
             uniform int offset = 3 * (y * w + x);
-            for (uniform int p = 0; p < programCount; p += 4, ++offset) {
+            for (uniform int p = 0; p < programCount; p += 4, offset += 3) {
                 // Get the four sample values for this pixel
                 uniform float sumret = retArray[p] + retArray[p+1] + retArray[p+2] +
                     retArray[p+3];
@@ -315,3 +319,15 @@ export void ao_ispc(uniform int w, uniform int h, uniform int nsubsamples,
                     uniform float image[]) {
     ao_scanlines(0, h, w, h, nsubsamples, image);
 }
+
+
+static void task ao_task(uniform int width, uniform int height, 
+                         uniform int nsubsamples, uniform float image[]) {
+    ao_scanlines(taskIndex, taskIndex+1, width, height, nsubsamples, image);
+}
+
+
+export void ao_ispc_tasks(uniform int w, uniform int h, uniform int nsubsamples, 
+                          uniform float image[]) {
+    launch[h] < ao_task(w, h, nsubsamples, image) >;
+}

examples/aobench_instrumented/aobench_instrumented.vcxproj

@@ -21,22 +21,23 @@
   <ItemGroup>
     <ClCompile Include="ao.cpp" />
     <ClCompile Include="instrument.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="ao.ispc">
       <FileType>Document</FileType>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --instrument
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename)_instrumented.obj -h $(TargetDir)%(Filename)_instrumented_ispc.h --arch=x86 --instrument --target=sse2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --instrument
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename)_instrumented.obj -h $(TargetDir)%(Filename)_instrumented_ispc.h --instrument --target=sse2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj</Outputs>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --instrument
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename)_instrumented.obj;$(TargetDir)%(Filename)_instrumented_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename)_instrumented.obj;$(TargetDir)%(Filename)_instrumented_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename)_instrumented.obj -h $(TargetDir)%(Filename)_instrumented_ispc.h --arch=x86 --instrument --target=sse2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --instrument
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename)_instrumented.obj -h $(TargetDir)%(Filename)_instrumented_ispc.h --instrument --target=sse2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename)_instrumented.obj;$(TargetDir)%(Filename)_instrumented_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename)_instrumented.obj;$(TargetDir)%(Filename)_instrumented_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <PropertyGroup Label="Globals">
@@ -85,15 +86,23 @@
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+    <PreBuildEventUseInBuild>true</PreBuildEventUseInBuild>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+    <PreBuildEventUseInBuild>true</PreBuildEventUseInBuild>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+    <PreBuildEventUseInBuild>true</PreBuildEventUseInBuild>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+    <PreBuildEventUseInBuild>true</PreBuildEventUseInBuild>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -101,7 +110,8 @@
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;_CRT_SECURE_NO_WARNINGS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -114,7 +124,8 @@
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;_CRT_SECURE_NO_WARNINGS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -129,7 +140,8 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;_CRT_SECURE_NO_WARNINGS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -146,7 +158,8 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;_CRT_SECURE_NO_WARNINGS;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -158,4 +171,4 @@
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

examples/deferred/Makefile

@@ -0,0 +1,38 @@
+
+ARCH = $(shell uname)
+
+TASK_CXX=../tasksys.cpp
+TASK_LIB=-lpthread
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
+CXX=g++
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
+ISPC=ispc
+ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx-x2 --arch=x86-64 --math-lib=fast
+
+OBJS=objs/main.o objs/common.o objs/kernels_ispc.o objs/kernels_ispc_sse2.o \
+	objs/kernels_ispc_sse4.o objs/kernels_ispc_avx.o \
+	objs/dynamic_c.o objs/dynamic_cilk.o
+
+default: deferred_shading
+
+.PHONY: dirs clean
+.PRECIOUS: objs/kernels_ispc.h
+
+dirs:
+	/bin/mkdir -p objs/
+
+clean:
+	/bin/rm -rf objs *~ deferred_shading
+
+deferred_shading: dirs $(OBJS) $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) $(TASK_OBJ) -lm $(TASK_LIB)
+
+objs/%.o: %.cpp objs/kernels_ispc.h deferred.h
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
+	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/deferred/common.cpp

@@ -0,0 +1,209 @@
+/*
+  Copyright (c) 2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#ifdef _MSC_VER
+#define _CRT_SECURE_NO_WARNINGS
+#define ISPC_IS_WINDOWS
+#elif defined(__linux__)
+#define ISPC_IS_LINUX
+#elif defined(__APPLE__)
+#define ISPC_IS_APPLE
+#endif
+
+#include <fcntl.h>
+#include <float.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+#include <stdint.h>
+#include <algorithm>
+#include <assert.h>
+#include <vector>
+#ifdef ISPC_IS_WINDOWS
+  #define WIN32_LEAN_AND_MEAN
+  #include <windows.h>
+#endif
+#ifdef ISPC_IS_LINUX
+  #include <malloc.h>
+#endif
+#include "deferred.h"
+#include "../timing.h"
+
+///////////////////////////////////////////////////////////////////////////
+
+static void *
+lAlignedMalloc(size_t size, int32_t alignment) {
+#ifdef ISPC_IS_WINDOWS
+    return _aligned_malloc(size, alignment);
+#endif
+#ifdef ISPC_IS_LINUX
+    return memalign(alignment, size);
+#endif
+#ifdef ISPC_IS_APPLE
+    void *mem = malloc(size + (alignment-1) + sizeof(void*));
+    char *amem = ((char*)mem) + sizeof(void*);
+    amem = amem + uint32_t(alignment - (reinterpret_cast<uint64_t>(amem) &
+                                        (alignment - 1)));
+    ((void**)amem)[-1] = mem;
+    return amem;
+#endif
+}
+
+
+static void
+lAlignedFree(void *ptr) {
+#ifdef ISPC_IS_WINDOWS
+    _aligned_free(ptr);
+#endif
+#ifdef ISPC_IS_LINUX
+    free(ptr);
+#endif
+#ifdef ISPC_IS_APPLE
+    free(((void**)ptr)[-1]);
+#endif
+}
+
+
+Framebuffer::Framebuffer(int width, int height) {
+    nPixels = width*height;
+    r = (uint8_t *)lAlignedMalloc(nPixels, ALIGNMENT_BYTES);
+    g = (uint8_t *)lAlignedMalloc(nPixels, ALIGNMENT_BYTES);
+    b = (uint8_t *)lAlignedMalloc(nPixels, ALIGNMENT_BYTES);
+}
+
+
+Framebuffer::~Framebuffer() {
+    lAlignedFree(r);
+    lAlignedFree(g);
+    lAlignedFree(b);
+}
+
+
+void
+Framebuffer::clear() {
+    memset(r, 0, nPixels);
+    memset(g, 0, nPixels);
+    memset(b, 0, nPixels);
+}
+
+
+InputData *
+CreateInputDataFromFile(const char *path) {
+    FILE *in = fopen(path, "rb");
+    if (!in) return 0;
+
+    InputData *input = new InputData;
+
+    // Load header
+    if (fread(&input->header, sizeof(ispc::InputHeader), 1, in) != 1) {
+        fprintf(stderr, "Preumature EOF reading file \"%s\"\n", path);
+        return NULL;
+    }
+
+    // Load data chunk and update pointers
+    input->chunk = (uint8_t *)lAlignedMalloc(input->header.inputDataChunkSize, 
+                                             ALIGNMENT_BYTES);
+    if (fread(input->chunk, input->header.inputDataChunkSize, 1, in) != 1) {
+        fprintf(stderr, "Preumature EOF reading file \"%s\"\n", path);
+        return NULL;
+    }
+    
+    input->arrays.zBuffer =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaZBuffer]];
+    input->arrays.normalEncoded_x =
+        (uint16_t *)&input->chunk[input->header.inputDataArrayOffsets[idaNormalEncoded_x]];
+    input->arrays.normalEncoded_y =
+        (uint16_t *)&input->chunk[input->header.inputDataArrayOffsets[idaNormalEncoded_y]];
+    input->arrays.specularAmount =
+        (uint16_t *)&input->chunk[input->header.inputDataArrayOffsets[idaSpecularAmount]];
+    input->arrays.specularPower =
+        (uint16_t *)&input->chunk[input->header.inputDataArrayOffsets[idaSpecularPower]];
+    input->arrays.albedo_x =
+        (uint8_t *)&input->chunk[input->header.inputDataArrayOffsets[idaAlbedo_x]];
+    input->arrays.albedo_y =
+        (uint8_t *)&input->chunk[input->header.inputDataArrayOffsets[idaAlbedo_y]];
+    input->arrays.albedo_z =
+        (uint8_t *)&input->chunk[input->header.inputDataArrayOffsets[idaAlbedo_z]];
+    input->arrays.lightPositionView_x =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightPositionView_x]];
+    input->arrays.lightPositionView_y =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightPositionView_y]];
+    input->arrays.lightPositionView_z =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightPositionView_z]];
+    input->arrays.lightAttenuationBegin =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightAttenuationBegin]];
+    input->arrays.lightColor_x =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightColor_x]];
+    input->arrays.lightColor_y =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightColor_y]];
+    input->arrays.lightColor_z =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightColor_z]];
+    input->arrays.lightAttenuationEnd =
+        (float *)&input->chunk[input->header.inputDataArrayOffsets[idaLightAttenuationEnd]];
+
+    fclose(in);
+    return input;
+}
+
+
+void DeleteInputData(InputData *input) {
+    lAlignedFree(input->chunk);
+}
+
+
+void WriteFrame(const char *filename, const InputData *input,
+                const Framebuffer &framebuffer) {
+    // Deswizzle and copy to RGBA output
+    // Doesn't need to be fast... only happens once
+    size_t imageBytes = 3 * input->header.framebufferWidth * 
+        input->header.framebufferHeight;
+    uint8_t* framebufferAOS = (uint8_t *)lAlignedMalloc(imageBytes, ALIGNMENT_BYTES);
+    memset(framebufferAOS, 0, imageBytes);
+
+    for (int i = 0; i < input->header.framebufferWidth * 
+                        input->header.framebufferHeight; ++i) {
+        framebufferAOS[3 * i + 0] = framebuffer.r[i];
+        framebufferAOS[3 * i + 1] = framebuffer.g[i];
+        framebufferAOS[3 * i + 2] = framebuffer.b[i];
+    }
+    
+    // Write out simple PPM file
+    FILE *out = fopen(filename, "wb");
+    fprintf(out, "P6 %d %d 255\n", input->header.framebufferWidth, 
+            input->header.framebufferHeight);
+    fwrite(framebufferAOS, imageBytes, 1, out);
+
+    lAlignedFree(framebufferAOS);
+}

examples/deferred/deferred.h

@@ -0,0 +1,108 @@
+/*
+  Copyright (c) 2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#ifndef DEFERRED_H
+#define DEFERRED_H
+
+// Currently tile widths must be a multiple of SIMD width (i.e. 8 for ispc sse4x2)!
+#define MIN_TILE_WIDTH 16
+#define MIN_TILE_HEIGHT 16
+#define MAX_LIGHTS 1024
+
+enum InputDataArraysEnum {
+    idaZBuffer = 0,
+    idaNormalEncoded_x,
+    idaNormalEncoded_y,
+    idaSpecularAmount,
+    idaSpecularPower,
+    idaAlbedo_x,
+    idaAlbedo_y,
+    idaAlbedo_z,
+    idaLightPositionView_x,
+    idaLightPositionView_y,
+    idaLightPositionView_z,
+    idaLightAttenuationBegin,
+    idaLightColor_x,
+    idaLightColor_y,
+    idaLightColor_z,
+    idaLightAttenuationEnd,
+
+    idaNum
+};
+
+#ifndef ISPC
+
+#include <stdint.h>
+#include "kernels_ispc.h"
+
+#define ALIGNMENT_BYTES 64
+
+#define MAX_LIGHTS 1024
+
+#define VISUALIZE_LIGHT_COUNT 0
+
+struct InputData
+{
+    ispc::InputHeader header;
+    ispc::InputDataArrays arrays;
+    uint8_t *chunk;
+};
+
+
+struct Framebuffer {
+    Framebuffer(int width, int height);
+    ~Framebuffer();
+
+    void clear();
+
+    uint8_t *r, *g, *b;
+
+private:
+    int nPixels;
+    Framebuffer(const Framebuffer &);
+    Framebuffer &operator=(const Framebuffer *);
+};
+
+
+InputData *CreateInputDataFromFile(const char *path);
+void DeleteInputData(InputData *input);
+void WriteFrame(const char *filename, const InputData *input,
+                const Framebuffer &framebuffer);
+void InitDynamicC(InputData *input);
+void InitDynamicCilk(InputData *input);
+void DispatchDynamicC(InputData *input, Framebuffer *framebuffer);
+void DispatchDynamicCilk(InputData *input, Framebuffer *framebuffer);
+
+#endif // !ISPC
+
+#endif // DEFERRED_H

examples/deferred/deferred_shading.vcxproj

@@ -0,0 +1,178 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+  <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug|Win32">
+      <Configuration>Debug</Configuration>
+      <Platform>Win32</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Debug|x64">
+      <Configuration>Debug</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|Win32">
+      <Configuration>Release</Configuration>
+      <Platform>Win32</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|x64">
+      <Configuration>Release</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+  </ItemGroup>
+  <PropertyGroup Label="Globals">
+    <ProjectGuid>{87f53c53-957e-4e91-878a-bc27828fb9eb}</ProjectGuid>
+    <Keyword>Win32Proj</Keyword>
+    <RootNamespace>mandelbrot</RootNamespace>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>true</UseDebugLibraries>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>true</UseDebugLibraries>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>false</UseDebugLibraries>
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>false</UseDebugLibraries>
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
+  <ImportGroup Label="ExtensionSettings">
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <PropertyGroup Label="UserMacros" />
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <Optimization>MaxSpeed</Optimization>
+      <FunctionLevelLinking>true</FunctionLevelLinking>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <OptimizeReferences>true</OptimizeReferences>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <Optimization>MaxSpeed</Optimization>
+      <FunctionLevelLinking>true</FunctionLevelLinking>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <OptimizeReferences>true</OptimizeReferences>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <ClCompile Include="common.cpp" />
+    <ClCompile Include="dynamic_c.cpp" />
+    <ClCompile Include="dynamic_cilk.cpp" />
+    <ClCompile Include="main.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
+  </ItemGroup>
+  <ItemGroup>
+    <CustomBuild Include="kernels.ispc">
+      <FileType>Document</FileType>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
+</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
+</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+    </CustomBuild>
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+  </ImportGroup>
+</Project>

examples/deferred/dynamic_c.cpp

@@ -0,0 +1,870 @@
+/*
+  Copyright (c) 2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#include "deferred.h"
+#include "kernels_ispc.h"
+#include <algorithm>
+#include <stdint.h>
+#include <assert.h>
+#include <math.h>
+
+#ifdef _MSC_VER
+#define ISPC_IS_WINDOWS
+#elif defined(__linux__)
+#define ISPC_IS_LINUX
+#elif defined(__APPLE__)
+#define ISPC_IS_APPLE
+#endif
+
+#ifdef ISPC_IS_LINUX
+#include <malloc.h>
+#endif // ISPC_IS_LINUX
+
+// Currently tile widths must be a multiple of SIMD width (i.e. 8 for ispc sse4x2)!
+#define MIN_TILE_WIDTH 16
+#define MIN_TILE_HEIGHT 16
+
+
+#define DYNAMIC_TREE_LEVELS 5
+// If this is set to 1 then the result will be identical to the static version
+#define DYNAMIC_MIN_LIGHTS_TO_SUBDIVIDE 1
+
+static void *
+lAlignedMalloc(size_t size, int32_t alignment) {
+#ifdef ISPC_IS_WINDOWS
+    return _aligned_malloc(size, alignment);
+#endif
+#ifdef ISPC_IS_LINUX
+    return memalign(alignment, size);
+#endif
+#ifdef ISPC_IS_APPLE
+    void *mem = malloc(size + (alignment-1) + sizeof(void*));
+    char *amem = ((char*)mem) + sizeof(void*);
+    amem = amem + uint32_t(alignment - (reinterpret_cast<uint64_t>(amem) &
+                                        (alignment - 1)));
+    ((void**)amem)[-1] = mem;
+    return amem;
+#endif
+}
+
+
+static void
+lAlignedFree(void *ptr) {
+#ifdef ISPC_IS_WINDOWS
+    _aligned_free(ptr);
+#endif
+#ifdef ISPC_IS_LINUX
+    free(ptr);
+#endif
+#ifdef ISPC_IS_APPLE
+    free(((void**)ptr)[-1]);
+#endif
+}
+
+
+static void
+ComputeZBounds(int tileStartX, int tileEndX,
+               int tileStartY, int tileEndY,
+               // G-buffer data
+               float zBuffer[],
+               int gBufferWidth,
+               // Camera data
+               float cameraProj_33, float cameraProj_43,
+               float cameraNear, float cameraFar,
+               // Output
+               float *minZ, float *maxZ)
+{
+    // Find Z bounds
+    float laneMinZ = cameraFar;
+    float laneMaxZ = cameraNear;
+    for (int y = tileStartY; y < tileEndY; ++y) {
+        for (int x = tileStartX; x < tileEndX; ++x) {
+            // Unproject depth buffer Z value into view space
+            float z = zBuffer[(y * gBufferWidth + x)];
+            float viewSpaceZ = cameraProj_43 / (z - cameraProj_33);
+
+            // Work out Z bounds for our samples
+            // Avoid considering skybox/background or otherwise invalid pixels
+            if ((viewSpaceZ < cameraFar) && (viewSpaceZ >= cameraNear)) {
+                laneMinZ = std::min(laneMinZ, viewSpaceZ);
+                laneMaxZ = std::max(laneMaxZ, viewSpaceZ);
+            }
+        }
+    }
+    *minZ = laneMinZ;
+    *maxZ = laneMaxZ;
+}
+
+
+static void
+ComputeZBoundsRow(int tileY, int tileWidth, int tileHeight,
+                  int numTilesX, int numTilesY,
+                  // G-buffer data
+                  float zBuffer[],
+                  int gBufferWidth,
+                  // Camera data
+                  float cameraProj_33, float cameraProj_43,
+                  float cameraNear, float cameraFar,
+                  // Output
+                  float minZArray[],
+                  float maxZArray[])
+{
+    for (int tileX = 0; tileX < numTilesX; ++tileX) {
+        float minZ, maxZ;
+        ComputeZBounds(tileX * tileWidth, tileX * tileWidth + tileWidth,
+                       tileY * tileHeight, tileY * tileHeight + tileHeight,
+                       zBuffer, gBufferWidth, cameraProj_33, cameraProj_43, 
+                       cameraNear, cameraFar, &minZ, &maxZ);
+        minZArray[tileX] = minZ;
+        maxZArray[tileX] = maxZ;
+    }
+}
+
+
+class MinMaxZTree
+{
+public:
+    // Currently (min) tile dimensions must divide gBuffer dimensions evenly
+    // Levels must be small enough that neither dimension goes below one tile
+    MinMaxZTree(
+        int tileWidth, int tileHeight, int levels,
+        int gBufferWidth, int gBufferHeight)
+        : mTileWidth(tileWidth), mTileHeight(tileHeight), mLevels(levels)
+    {
+        mNumTilesX = gBufferWidth / mTileWidth;
+        mNumTilesY = gBufferHeight / mTileHeight;
+        
+        // Allocate arrays
+        mMinZArrays = (float **)lAlignedMalloc(sizeof(float *) * mLevels, 16);
+        mMaxZArrays = (float **)lAlignedMalloc(sizeof(float *) * mLevels, 16);
+        for (int i = 0; i < mLevels; ++i) {
+            int x = NumTilesX(i);
+            int y = NumTilesY(i);
+            assert(x > 0);
+            assert(y > 0);
+            // NOTE: If the following two asserts fire it probably means that
+            // the base tile dimensions do not evenly divide the G-buffer dimensions
+            assert(x * (mTileWidth << i) >= gBufferWidth);
+            assert(y * (mTileHeight << i) >= gBufferHeight);
+            mMinZArrays[i] = (float *)lAlignedMalloc(sizeof(float) * x * y, 16);
+            mMaxZArrays[i] = (float *)lAlignedMalloc(sizeof(float) * x * y, 16);
+        }
+    }
+
+    void Update(float *zBuffer, int gBufferPitchInElements,
+        float cameraProj_33, float cameraProj_43,
+        float cameraNear, float cameraFar)
+    {
+        for (int tileY = 0; tileY < mNumTilesY; ++tileY) {
+            ComputeZBoundsRow(tileY, mTileWidth, mTileHeight, mNumTilesX, mNumTilesY,
+                              zBuffer, gBufferPitchInElements,
+                              cameraProj_33, cameraProj_43, cameraNear, cameraFar,
+                              mMinZArrays[0] + (tileY * mNumTilesX),
+                              mMaxZArrays[0] + (tileY * mNumTilesX));
+        }
+
+        // Generate other levels
+        for (int level = 1; level < mLevels; ++level) {
+            int destTilesX = NumTilesX(level);
+            int destTilesY = NumTilesY(level);
+            int srcLevel = level - 1;
+            int srcTilesX = NumTilesX(srcLevel);
+            int srcTilesY = NumTilesY(srcLevel);
+            for (int y = 0; y < destTilesY; ++y) {
+                for (int x = 0; x < destTilesX; ++x) {
+                    int srcX = x << 1;
+                    int srcY = y << 1;
+                    // NOTE: Ugly branches to deal with non-multiple dimensions at some levels
+                    // TODO: SSE branchless min/max is probably better...
+                    float minZ = mMinZArrays[srcLevel][(srcY) * srcTilesX + (srcX)];
+                    float maxZ = mMaxZArrays[srcLevel][(srcY) * srcTilesX + (srcX)];
+                    if (srcX + 1 < srcTilesX) {
+                        minZ = std::min(minZ, mMinZArrays[srcLevel][(srcY) * srcTilesX + 
+                                                                    (srcX + 1)]);
+                        maxZ = std::max(maxZ, mMaxZArrays[srcLevel][(srcY) * srcTilesX +
+                                                                    (srcX + 1)]);
+                        if (srcY + 1 < srcTilesY) {
+                            minZ = std::min(minZ, mMinZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                        (srcX + 1)]);
+                            maxZ = std::max(maxZ, mMaxZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                        (srcX + 1)]);
+                        }
+                    }
+                    if (srcY + 1 < srcTilesY) {
+                        minZ = std::min(minZ, mMinZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                    (srcX    )]);
+                        maxZ = std::max(maxZ, mMaxZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                    (srcX    )]);
+                    }
+                    mMinZArrays[level][y * destTilesX + x] = minZ;
+                    mMaxZArrays[level][y * destTilesX + x] = maxZ;
+                }
+            }
+        }
+    }
+
+    ~MinMaxZTree() {
+        for (int i = 0; i < mLevels; ++i) {
+            lAlignedFree(mMinZArrays[i]);
+            lAlignedFree(mMaxZArrays[i]);
+        }
+        lAlignedFree(mMinZArrays);
+        lAlignedFree(mMaxZArrays); 
+    }
+
+    int Levels() const { return mLevels; }
+
+    // These round UP, so beware that the last tile for a given level may not be completely full
+    // TODO: Verify this...
+    int NumTilesX(int level = 0) const { return (mNumTilesX + (1 << level) - 1) >> level; }
+    int NumTilesY(int level = 0) const { return (mNumTilesY + (1 << level) - 1) >> level; }
+    int TileWidth(int level = 0) const { return (mTileWidth << level); }
+    int TileHeight(int level = 0) const { return (mTileHeight << level); }
+
+    float MinZ(int level, int tileX, int tileY) const {
+        return mMinZArrays[level][tileY * NumTilesX(level) + tileX];
+    }
+    float MaxZ(int level, int tileX, int tileY) const {
+        return mMaxZArrays[level][tileY * NumTilesX(level) + tileX];
+    }
+
+private:
+    int mTileWidth;
+    int mTileHeight;
+    int mLevels;
+    int mNumTilesX;
+    int mNumTilesY;
+
+    // One array for each "level" in the tree
+    float **mMinZArrays;
+    float **mMaxZArrays;
+};
+
+static MinMaxZTree *gMinMaxZTree = 0;
+
+void InitDynamicC(InputData *input) {
+    gMinMaxZTree = 
+        new MinMaxZTree(MIN_TILE_WIDTH, MIN_TILE_HEIGHT, DYNAMIC_TREE_LEVELS,
+                        input->header.framebufferWidth, 
+                        input->header.framebufferHeight);
+}
+
+
+/* We're going to split a tile into 4 sub-tiles.  This function
+   reclassifies the tile's lights with respect to the sub-tiles. */
+static void
+SplitTileMinMax(
+    int tileMidX, int tileMidY,
+    // Subtile data (00, 10, 01, 11)
+    float subtileMinZ[],
+    float subtileMaxZ[],
+    // G-buffer data
+    int gBufferWidth, int gBufferHeight,
+    // Camera data
+    float cameraProj_11, float cameraProj_22,
+    // Light Data
+    int lightIndices[],
+    int numLights,
+    float light_positionView_x_array[],
+    float light_positionView_y_array[],
+    float light_positionView_z_array[],
+    float light_attenuationEnd_array[],
+    // Outputs
+    int subtileIndices[],
+    int subtileIndicesPitch,
+    int subtileNumLights[]
+    )
+{
+    float gBufferScale_x = 0.5f * (float)gBufferWidth;
+    float gBufferScale_y = 0.5f * (float)gBufferHeight;
+        
+    float frustumPlanes_xy[2] = { -(cameraProj_11 * gBufferScale_x),
+                                   (cameraProj_22 * gBufferScale_y) };
+    float frustumPlanes_z[2] = { tileMidX - gBufferScale_x,
+                                 tileMidY - gBufferScale_y };
+
+    for (int i = 0; i < 2; ++i) {
+        // Normalize
+        float norm = 1.f / sqrtf(frustumPlanes_xy[i] * frustumPlanes_xy[i] + 
+                                 frustumPlanes_z[i] * frustumPlanes_z[i]);
+        frustumPlanes_xy[i] *= norm;
+        frustumPlanes_z[i] *= norm;
+    }
+
+    // Initialize
+    int subtileLightOffset[4];
+    subtileLightOffset[0] = 0 * subtileIndicesPitch;
+    subtileLightOffset[1] = 1 * subtileIndicesPitch;
+    subtileLightOffset[2] = 2 * subtileIndicesPitch;
+    subtileLightOffset[3] = 3 * subtileIndicesPitch;
+
+    for (int i = 0; i < numLights; ++i) {
+        int lightIndex = lightIndices[i];
+
+        float light_positionView_x = light_positionView_x_array[lightIndex];
+        float light_positionView_y = light_positionView_y_array[lightIndex];
+        float light_positionView_z = light_positionView_z_array[lightIndex];
+        float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
+        float light_attenuationEndNeg = -light_attenuationEnd;
+        
+        // Test lights again against subtile z bounds
+        bool inFrustum[4];
+        inFrustum[0] = (light_positionView_z - subtileMinZ[0] >= light_attenuationEndNeg) &&
+            (subtileMaxZ[0] - light_positionView_z >= light_attenuationEndNeg);
+        inFrustum[1] = (light_positionView_z - subtileMinZ[1] >= light_attenuationEndNeg) && 
+            (subtileMaxZ[1] - light_positionView_z >= light_attenuationEndNeg);
+        inFrustum[2] = (light_positionView_z - subtileMinZ[2] >= light_attenuationEndNeg) && 
+            (subtileMaxZ[2] - light_positionView_z >= light_attenuationEndNeg);
+        inFrustum[3] = (light_positionView_z - subtileMinZ[3] >= light_attenuationEndNeg) && 
+            (subtileMaxZ[3] - light_positionView_z >= light_attenuationEndNeg);
+
+        float dx = light_positionView_z * frustumPlanes_z[0] + 
+            light_positionView_x * frustumPlanes_xy[0];
+        float dy = light_positionView_z * frustumPlanes_z[1] +
+            light_positionView_y * frustumPlanes_xy[1];
+        
+        if (fabsf(dx) > light_attenuationEnd) {
+            bool positiveX = dx > 0.0f;
+            inFrustum[0] = inFrustum[0] &&  positiveX;    // 00 subtile
+            inFrustum[1] = inFrustum[1] && !positiveX;    // 10 subtile
+            inFrustum[2] = inFrustum[2] &&  positiveX;    // 01 subtile
+            inFrustum[3] = inFrustum[3] && !positiveX;    // 11 subtile
+        }
+        if (fabsf(dy) > light_attenuationEnd) {
+            bool positiveY = dy > 0.0f;
+            inFrustum[0] = inFrustum[0] &&  positiveY;    // 00 subtile
+            inFrustum[1] = inFrustum[1] &&  positiveY;    // 10 subtile
+            inFrustum[2] = inFrustum[2] && !positiveY;    // 01 subtile
+            inFrustum[3] = inFrustum[3] && !positiveY;    // 11 subtile
+        }
+
+        if (inFrustum[0])
+            subtileIndices[subtileLightOffset[0]++] = lightIndex;
+        if (inFrustum[1])
+            subtileIndices[subtileLightOffset[1]++] = lightIndex;
+        if (inFrustum[2])
+            subtileIndices[subtileLightOffset[2]++] = lightIndex;
+        if (inFrustum[3])
+            subtileIndices[subtileLightOffset[3]++] = lightIndex;
+    }
+
+    subtileNumLights[0] = subtileLightOffset[0] - 0 * subtileIndicesPitch;
+    subtileNumLights[1] = subtileLightOffset[1] - 1 * subtileIndicesPitch;
+    subtileNumLights[2] = subtileLightOffset[2] - 2 * subtileIndicesPitch;
+    subtileNumLights[3] = subtileLightOffset[3] - 3 * subtileIndicesPitch;
+}
+
+
+static inline float
+dot3(float x, float y, float z, float a, float b, float c) {
+    return (x*a + y*b + z*c);
+}
+
+
+static inline void
+normalize3(float x, float y, float z, float &ox, float &oy, float &oz) {
+    float n = 1.f / sqrtf(x*x + y*y + z*z);
+    ox = x * n;
+    oy = y * n;
+    oz = z * n;
+}
+
+
+static inline float
+Unorm8ToFloat32(uint8_t u) {
+    return (float)u * (1.0f / 255.0f);
+}
+
+
+static inline uint8_t
+Float32ToUnorm8(float f) {
+    return (uint8_t)(f * 255.0f);
+}
+
+
+static inline float
+half_to_float_fast(uint16_t h) {
+    uint32_t hs = h & (int32_t)0x8000u;  // Pick off sign bit
+    uint32_t he = h & (int32_t)0x7C00u;  // Pick off exponent bits
+    uint32_t hm = h & (int32_t)0x03FFu;  // Pick off mantissa bits
+
+    // sign
+    uint32_t xs = ((uint32_t) hs) << 16; 
+    // Exponent: unbias the halfp, then bias the single
+    int32_t xes = ((int32_t) (he >> 10)) - 15 + 127; 
+    // Exponent
+    uint32_t xe = (uint32_t) (xes << 23);
+    // Mantissa
+    uint32_t xm = ((uint32_t) hm) << 13; 
+
+    uint32_t bits = (xs | xe | xm);
+    float *fp = reinterpret_cast<float *>(&bits);
+    return *fp;
+}
+
+
+static void
+ShadeTileC(
+    int32_t tileStartX, int32_t tileEndX,
+    int32_t tileStartY, int32_t tileEndY,
+    int32_t gBufferWidth, int32_t gBufferHeight,
+    const ispc::InputDataArrays &inputData,
+    // Camera data
+    float cameraProj_11, float cameraProj_22,
+    float cameraProj_33, float cameraProj_43,
+    // Light list
+    int32_t tileLightIndices[],
+    int32_t tileNumLights,
+    // UI
+    bool visualizeLightCount,
+    // Output
+    uint8_t framebuffer_r[],
+    uint8_t framebuffer_g[],
+    uint8_t framebuffer_b[]
+    )
+{
+    if (tileNumLights == 0 || visualizeLightCount) {
+        uint8_t c = (uint8_t)(std::min(tileNumLights << 2, 255));
+        for (int32_t y = tileStartY; y < tileEndY; ++y) {
+            for (int32_t x = tileStartX; x < tileEndX; ++x) {
+                int32_t framebufferIndex = (y * gBufferWidth + x);
+                framebuffer_r[framebufferIndex] = c;
+                framebuffer_g[framebufferIndex] = c;
+                framebuffer_b[framebufferIndex] = c;
+            }
+        }
+    } else {
+        float twoOverGBufferWidth = 2.0f / gBufferWidth;
+        float twoOverGBufferHeight = 2.0f / gBufferHeight;
+        
+        for (int32_t y = tileStartY; y < tileEndY; ++y) {
+            float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
+
+            for (int32_t x = tileStartX; x < tileEndX; ++x) {
+                int32_t gBufferOffset = y * gBufferWidth + x;
+                
+                // Reconstruct position and (negative) view vector from G-buffer
+                float surface_positionView_x, surface_positionView_y, surface_positionView_z;
+                float Vneg_x, Vneg_y, Vneg_z;
+
+                float z = inputData.zBuffer[gBufferOffset];
+
+                // Compute screen/clip-space position
+                // NOTE: Mind DX11 viewport transform and pixel center!
+                float positionScreen_x = (0.5f + (float)(x)) * 
+                    twoOverGBufferWidth - 1.0f;
+
+                // Unproject depth buffer Z value into view space
+                surface_positionView_z = cameraProj_43 / (z - cameraProj_33);
+                surface_positionView_x = positionScreen_x * surface_positionView_z / 
+                    cameraProj_11;
+                surface_positionView_y = positionScreen_y * surface_positionView_z / 
+                    cameraProj_22;
+                
+                // We actually end up with a vector pointing *at* the
+                // surface (i.e. the negative view vector)
+                normalize3(surface_positionView_x, surface_positionView_y, 
+                           surface_positionView_z, Vneg_x, Vneg_y, Vneg_z);
+
+                // Reconstruct normal from G-buffer
+                float surface_normal_x, surface_normal_y, surface_normal_z;
+                float normal_x = half_to_float_fast(inputData.normalEncoded_x[gBufferOffset]);
+                float normal_y = half_to_float_fast(inputData.normalEncoded_y[gBufferOffset]);
+                    
+                float f = (normal_x - normal_x * normal_x) + (normal_y - normal_y * normal_y);
+                float m = sqrtf(4.0f * f - 1.0f);
+                    
+                surface_normal_x = m * (4.0f * normal_x - 2.0f);
+                surface_normal_y = m * (4.0f * normal_y - 2.0f);
+                surface_normal_z = 3.0f - 8.0f * f;
+
+                // Load other G-buffer parameters
+                float surface_specularAmount = 
+                    half_to_float_fast(inputData.specularAmount[gBufferOffset]);
+                float surface_specularPower  = 
+                    half_to_float_fast(inputData.specularPower[gBufferOffset]);
+                float surface_albedo_x = Unorm8ToFloat32(inputData.albedo_x[gBufferOffset]);
+                float surface_albedo_y = Unorm8ToFloat32(inputData.albedo_y[gBufferOffset]);
+                float surface_albedo_z = Unorm8ToFloat32(inputData.albedo_z[gBufferOffset]);
+                
+                float lit_x = 0.0f;
+                float lit_y = 0.0f;
+                float lit_z = 0.0f;
+                for (int32_t tileLightIndex = 0; tileLightIndex < tileNumLights; 
+                     ++tileLightIndex) {
+                    int32_t lightIndex = tileLightIndices[tileLightIndex];
+                                        
+                    // Gather light data relevant to initial culling
+                    float light_positionView_x = 
+                        inputData.lightPositionView_x[lightIndex];
+                    float light_positionView_y = 
+                        inputData.lightPositionView_y[lightIndex];
+                    float light_positionView_z = 
+                        inputData.lightPositionView_z[lightIndex];
+                    float light_attenuationEnd = 
+                        inputData.lightAttenuationEnd[lightIndex];
+                    
+                    // Compute light vector
+                    float L_x = light_positionView_x - surface_positionView_x;
+                    float L_y = light_positionView_y - surface_positionView_y;
+                    float L_z = light_positionView_z - surface_positionView_z;
+
+                    float distanceToLight2 = dot3(L_x, L_y, L_z, L_x, L_y, L_z);
+                    
+                    // Clip at end of attenuation
+                    float light_attenutaionEnd2 = light_attenuationEnd * light_attenuationEnd;
+
+                    if (distanceToLight2 < light_attenutaionEnd2) {                    
+                        float distanceToLight = sqrtf(distanceToLight2);
+
+                        float distanceToLightRcp = 1.f / distanceToLight;
+                        L_x *= distanceToLightRcp;
+                        L_y *= distanceToLightRcp;
+                        L_z *= distanceToLightRcp;
+
+                        // Start computing brdf
+                        float NdotL = dot3(surface_normal_x, surface_normal_y, 
+                                           surface_normal_z, L_x, L_y, L_z);
+                    
+                        // Clip back facing
+                        if (NdotL > 0.0f) {
+                            float light_attenuationBegin = 
+                                inputData.lightAttenuationBegin[lightIndex];
+
+                            // Light distance attenuation (linstep)
+                            float lightRange = (light_attenuationEnd - light_attenuationBegin);
+                            float falloffPosition = (light_attenuationEnd - distanceToLight);
+                            float attenuation = std::min(falloffPosition / lightRange, 1.0f);
+
+                            float H_x = (L_x - Vneg_x);
+                            float H_y = (L_y - Vneg_y);
+                            float H_z = (L_z - Vneg_z);
+                            normalize3(H_x, H_y, H_z, H_x, H_y, H_z);
+                    
+                            float NdotH = dot3(surface_normal_x, surface_normal_y, 
+                                               surface_normal_z, H_x, H_y, H_z);
+                            NdotH = std::max(NdotH, 0.0f);
+
+                            float specular = powf(NdotH, surface_specularPower);
+                            float specularNorm = (surface_specularPower + 2.0f) * 
+                                (1.0f / 8.0f);
+                            float specularContrib = surface_specularAmount * 
+                                specularNorm * specular;
+
+                            float k = attenuation * NdotL * (1.0f + specularContrib);
+                    
+                            float light_color_x = inputData.lightColor_x[lightIndex];
+                            float light_color_y = inputData.lightColor_y[lightIndex];
+                            float light_color_z = inputData.lightColor_z[lightIndex];
+
+                            float lightContrib_x = surface_albedo_x * light_color_x;
+                            float lightContrib_y = surface_albedo_y * light_color_y;
+                            float lightContrib_z = surface_albedo_z * light_color_z;
+
+                            lit_x += lightContrib_x * k;
+                            lit_y += lightContrib_y * k;
+                            lit_z += lightContrib_z * k;
+                        }
+                    }
+                }
+
+                // Gamma correct
+                float gamma = 1.0 / 2.2f;
+                lit_x = powf(std::min(std::max(lit_x, 0.0f), 1.0f), gamma);
+                lit_y = powf(std::min(std::max(lit_y, 0.0f), 1.0f), gamma);
+                lit_z = powf(std::min(std::max(lit_z, 0.0f), 1.0f), gamma);
+                
+                framebuffer_r[gBufferOffset] = Float32ToUnorm8(lit_x);
+                framebuffer_g[gBufferOffset] = Float32ToUnorm8(lit_y);
+                framebuffer_b[gBufferOffset] = Float32ToUnorm8(lit_z);
+            }
+        }
+    }
+}
+
+
+void
+ShadeDynamicTileRecurse(InputData *input, int level, int tileX, int tileY, 
+                        int *lightIndices, int numLights, 
+                        Framebuffer *framebuffer) {
+    const MinMaxZTree *minMaxZTree = gMinMaxZTree;
+    
+    // If we few enough lights or this is the base case (last level), shade
+    // this full tile directly
+    if (level == 0 || numLights < DYNAMIC_MIN_LIGHTS_TO_SUBDIVIDE) {
+        int width = minMaxZTree->TileWidth(level);
+        int height = minMaxZTree->TileHeight(level);
+        int startX = tileX * width;
+        int startY = tileY * height;
+        int endX = std::min(input->header.framebufferWidth, startX + width);
+        int endY = std::min(input->header.framebufferHeight, startY + height);
+        
+        // Skip entirely offscreen tiles
+        if (endX > startX && endY > startY) {
+            ShadeTileC(startX, endX, startY, endY,
+                       input->header.framebufferWidth, input->header.framebufferHeight,
+                       input->arrays,
+                       input->header.cameraProj[0][0], input->header.cameraProj[1][1], 
+                       input->header.cameraProj[2][2], input->header.cameraProj[3][2],
+                       lightIndices, numLights, VISUALIZE_LIGHT_COUNT, 
+                       framebuffer->r, framebuffer->g, framebuffer->b);
+        }
+    } 
+    else {
+        // Otherwise, subdivide and 4-way recurse using X and Y splitting planes
+        // Move down a level in the tree
+        --level;
+        tileX <<= 1;
+        tileY <<= 1;
+        int width = minMaxZTree->TileWidth(level);
+        int height = minMaxZTree->TileHeight(level);
+
+        // Work out splitting coords
+        int midX = (tileX + 1) * width;
+        int midY = (tileY + 1) * height;
+
+        // Read subtile min/max data
+        // NOTE: We must be sure to handle out-of-bounds access here since
+        // sometimes we'll only have 1 or 2 subtiles for non-pow-2
+        // framebuffer sizes.
+        bool rightTileExists = (tileX + 1 < minMaxZTree->NumTilesX(level));
+        bool bottomTileExists = (tileY + 1 < minMaxZTree->NumTilesY(level));
+
+        // NOTE: Order is 00, 10, 01, 11
+        // Set defaults up to cull all lights if the tile doesn't exist (offscreen)
+        float minZ[4] = {input->header.cameraFar, input->header.cameraFar, 
+                         input->header.cameraFar, input->header.cameraFar};
+        float maxZ[4] = {input->header.cameraNear, input->header.cameraNear, 
+                         input->header.cameraNear, input->header.cameraNear};
+
+        minZ[0] = minMaxZTree->MinZ(level, tileX, tileY);
+        maxZ[0] = minMaxZTree->MaxZ(level, tileX, tileY);
+        if (rightTileExists) {
+            minZ[1] = minMaxZTree->MinZ(level, tileX + 1, tileY);
+            maxZ[1] = minMaxZTree->MaxZ(level, tileX + 1, tileY);
+            if (bottomTileExists) {
+                minZ[3] = minMaxZTree->MinZ(level, tileX + 1, tileY + 1);
+                maxZ[3] = minMaxZTree->MaxZ(level, tileX + 1, tileY + 1);
+            }
+        }
+        if (bottomTileExists) {
+            minZ[2] = minMaxZTree->MinZ(level, tileX, tileY + 1);
+            maxZ[2] = minMaxZTree->MaxZ(level, tileX, tileY + 1);
+        }
+
+        // Cull lights into subtile lists
+#ifdef ISPC_IS_WINDOWS
+        __declspec(align(ALIGNMENT_BYTES)) 
+#endif
+            int subtileLightIndices[4][MAX_LIGHTS]
+#ifndef ISPC_IS_WINDOWS
+            __attribute__ ((aligned(ALIGNMENT_BYTES)))
+#endif
+;
+        int subtileNumLights[4];
+        SplitTileMinMax(midX, midY, minZ, maxZ,
+            input->header.framebufferWidth, input->header.framebufferHeight, 
+            input->header.cameraProj[0][0], input->header.cameraProj[1][1],
+            lightIndices, numLights, input->arrays.lightPositionView_x, 
+            input->arrays.lightPositionView_y, input->arrays.lightPositionView_z, 
+            input->arrays.lightAttenuationEnd,
+            subtileLightIndices[0], MAX_LIGHTS, subtileNumLights);
+        
+        // Recurse into subtiles
+        ShadeDynamicTileRecurse(input, level, tileX    , tileY, 
+                                subtileLightIndices[0], subtileNumLights[0],
+                                framebuffer);
+        ShadeDynamicTileRecurse(input, level, tileX + 1, tileY,
+                                subtileLightIndices[1], subtileNumLights[1],
+                                framebuffer);
+        ShadeDynamicTileRecurse(input, level, tileX    , tileY + 1,
+                                subtileLightIndices[2], subtileNumLights[2],
+                                framebuffer);
+        ShadeDynamicTileRecurse(input, level, tileX + 1, tileY + 1,
+                                subtileLightIndices[3], subtileNumLights[3],
+                                framebuffer);
+    }
+}
+
+
+static int
+IntersectLightsWithTileMinMax(
+    int tileStartX, int tileEndX,
+    int tileStartY, int tileEndY,
+    // Tile data
+    float minZ,
+    float maxZ,
+    // G-buffer data
+    int gBufferWidth, int gBufferHeight,
+    // Camera data
+    float cameraProj_11, float cameraProj_22,
+    // Light Data
+    int numLights,
+    float light_positionView_x_array[],
+    float light_positionView_y_array[],
+    float light_positionView_z_array[],
+    float light_attenuationEnd_array[],
+    // Output
+    int tileLightIndices[]
+    )
+{
+    float gBufferScale_x = 0.5f * (float)gBufferWidth;
+    float gBufferScale_y = 0.5f * (float)gBufferHeight;
+        
+    float frustumPlanes_xy[4];
+    float frustumPlanes_z[4];
+
+    // This one is totally constant over the whole screen... worth pulling it up at all?
+    float frustumPlanes_xy_v[4] = { -(cameraProj_11 * gBufferScale_x),
+                                    (cameraProj_11 * gBufferScale_x),
+                                    (cameraProj_22 * gBufferScale_y),
+                                    -(cameraProj_22 * gBufferScale_y) };
+    
+    float frustumPlanes_z_v[4] = {  tileEndX - gBufferScale_x,
+                                    -tileStartX + gBufferScale_x,
+                                    tileEndY - gBufferScale_y,
+                                    -tileStartY + gBufferScale_y };
+
+    for (int i = 0; i < 4; ++i) {
+        float norm = 1.f / sqrtf(frustumPlanes_xy_v[i] * frustumPlanes_xy_v[i] + 
+                                 frustumPlanes_z_v[i] * frustumPlanes_z_v[i]);
+        frustumPlanes_xy_v[i] *= norm;
+        frustumPlanes_z_v[i] *= norm;
+
+        frustumPlanes_xy[i] = frustumPlanes_xy_v[i];
+        frustumPlanes_z[i] = frustumPlanes_z_v[i];
+    }
+
+    int tileNumLights = 0;
+
+    for (int lightIndex = 0; lightIndex < numLights; ++lightIndex) {
+        float light_positionView_z = light_positionView_z_array[lightIndex];
+        float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
+        float light_attenuationEndNeg = -light_attenuationEnd;
+
+        float d = light_positionView_z - minZ;
+        bool inFrustum = (d >= light_attenuationEndNeg);
+
+        d = maxZ - light_positionView_z;
+        inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+        
+        if (!inFrustum) 
+            continue;
+
+        float light_positionView_x = light_positionView_x_array[lightIndex];
+        float light_positionView_y = light_positionView_y_array[lightIndex];
+
+        d = light_positionView_z * frustumPlanes_z[0] + 
+            light_positionView_x * frustumPlanes_xy[0];
+        inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+
+        d = light_positionView_z * frustumPlanes_z[1] + 
+            light_positionView_x * frustumPlanes_xy[1];
+        inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+
+        d = light_positionView_z * frustumPlanes_z[2] + 
+            light_positionView_y * frustumPlanes_xy[2];
+        inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+
+        d = light_positionView_z * frustumPlanes_z[3] + 
+            light_positionView_y * frustumPlanes_xy[3];
+        inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+        
+        // Pack and store intersecting lights
+        if (inFrustum)
+            tileLightIndices[tileNumLights++] = lightIndex;
+    }
+
+    return tileNumLights;
+}
+
+
+void
+ShadeDynamicTile(InputData *input, int level, int tileX, int tileY,
+                 Framebuffer *framebuffer) {
+    const MinMaxZTree *minMaxZTree = gMinMaxZTree;
+
+    // Get Z min/max for this tile
+    int width = minMaxZTree->TileWidth(level);
+    int height = minMaxZTree->TileHeight(level);
+    float minZ = minMaxZTree->MinZ(level, tileX, tileY);
+    float maxZ = minMaxZTree->MaxZ(level, tileX, tileY);
+
+    int startX = tileX * width;
+    int startY = tileY * height;
+    int endX = std::min(input->header.framebufferWidth, startX + width);
+    int endY = std::min(input->header.framebufferHeight, startY + height);
+
+    // This is a root tile, so first do a full 6-plane cull
+#ifdef ISPC_IS_WINDOWS
+    __declspec(align(ALIGNMENT_BYTES)) 
+#endif
+        int lightIndices[MAX_LIGHTS]
+#ifndef ISPC_IS_WINDOWS
+        __attribute__ ((aligned(ALIGNMENT_BYTES)))
+#endif
+;
+    int numLights = IntersectLightsWithTileMinMax(
+        startX, endX, startY, endY,    minZ, maxZ,
+        input->header.framebufferWidth, input->header.framebufferHeight,
+        input->header.cameraProj[0][0], input->header.cameraProj[1][1],
+        MAX_LIGHTS, input->arrays.lightPositionView_x, 
+        input->arrays.lightPositionView_y, input->arrays.lightPositionView_z, 
+        input->arrays.lightAttenuationEnd, lightIndices);
+
+    // Now kick off the recursive process for this tile
+    ShadeDynamicTileRecurse(input, level, tileX, tileY, lightIndices, 
+                            numLights, framebuffer);
+}
+
+
+void
+DispatchDynamicC(InputData *input, Framebuffer *framebuffer)
+{
+    MinMaxZTree *minMaxZTree = gMinMaxZTree;
+        
+    // Update min/max Z tree
+    minMaxZTree->Update(input->arrays.zBuffer, input->header.framebufferWidth,
+        input->header.cameraProj[2][2], input->header.cameraProj[3][2], 
+        input->header.cameraNear, input->header.cameraFar);
+
+    int rootLevel = minMaxZTree->Levels() - 1;
+    int rootTilesX = minMaxZTree->NumTilesX(rootLevel);
+    int rootTilesY = minMaxZTree->NumTilesY(rootLevel);
+    int rootTiles = rootTilesX * rootTilesY;
+    for (int g = 0; g < rootTiles; ++g) {
+        uint32_t tileY = g / rootTilesX;
+        uint32_t tileX = g % rootTilesX;
+        ShadeDynamicTile(input, rootLevel, tileX, tileY, framebuffer);
+    }
+}

examples/deferred/dynamic_cilk.cpp

@@ -0,0 +1,398 @@
+/*
+  Copyright (c) 2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#ifdef __cilk
+
+#include "deferred.h"
+#include "kernels_ispc.h"
+#include <algorithm>
+#include <assert.h>
+
+#ifdef _MSC_VER
+#define ISPC_IS_WINDOWS
+#elif defined(__linux__)
+#define ISPC_IS_LINUX
+#elif defined(__APPLE__)
+#define ISPC_IS_APPLE
+#endif
+
+#ifdef ISPC_IS_LINUX
+#include <malloc.h>
+#endif // ISPC_IS_LINUX
+
+// Currently tile widths must be a multiple of SIMD width (i.e. 8 for ispc sse4x2)!
+#define MIN_TILE_WIDTH 16
+#define MIN_TILE_HEIGHT 16
+
+
+#define DYNAMIC_TREE_LEVELS 5
+// If this is set to 1 then the result will be identical to the static version
+#define DYNAMIC_MIN_LIGHTS_TO_SUBDIVIDE 1
+
+static void *
+lAlignedMalloc(size_t size, int32_t alignment) {
+#ifdef ISPC_IS_WINDOWS
+    return _aligned_malloc(size, alignment);
+#endif
+#ifdef ISPC_IS_LINUX
+    return memalign(alignment, size);
+#endif
+#ifdef ISPC_IS_APPLE
+    void *mem = malloc(size + (alignment-1) + sizeof(void*));
+    char *amem = ((char*)mem) + sizeof(void*);
+    amem = amem + uint32_t(alignment - (reinterpret_cast<uint64_t>(amem) &
+                                        (alignment - 1)));
+    ((void**)amem)[-1] = mem;
+    return amem;
+#endif
+}
+
+
+static void
+lAlignedFree(void *ptr) {
+#ifdef ISPC_IS_WINDOWS
+    _aligned_free(ptr);
+#endif
+#ifdef ISPC_IS_LINUX
+    free(ptr);
+#endif
+#ifdef ISPC_IS_APPLE
+    free(((void**)ptr)[-1]);
+#endif
+}
+
+
+class MinMaxZTreeCilk
+{
+public:
+    // Currently (min) tile dimensions must divide gBuffer dimensions evenly
+    // Levels must be small enough that neither dimension goes below one tile
+    MinMaxZTreeCilk(
+        int tileWidth, int tileHeight, int levels,
+        int gBufferWidth, int gBufferHeight)
+        : mTileWidth(tileWidth), mTileHeight(tileHeight), mLevels(levels)
+    {
+        mNumTilesX = gBufferWidth / mTileWidth;
+        mNumTilesY = gBufferHeight / mTileHeight;
+        
+        // Allocate arrays
+        mMinZArrays = (float **)lAlignedMalloc(sizeof(float *) * mLevels, 16);
+        mMaxZArrays = (float **)lAlignedMalloc(sizeof(float *) * mLevels, 16);
+        for (int i = 0; i < mLevels; ++i) {
+            int x = NumTilesX(i);
+            int y = NumTilesY(i);
+            assert(x > 0);
+            assert(y > 0);
+            // NOTE: If the following two asserts fire it probably means that
+            // the base tile dimensions do not evenly divide the G-buffer dimensions
+            assert(x * (mTileWidth << i) >= gBufferWidth);
+            assert(y * (mTileHeight << i) >= gBufferHeight);
+            mMinZArrays[i] = (float *)lAlignedMalloc(sizeof(float) * x * y, 16);
+            mMaxZArrays[i] = (float *)lAlignedMalloc(sizeof(float) * x * y, 16);
+        }
+    }
+
+    void Update(float *zBuffer, int gBufferPitchInElements,
+        float cameraProj_33, float cameraProj_43,
+        float cameraNear, float cameraFar)
+    {
+        // Compute level 0 in parallel. Outer loops is here since we use Cilk
+        _Cilk_for (int tileY = 0; tileY < mNumTilesY; ++tileY) {
+            ispc::ComputeZBoundsRow(tileY,
+                mTileWidth, mTileHeight, mNumTilesX, mNumTilesY,
+                zBuffer, gBufferPitchInElements,
+                cameraProj_33, cameraProj_43, cameraNear, cameraFar,
+                mMinZArrays[0] + (tileY * mNumTilesX),
+                mMaxZArrays[0] + (tileY * mNumTilesX));
+        }
+
+        // Generate other levels
+        // NOTE: We currently don't use ispc here since it's sort of an
+        // awkward gather-based reduction Using SSE odd pack/unpack
+        // instructions might actually work here when we need to optimize
+        for (int level = 1; level < mLevels; ++level) {
+            int destTilesX = NumTilesX(level);
+            int destTilesY = NumTilesY(level);
+            int srcLevel = level - 1;
+            int srcTilesX = NumTilesX(srcLevel);
+            int srcTilesY = NumTilesY(srcLevel);
+            _Cilk_for (int y = 0; y < destTilesY; ++y) {
+                for (int x = 0; x < destTilesX; ++x) {
+                    int srcX = x << 1;
+                    int srcY = y << 1;
+                    // NOTE: Ugly branches to deal with non-multiple dimensions at some levels
+                    // TODO: SSE branchless min/max is probably better...
+                    float minZ = mMinZArrays[srcLevel][(srcY) * srcTilesX + (srcX)];
+                    float maxZ = mMaxZArrays[srcLevel][(srcY) * srcTilesX + (srcX)];
+                    if (srcX + 1 < srcTilesX) {
+                        minZ = std::min(minZ, mMinZArrays[srcLevel][(srcY) * srcTilesX + 
+                                                                    (srcX + 1)]);
+                        maxZ = std::max(maxZ, mMaxZArrays[srcLevel][(srcY) * srcTilesX +
+                                                                    (srcX + 1)]);
+                        if (srcY + 1 < srcTilesY) {
+                            minZ = std::min(minZ, mMinZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                        (srcX + 1)]);
+                            maxZ = std::max(maxZ, mMaxZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                        (srcX + 1)]);
+                        }
+                    }
+                    if (srcY + 1 < srcTilesY) {
+                        minZ = std::min(minZ, mMinZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                    (srcX    )]);
+                        maxZ = std::max(maxZ, mMaxZArrays[srcLevel][(srcY + 1) * srcTilesX +
+                                                                    (srcX    )]);
+                    }
+                    mMinZArrays[level][y * destTilesX + x] = minZ;
+                    mMaxZArrays[level][y * destTilesX + x] = maxZ;
+                }
+            }
+        }
+    }
+
+    ~MinMaxZTreeCilk() {
+        for (int i = 0; i < mLevels; ++i) {
+            lAlignedFree(mMinZArrays[i]);
+            lAlignedFree(mMaxZArrays[i]);
+        }
+        lAlignedFree(mMinZArrays);
+        lAlignedFree(mMaxZArrays); 
+    }
+
+    int Levels() const { return mLevels; }
+
+    // These round UP, so beware that the last tile for a given level may not be completely full
+    // TODO: Verify this...
+    int NumTilesX(int level = 0) const { return (mNumTilesX + (1 << level) - 1) >> level; }
+    int NumTilesY(int level = 0) const { return (mNumTilesY + (1 << level) - 1) >> level; }
+    int TileWidth(int level = 0) const { return (mTileWidth << level); }
+    int TileHeight(int level = 0) const { return (mTileHeight << level); }
+
+    float MinZ(int level, int tileX, int tileY) const {
+        return mMinZArrays[level][tileY * NumTilesX(level) + tileX];
+    }
+    float MaxZ(int level, int tileX, int tileY) const {
+        return mMaxZArrays[level][tileY * NumTilesX(level) + tileX];
+    }
+
+private:
+    int mTileWidth;
+    int mTileHeight;
+    int mLevels;
+    int mNumTilesX;
+    int mNumTilesY;
+
+    // One array for each "level" in the tree
+    float **mMinZArrays;
+    float **mMaxZArrays;
+};
+
+static MinMaxZTreeCilk *gMinMaxZTreeCilk = 0;
+
+void InitDynamicCilk(InputData *input) {
+    gMinMaxZTreeCilk = 
+        new MinMaxZTreeCilk(MIN_TILE_WIDTH, MIN_TILE_HEIGHT, DYNAMIC_TREE_LEVELS,
+                            input->header.framebufferWidth, 
+                            input->header.framebufferHeight);
+}
+
+
+static void
+ShadeDynamicTileRecurse(InputData *input, int level, int tileX, int tileY, 
+                        int *lightIndices, int numLights, 
+                        Framebuffer *framebuffer) {
+    const MinMaxZTreeCilk *minMaxZTree = gMinMaxZTreeCilk;
+    
+    // If we few enough lights or this is the base case (last level), shade
+    // this full tile directly
+    if (level == 0 || numLights < DYNAMIC_MIN_LIGHTS_TO_SUBDIVIDE) {
+        int width = minMaxZTree->TileWidth(level);
+        int height = minMaxZTree->TileHeight(level);
+        int startX = tileX * width;
+        int startY = tileY * height;
+        int endX = std::min(input->header.framebufferWidth, startX + width);
+        int endY = std::min(input->header.framebufferHeight, startY + height);
+        
+        // Skip entirely offscreen tiles
+        if (endX > startX && endY > startY) {
+            ispc::ShadeTile(
+                startX, endX, startY, endY,
+                input->header.framebufferWidth, input->header.framebufferHeight,
+                &input->arrays,
+                input->header.cameraProj[0][0], input->header.cameraProj[1][1], 
+                input->header.cameraProj[2][2], input->header.cameraProj[3][2],
+                lightIndices, numLights, VISUALIZE_LIGHT_COUNT, 
+                framebuffer->r, framebuffer->g, framebuffer->b);
+        }
+    } 
+    else {
+        // Otherwise, subdivide and 4-way recurse using X and Y splitting planes
+        // Move down a level in the tree
+        --level;
+        tileX <<= 1;
+        tileY <<= 1;
+        int width = minMaxZTree->TileWidth(level);
+        int height = minMaxZTree->TileHeight(level);
+
+        // Work out splitting coords
+        int midX = (tileX + 1) * width;
+        int midY = (tileY + 1) * height;
+
+        // Read subtile min/max data
+        // NOTE: We must be sure to handle out-of-bounds access here since
+        // sometimes we'll only have 1 or 2 subtiles for non-pow-2
+        // framebuffer sizes.
+        bool rightTileExists = (tileX + 1 < minMaxZTree->NumTilesX(level));
+        bool bottomTileExists = (tileY + 1 < minMaxZTree->NumTilesY(level));
+
+        // NOTE: Order is 00, 10, 01, 11
+        // Set defaults up to cull all lights if the tile doesn't exist (offscreen)
+        float minZ[4] = {input->header.cameraFar, input->header.cameraFar, 
+                         input->header.cameraFar, input->header.cameraFar};
+        float maxZ[4] = {input->header.cameraNear, input->header.cameraNear, 
+                         input->header.cameraNear, input->header.cameraNear};
+
+        minZ[0] = minMaxZTree->MinZ(level, tileX, tileY);
+        maxZ[0] = minMaxZTree->MaxZ(level, tileX, tileY);
+        if (rightTileExists) {
+            minZ[1] = minMaxZTree->MinZ(level, tileX + 1, tileY);
+            maxZ[1] = minMaxZTree->MaxZ(level, tileX + 1, tileY);
+            if (bottomTileExists) {
+                minZ[3] = minMaxZTree->MinZ(level, tileX + 1, tileY + 1);
+                maxZ[3] = minMaxZTree->MaxZ(level, tileX + 1, tileY + 1);
+            }
+        }
+        if (bottomTileExists) {
+            minZ[2] = minMaxZTree->MinZ(level, tileX, tileY + 1);
+            maxZ[2] = minMaxZTree->MaxZ(level, tileX, tileY + 1);
+        }
+
+        // Cull lights into subtile lists
+#ifdef ISPC_IS_WINDOWS
+        __declspec(align(ALIGNMENT_BYTES)) 
+#endif
+            int subtileLightIndices[4][MAX_LIGHTS]
+#ifndef ISPC_IS_WINDOWS
+            __attribute__ ((aligned(ALIGNMENT_BYTES)))
+#endif
+;
+        int subtileNumLights[4];
+        ispc::SplitTileMinMax(midX, midY, minZ, maxZ,
+            input->header.framebufferWidth, input->header.framebufferHeight, 
+            input->header.cameraProj[0][0], input->header.cameraProj[1][1],
+            lightIndices, numLights, input->arrays.lightPositionView_x, 
+            input->arrays.lightPositionView_y, input->arrays.lightPositionView_z, 
+            input->arrays.lightAttenuationEnd,
+            subtileLightIndices[0], MAX_LIGHTS, subtileNumLights);
+        
+        // Recurse into subtiles
+        _Cilk_spawn ShadeDynamicTileRecurse(input, level, tileX    , tileY, 
+                                            subtileLightIndices[0], subtileNumLights[0],
+                                            framebuffer);
+        _Cilk_spawn ShadeDynamicTileRecurse(input, level, tileX + 1, tileY,
+                                            subtileLightIndices[1], subtileNumLights[1],
+                                            framebuffer);
+        _Cilk_spawn ShadeDynamicTileRecurse(input, level, tileX    , tileY + 1,
+                                            subtileLightIndices[2], subtileNumLights[2],
+                                            framebuffer);
+        ShadeDynamicTileRecurse(input, level, tileX + 1, tileY + 1,
+                                subtileLightIndices[3], subtileNumLights[3],
+                                framebuffer);
+    }
+}
+
+
+static void
+ShadeDynamicTile(InputData *input, int level, int tileX, int tileY,
+                 Framebuffer *framebuffer) {
+    const MinMaxZTreeCilk *minMaxZTree = gMinMaxZTreeCilk;
+
+    // Get Z min/max for this tile
+    int width = minMaxZTree->TileWidth(level);
+    int height = minMaxZTree->TileHeight(level);
+    float minZ = minMaxZTree->MinZ(level, tileX, tileY);
+    float maxZ = minMaxZTree->MaxZ(level, tileX, tileY);
+
+    int startX = tileX * width;
+    int startY = tileY * height;
+    int endX = std::min(input->header.framebufferWidth, startX + width);
+    int endY = std::min(input->header.framebufferHeight, startY + height);
+
+    // This is a root tile, so first do a full 6-plane cull
+#ifdef ISPC_IS_WINDOWS
+    __declspec(align(ALIGNMENT_BYTES)) 
+#endif
+        int lightIndices[MAX_LIGHTS]
+#ifndef ISPC_IS_WINDOWS
+        __attribute__ ((aligned(ALIGNMENT_BYTES)))
+#endif
+;
+    int numLights = ispc::IntersectLightsWithTileMinMax(
+        startX, endX, startY, endY,    minZ, maxZ,
+        input->header.framebufferWidth, input->header.framebufferHeight,
+        input->header.cameraProj[0][0], input->header.cameraProj[1][1],
+        MAX_LIGHTS, input->arrays.lightPositionView_x, 
+        input->arrays.lightPositionView_y, input->arrays.lightPositionView_z, 
+        input->arrays.lightAttenuationEnd, lightIndices);
+
+    // Now kick off the recursive process for this tile
+    ShadeDynamicTileRecurse(input, level, tileX, tileY, lightIndices, 
+                            numLights, framebuffer);
+}
+
+
+void
+DispatchDynamicCilk(InputData *input, Framebuffer *framebuffer)
+{
+    MinMaxZTreeCilk *minMaxZTree = gMinMaxZTreeCilk;
+        
+    // Update min/max Z tree
+    minMaxZTree->Update(input->arrays.zBuffer, input->header.framebufferWidth,
+        input->header.cameraProj[2][2], input->header.cameraProj[3][2], 
+        input->header.cameraNear, input->header.cameraFar);
+
+    // Launch the "root" tiles.  Ideally these should at least fill the
+    // machine... at the moment we have a static number of "levels" to the
+    // mip tree but it might make sense to compute it based on the width of
+    // the machine.
+    int rootLevel = minMaxZTree->Levels() - 1;
+    int rootTilesX = minMaxZTree->NumTilesX(rootLevel);
+    int rootTilesY = minMaxZTree->NumTilesY(rootLevel);
+    int rootTiles = rootTilesX * rootTilesY;
+    _Cilk_for (int g = 0; g < rootTiles; ++g) {
+        uint32_t tileY = g / rootTilesX;
+        uint32_t tileX = g % rootTilesX;
+        ShadeDynamicTile(input, rootLevel, tileX, tileY, framebuffer);
+    }
+}
+
+#endif // __cilk

examples/deferred/kernels.ispc

@@ -0,0 +1,700 @@
+/*
+  Copyright (c) 2010-2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#include "deferred.h"
+
+struct InputDataArrays
+{
+    uniform float * uniform zBuffer;
+    uniform unsigned int16 * uniform normalEncoded_x; // half float
+    uniform unsigned int16 * uniform normalEncoded_y; // half float
+    uniform unsigned int16 * uniform specularAmount; // half float
+    uniform unsigned int16 * uniform specularPower; // half float
+    uniform unsigned int8 * uniform albedo_x; // unorm8
+    uniform unsigned int8 * uniform albedo_y; // unorm8
+    uniform unsigned int8 * uniform albedo_z; // unorm8
+    uniform float * uniform lightPositionView_x;
+    uniform float * uniform lightPositionView_y;
+    uniform float * uniform lightPositionView_z;
+    uniform float * uniform lightAttenuationBegin;
+    uniform float * uniform lightColor_x;
+    uniform float * uniform lightColor_y;
+    uniform float * uniform lightColor_z;
+    uniform float * uniform lightAttenuationEnd;
+};
+
+struct InputHeader
+{
+    uniform float cameraProj[4][4];
+    uniform float cameraNear;
+    uniform float cameraFar;
+
+    uniform int32 framebufferWidth;
+    uniform int32 framebufferHeight;
+    uniform int32 numLights;
+    uniform int32 inputDataChunkSize;
+    uniform int32 inputDataArrayOffsets[idaNum];
+};
+
+
+///////////////////////////////////////////////////////////////////////////
+// Common utility routines
+
+static inline float
+dot3(float x, float y, float z, float a, float b, float c) {
+    return (x*a + y*b + z*c);
+}
+
+
+static inline void
+normalize3(float x, float y, float z, float &ox, float &oy, float &oz) {
+    float n = rsqrt(x*x + y*y + z*z);
+    ox = x * n;
+    oy = y * n;
+    oz = z * n;
+}
+
+
+static inline float
+Unorm8ToFloat32(unsigned int8 u) {
+    return (float)u * (1.0f / 255.0f);
+}
+
+
+static inline unsigned int8
+Float32ToUnorm8(float f) {
+    return (unsigned int8)(f * 255.0f);
+}
+
+
+static void
+ComputeZBounds(
+    uniform int32 tileStartX, uniform int32 tileEndX,
+    uniform int32 tileStartY, uniform int32 tileEndY,
+    // G-buffer data
+    uniform float zBuffer[],
+    uniform int32 gBufferWidth,
+    // Camera data
+    uniform float cameraProj_33, uniform float cameraProj_43,
+    uniform float cameraNear, uniform float cameraFar,
+    // Output
+    uniform float &minZ,
+    uniform float &maxZ
+    )
+{
+    // Find Z bounds
+    float laneMinZ = cameraFar;
+    float laneMaxZ = cameraNear;
+    for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
+        foreach (x = tileStartX ... tileEndX) {
+            // Unproject depth buffer Z value into view space
+            float z = zBuffer[y * gBufferWidth + x];
+            float viewSpaceZ = cameraProj_43 / (z - cameraProj_33);
+
+            // Work out Z bounds for our samples
+            // Avoid considering skybox/background or otherwise invalid pixels
+            if ((viewSpaceZ < cameraFar) && (viewSpaceZ >= cameraNear)) {
+                laneMinZ = min(laneMinZ, viewSpaceZ);
+                laneMaxZ = max(laneMaxZ, viewSpaceZ);
+            }
+        }
+    }
+    minZ = reduce_min(laneMinZ);
+    maxZ = reduce_max(laneMaxZ);
+}
+
+
+export uniform int32
+IntersectLightsWithTileMinMax(
+    uniform int32 tileStartX, uniform int32 tileEndX,
+    uniform int32 tileStartY, uniform int32 tileEndY,
+    // Tile data
+    uniform float minZ,
+    uniform float maxZ,
+    // G-buffer data
+    uniform int32 gBufferWidth, uniform int32 gBufferHeight,
+    // Camera data
+    uniform float cameraProj_11, uniform float cameraProj_22,
+    // Light Data
+    uniform int32 numLights,
+    uniform float light_positionView_x_array[],
+    uniform float light_positionView_y_array[],
+    uniform float light_positionView_z_array[],
+    uniform float light_attenuationEnd_array[],
+    // Output
+    uniform int32 tileLightIndices[]
+    )
+{
+    uniform float gBufferScale_x = 0.5f * (float)gBufferWidth;
+    uniform float gBufferScale_y = 0.5f * (float)gBufferHeight;
+        
+    // Parallize across frustum planes.
+    // We really only have four side planes here, but write the code to
+    // handle programCount > 4 robustly
+    uniform float frustumPlanes_xy[programCount];
+    uniform float frustumPlanes_z[programCount];
+
+    // TODO: If programIndex < 4 here? Don't care about masking off the
+    // rest but if interleaving ("x2" modes) the other lanes should ideally
+    // not be emitted...
+    {
+        // This one is totally constant over the whole screen... worth pulling it up at all?
+        float frustumPlanes_xy_v;
+        frustumPlanes_xy_v = insert(frustumPlanes_xy_v, 0, -(cameraProj_11 * gBufferScale_x));
+        frustumPlanes_xy_v = insert(frustumPlanes_xy_v, 1,  (cameraProj_11 * gBufferScale_x));
+        frustumPlanes_xy_v = insert(frustumPlanes_xy_v, 2,  (cameraProj_22 * gBufferScale_y));
+        frustumPlanes_xy_v = insert(frustumPlanes_xy_v, 3, -(cameraProj_22 * gBufferScale_y));
+    
+        float frustumPlanes_z_v;
+        frustumPlanes_z_v = insert(frustumPlanes_z_v, 0,  tileEndX - gBufferScale_x);
+        frustumPlanes_z_v = insert(frustumPlanes_z_v, 1, -tileStartX + gBufferScale_x);
+        frustumPlanes_z_v = insert(frustumPlanes_z_v, 2,  tileEndY - gBufferScale_y);
+        frustumPlanes_z_v = insert(frustumPlanes_z_v, 3, -tileStartY + gBufferScale_y);
+
+        // Normalize
+        float norm = rsqrt(frustumPlanes_xy_v * frustumPlanes_xy_v + 
+                           frustumPlanes_z_v * frustumPlanes_z_v);
+            frustumPlanes_xy_v *= norm;
+            frustumPlanes_z_v *= norm;
+
+        // Save out for uniform use later
+        frustumPlanes_xy[programIndex] = frustumPlanes_xy_v;
+        frustumPlanes_z[programIndex] = frustumPlanes_z_v;
+    }
+
+    uniform int32 tileNumLights = 0;
+
+    foreach (lightIndex = 0 ... numLights) {
+        float light_positionView_z = light_positionView_z_array[lightIndex];
+        float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
+        float light_attenuationEndNeg = -light_attenuationEnd;
+
+        float d = light_positionView_z - minZ;
+        bool inFrustum = (d >= light_attenuationEndNeg);
+
+        d = maxZ - light_positionView_z;
+        inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+        
+        // This seems better than cif(!inFrustum) ccontinue; here since we
+        // don't actually need to mask the rest of this function - this is
+        // just a greedy early-out.  Could also structure all of this as
+        // nested if() statements, but this a bit easier to read
+        if (any(inFrustum)) {
+            float light_positionView_x = light_positionView_x_array[lightIndex];
+            float light_positionView_y = light_positionView_y_array[lightIndex];
+
+            d = light_positionView_z * frustumPlanes_z[0] + 
+                light_positionView_x * frustumPlanes_xy[0];
+            inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+
+            d = light_positionView_z * frustumPlanes_z[1] + 
+                light_positionView_x * frustumPlanes_xy[1];
+            inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+
+            d = light_positionView_z * frustumPlanes_z[2] + 
+                light_positionView_y * frustumPlanes_xy[2];
+            inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+
+            d = light_positionView_z * frustumPlanes_z[3] + 
+                light_positionView_y * frustumPlanes_xy[3];
+            inFrustum = inFrustum && (d >= light_attenuationEndNeg);
+        
+            // Pack and store intersecting lights
+            cif (inFrustum) {
+                tileNumLights += packed_store_active(&tileLightIndices[tileNumLights], 
+                                                     lightIndex);
+            }
+        }
+    }
+
+    return tileNumLights;
+}
+
+
+static uniform int32
+IntersectLightsWithTile(
+    uniform int32 tileStartX, uniform int32 tileEndX,
+    uniform int32 tileStartY, uniform int32 tileEndY,
+    uniform int32 gBufferWidth, uniform int32 gBufferHeight,
+    // G-buffer data
+    uniform float zBuffer[],
+    // Camera data
+    uniform float cameraProj_11, uniform float cameraProj_22,
+    uniform float cameraProj_33, uniform float cameraProj_43,
+    uniform float cameraNear, uniform float cameraFar,
+    // Light Data
+    uniform int32 numLights,
+    uniform float light_positionView_x_array[],
+    uniform float light_positionView_y_array[],
+    uniform float light_positionView_z_array[],
+    uniform float light_attenuationEnd_array[],
+    // Output
+    uniform int32 tileLightIndices[]
+    )
+{
+    uniform float minZ, maxZ;
+    ComputeZBounds(tileStartX, tileEndX, tileStartY, tileEndY,
+        zBuffer, gBufferWidth, cameraProj_33, cameraProj_43, cameraNear, cameraFar,
+        minZ, maxZ);
+
+    uniform int32 tileNumLights = IntersectLightsWithTileMinMax(
+        tileStartX, tileEndX, tileStartY, tileEndY, minZ, maxZ,
+        gBufferWidth, gBufferHeight, cameraProj_11, cameraProj_22,
+        MAX_LIGHTS, light_positionView_x_array, light_positionView_y_array, 
+        light_positionView_z_array, light_attenuationEnd_array,
+        tileLightIndices);
+
+    return tileNumLights;
+}
+
+
+export void
+ShadeTile(
+    uniform int32 tileStartX, uniform int32 tileEndX,
+    uniform int32 tileStartY, uniform int32 tileEndY,
+    uniform int32 gBufferWidth, uniform int32 gBufferHeight,
+    uniform InputDataArrays &inputData,
+    // Camera data
+    uniform float cameraProj_11, uniform float cameraProj_22,
+    uniform float cameraProj_33, uniform float cameraProj_43,
+    // Light list
+    uniform int32 tileLightIndices[],
+    uniform int32 tileNumLights,
+    // UI
+    uniform bool visualizeLightCount,
+    // Output
+    uniform unsigned int8 framebuffer_r[],
+    uniform unsigned int8 framebuffer_g[],
+    uniform unsigned int8 framebuffer_b[]
+    )
+{
+    if (tileNumLights == 0 || visualizeLightCount) {
+        uniform unsigned int8 c = (unsigned int8)(min(tileNumLights << 2, 255));
+        for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
+            foreach (x = tileStartX ... tileEndX) {
+                int32 framebufferIndex = (y * gBufferWidth + x);
+                framebuffer_r[framebufferIndex] = c;
+                framebuffer_g[framebufferIndex] = c;
+                framebuffer_b[framebufferIndex] = c;
+            }
+        }
+    } else {
+        uniform float twoOverGBufferWidth = 2.0f / gBufferWidth;
+        uniform float twoOverGBufferHeight = 2.0f / gBufferHeight;
+        
+        for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
+            uniform float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
+
+            foreach (x = tileStartX ... tileEndX) {
+                int32 gBufferOffset = y * gBufferWidth + x;
+                
+                // Reconstruct position and (negative) view vector from G-buffer
+                float surface_positionView_x, surface_positionView_y, surface_positionView_z;
+                float Vneg_x, Vneg_y, Vneg_z;
+
+                float z = inputData.zBuffer[gBufferOffset];
+
+                // Compute screen/clip-space position
+                // NOTE: Mind DX11 viewport transform and pixel center!
+                float positionScreen_x = (0.5f + (float)(x)) * 
+                    twoOverGBufferWidth - 1.0f;
+
+                // Unproject depth buffer Z value into view space
+                surface_positionView_z = cameraProj_43 / (z - cameraProj_33);
+                surface_positionView_x = positionScreen_x * surface_positionView_z / 
+                    cameraProj_11;
+                surface_positionView_y = positionScreen_y * surface_positionView_z / 
+                    cameraProj_22;
+                
+                // We actually end up with a vector pointing *at* the
+                // surface (i.e. the negative view vector)
+                normalize3(surface_positionView_x, surface_positionView_y, 
+                           surface_positionView_z, Vneg_x, Vneg_y, Vneg_z);
+
+                // Reconstruct normal from G-buffer
+                float surface_normal_x, surface_normal_y, surface_normal_z;
+                float normal_x = half_to_float_fast(inputData.normalEncoded_x[gBufferOffset]);
+                float normal_y = half_to_float_fast(inputData.normalEncoded_y[gBufferOffset]);
+                    
+                float f = (normal_x - normal_x * normal_x) + (normal_y - normal_y * normal_y);
+                float m = sqrt(4.0f * f - 1.0f);
+                    
+                surface_normal_x = m * (4.0f * normal_x - 2.0f);
+                surface_normal_y = m * (4.0f * normal_y - 2.0f);
+                surface_normal_z = 3.0f - 8.0f * f;
+
+                // Load other G-buffer parameters
+                float surface_specularAmount = 
+                    half_to_float_fast(inputData.specularAmount[gBufferOffset]);
+                float surface_specularPower  = 
+                    half_to_float_fast(inputData.specularPower[gBufferOffset]);
+                float surface_albedo_x = Unorm8ToFloat32(inputData.albedo_x[gBufferOffset]);
+                float surface_albedo_y = Unorm8ToFloat32(inputData.albedo_y[gBufferOffset]);
+                float surface_albedo_z = Unorm8ToFloat32(inputData.albedo_z[gBufferOffset]);
+                
+                float lit_x = 0.0f;
+                float lit_y = 0.0f;
+                float lit_z = 0.0f;
+                for (uniform int32 tileLightIndex = 0; tileLightIndex < tileNumLights; 
+                     ++tileLightIndex) {
+                    uniform int32 lightIndex = tileLightIndices[tileLightIndex];
+                                        
+                    // Gather light data relevant to initial culling
+                    uniform float light_positionView_x = 
+                        inputData.lightPositionView_x[lightIndex];
+                    uniform float light_positionView_y = 
+                        inputData.lightPositionView_y[lightIndex];
+                    uniform float light_positionView_z = 
+                        inputData.lightPositionView_z[lightIndex];
+                    uniform float light_attenuationEnd = 
+                        inputData.lightAttenuationEnd[lightIndex];
+                    
+                    // Compute light vector
+                    float L_x = light_positionView_x - surface_positionView_x;
+                    float L_y = light_positionView_y - surface_positionView_y;
+                    float L_z = light_positionView_z - surface_positionView_z;
+
+                    float distanceToLight2 = dot3(L_x, L_y, L_z, L_x, L_y, L_z);
+                    
+                    // Clip at end of attenuation
+                    float light_attenutaionEnd2 = light_attenuationEnd * light_attenuationEnd;
+
+                    cif (distanceToLight2 < light_attenutaionEnd2) {                    
+                        float distanceToLight = sqrt(distanceToLight2);
+
+                        // HLSL "rcp" is allowed to be fairly inaccurate
+                        float distanceToLightRcp = rcp(distanceToLight);
+                        L_x *= distanceToLightRcp;
+                        L_y *= distanceToLightRcp;
+                        L_z *= distanceToLightRcp;
+
+                        // Start computing brdf
+                        float NdotL = dot3(surface_normal_x, surface_normal_y, 
+                                           surface_normal_z, L_x, L_y, L_z);
+                    
+                        // Clip back facing
+                        cif (NdotL > 0.0f) {
+                            uniform float light_attenuationBegin = 
+                                inputData.lightAttenuationBegin[lightIndex];
+
+                            // Light distance attenuation (linstep)
+                            float lightRange = (light_attenuationEnd - light_attenuationBegin);
+                            float falloffPosition = (light_attenuationEnd - distanceToLight);
+                            float attenuation = min(falloffPosition / lightRange, 1.0f);
+
+                            float H_x = (L_x - Vneg_x);
+                            float H_y = (L_y - Vneg_y);
+                            float H_z = (L_z - Vneg_z);
+                            normalize3(H_x, H_y, H_z, H_x, H_y, H_z);
+                    
+                            float NdotH = dot3(surface_normal_x, surface_normal_y, 
+                                               surface_normal_z, H_x, H_y, H_z);
+                            NdotH = max(NdotH, 0.0f);
+
+                            float specular = pow(NdotH, surface_specularPower);
+                            float specularNorm = (surface_specularPower + 2.0f) * 
+                                (1.0f / 8.0f);
+                            float specularContrib = surface_specularAmount * 
+                                specularNorm * specular;
+
+                            float k = attenuation * NdotL * (1.0f + specularContrib);
+                    
+                            uniform float light_color_x = inputData.lightColor_x[lightIndex];
+                            uniform float light_color_y = inputData.lightColor_y[lightIndex];
+                            uniform float light_color_z = inputData.lightColor_z[lightIndex];
+
+                            float lightContrib_x = surface_albedo_x * light_color_x;
+                            float lightContrib_y = surface_albedo_y * light_color_y;
+                            float lightContrib_z = surface_albedo_z * light_color_z;
+
+                            lit_x += lightContrib_x * k;
+                            lit_y += lightContrib_y * k;
+                            lit_z += lightContrib_z * k;
+                        }
+                    }
+                }
+
+                // Gamma correct
+                // These pows are pretty slow right now, but we can do
+                // something faster if really necessary to squeeze every
+                // last bit of performance out of it
+                float gamma = 1.0 / 2.2f;
+                lit_x = pow(clamp(lit_x, 0.0f, 1.0f), gamma);
+                lit_y = pow(clamp(lit_y, 0.0f, 1.0f), gamma);
+                lit_z = pow(clamp(lit_z, 0.0f, 1.0f), gamma);
+                
+                framebuffer_r[gBufferOffset] = Float32ToUnorm8(lit_x);
+                framebuffer_g[gBufferOffset] = Float32ToUnorm8(lit_y);
+                framebuffer_b[gBufferOffset] = Float32ToUnorm8(lit_z);
+            }
+        }
+    }
+}
+
+
+///////////////////////////////////////////////////////////////////////////
+// Static decomposition
+
+task void
+RenderTile(uniform int num_groups_x, uniform int num_groups_y,
+           uniform InputHeader &inputHeader,
+           uniform InputDataArrays &inputData,
+           uniform int visualizeLightCount,
+           // Output
+           uniform unsigned int8 framebuffer_r[],
+           uniform unsigned int8 framebuffer_g[],
+           uniform unsigned int8 framebuffer_b[]) {
+    uniform int32 group_y = taskIndex / num_groups_x;
+    uniform int32 group_x = taskIndex % num_groups_x;
+    uniform int32 tile_start_x = group_x * MIN_TILE_WIDTH;
+    uniform int32 tile_start_y = group_y * MIN_TILE_HEIGHT;
+    uniform int32 tile_end_x = tile_start_x + MIN_TILE_WIDTH;
+    uniform int32 tile_end_y = tile_start_y + MIN_TILE_HEIGHT;
+
+    uniform int framebufferWidth = inputHeader.framebufferWidth;
+    uniform int framebufferHeight = inputHeader.framebufferHeight;
+    uniform float cameraProj_00 = inputHeader.cameraProj[0][0];
+    uniform float cameraProj_11 = inputHeader.cameraProj[1][1];
+    uniform float cameraProj_22 = inputHeader.cameraProj[2][2];
+    uniform float cameraProj_32 = inputHeader.cameraProj[3][2];
+
+    // Light intersection: figure out which lights illuminate this tile.
+    uniform int tileLightIndices[MAX_LIGHTS];  // Light list for the tile
+    uniform int numTileLights = 
+        IntersectLightsWithTile(tile_start_x, tile_end_x, 
+                                tile_start_y, tile_end_y,
+                                framebufferWidth, framebufferHeight,
+                                inputData.zBuffer,
+                                cameraProj_00, cameraProj_11,
+                                cameraProj_22, cameraProj_32,
+                                inputHeader.cameraNear, inputHeader.cameraFar,
+                                MAX_LIGHTS,
+                                inputData.lightPositionView_x, 
+                                inputData.lightPositionView_y, 
+                                inputData.lightPositionView_z, 
+                                inputData.lightAttenuationEnd,
+                                tileLightIndices);
+
+    // And now shade the tile, using the lights in tileLightIndices
+    ShadeTile(tile_start_x, tile_end_x, tile_start_y, tile_end_y,
+              framebufferWidth, framebufferHeight, inputData,
+              cameraProj_00, cameraProj_11, cameraProj_22, cameraProj_32,
+              tileLightIndices, numTileLights, visualizeLightCount, 
+              framebuffer_r, framebuffer_g, framebuffer_b);
+}
+
+
+export void
+RenderStatic(uniform InputHeader &inputHeader,
+             uniform InputDataArrays &inputData,
+             uniform int visualizeLightCount,
+             // Output
+             uniform unsigned int8 framebuffer_r[],
+             uniform unsigned int8 framebuffer_g[],
+             uniform unsigned int8 framebuffer_b[]) {
+    uniform int num_groups_x = (inputHeader.framebufferWidth + 
+                                MIN_TILE_WIDTH - 1) / MIN_TILE_WIDTH;
+    uniform int num_groups_y = (inputHeader.framebufferHeight + 
+                                MIN_TILE_HEIGHT - 1) / MIN_TILE_HEIGHT;
+    uniform int num_groups = num_groups_x * num_groups_y;
+
+    // Launch a task to render each tile, each of which is MIN_TILE_WIDTH
+    // by MIN_TILE_HEIGHT pixels.
+    launch[num_groups] < RenderTile(num_groups_x, num_groups_y,
+                                    inputHeader, inputData, visualizeLightCount,
+                                    framebuffer_r, framebuffer_g, framebuffer_b) >;
+}
+
+
+///////////////////////////////////////////////////////////////////////////
+// Routines for dynamic decomposition path
+
+// This computes the z min/max range for a whole row worth of tiles.
+export void
+ComputeZBoundsRow(
+    uniform int32 tileY,
+    uniform int32 tileWidth, uniform int32 tileHeight,
+    uniform int32 numTilesX, uniform int32 numTilesY,
+    // G-buffer data
+    uniform float zBuffer[],
+    uniform int32 gBufferWidth,
+    // Camera data
+    uniform float cameraProj_33, uniform float cameraProj_43,
+    uniform float cameraNear, uniform float cameraFar,
+    // Output
+    uniform float minZArray[],
+    uniform float maxZArray[]
+    )
+{
+    for (uniform int32 tileX = 0; tileX < numTilesX; ++tileX) {
+        uniform float minZ, maxZ;
+        ComputeZBounds(
+            tileX * tileWidth, tileX * tileWidth + tileWidth,
+            tileY * tileHeight, tileY * tileHeight + tileHeight,
+            zBuffer, gBufferWidth,
+            cameraProj_33, cameraProj_43, cameraNear, cameraFar,
+            minZ, maxZ);
+        minZArray[tileX] = minZ;
+        maxZArray[tileX] = maxZ;
+    }
+}
+
+
+// Reclassifies the lights with respect to four sub-tiles when we refine a tile.
+// numLights need not be a multiple of programCount here, but the input and output arrays
+// should be able to handle programCount-sized load/stores.
+export void
+SplitTileMinMax(
+    uniform int32 tileMidX, uniform int32 tileMidY,
+    // Subtile data (00, 10, 01, 11)
+    uniform float subtileMinZ[],
+    uniform float subtileMaxZ[],
+    // G-buffer data
+    uniform int32 gBufferWidth, uniform int32 gBufferHeight,
+    // Camera data
+    uniform float cameraProj_11, uniform float cameraProj_22,
+    // Light Data
+    uniform int32 lightIndices[],
+    uniform int32 numLights,
+    uniform float light_positionView_x_array[],
+    uniform float light_positionView_y_array[],
+    uniform float light_positionView_z_array[],
+    uniform float light_attenuationEnd_array[],
+    // Outputs
+    // TODO: ISPC doesn't currently like multidimensionsal arrays so we'll do the
+    // indexing math ourselves
+    uniform int32 subtileIndices[],
+    uniform int32 subtileIndicesPitch,
+    uniform int32 subtileNumLights[]
+    )
+{
+    uniform float gBufferScale_x = 0.5f * (float)gBufferWidth;
+    uniform float gBufferScale_y = 0.5f * (float)gBufferHeight;
+        
+    // Parallize across frustum planes
+    // Only have 2 frustum split planes here so may not be worth it, but
+    // we'll do it for now for consistency
+    uniform float frustumPlanes_xy[programCount];
+    uniform float frustumPlanes_z[programCount];
+
+    // This one is totally constant over the whole screen... worth pulling it up at all?
+    float frustumPlanes_xy_v;
+    frustumPlanes_xy_v = insert(frustumPlanes_xy_v, 0, -(cameraProj_11 * gBufferScale_x));
+    frustumPlanes_xy_v = insert(frustumPlanes_xy_v, 1,  (cameraProj_22 * gBufferScale_y));
+    
+    float frustumPlanes_z_v;
+    frustumPlanes_z_v = insert(frustumPlanes_z_v, 0, tileMidX - gBufferScale_x);
+    frustumPlanes_z_v = insert(frustumPlanes_z_v, 1, tileMidY - gBufferScale_y);
+
+    // Normalize
+    float norm = rsqrt(frustumPlanes_xy_v * frustumPlanes_xy_v + 
+                       frustumPlanes_z_v * frustumPlanes_z_v);
+    frustumPlanes_xy_v *= norm;
+    frustumPlanes_z_v *= norm;
+
+    // Save out for uniform use later
+    frustumPlanes_xy[programIndex] = frustumPlanes_xy_v;
+    frustumPlanes_z[programIndex] = frustumPlanes_z_v;
+
+    // Initialize
+    uniform int32 subtileLightOffset[4];
+    subtileLightOffset[0] = 0 * subtileIndicesPitch;
+    subtileLightOffset[1] = 1 * subtileIndicesPitch;
+    subtileLightOffset[2] = 2 * subtileIndicesPitch;
+    subtileLightOffset[3] = 3 * subtileIndicesPitch;
+
+    foreach (i = 0 ... numLights) {
+        int32 lightIndex = lightIndices[i];
+
+        float light_positionView_x = light_positionView_x_array[lightIndex];
+        float light_positionView_y = light_positionView_y_array[lightIndex];
+        float light_positionView_z = light_positionView_z_array[lightIndex];
+        float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
+        float light_attenuationEndNeg = -light_attenuationEnd;
+        
+        // Test lights again subtile z bounds
+        bool inFrustum[4];
+        inFrustum[0] = (light_positionView_z - subtileMinZ[0] >= light_attenuationEndNeg) &&
+            (subtileMaxZ[0] - light_positionView_z >= light_attenuationEndNeg);
+        inFrustum[1] = (light_positionView_z - subtileMinZ[1] >= light_attenuationEndNeg) && 
+            (subtileMaxZ[1] - light_positionView_z >= light_attenuationEndNeg);
+        inFrustum[2] = (light_positionView_z - subtileMinZ[2] >= light_attenuationEndNeg) && 
+            (subtileMaxZ[2] - light_positionView_z >= light_attenuationEndNeg);
+        inFrustum[3] = (light_positionView_z - subtileMinZ[3] >= light_attenuationEndNeg) && 
+            (subtileMaxZ[3] - light_positionView_z >= light_attenuationEndNeg);
+
+        float dx = light_positionView_z * frustumPlanes_z[0] + 
+            light_positionView_x * frustumPlanes_xy[0];
+        float dy = light_positionView_z * frustumPlanes_z[1] +
+            light_positionView_y * frustumPlanes_xy[1];
+        
+        cif (abs(dx) > light_attenuationEnd) {
+            bool positiveX = dx > 0.0f;
+            inFrustum[0] = inFrustum[0] &&  positiveX;    // 00 subtile
+            inFrustum[1] = inFrustum[1] && !positiveX;    // 10 subtile
+            inFrustum[2] = inFrustum[2] &&  positiveX;    // 01 subtile
+            inFrustum[3] = inFrustum[3] && !positiveX;    // 11 subtile
+        }
+        cif (abs(dy) > light_attenuationEnd) {
+            bool positiveY = dy > 0.0f;
+            inFrustum[0] = inFrustum[0] &&  positiveY;    // 00 subtile
+            inFrustum[1] = inFrustum[1] &&  positiveY;    // 10 subtile
+            inFrustum[2] = inFrustum[2] && !positiveY;    // 01 subtile
+            inFrustum[3] = inFrustum[3] && !positiveY;    // 11 subtile
+        }
+
+        // Pack and store intersecting lights
+        // TODO: Experiment with a loop here instead
+        cif (inFrustum[0])
+            subtileLightOffset[0] += 
+            packed_store_active(&subtileIndices[subtileLightOffset[0]],
+                                lightIndex);
+        cif (inFrustum[1])
+            subtileLightOffset[1] += 
+            packed_store_active(&subtileIndices[subtileLightOffset[1]],
+                                lightIndex);
+        cif (inFrustum[2])
+            subtileLightOffset[2] += 
+            packed_store_active(&subtileIndices[subtileLightOffset[2]], 
+                                lightIndex);
+        cif (inFrustum[3])
+            subtileLightOffset[3] += 
+            packed_store_active(&subtileIndices[subtileLightOffset[3]], 
+                                lightIndex);
+    }
+
+    subtileNumLights[0] = subtileLightOffset[0] - 0 * subtileIndicesPitch;
+    subtileNumLights[1] = subtileLightOffset[1] - 1 * subtileIndicesPitch;
+    subtileNumLights[2] = subtileLightOffset[2] - 2 * subtileIndicesPitch;
+    subtileNumLights[3] = subtileLightOffset[3] - 3 * subtileIndicesPitch;
+}

examples/deferred/main.cpp

@@ -0,0 +1,139 @@
+/*
+  Copyright (c) 2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#ifdef _MSC_VER
+#define ISPC_IS_WINDOWS
+#define NOMINMAX
+#elif defined(__linux__)
+#define ISPC_IS_LINUX
+#elif defined(__APPLE__)
+#define ISPC_IS_APPLE
+#endif
+
+#include <fcntl.h>
+#include <float.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/types.h>
+#include <stdint.h>
+#include <algorithm>
+#include <assert.h>
+#include <vector>
+#ifdef ISPC_IS_WINDOWS
+  #define WIN32_LEAN_AND_MEAN
+  #include <windows.h>
+#endif
+#include "deferred.h"
+#include "kernels_ispc.h"
+#include "../timing.h"
+
+///////////////////////////////////////////////////////////////////////////
+
+int main(int argc, char** argv) {
+    if (argc != 2) {
+        printf("usage: deferred_shading <input_file (e.g. data/pp1280x720.bin)>\n");
+        return 1;
+    }
+
+    InputData *input = CreateInputDataFromFile(argv[1]);
+    if (!input) {
+        printf("Failed to load input file \"%s\"!\n", argv[1]);
+        return 1;
+    }
+
+    Framebuffer framebuffer(input->header.framebufferWidth,
+                            input->header.framebufferHeight);
+
+    InitDynamicC(input);
+#ifdef __cilk
+    InitDynamicCilk(input);
+#endif // __cilk
+
+    int nframes = 5;
+    double ispcCycles = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        framebuffer.clear();
+        reset_and_start_timer();
+        for (int j = 0; j < nframes; ++j)
+            ispc::RenderStatic(&input->header, &input->arrays, 
+                               VISUALIZE_LIGHT_COUNT,
+                               framebuffer.r, framebuffer.g, framebuffer.b);
+        double mcycles = get_elapsed_mcycles() / nframes;
+        ispcCycles = std::min(ispcCycles, mcycles);
+    }
+    printf("[ispc static + tasks]:\t\t[%.3f] million cycles to render "
+           "%d x %d image\n", ispcCycles,
+           input->header.framebufferWidth, input->header.framebufferHeight);
+    WriteFrame("deferred-ispc-static.ppm", input, framebuffer);
+
+#ifdef __cilk
+    double dynamicCilkCycles = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        framebuffer.clear();
+        reset_and_start_timer();
+        for (int j = 0; j < nframes; ++j)
+            DispatchDynamicCilk(input, &framebuffer);
+        double mcycles = get_elapsed_mcycles() / nframes;
+        dynamicCilkCycles = std::min(dynamicCilkCycles, mcycles);
+    }
+    printf("[ispc + Cilk dynamic]:\t\t[%.3f] million cycles to render image\n", 
+           dynamicCilkCycles);
+    WriteFrame("deferred-ispc-dynamic.ppm", input, framebuffer);
+#endif // __cilk
+
+    double serialCycles = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        framebuffer.clear();
+        reset_and_start_timer();
+        for (int j = 0; j < nframes; ++j)
+            DispatchDynamicC(input, &framebuffer);
+        double mcycles = get_elapsed_mcycles() / nframes;
+        serialCycles = std::min(serialCycles, mcycles);
+    }
+    printf("[C++ serial dynamic, 1 core]:\t[%.3f] million cycles to render image\n", 
+           serialCycles);
+    WriteFrame("deferred-serial-dynamic.ppm", input, framebuffer);
+
+#ifdef __cilk
+    printf("\t\t\t\t(%.2fx speedup from static ISPC, %.2fx from Cilk+ISPC)\n", 
+           serialCycles/ispcCycles, serialCycles/dynamicCilkCycles);
+#else
+    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", serialCycles/ispcCycles);
+#endif // __cilk
+
+    DeleteInputData(input);
+
+    return 0;
+}

examples/examples.sln

@@ -15,6 +15,14 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "mandelbrot_tasks", "mandelb
 EndProject
 Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "aobench_instrumented", "aobench_instrumented\aobench_instrumented.vcxproj", "{B3B4AE3D-6D5A-4CF9-AF5B-43CF2131B958}"
 EndProject
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "noise", "noise\noise.vcxproj", "{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}"
+EndProject
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "volume", "volume_rendering\volume.vcxproj", "{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}"
+EndProject
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "stencil", "stencil\stencil.vcxproj", "{2EF070A1-F62F-4E6A-944B-88D140945C3C}"
+EndProject
+Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "deferred_shading", "deferred\deferred_shading.vcxproj", "{87F53C53-957E-4E91-878A-BC27828FB9EB}"
+EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
 		Debug|Win32 = Debug|Win32
@@ -79,6 +87,38 @@ Global
 		{B3B4AE3D-6D5A-4CF9-AF5B-43CF2131B958}.Release|Win32.Build.0 = Release|Win32
 		{B3B4AE3D-6D5A-4CF9-AF5B-43CF2131B958}.Release|x64.ActiveCfg = Release|x64
 		{B3B4AE3D-6D5A-4CF9-AF5B-43CF2131B958}.Release|x64.Build.0 = Release|x64
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Debug|Win32.ActiveCfg = Debug|Win32
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Debug|Win32.Build.0 = Debug|Win32
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Debug|x64.ActiveCfg = Debug|x64
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Debug|x64.Build.0 = Debug|x64
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Release|Win32.ActiveCfg = Release|Win32
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Release|Win32.Build.0 = Release|Win32
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Release|x64.ActiveCfg = Release|x64
+		{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}.Release|x64.Build.0 = Release|x64
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Debug|Win32.ActiveCfg = Debug|Win32
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Debug|Win32.Build.0 = Debug|Win32
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Debug|x64.ActiveCfg = Debug|x64
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Debug|x64.Build.0 = Debug|x64
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Release|Win32.ActiveCfg = Release|Win32
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Release|Win32.Build.0 = Release|Win32
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Release|x64.ActiveCfg = Release|x64
+		{DEE5733A-E93E-449D-9114-9BFFCAEB4DF9}.Release|x64.Build.0 = Release|x64
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Debug|Win32.ActiveCfg = Debug|Win32
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Debug|Win32.Build.0 = Debug|Win32
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Debug|x64.ActiveCfg = Debug|x64
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Debug|x64.Build.0 = Debug|x64
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Release|Win32.ActiveCfg = Release|Win32
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Release|Win32.Build.0 = Release|Win32
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Release|x64.ActiveCfg = Release|x64
+		{2EF070A1-F62F-4E6A-944B-88D140945C3C}.Release|x64.Build.0 = Release|x64
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Debug|Win32.ActiveCfg = Debug|Win32
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Debug|Win32.Build.0 = Debug|Win32
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Debug|x64.ActiveCfg = Debug|x64
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Debug|x64.Build.0 = Debug|x64
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Release|Win32.ActiveCfg = Release|Win32
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Release|Win32.Build.0 = Release|Win32
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Release|x64.ActiveCfg = Release|x64
+		{87F53C53-957E-4E91-878A-BC27828FB9EB}.Release|x64.Build.0 = Release|x64
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE

examples/mandelbrot/.gitignore

@@ -0,0 +1,3 @@
+mandelbrot
+*.ppm
+objs

examples/mandelbrot/Makefile

@@ -1,8 +1,8 @@
 
-CXX=g++
+CXX=g++ -m64
 CXXFLAGS=-Iobjs/ -O3 -Wall
 ISPC=ispc
-ISPCFLAGS=-O2 --target=sse4x2
+ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx-x2 --arch=x86-64
 
 default: mandelbrot
 
@@ -14,13 +14,17 @@ dirs:
 clean:
 	/bin/rm -rf objs *~ mandelbrot
 
-mandelbrot: dirs objs/mandelbrot.o objs/mandelbrot_serial.o objs/mandelbrot_ispc.o
-	$(CXX) $(CXXFLAGS) -o $@ objs/mandelbrot.o objs/mandelbrot_ispc.o objs/mandelbrot_serial.o -lm
+OBJS=objs/mandelbrot.o objs/mandelbrot_serial.o objs/mandelbrot_ispc_sse2.o \
+	objs/mandelbrot_ispc_sse4.o objs/mandelbrot_ispc_avx.o \
+	objs/mandelbrot_ispc.o
+
+mandelbrot: dirs $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm
 
 objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
 objs/mandelbrot.o: objs/mandelbrot_ispc.h 
 
-objs/%_ispc.h objs/%_ispc.o: %.ispc
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
 	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/mandelbrot/mandelbrot.cpp

@@ -63,6 +63,7 @@ writePPM(int *buf, int width, int height, const char *fn) {
             fputc(c, fp);
     }
     fclose(fp);
+    printf("Wrote image file %s\n", fn);
 }
 
 

examples/mandelbrot/mandelbrot.ispc

@@ -51,7 +51,7 @@ export void mandelbrot_ispc(uniform float x0, uniform float y0,
                             uniform float x1, uniform float y1,
                             uniform int width, uniform int height, 
                             uniform int maxIterations,
-                            reference uniform int output[])
+                            uniform int output[])
 {
     float dx = (x1 - x0) / width;
     float dy = (y1 - y0) / height;
@@ -60,16 +60,16 @@ export void mandelbrot_ispc(uniform float x0, uniform float y0,
         // Note that we'll be doing programCount computations in parallel,
         // so increment i by that much.  This assumes that width evenly
         // divides programCount.
-        for (uniform int i = 0; i < width; i += programCount) {
+        foreach (i = 0 ... width) {
             // Figure out the position on the complex plane to compute the
             // number of iterations at.  Note that the x values are
             // different across different program instances, since its
             // initializer incorporates the value of the programIndex
             // variable.
-            float x = x0 + (programIndex + i) * dx;
+            float x = x0 + i * dx;
             float y = y0 + j * dy;
 
-            int index = j * width + i + programIndex;
+            int index = j * width + i;
             output[index] = mandel(x, y, maxIterations);
         }
     }

examples/mandelbrot/mandelbrot.vcxproj

@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup Label="ProjectConfigurations">
     <ProjectConfiguration Include="Debug|Win32">
@@ -64,15 +64,19 @@
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -81,6 +85,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -94,6 +101,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -109,6 +119,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -126,6 +138,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -141,18 +155,18 @@
   <ItemGroup>
     <CustomBuild Include="mandelbrot.ispc">
       <FileType>Document</FileType>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />

examples/mandelbrot/mandelbrot_serial.cpp

@@ -36,7 +36,7 @@ static int mandel(float c_re, float c_im, int count) {
     float z_re = c_re, z_im = c_im;
     int i;
     for (i = 0; i < count; ++i) {
-        if (z_re * z_re + z_im * z_im > 4.)
+        if (z_re * z_re + z_im * z_im > 4.f)
             break;
 
         float new_re = z_re*z_re - z_im*z_im;

examples/mandelbrot_tasks/.gitignore

@@ -0,0 +1,2 @@
+mandelbrot
+*.ppm

examples/mandelbrot_tasks/Makefile

@@ -1,20 +1,18 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
-
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=tasks_gcd.cpp
-  TASK_LIB=
-endif
-
-TASK_OBJ=$(addprefix objs/, $(TASK_CXX:.cpp=.o))
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
 CXX=g++
-CXXFLAGS=-Iobjs/ -O3 -Wall
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
 ISPC=ispc
-ISPCFLAGS=-O2 --target=sse4x2
+ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx-x2 --arch=x86-64
+
+OBJS=objs/mandelbrot.o objs/mandelbrot_serial.o $(TASK_OBJ) \
+	objs/mandelbrot_ispc.o objs/mandelbrot_ispc_sse2.o \
+	objs/mandelbrot_ispc_sse4.o objs/mandelbrot_ispc_avx.o 
 
 default: mandelbrot
 
@@ -26,13 +24,16 @@ dirs:
 clean:
 	/bin/rm -rf objs *~ mandelbrot
 
-mandelbrot: dirs objs/mandelbrot.o objs/mandelbrot_serial.o objs/mandelbrot_ispc.o $(TASK_OBJ)
-	$(CXX) $(CXXFLAGS) -o $@ objs/mandelbrot.o objs/mandelbrot_ispc.o objs/mandelbrot_serial.o $(TASK_OBJ) -lm $(TASK_LIB)
+mandelbrot: dirs $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm $(TASK_LIB)
 
 objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
 objs/mandelbrot.o: objs/mandelbrot_ispc.h 
 
-objs/%_ispc.h objs/%_ispc.o: %.ispc
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
 	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/mandelbrot_tasks/mandelbrot.cpp

@@ -40,6 +40,7 @@
 
 #include <stdio.h>
 #include <algorithm>
+#include <string.h>
 #include "../timing.h"
 #include "mandelbrot_ispc.h"
 using namespace ispc;
@@ -63,10 +64,16 @@ writePPM(int *buf, int width, int height, const char *fn) {
             fputc(c, fp);
     }
     fclose(fp);
+    printf("Wrote image file %s\n", fn);
 }
 
 
-int main() {
+static void usage() {
+    fprintf(stderr, "usage: mandelbrot [--scale=<factor>]\n");
+    exit(1);
+}
+
+int main(int argc, char *argv[]) {
     unsigned int width = 1536;
     unsigned int height = 1024;
     float x0 = -2;
@@ -74,8 +81,24 @@ int main() {
     float y0 = -1;
     float y1 = 1;
 
-    extern void TasksInit();
-    TasksInit();
+    if (argc == 1)
+        ;
+    else if (argc == 2) {
+        if (strncmp(argv[1], "--scale=", 8) == 0) {
+            float scale = atof(argv[1] + 8);
+            if (scale == 0.f)
+                usage();
+            width *= scale;
+            height *= scale;
+            // round up to multiples of 16
+            width = (width + 0xf) & ~0xf;
+            height = (height + 0xf) & ~0xf;
+        }
+        else 
+            usage();
+    }
+    else
+        usage();
 
     int maxIterations = 512;
     int *buf = new int[width*height];
@@ -86,6 +109,9 @@ int main() {
     //
     double minISPC = 1e30;
     for (int i = 0; i < 3; ++i) {
+        // Clear out the buffer
+        for (unsigned int i = 0; i < width * height; ++i)
+            buf[i] = 0;
         reset_and_start_timer();
         mandelbrot_ispc(x0, y0, x1, y1, width, height, maxIterations, buf);
         double dt = get_elapsed_mcycles();
@@ -95,9 +121,6 @@ int main() {
     printf("[mandelbrot ispc+tasks]:\t[%.3f] million cycles\n", minISPC);
     writePPM(buf, width, height, "mandelbrot-ispc.ppm");
 
-    // Clear out the buffer
-    for (unsigned int i = 0; i < width * height; ++i)
-        buf[i] = 0;
 
     // 
     // And run the serial implementation 3 times, again reporting the
@@ -105,6 +128,9 @@ int main() {
     //
     double minSerial = 1e30;
     for (int i = 0; i < 3; ++i) {
+        // Clear out the buffer
+        for (unsigned int i = 0; i < width * height; ++i)
+            buf[i] = 0;
         reset_and_start_timer();
         mandelbrot_serial(x0, y0, x1, y1, width, height, maxIterations, buf);
         double dt = get_elapsed_mcycles();

examples/mandelbrot_tasks/mandelbrot.ispc

@@ -53,34 +53,46 @@ mandel(float c_re, float c_im, int count) {
    [ystart,yend).
  */
 task void
-mandelbrot_scanlines(uniform int ystart, uniform int yend,
+mandelbrot_scanlines(uniform int ybase, uniform int span,
                      uniform float x0, uniform float dx, 
                      uniform float y0, uniform float dy,
                      uniform int width, uniform int maxIterations,
-                     reference uniform int output[]) {
-    for (uniform int j = ystart; j < yend; ++j) {
-        for (uniform int i = 0; i < width; i += programCount) {
-            float x = x0 + (programIndex + i) * dx;
-            float y = y0 + j * dy;
+                     uniform int output[]) {
+    uniform int ystart = ybase + taskIndex * span;
+    uniform int yend = ystart + span;
 
-            int index = j * width + i + programIndex;
-            output[index] = mandel(x, y, maxIterations);
-        }
+    foreach (yi = ystart ... yend, xi = 0 ... width) {
+        float x = x0 + xi * dx;
+        float y = y0 + yi * dy;
+
+        int index = yi * width + xi;
+        output[index] = mandel(x, y, maxIterations);
     }
 }
                                
 
+task void
+mandelbrot_chunk(uniform float x0, uniform float dx,
+                 uniform float y0, uniform float dy,
+                 uniform int width, uniform int height,
+                 uniform int maxIterations, uniform int output[]) {
+    uniform int ystart = taskIndex * (height/taskCount);
+    uniform int yend = (taskIndex+1) * (height/taskCount);
+    uniform int span = 1;
+
+    launch[(yend-ystart)/span] < mandelbrot_scanlines(ystart, span, x0, dx, y0, dy,
+                                                      width, maxIterations, output) >;
+}
+
+
 export void
 mandelbrot_ispc(uniform float x0, uniform float y0, 
                 uniform float x1, uniform float y1,
                 uniform int width, uniform int height, 
-                uniform int maxIterations, reference uniform int output[]) {
+                uniform int maxIterations, uniform int output[]) {
     uniform float dx = (x1 - x0) / width;
     uniform float dy = (y1 - y0) / height;
 
-    /* Launch task to compute results for spans of 'span' scanlines. */
-    uniform int span = 2;
-    for (uniform int j = 0; j < height; j += span)
-        launch < mandelbrot_scanlines(j, j+span, x0, dx, y0, dy, width,
-                                      maxIterations, output) >;
+    launch[32] < mandelbrot_chunk(x0, dx, y0, dy, width, height,
+                                  maxIterations, output) >;
 }

examples/mandelbrot_tasks/mandelbrot_serial.cpp

@@ -36,7 +36,7 @@ static int mandel(float c_re, float c_im, int count) {
     float z_re = c_re, z_im = c_im;
     int i;
     for (i = 0; i < count; ++i) {
-        if (z_re * z_re + z_im * z_im > 4.)
+        if (z_re * z_re + z_im * z_im > 4.f)
             break;
 
         float new_re = z_re*z_re - z_im*z_im;

examples/mandelbrot_tasks/mandelbrot_tasks.vcxproj

@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup Label="ProjectConfigurations">
     <ProjectConfiguration Include="Debug|Win32">
@@ -64,15 +64,19 @@
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -81,6 +85,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -94,6 +101,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -109,6 +119,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -126,6 +138,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -137,26 +151,26 @@
   <ItemGroup>
     <ClCompile Include="mandelbrot.cpp" />
     <ClCompile Include="mandelbrot_serial.cpp" />
-    <ClCompile Include="tasks_concrt.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="mandelbrot.ispc">
       <FileType>Document</FileType>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

examples/mandelbrot_tasks/tasks_concrt.cpp

@@ -1,115 +0,0 @@
-/*
-  Copyright (c) 2010-2011, Intel Corporation
-  All rights reserved.
-
-  Redistribution and use in source and binary forms, with or without
-  modification, are permitted provided that the following conditions are
-  met:
-
-    * Redistributions of source code must retain the above copyright
-      notice, this list of conditions and the following disclaimer.
-
-    * Redistributions in binary form must reproduce the above copyright
-      notice, this list of conditions and the following disclaimer in the
-      documentation and/or other materials provided with the distribution.
-
-    * Neither the name of Intel Corporation nor the names of its
-      contributors may be used to endorse or promote products derived from
-      this software without specific prior written permission.
-
-
-   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
-   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
-   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
-   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
-   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
-   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
-   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
-   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
-   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
-   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
-*/
-
-/* Simple task system implementation for ispc based on Microsoft's
-   Concurrency Runtime. */
-
-#include <windows.h>
-#include <concrt.h>
-using namespace Concurrency;
-#include <assert.h>
-#include <stdio.h>
-
-// ispc expects these functions to have C linkage / not be mangled
-extern "C" { 
-    void ISPCLaunch(void *f, void *data);
-    void ISPCSync();
-}
-
-typedef void (*TaskFuncType)(void *, int, int);
-
-struct TaskInfo {
-    TaskFuncType ispcFunc;
-    void *ispcData;
-};
-
-// This is a simple implementation that just aborts if more than MAX_TASKS
-// are launched.  It could easily be extended to be more general...
-
-#define MAX_TASKS 4096
-static int taskOffset;
-static TaskInfo taskInfo[MAX_TASKS];
-static event *events[MAX_TASKS];
-static CRITICAL_SECTION criticalSection;
-
-void
-TasksInit() {
-    InitializeCriticalSection(&criticalSection);
-    for (int i = 0; i < MAX_TASKS; ++i)
-        events[i] = new event;
-}
-
-
-void __cdecl
-lRunTask(LPVOID param) {
-    TaskInfo *ti = (TaskInfo *)param;
-    
-    // Actually run the task. 
-    // FIXME: like the tasks_gcd.cpp implementation, this is passing bogus
-    // values for the threadIndex and threadCount builtins, which in turn
-    // will cause bugs in code that uses those.  FWIW this example doesn't
-    // use them...
-    int threadIndex = 0;
-    int threadCount = 1;
-    ti->ispcFunc(ti->ispcData, threadIndex, threadCount);
-
-    // Signal the event that this task is done
-    int taskNum = ti - &taskInfo[0];
-    events[taskNum]->set();
-}
-
-
-void
-ISPCLaunch(void *func, void *data) {
-    // Get a TaskInfo struct for this task
-    EnterCriticalSection(&criticalSection);
-    TaskInfo *ti = &taskInfo[taskOffset++];
-    assert(taskOffset < MAX_TASKS);
-    LeaveCriticalSection(&criticalSection);
-
-    // And pass it on to the Concurrency Runtime...
-    ti->ispcFunc = (TaskFuncType)func;
-    ti->ispcData = data;
-    CurrentScheduler::ScheduleTask(lRunTask, ti);
-}
-
-
-void ISPCSync() {
-    event::wait_for_multiple(&events[0], taskOffset, true, 
-                             COOPERATIVE_TIMEOUT_INFINITE);
-
-    for (int i = 0; i < taskOffset; ++i)
-        events[i]->reset();
-
-    taskOffset = 0;
-}

examples/mandelbrot_tasks/tasks_pthreads.cpp

@@ -1,285 +0,0 @@
-/*
-  Copyright (c) 2010-2011, Intel Corporation
-  All rights reserved.
-
-  Redistribution and use in source and binary forms, with or without
-  modification, are permitted provided that the following conditions are
-  met:
-
-    * Redistributions of source code must retain the above copyright
-      notice, this list of conditions and the following disclaimer.
-
-    * Redistributions in binary form must reproduce the above copyright
-      notice, this list of conditions and the following disclaimer in the
-      documentation and/or other materials provided with the distribution.
-
-    * Neither the name of Intel Corporation nor the names of its
-      contributors may be used to endorse or promote products derived from
-      this software without specific prior written permission.
-
-
-   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
-   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
-   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
-   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
-   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
-   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
-   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
-   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
-   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
-   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
-   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
-*/
-
-#include <pthread.h>
-#include <semaphore.h>
-#include <string.h>
-#include <unistd.h>
-#include <assert.h>
-#include <stdio.h>
-#include <fcntl.h>
-#include <sys/types.h>
-#include <sys/stat.h>
-#include <sys/param.h>
-#include <sys/sysctl.h>
-#include <stdint.h>
-#include <stdlib.h>
-#include <errno.h>
-#include <vector>
-
-// ispc expects these functions to have C linkage / not be mangled
-extern "C" { 
-    void ISPCLaunch(void *f, void *data);
-    void ISPCSync();
-}
-
-
-static int nThreads;
-static pthread_t *threads;
-static pthread_mutex_t taskQueueMutex;
-static std::vector<std::pair<void *, void *> > taskQueue;
-static sem_t *workerSemaphore;
-static uint32_t numUnfinishedTasks;
-static pthread_mutex_t tasksRunningConditionMutex;
-static pthread_cond_t tasksRunningCondition;
-
-static void *lTaskEntry(void *arg);
-
-/** Figure out how many CPU cores there are in the system
- */
-static int
-lNumCPUCores() {
-#if defined(__linux__)
-    return sysconf(_SC_NPROCESSORS_ONLN);
-#else
-    // Mac
-    int mib[2];
-    mib[0] = CTL_HW;
-    size_t length = 2;
-    if (sysctlnametomib("hw.logicalcpu", mib, &length) == -1) {
-        fprintf(stderr, "sysctlnametomib() filed.  Guessing 2 cores.");
-        return 2;
-    }
-    assert(length == 2);
-
-    int nCores = 0;
-    size_t size = sizeof(nCores);
-
-    if (sysctl(mib, 2, &nCores, &size, NULL, 0) == -1) {
-        fprintf(stderr, "sysctl() to find number of cores present failed.  Guessing 2.");
-        return 2;
-    }
-    return nCores;
-#endif
-}
-
-void
-TasksInit() {
-    nThreads = lNumCPUCores();
-
-    threads = new pthread_t[nThreads];
-
-    int err;
-    if ((err = pthread_mutex_init(&taskQueueMutex, NULL)) != 0) {
-        fprintf(stderr, "Error creating mutex: %s\n", strerror(err));
-        exit(1);
-    }
-
-    char name[32];
-    sprintf(name, "mandelbrot.%d", (int)getpid());
-    workerSemaphore = sem_open(name, O_CREAT, S_IRUSR|S_IWUSR, 0);
-    if (!workerSemaphore) {
-        fprintf(stderr, "Error creating semaphore: %s\n", strerror(err));
-        exit(1);
-    }
-
-    if ((err = pthread_cond_init(&tasksRunningCondition, NULL)) != 0) {
-        fprintf(stderr, "Error creating condition variable: %s\n", strerror(err));
-        exit(1);
-    }
-
-    if ((err = pthread_mutex_init(&tasksRunningConditionMutex, NULL)) != 0) {
-        fprintf(stderr, "Error creating mutex: %s\n", strerror(err));
-        exit(1);
-    }
-
-    for (int i = 0; i < nThreads; ++i) {
-        err = pthread_create(&threads[i], NULL, &lTaskEntry, reinterpret_cast<void *>(i));
-        if (err != 0) {
-            fprintf(stderr, "Error creating pthread %d: %s\n", i, strerror(err));
-            exit(1);
-        }
-    }
-}
-
-
-void
-ISPCLaunch(void *f, void *d) {
-    //
-    // Acquire mutex, add task
-    //
-    int err;
-    if ((err = pthread_mutex_lock(&taskQueueMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    taskQueue.push_back(std::make_pair(f, d));
-
-    if ((err = pthread_mutex_unlock(&taskQueueMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    //
-    // Update count of number of tasks left to run
-    //
-    if ((err = pthread_mutex_lock(&tasksRunningConditionMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    ++numUnfinishedTasks;
-
-    if ((err = pthread_mutex_unlock(&tasksRunningConditionMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    //
-    // Post to the worker semaphore to wake up worker threads that are
-    // sleeping waiting for tasks to show up
-    //
-    if ((err = sem_post(workerSemaphore)) != 0) {
-        fprintf(stderr, "Error from sem_post: %s\n", strerror(err));
-        exit(1);
-    }
-}
-
-
-static void *
-lTaskEntry(void *arg) {
-    int threadIndex = int(reinterpret_cast<int64_t>(arg));
-    int threadCount = nThreads;
-
-    while (true) {
-        int err;
-        if ((err = sem_wait(workerSemaphore)) != 0) {
-            fprintf(stderr, "Error from sem_wait: %s\n", strerror(err));
-            exit(1);
-        }
-
-        std::pair<void *, void *> myTask;
-        //
-        // Acquire mutex, get task
-        //
-        if ((err = pthread_mutex_lock(&taskQueueMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-            exit(1);
-        }
-        if (taskQueue.size() == 0) {
-            //
-            // Task queue is empty, go back and wait on the semaphore
-            //
-            if ((err = pthread_mutex_unlock(&taskQueueMutex)) != 0) {
-                fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
-                exit(1);
-            }
-            continue;
-        }
-
-        myTask = taskQueue.back();
-        taskQueue.pop_back();
-
-        if ((err = pthread_mutex_unlock(&taskQueueMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
-            exit(1);
-        }
-
-        //
-        // Do work for _myTask_
-        //
-        typedef void (*TaskFunType)(void *, int, int);
-        TaskFunType func = (TaskFunType)myTask.first;
-        func(myTask.second, threadIndex, threadCount);
-
-        //
-        // Decrement the number of unfinished tasks counter
-        //
-        if ((err = pthread_mutex_lock(&tasksRunningConditionMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-            exit(1);
-        }
-
-        int unfinished = --numUnfinishedTasks;
-        if (unfinished == 0) {
-            //
-            // Signal the "no more tasks are running" condition if all of
-            // them are done.
-            //
-            int err;
-            if ((err = pthread_cond_signal(&tasksRunningCondition)) != 0) {
-                fprintf(stderr, "Error from pthread_cond_signal: %s\n", strerror(err));
-                exit(1);
-            }
-        }
-
-        if ((err = pthread_mutex_unlock(&tasksRunningConditionMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-            exit(1);
-        }
-    }
-
-    pthread_exit(NULL);
-    return 0;
-}
-
-
-void ISPCSync() {
-    int err;
-    if ((err = pthread_mutex_lock(&tasksRunningConditionMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    // As long as there are tasks running, wait on the condition variable;
-    // doing so causes this thread to go to sleep until someone signals on
-    // the tasksRunningCondition condition variable.
-    while (numUnfinishedTasks > 0) {
-        if ((err = pthread_cond_wait(&tasksRunningCondition, 
-                                     &tasksRunningConditionMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_cond_wait: %s\n", strerror(err));
-            exit(1);
-        }
-    }
-    
-    // We acquire ownership of the condition variable mutex when the above
-    // pthread_cond_wait returns.
-    // FIXME: is there a lurking issue here if numUnfinishedTasks gets back
-    // to zero by the time we get to ISPCSync() and thence we're trying to
-    // unlock a mutex we don't have a lock on?
-    if ((err = pthread_mutex_unlock(&tasksRunningConditionMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-}

examples/noise/.gitignore

@@ -0,0 +1,3 @@
+noise
+*.ppm
+objs

examples/noise/Makefile

@@ -0,0 +1,29 @@
+
+CXX=g++ -m64
+CXXFLAGS=-Iobjs/ -O3 -Wall
+ISPC=ispc
+ISPCFLAGS=-O2 --target=sse2,sse4,avx-x2 --arch=x86-64
+
+OBJS=objs/noise.o objs/noise_serial.o objs/noise_ispc.o objs/noise_ispc_sse2.o \
+	objs/noise_ispc_sse4.o objs/noise_ispc_avx.o 
+
+default: noise
+
+.PHONY: dirs clean
+
+dirs:
+	/bin/mkdir -p objs/
+
+clean:
+	/bin/rm -rf objs *~ noise
+
+noise: dirs $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm
+
+objs/%.o: %.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/noise.o: objs/noise_ispc.h 
+
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
+	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/noise/noise.cpp

@@ -0,0 +1,115 @@
+/*
+  Copyright (c) 2010-2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#ifdef _MSC_VER
+#define _CRT_SECURE_NO_WARNINGS
+#define NOMINMAX
+#pragma warning (disable: 4244)
+#pragma warning (disable: 4305)
+#endif
+
+#include <stdio.h>
+#include <algorithm>
+#include "../timing.h"
+#include "noise_ispc.h"
+using namespace ispc;
+
+extern void noise_serial(float x0, float y0, float x1, float y1,
+                         int width, int height, float output[]);
+
+/* Write a PPM image file with the image */
+static void
+writePPM(float *buf, int width, int height, const char *fn) {
+    FILE *fp = fopen(fn, "wb");
+    fprintf(fp, "P6\n");
+    fprintf(fp, "%d %d\n", width, height);
+    fprintf(fp, "255\n");
+    for (int i = 0; i < width*height; ++i) {
+        float v = buf[i] * 255.f;
+        if (v < 0) v = 0;
+        if (v > 255) v = 255;
+        for (int j = 0; j < 3; ++j)
+            fputc((char)v, fp);
+    }
+    fclose(fp);
+}
+
+
+int main() {
+    unsigned int width = 768;
+    unsigned int height = 768;
+    float x0 = -10;
+    float x1 = 10;
+    float y0 = -10;
+    float y1 = 10;
+
+    float *buf = new float[width*height];
+
+    //
+    // Compute the image using the ispc implementation; report the minimum
+    // time of three runs.
+    //
+    double minISPC = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        noise_ispc(x0, y0, x1, y1, width, height, buf);
+        double dt = get_elapsed_mcycles();
+        minISPC = std::min(minISPC, dt);
+    }
+
+    printf("[noise ispc]:\t\t\t[%.3f] million cycles\n", minISPC);
+    writePPM(buf, width, height, "noise-ispc.ppm");
+
+    // Clear out the buffer
+    for (unsigned int i = 0; i < width * height; ++i)
+        buf[i] = 0;
+
+    // 
+    // And run the serial implementation 3 times, again reporting the
+    // minimum time.
+    //
+    double minSerial = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        noise_serial(x0, y0, x1, y1, width, height, buf);
+        double dt = get_elapsed_mcycles();
+        minSerial = std::min(minSerial, dt);
+    }
+
+    printf("[noise serial]:\t\t\t[%.3f] millon cycles\n", minSerial);
+    writePPM(buf, width, height, "noise-serial.ppm");
+
+    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", minSerial/minISPC);
+
+    return 0;
+}

examples/noise/noise.ispc

@@ -0,0 +1,164 @@
+/*
+  Copyright (c) 2010-2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#define NOISE_PERM_SIZE 256
+
+static uniform int NoisePerm[2 * NOISE_PERM_SIZE] = {
+    151, 160, 137, 91, 90, 15, 131, 13, 201, 95, 96, 53, 194, 233, 7, 225, 140,
+    36, 103, 30, 69, 142, 8, 99, 37, 240, 21, 10, 23, 190, 6, 148, 247, 120,
+    234, 75, 0, 26, 197, 62, 94, 252, 219, 203, 117, 35, 11, 32, 57, 177, 33,
+    88, 237, 149, 56, 87, 174, 20, 125, 136, 171, 168,  68, 175, 74, 165, 71, 
+    134, 139, 48, 27, 166, 77, 146, 158, 231, 83, 111, 229, 122, 60, 211, 133, 
+    230, 220, 105, 92, 41, 55, 46, 245, 40, 244, 102, 143, 54, 65, 25, 63, 161,
+    1, 216, 80, 73, 209, 76, 132, 187, 208,  89, 18, 169, 200, 196, 135, 130, 
+    116, 188, 159, 86, 164, 100, 109, 198, 173, 186,  3, 64, 52, 217, 226, 250,
+    124, 123, 5, 202, 38, 147, 118, 126, 255, 82, 85, 212, 207, 206, 59, 227, 
+    47, 16, 58, 17, 182, 189, 28, 42, 223, 183, 170, 213, 119, 248, 152,  2, 44,
+    154, 163, 70, 221, 153, 101, 155, 167,  43, 172, 9, 129, 22, 39, 253,  19, 
+    98, 108, 110, 79, 113, 224, 232, 178, 185,  112, 104, 218, 246, 97, 228, 251,
+    34, 242, 193, 238, 210, 144, 12, 191, 179, 162, 241, 81, 51, 145, 235, 249,
+    14, 239, 107, 49, 192, 214,  31, 181, 199, 106, 157, 184, 84, 204, 176, 115,
+    121, 50, 45, 127,  4, 150, 254, 138, 236, 205, 93, 222, 114, 67, 29, 24, 72, 
+    243, 141, 128, 195, 78, 66, 215, 61, 156, 180, 151, 160, 137, 91, 90, 15,
+    131, 13, 201, 95, 96, 53, 194, 233, 7, 225, 140, 36, 103, 30, 69, 142, 8, 99,
+    37, 240, 21, 10, 23, 190, 6, 148, 247, 120, 234, 75, 0, 26, 197, 62, 94, 252,
+    219, 203, 117, 35, 11, 32, 57, 177, 33, 88, 237, 149, 56, 87, 174, 20, 125, 
+    136, 171, 168,  68, 175, 74, 165, 71, 134, 139, 48, 27, 166, 77, 146, 158,
+    231, 83, 111, 229, 122, 60, 211, 133, 230, 220, 105, 92, 41, 55, 46, 245,
+    40, 244, 102, 143, 54,  65, 25, 63, 161,  1, 216, 80, 73, 209, 76, 132, 187,
+    208,  89, 18, 169, 200, 196, 135, 130, 116, 188, 159, 86, 164, 100, 109, 
+    198, 173, 186,  3, 64, 52, 217, 226, 250, 124, 123, 5, 202, 38, 147, 118,
+    126, 255, 82, 85, 212, 207, 206, 59, 227, 47, 16, 58, 17, 182, 189, 28, 42,
+    223, 183, 170, 213, 119, 248, 152,  2, 44, 154, 163, 70, 221, 153, 101, 155, 
+    167,  43, 172, 9, 129, 22, 39, 253,  19, 98, 108, 110, 79, 113, 224, 232,
+    178, 185,  112, 104, 218, 246, 97, 228, 251, 34, 242, 193, 238, 210, 144,
+    12, 191, 179, 162, 241,  81, 51, 145, 235, 249, 14, 239, 107, 49, 192, 214,
+    31, 181, 199, 106, 157, 184,  84, 204, 176, 115, 121, 50, 45, 127,  4, 150,
+    254, 138, 236, 205, 93, 222, 114, 67, 29, 24, 72, 243, 141, 128, 195, 78, 
+    66, 215, 61, 156, 180
+};
+
+
+inline float SmoothStep(float low, float high, float value) {
+    float v = clamp((value - low) / (high - low), 0.f, 1.f);
+    return v * v * (-2.f * v  + 3.f);
+}
+
+
+inline int Floor2Int(float val) {
+    return (int)floor(val);
+}
+
+
+inline float Grad(int x, int y, int z, float dx, float dy, float dz) {
+    int h = NoisePerm[NoisePerm[NoisePerm[x]+y]+z];
+    h &= 15;
+    float u = h<8 || h==12 || h==13 ? dx : dy;
+    float v = h<4 || h==12 || h==13 ? dy : dz;
+    return ((h&1) ? -u : u) + ((h&2) ? -v : v);
+}
+
+
+inline float NoiseWeight(float t) {
+    float t3 = t*t*t;
+    float t4 = t3*t;
+    return 6.f*t4*t - 15.f*t4 + 10.f*t3;
+}
+
+
+inline float Lerp(float t, float low, float high) {
+    return (1. - t) * low + t * high;
+}
+
+
+static float Noise(float x, float y, float z) {
+    // Compute noise cell coordinates and offsets
+    int ix = Floor2Int(x), iy = Floor2Int(y), iz = Floor2Int(z);
+    float dx = x - ix, dy = y - iy, dz = z - iz;
+
+    // Compute gradient weights
+    ix &= (NOISE_PERM_SIZE-1);
+    iy &= (NOISE_PERM_SIZE-1);
+    iz &= (NOISE_PERM_SIZE-1);
+    float w000 = Grad(ix,   iy,   iz,   dx,   dy,   dz);
+    float w100 = Grad(ix+1, iy,   iz,   dx-1, dy,   dz);
+    float w010 = Grad(ix,   iy+1, iz,   dx,   dy-1, dz);
+    float w110 = Grad(ix+1, iy+1, iz,   dx-1, dy-1, dz);
+    float w001 = Grad(ix,   iy,   iz+1, dx,   dy,   dz-1);
+    float w101 = Grad(ix+1, iy,   iz+1, dx-1, dy,   dz-1);
+    float w011 = Grad(ix,   iy+1, iz+1, dx,   dy-1, dz-1);
+    float w111 = Grad(ix+1, iy+1, iz+1, dx-1, dy-1, dz-1);
+
+    // Compute trilinear interpolation of weights
+    float wx = NoiseWeight(dx), wy = NoiseWeight(dy), wz = NoiseWeight(dz);
+    float x00 = Lerp(wx, w000, w100);
+    float x10 = Lerp(wx, w010, w110);
+    float x01 = Lerp(wx, w001, w101);
+    float x11 = Lerp(wx, w011, w111);
+    float y0 = Lerp(wy, x00, x10);
+    float y1 = Lerp(wy, x01, x11);
+    return Lerp(wz, y0, y1);
+}
+
+
+static float Turbulence(float x, float y, float z, uniform int octaves) {
+    float omega = 0.6;
+
+    float sum = 0., lambda = 1., o = 1.;
+    for (uniform int i = 0; i < octaves; ++i) {
+        sum += abs(o * Noise(lambda * x, lambda * y, lambda * z));
+        lambda *= 1.99f;
+        o *= omega;
+    }
+    return sum * 0.5;
+}
+
+
+export void noise_ispc(uniform float x0, uniform float y0, uniform float x1, 
+                       uniform float y1, uniform int width, uniform int height, 
+                       uniform float output[])
+{
+    uniform float dx = (x1 - x0) / width;
+    uniform float dy = (y1 - y0) / height;
+
+    for (uniform int j = 0; j < height; j++) {
+        for (uniform int i = 0; i < width; i += programCount) {
+            float x = x0 + (i + programIndex) * dx;
+            float y = y0 + j * dy;
+
+            int index = (j * width + i + programIndex);
+            output[index] = Turbulence(x, y, 0.6, 8);
+        }
+    }
+}
+

examples/noise/noise.vcxproj

@@ -0,0 +1,175 @@
+<?xml version="1.0" encoding="utf-8"?>
+<Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+  <ItemGroup Label="ProjectConfigurations">
+    <ProjectConfiguration Include="Debug|Win32">
+      <Configuration>Debug</Configuration>
+      <Platform>Win32</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Debug|x64">
+      <Configuration>Debug</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|Win32">
+      <Configuration>Release</Configuration>
+      <Platform>Win32</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|x64">
+      <Configuration>Release</Configuration>
+      <Platform>x64</Platform>
+    </ProjectConfiguration>
+  </ItemGroup>
+  <PropertyGroup Label="Globals">
+    <ProjectGuid>{0E0886D8-8B5E-4EAF-9A21-91E63DAF81FD}</ProjectGuid>
+    <Keyword>Win32Proj</Keyword>
+    <RootNamespace>noise</RootNamespace>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>true</UseDebugLibraries>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>true</UseDebugLibraries>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>false</UseDebugLibraries>
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="Configuration">
+    <ConfigurationType>Application</ConfigurationType>
+    <UseDebugLibraries>false</UseDebugLibraries>
+    <WholeProgramOptimization>true</WholeProgramOptimization>
+    <CharacterSet>Unicode</CharacterSet>
+  </PropertyGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
+  <ImportGroup Label="ExtensionSettings">
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <ImportGroup Label="PropertySheets" Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <ImportGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'" Label="PropertySheets">
+    <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" Label="LocalAppDataPlatform" />
+  </ImportGroup>
+  <PropertyGroup Label="UserMacros" />
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
+  </PropertyGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <Optimization>MaxSpeed</Optimization>
+      <FunctionLevelLinking>true</FunctionLevelLinking>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <OptimizeReferences>true</OptimizeReferences>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <ClCompile>
+      <WarningLevel>Level3</WarningLevel>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <Optimization>MaxSpeed</Optimization>
+      <FunctionLevelLinking>true</FunctionLevelLinking>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <OptimizeReferences>true</OptimizeReferences>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <ClCompile Include="noise.cpp" />
+    <ClCompile Include="noise_serial.cpp" />
+  </ItemGroup>
+  <ItemGroup>
+    <CustomBuild Include="noise.ispc">
+      <FileType>Document</FileType>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4,avx-x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4,avx-x2
+</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4,avx-x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4,avx-x2
+</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+    </CustomBuild>
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+  </ImportGroup>
+</Project>

examples/noise/noise_serial.cpp

@@ -0,0 +1,170 @@
+/*
+  Copyright (c) 2010-2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#include <math.h>
+
+#define NOISE_PERM_SIZE 256
+
+static int NoisePerm[2 * NOISE_PERM_SIZE] = {
+    151, 160, 137, 91, 90, 15, 131, 13, 201, 95, 96, 53, 194, 233, 7, 225, 140,
+    36, 103, 30, 69, 142, 8, 99, 37, 240, 21, 10, 23, 190, 6, 148, 247, 120,
+    234, 75, 0, 26, 197, 62, 94, 252, 219, 203, 117, 35, 11, 32, 57, 177, 33,
+    88, 237, 149, 56, 87, 174, 20, 125, 136, 171, 168,  68, 175, 74, 165, 71, 
+    134, 139, 48, 27, 166, 77, 146, 158, 231, 83, 111, 229, 122, 60, 211, 133, 
+    230, 220, 105, 92, 41, 55, 46, 245, 40, 244, 102, 143, 54, 65, 25, 63, 161,
+    1, 216, 80, 73, 209, 76, 132, 187, 208,  89, 18, 169, 200, 196, 135, 130, 
+    116, 188, 159, 86, 164, 100, 109, 198, 173, 186,  3, 64, 52, 217, 226, 250,
+    124, 123, 5, 202, 38, 147, 118, 126, 255, 82, 85, 212, 207, 206, 59, 227, 
+    47, 16, 58, 17, 182, 189, 28, 42, 223, 183, 170, 213, 119, 248, 152,  2, 44,
+    154, 163, 70, 221, 153, 101, 155, 167,  43, 172, 9, 129, 22, 39, 253,  19, 
+    98, 108, 110, 79, 113, 224, 232, 178, 185,  112, 104, 218, 246, 97, 228, 251,
+    34, 242, 193, 238, 210, 144, 12, 191, 179, 162, 241, 81, 51, 145, 235, 249,
+    14, 239, 107, 49, 192, 214,  31, 181, 199, 106, 157, 184, 84, 204, 176, 115,
+    121, 50, 45, 127,  4, 150, 254, 138, 236, 205, 93, 222, 114, 67, 29, 24, 72, 
+    243, 141, 128, 195, 78, 66, 215, 61, 156, 180, 151, 160, 137, 91, 90, 15,
+    131, 13, 201, 95, 96, 53, 194, 233, 7, 225, 140, 36, 103, 30, 69, 142, 8, 99,
+    37, 240, 21, 10, 23, 190, 6, 148, 247, 120, 234, 75, 0, 26, 197, 62, 94, 252,
+    219, 203, 117, 35, 11, 32, 57, 177, 33, 88, 237, 149, 56, 87, 174, 20, 125, 
+    136, 171, 168,  68, 175, 74, 165, 71, 134, 139, 48, 27, 166, 77, 146, 158,
+    231, 83, 111, 229, 122, 60, 211, 133, 230, 220, 105, 92, 41, 55, 46, 245,
+    40, 244, 102, 143, 54,  65, 25, 63, 161,  1, 216, 80, 73, 209, 76, 132, 187,
+    208,  89, 18, 169, 200, 196, 135, 130, 116, 188, 159, 86, 164, 100, 109, 
+    198, 173, 186,  3, 64, 52, 217, 226, 250, 124, 123, 5, 202, 38, 147, 118,
+    126, 255, 82, 85, 212, 207, 206, 59, 227, 47, 16, 58, 17, 182, 189, 28, 42,
+    223, 183, 170, 213, 119, 248, 152,  2, 44, 154, 163, 70, 221, 153, 101, 155, 
+    167,  43, 172, 9, 129, 22, 39, 253,  19, 98, 108, 110, 79, 113, 224, 232,
+    178, 185,  112, 104, 218, 246, 97, 228, 251, 34, 242, 193, 238, 210, 144,
+    12, 191, 179, 162, 241,  81, 51, 145, 235, 249, 14, 239, 107, 49, 192, 214,
+    31, 181, 199, 106, 157, 184,  84, 204, 176, 115, 121, 50, 45, 127,  4, 150,
+    254, 138, 236, 205, 93, 222, 114, 67, 29, 24, 72, 243, 141, 128, 195, 78, 
+    66, 215, 61, 156, 180
+};
+
+
+inline float Clamp(float v, float low, float high) {
+    return v < low ? low : ((v > high) ? high : v);
+}
+
+
+inline float SmoothStep(float low, float high, float value) {
+    float v = Clamp((value - low) / (high - low), 0.f, 1.f);
+    return v * v * (-2.f * v  + 3.f);
+}
+
+
+inline int Floor2Int(float val) {
+    return (int)floorf(val);
+}
+
+
+inline float Grad(int x, int y, int z, float dx, float dy, float dz) {
+    int h = NoisePerm[NoisePerm[NoisePerm[x]+y]+z];
+    h &= 15;
+    float u = h<8 || h==12 || h==13 ? dx : dy;
+    float v = h<4 || h==12 || h==13 ? dy : dz;
+    return ((h&1) ? -u : u) + ((h&2) ? -v : v);
+}
+
+
+inline float NoiseWeight(float t) {
+    float t3 = t*t*t;
+    float t4 = t3*t;
+    return 6.f*t4*t - 15.f*t4 + 10.f*t3;
+}
+
+
+inline float Lerp(float t, float low, float high) {
+    return (1.f - t) * low + t * high;
+}
+
+
+static float Noise(float x, float y, float z) {
+    // Compute noise cell coordinates and offsets
+    int ix = Floor2Int(x), iy = Floor2Int(y), iz = Floor2Int(z);
+    float dx = x - ix, dy = y - iy, dz = z - iz;
+
+    // Compute gradient weights
+    ix &= (NOISE_PERM_SIZE-1);
+    iy &= (NOISE_PERM_SIZE-1);
+    iz &= (NOISE_PERM_SIZE-1);
+    float w000 = Grad(ix,   iy,   iz,   dx,   dy,   dz);
+    float w100 = Grad(ix+1, iy,   iz,   dx-1, dy,   dz);
+    float w010 = Grad(ix,   iy+1, iz,   dx,   dy-1, dz);
+    float w110 = Grad(ix+1, iy+1, iz,   dx-1, dy-1, dz);
+    float w001 = Grad(ix,   iy,   iz+1, dx,   dy,   dz-1);
+    float w101 = Grad(ix+1, iy,   iz+1, dx-1, dy,   dz-1);
+    float w011 = Grad(ix,   iy+1, iz+1, dx,   dy-1, dz-1);
+    float w111 = Grad(ix+1, iy+1, iz+1, dx-1, dy-1, dz-1);
+
+    // Compute trilinear interpolation of weights
+    float wx = NoiseWeight(dx), wy = NoiseWeight(dy), wz = NoiseWeight(dz);
+    float x00 = Lerp(wx, w000, w100);
+    float x10 = Lerp(wx, w010, w110);
+    float x01 = Lerp(wx, w001, w101);
+    float x11 = Lerp(wx, w011, w111);
+    float y0 = Lerp(wy, x00, x10);
+    float y1 = Lerp(wy, x01, x11);
+    return Lerp(wz, y0, y1);
+}
+
+
+static float Turbulence(float x, float y, float z, int octaves) {
+    float omega = 0.6;
+
+    float sum = 0., lambda = 1., o = 1.;
+    for (int i = 0; i < octaves; ++i) {
+        sum += fabsf(o * Noise(lambda * x, lambda * y, lambda * z));
+        lambda *= 1.99f;
+        o *= omega;
+    }
+    return sum * 0.5f;
+}
+
+
+void noise_serial(float x0, float y0, float x1, float y1,
+                  int width, int height, float output[])
+{
+    float dx = (x1 - x0) / width;
+    float dy = (y1 - y0) / height;
+
+    for (int j = 0; j < height; j++) {
+        for (int i = 0; i < width; ++i) {
+            float x = x0 + i * dx;
+            float y = y0 + j * dy;
+
+            int index = (j * width + i);
+            output[index] = Turbulence(x, y, 0.6f, 8);
+        }
+    }
+}
+

examples/options/.gitignore

@@ -0,0 +1 @@
+options

examples/options/Makefile

@@ -1,8 +1,17 @@
 
-CXX=g++
+TASK_CXX=../tasksys.cpp
+TASK_LIB=-lpthread
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
+
+CXX=g++ -m64
 CXXFLAGS=-Iobjs/ -g -Wall
 ISPC=ispc
-ISPCFLAGS=-O2 --target=sse4x2
+ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx-x2 --arch=x86-64
+
+OBJS=objs/options.o objs/options_serial.o objs/options_ispc.o \
+	objs/options_ispc_sse2.o objs/options_ispc_sse4.o \
+	objs/options_ispc_avx.o $(TASK_OBJ)
 
 default: options
 
@@ -14,13 +23,16 @@ dirs:
 clean:
 	/bin/rm -rf objs *~ options
 
-options: dirs objs/options.o objs/options_serial.o objs/options_ispc.o
-	$(CXX) $(CXXFLAGS) -o $@ objs/options.o objs/options_ispc.o objs/options_serial.o -lm
+options: dirs $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm $(TASK_LIB)
 
 objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
 objs/options.o: objs/options_ispc.h options_defs.h
 
-objs/%_ispc.h objs/%_ispc.o: %.ispc options_defs.h
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc options_defs.h
 	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/options/options.cpp

@@ -31,15 +31,14 @@
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
 */
 
+#define NOMINMAX
+
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 #include <assert.h>
 #include <math.h>
 #include <algorithm>
-#ifndef __APPLE__
-#include <malloc.h>
-#endif // !__APPLE__
 using std::max;
 
 #include "options_defs.h"
@@ -48,23 +47,6 @@ using std::max;
 #include "options_ispc.h"
 using namespace ispc;
 
-// Allocate memory with 64-byte alignment.
-float *AllocFloats(int count) {
-    int size = count * sizeof(float);
-#if defined(_WIN32) || defined(_WIN64)
-    return (float *)_aligned_malloc(size, 64);
-#elif defined (__APPLE__)
-    // Allocate excess memory to ensure an aligned pointer can be returned
-    void *mem = malloc(size + (64-1) + sizeof(void*));
-    char *amem = ((char*)mem) + sizeof(void*);
-    amem += 64 - (reinterpret_cast<uint64_t>(amem) & (64 - 1));
-    ((void**)amem)[-1] = mem;
-    return (float *)amem;
-#else
-    return (float *)memalign(64, size);
-#endif
-}
-
 extern void black_scholes_serial(float Sa[], float Xa[], float Ta[], 
                                  float ra[], float va[], 
                                  float result[], int count);
@@ -73,17 +55,32 @@ extern void binomial_put_serial(float Sa[], float Xa[], float Ta[],
                                 float ra[], float va[], 
                                 float result[], int count);
 
-int main() {
-    // Pointers passed to ispc code must have alignment of the target's
-    // vector width at minimum.
-    float *S = AllocFloats(N_OPTIONS);
-    float *X = AllocFloats(N_OPTIONS);
-    float *T = AllocFloats(N_OPTIONS);
-    float *r = AllocFloats(N_OPTIONS);
-    float *v = AllocFloats(N_OPTIONS);
-    float *result = AllocFloats(N_OPTIONS);
+static void usage() {
+    printf("usage: options [--count=<num options>]\n");
+}
 
-    for (int i = 0; i < N_OPTIONS; ++i) {
+
+int main(int argc, char *argv[]) {
+    int nOptions = 128*1024;
+
+    for (int i = 1; i < argc; ++i) {
+        if (strncmp(argv[i], "--count=", 8) == 0) {
+            nOptions = atoi(argv[i] + 8);
+            if (nOptions <= 0) {
+                usage();
+                exit(1);
+            }
+        }
+    }
+
+    float *S = new float[nOptions];
+    float *X = new float[nOptions];
+    float *T = new float[nOptions];
+    float *r = new float[nOptions];
+    float *v = new float[nOptions];
+    float *result = new float[nOptions];
+
+    for (int i = 0; i < nOptions; ++i) {
         S[i] = 100;  // stock price
         X[i] = 98;   // option strike price
         T[i] = 2;    // time (years)
@@ -91,61 +88,109 @@ int main() {
         v[i] = 5;    // volatility
     }
 
+    double sum;
+
     //
     // Binomial options pricing model, ispc implementation
     //
-    reset_and_start_timer();
-    binomial_put_ispc(S, X, T, r, v, result, N_OPTIONS);
-    double binomial_ispc = get_elapsed_mcycles();
-    float sum = 0.f;
-    for (int i = 0; i < N_OPTIONS; ++i)
-        sum += result[i];
-    printf("[binomial ispc]:\t\t[%.3f] million cycles (avg %f)\n", 
-           binomial_ispc, sum / N_OPTIONS);
+    double binomial_ispc = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        binomial_put_ispc(S, X, T, r, v, result, nOptions);
+        double dt = get_elapsed_mcycles();
+        sum = 0.;
+        for (int i = 0; i < nOptions; ++i)
+            sum += result[i];
+        binomial_ispc = std::min(binomial_ispc, dt);
+    }
+    printf("[binomial ispc, 1 thread]:\t[%.3f] million cycles (avg %f)\n", 
+           binomial_ispc, sum / nOptions);
+
+    //
+    // Binomial options pricing model, ispc implementation, tasks
+    //
+    double binomial_tasks = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        binomial_put_ispc_tasks(S, X, T, r, v, result, nOptions);
+        double dt = get_elapsed_mcycles();
+        sum = 0.;
+        for (int i = 0; i < nOptions; ++i)
+            sum += result[i];
+        binomial_tasks = std::min(binomial_tasks, dt);
+    }
+    printf("[binomial ispc, tasks]:\t\t[%.3f] million cycles (avg %f)\n", 
+           binomial_tasks, sum / nOptions);
 
     //
     // Binomial options, serial implementation
     //
-    reset_and_start_timer();
-    binomial_put_serial(S, X, T, r, v, result, N_OPTIONS);
-    double binomial_serial = get_elapsed_mcycles();
-    sum = 0.f;
-    for (int i = 0; i < N_OPTIONS; ++i)
-        sum += result[i];
-    printf("[binomial serial]:\t\t[%.3f] million cycles (avg %f)\n", 
-           binomial_serial, sum / N_OPTIONS);
-
-    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", binomial_serial / binomial_ispc);
-
-    //
-    // Black-Scholes options pricing model, ispc implementation
-    //
-    sum = 0.f;
-    reset_and_start_timer();
-    for (int a = 0; a < N_BLACK_SCHOLES_ROUNDS; ++a) {
-        black_scholes_ispc(S, X, T, r, v, result, N_OPTIONS);
-        for (int i = 0; i < N_OPTIONS; ++i)
+    double binomial_serial = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        binomial_put_serial(S, X, T, r, v, result, nOptions);
+        double dt = get_elapsed_mcycles();
+        sum = 0.;
+        for (int i = 0; i < nOptions; ++i)
             sum += result[i];
+        binomial_serial = std::min(binomial_serial, dt);
     }
-    double bs_ispc = get_elapsed_mcycles();
-    printf("[black-scholes ispc]:\t\t[%.3f] million cycles (avg %f)\n", 
-           bs_ispc, sum / (N_BLACK_SCHOLES_ROUNDS * N_OPTIONS));
+    printf("[binomial serial]:\t\t[%.3f] million cycles (avg %f)\n", 
+           binomial_serial, sum / nOptions);
+
+    printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n",
+           binomial_serial / binomial_ispc, binomial_serial / binomial_tasks);
+
+    //
+    // Black-Scholes options pricing model, ispc implementation, 1 thread
+    //
+    double bs_ispc = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        black_scholes_ispc(S, X, T, r, v, result, nOptions);
+        double dt = get_elapsed_mcycles();
+        sum = 0.;
+        for (int i = 0; i < nOptions; ++i)
+            sum += result[i];
+        bs_ispc = std::min(bs_ispc, dt);
+    }
+    printf("[black-scholes ispc, 1 thread]:\t[%.3f] million cycles (avg %f)\n", 
+           bs_ispc, sum / nOptions);
+
+    //
+    // Black-Scholes options pricing model, ispc implementation, tasks
+    //
+    double bs_ispc_tasks = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        black_scholes_ispc_tasks(S, X, T, r, v, result, nOptions);
+        double dt = get_elapsed_mcycles();
+        sum = 0.;
+        for (int i = 0; i < nOptions; ++i)
+            sum += result[i];
+        bs_ispc_tasks = std::min(bs_ispc_tasks, dt);
+    }
+    printf("[black-scholes ispc, tasks]:\t[%.3f] million cycles (avg %f)\n", 
+           bs_ispc_tasks, sum / nOptions);
 
     //
     // Black-Scholes options pricing model, serial implementation
     //
-    sum = 0.f;
-    reset_and_start_timer();
-    for (int a = 0; a < N_BLACK_SCHOLES_ROUNDS; ++a) {
-        black_scholes_serial(S, X, T, r, v, result, N_OPTIONS);
-        for (int i = 0; i < N_OPTIONS; ++i)
+    double bs_serial = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        black_scholes_serial(S, X, T, r, v, result, nOptions);
+        double dt = get_elapsed_mcycles();
+        sum = 0.;
+        for (int i = 0; i < nOptions; ++i)
             sum += result[i];
+        bs_serial = std::min(bs_serial, dt);
     }
-    double bs_serial = get_elapsed_mcycles();
     printf("[black-scholes serial]:\t\t[%.3f] million cycles (avg %f)\n", bs_serial, 
-           sum / (N_BLACK_SCHOLES_ROUNDS * N_OPTIONS));
+           sum / nOptions);
 
-    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", bs_serial / bs_ispc);
+    printf("\t\t\t\t(%.2fx speedup from ISPC, %.2fx speedup from ISPC + tasks)\n", 
+           bs_serial / bs_ispc, bs_serial / bs_ispc_tasks);
 
     return 0;
 }

examples/options/options.ispc

@@ -55,49 +55,100 @@ CND(float X) {
     return w;
 }
 
-export void
-black_scholes_ispc(uniform float Sa[], uniform float Xa[], uniform float Ta[],
-                   uniform float ra[], uniform float va[], 
-                   uniform float result[], uniform int count) {
-    for (uniform int i = 0; i < count; i += programCount) {
-        float S = Sa[i + programIndex], X = Xa[i + programIndex];
-        float T = Ta[i + programIndex], r = ra[i + programIndex];
-        float v = va[i + programIndex];
+task void
+bs_task(uniform float Sa[], uniform float Xa[], uniform float Ta[],
+        uniform float ra[], uniform float va[], 
+        uniform float result[], uniform int count) {
+    uniform int first = taskIndex * (count/taskCount);
+    uniform int last = min(count, (int)((taskIndex+1) * (count/taskCount)));
+
+    foreach (i = first ... last) {
+        float S = Sa[i], X = Xa[i], T = Ta[i], r = ra[i], v = va[i];
 
         float d1 = (log(S/X) + (r + v * v * .5f) * T) / (v * sqrt(T));
         float d2 = d1 - v * sqrt(T);
 
-        result[i + programIndex] = S * CND(d1) - X * exp(-r * T) * CND(d2);
+        result[i] = S * CND(d1) - X * exp(-r * T) * CND(d2);
     }
 }
 
+export void
+black_scholes_ispc_tasks(uniform float Sa[], uniform float Xa[], uniform float Ta[],
+                         uniform float ra[], uniform float va[], 
+                         uniform float result[], uniform int count) {
+    uniform int nTasks = max((int)64, (int)count/16384);
+    launch[nTasks] < bs_task(Sa, Xa, Ta, ra, va, result, count) >;
+}
+
+
+export void
+black_scholes_ispc(uniform float Sa[], uniform float Xa[], uniform float Ta[],
+                   uniform float ra[], uniform float va[], 
+                   uniform float result[], uniform int count) {
+    foreach (i = 0 ... count) {
+        float S = Sa[i], X = Xa[i], T = Ta[i], r = ra[i], v = va[i];
+
+        float d1 = (log(S/X) + (r + v * v * .5f) * T) / (v * sqrt(T));
+        float d2 = d1 - v * sqrt(T);
+
+        result[i] = S * CND(d1) - X * exp(-r * T) * CND(d2);
+    }
+}
+
+
+static inline float
+binomial_put(float S, float X, float T, float r, float v) {
+    float V[BINOMIAL_NUM];
+
+    float dt = T / BINOMIAL_NUM;
+    float u = exp(v * sqrt(dt));
+    float d = 1. / u;
+    float disc = exp(r * dt);
+    float Pu = (disc - d) / (u - d);
+
+    for (uniform int j = 0; j < BINOMIAL_NUM; ++j) {
+        float upow = pow(u, (float)(2*j-BINOMIAL_NUM));
+        V[j] = max(0., X - S * upow);
+    }
+
+    for (uniform int j = BINOMIAL_NUM-1; j >= 0; --j)
+        for (uniform int k = 0; k < j; ++k)
+            V[k] = ((1 - Pu) * V[k] + Pu * V[k + 1]) / disc;
+    return V[0];
+}
+
 
 export void
 binomial_put_ispc(uniform float Sa[], uniform float Xa[], uniform float Ta[], 
                   uniform float ra[], uniform float va[], 
                   uniform float result[], uniform int count) {
-    float V[BINOMIAL_NUM];
-
-    for (uniform int i = 0; i < count; i += programCount) {
-        float S = Sa[i + programIndex], X = Xa[i + programIndex];
-        float T = Ta[i + programIndex], r = ra[i + programIndex];
-        float v = va[i + programIndex];
-
-        float dt = T / BINOMIAL_NUM;
-        float u = exp(v * sqrt(dt));
-        float d = 1. / u;
-        float disc = exp(r * dt);
-        float Pu = (disc - d) / (u - d);
-
-        for (uniform int j = 0; j < BINOMIAL_NUM; ++j) {
-            float upow = pow(u, (float)(2*j-BINOMIAL_NUM));
-            V[j] = max(0., X - S * upow);
-        }
-
-        for (uniform int j = BINOMIAL_NUM-1; j >= 0; --j)
-            for (uniform int k = 0; k < j; ++k)
-                V[k] = ((1 - Pu) * V[k] + Pu * V[k + 1]) / disc;
-
-        result[i + programIndex] = V[0];
+    foreach (i = 0 ... count) {
+        float S = Sa[i], X = Xa[i], T = Ta[i], r = ra[i], v = va[i];
+        result[i] = binomial_put(S, X, T, r, v);
     }
 }
+
+
+task void
+binomial_task(uniform float Sa[], uniform float Xa[], 
+              uniform float Ta[], uniform float ra[], 
+              uniform float va[], uniform float result[], 
+              uniform int count) {
+    uniform int first = taskIndex * (count/taskCount);
+    uniform int last = min(count, (int)((taskIndex+1) * (count/taskCount)));
+
+    foreach (i = first ... last) {
+        float S = Sa[i], X = Xa[i], T = Ta[i], r = ra[i], v = va[i];
+        result[i] = binomial_put(S, X, T, r, v);
+    }
+}
+
+
+export void
+binomial_put_ispc_tasks(uniform float Sa[], uniform float Xa[], 
+                        uniform float Ta[], uniform float ra[], 
+                        uniform float va[], uniform float result[], 
+                        uniform int count) {
+    uniform int nTasks = max((int)64, (int)count/16384);
+    launch[nTasks] < binomial_task(Sa, Xa, Ta, ra, va, result, count) >;
+}

examples/options/options.vcxproj

@@ -64,15 +64,19 @@
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -81,7 +85,10 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4305</DisableSpecificWarnings>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -95,7 +102,10 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4305</DisableSpecificWarnings>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -111,7 +121,9 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4305</DisableSpecificWarnings>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -129,7 +141,9 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4305</DisableSpecificWarnings>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -141,22 +155,23 @@
   <ItemGroup>
     <ClCompile Include="options.cpp" />
     <ClCompile Include="options_serial.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="options.ispc">
       <FileType>Document</FileType>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx-x2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj;%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <ItemGroup>

examples/options/options_defs.h

@@ -35,8 +35,6 @@
 #define OPTIONS_DEFS_H 1
 
 #define BINOMIAL_NUM 64
-#define N_OPTIONS 65536
-#define N_BLACK_SCHOLES_ROUNDS 20
 
 
 #endif // OPTIONS_DEFS_H

examples/options/options_serial.cpp

@@ -47,7 +47,7 @@ static inline float
 CND(float X) {
     float L = fabsf(X);
 
-    float k = 1.0 / (1.0 + 0.2316419 * L);
+    float k = 1.f / (1.f + 0.2316419f * L);
     float k2 = k*k;
     float k3 = k2*k;
     float k4 = k2*k2;
@@ -59,7 +59,7 @@ CND(float X) {
     w *= invSqrt2Pi * expf(-L * L * .5f);
 
     if (X > 0.f)
-        w = 1.0 - w;
+        w = 1.f - w;
     return w;
 }
 
@@ -94,7 +94,7 @@ binomial_put_serial(float Sa[], float Xa[], float Ta[],
 
         float dt = T / BINOMIAL_NUM;
         float u = expf(v * sqrtf(dt));
-        float d = 1. / u;
+        float d = 1.f / u;
         float disc = expf(r * dt);
         float Pu = (disc - d) / (u - d);
 

examples/rt/.gitignore

@@ -0,0 +1,2 @@
+rt
+*.ppm

examples/rt/Makefile

@@ -1,8 +1,17 @@
 
+ARCH = $(shell uname)
+
+TASK_CXX=../tasksys.cpp
+TASK_LIB=-lpthread
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
 CXX=g++
-CXXFLAGS=-Iobjs/ -O3 -Wall
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
 ISPC=ispc
-ISPCFLAGS=-O2 --target=sse4x2
+ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx --arch=x86-64
+
+OBJS=objs/rt.o objs/rt_serial.o $(TASK_OBJ) objs/rt_ispc.o objs/rt_ispc_sse2.o \
+	objs/rt_ispc_sse4.o objs/rt_ispc_avx.o
 
 default: rt
 
@@ -14,11 +23,16 @@ dirs:
 clean:
 	/bin/rm -rf objs *~ rt
 
-rt: dirs objs/rt.o objs/rt_serial.o objs/rt_ispc.o
-	$(CXX) $(CXXFLAGS) -o $@ objs/rt.o objs/rt_ispc.o objs/rt_serial.o -lm
+rt: dirs $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm $(TASK_LIB)
 
-objs/%.o: %.cpp objs/rt_ispc.h
+objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
-objs/%_ispc.h objs/%_ispc.o: %.ispc
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/rt.o: objs/rt_ispc.h 
+
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
 	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/rt/rt.cpp

@@ -42,10 +42,8 @@
 #include <math.h>
 #include <algorithm>
 #include <assert.h>
+#include <string.h>
 #include <sys/types.h>
-#ifndef __APPLE__
-#include <malloc.h>
-#endif
 #include "../timing.h"
 #include "rt_ispc.h"
 
@@ -53,24 +51,8 @@ using namespace ispc;
 
 typedef unsigned int uint;
 
-template <typename T> 
-T *AllocAligned(int count) {
-    int size = count * sizeof(T);
-#if defined(_WIN32) || defined(_WIN64)
-    return (T *)_aligned_malloc(size, 64);
-#elif defined (__APPLE__)
-    // Allocate excess memory to ensure an aligned pointer can be returned
-    void *mem = malloc(size + (64-1) + sizeof(void*));
-    char *amem = ((char*)mem) + sizeof(void*);
-    amem += 64 - (reinterpret_cast<uint64_t>(amem) & (64 - 1));
-    ((void**)amem)[-1] = mem;
-    return (T *)amem;
-#else
-    return (T *)memalign(64, size);
-#endif
-}
-
-extern void raytrace_serial(int width, int height, const float raster2camera[4][4], 
+extern void raytrace_serial(int width, int height, int baseWidth, int baseHeight,
+                            const float raster2camera[4][4], 
                             const float camera2world[4][4], float image[],
                             int id[], const LinearBVHNode nodes[],
                             const Triangle triangles[]);
@@ -109,14 +91,32 @@ static void writeImage(int *idImage, float *depthImage, int width, int height,
         }
     }            
     fclose(f);
+    printf("Wrote image file %s\n", filename);
+}
+
+
+static void usage() {
+    fprintf(stderr, "rt [--scale=<factor>] <scene name base>\n");
+    exit(1);
 }
 
 
 int main(int argc, char *argv[]) {
-    if (argc != 2) {
-        fprintf(stderr, "usage: rt <filename base>\n");
-        exit(1);
+    float scale = 1.f;
+    const char *filename = NULL;
+    for (int i = 1; i < argc; ++i) {
+        if (strncmp(argv[i], "--scale=", 8) == 0) {
+            scale = atof(argv[i] + 8);
+            if (scale == 0.f)
+                usage();
+        }
+        else if (filename != NULL)
+            usage();
+        else
+            filename = argv[i];
     }
+    if (filename == NULL)
+        usage();
 
 #define READ(var, n)                                            \
     if (fread(&(var), sizeof(var), n, f) != (unsigned int)n) {  \
@@ -128,10 +128,10 @@ int main(int argc, char *argv[]) {
     // Read the camera specification information from the camera file
     //
     char fnbuf[1024];
-    sprintf(fnbuf, "%s.camera", argv[1]);
+    sprintf(fnbuf, "%s.camera", filename);
     FILE *f = fopen(fnbuf, "rb");
     if (!f) {
-        perror(argv[1]);
+        perror(fnbuf);
         return 1;
     }
 
@@ -139,20 +139,20 @@ int main(int argc, char *argv[]) {
     // Nothing fancy, and trouble if we run on a big-endian system, just
     // fread in the bits
     //
-    int width, height;
+    int baseWidth, baseHeight;
     float camera2world[4][4], raster2camera[4][4];
-    READ(width, 1);
-    READ(height, 1);
+    READ(baseWidth, 1);
+    READ(baseHeight, 1);
     READ(camera2world[0][0], 16);
     READ(raster2camera[0][0], 16);
 
     //
     // Read in the serialized BVH 
     //
-    sprintf(fnbuf, "%s.bvh", argv[1]);
+    sprintf(fnbuf, "%s.bvh", filename);
     f = fopen(fnbuf, "rb");
     if (!f) {
-        perror(argv[2]);
+        perror(fnbuf);
         return 1;
     }
 
@@ -161,46 +161,46 @@ int main(int argc, char *argv[]) {
     uint nNodes;
     READ(nNodes, 1);
 
-    LinearBVHNode *nodes = AllocAligned<LinearBVHNode>(nNodes);
+    LinearBVHNode *nodes = new LinearBVHNode[nNodes];
     for (unsigned int i = 0; i < nNodes; ++i) {
         // Each node is 6x floats for a boox, then an integer for an offset
         // to the second child node, then an integer that encodes the type
         // of node, the total number of int it if a leaf node, etc.
         float b[6];
         READ(b[0], 6);
-        nodes[i].bounds[0].v[0] = b[0];
-        nodes[i].bounds[0].v[1] = b[1];
-        nodes[i].bounds[0].v[2] = b[2];
-        nodes[i].bounds[1].v[0] = b[3];
-        nodes[i].bounds[1].v[1] = b[4];
-        nodes[i].bounds[1].v[2] = b[5];
+        nodes[i].bounds[0][0] = b[0];
+        nodes[i].bounds[0][1] = b[1];
+        nodes[i].bounds[0][2] = b[2];
+        nodes[i].bounds[1][0] = b[3];
+        nodes[i].bounds[1][1] = b[4];
+        nodes[i].bounds[1][2] = b[5];
         READ(nodes[i].offset, 1);
-        READ(nodes[i].primsAxis, 1);
+        READ(nodes[i].nPrimitives, 1);
+        READ(nodes[i].splitAxis, 1);
+        READ(nodes[i].pad, 1);
     }
 
     // And then read the triangles 
     uint nTris;
     READ(nTris, 1);
-    Triangle *triangles = AllocAligned<Triangle>(nTris);
+    Triangle *triangles = new Triangle[nTris];
     for (uint i = 0; i < nTris; ++i) {
         // 9x floats for the 3 vertices
         float v[9];
         READ(v[0], 9);
         float *vp = v;
         for (int j = 0; j < 3; ++j) {
-            triangles[i].p[j].v[0] = *vp++;
-            triangles[i].p[j].v[1] = *vp++;
-            triangles[i].p[j].v[2] = *vp++;
+            triangles[i].p[j][0] = *vp++;
+            triangles[i].p[j][1] = *vp++;
+            triangles[i].p[j][2] = *vp++;
         }
         // And create an object id
         triangles[i].id = i+1;
     }
     fclose(f);
 
-    // round image resolution up to multiple of 4 to makethings easy for
-    // the code that assigns pixels to ispc program instances
-    height = (height + 3) & ~3;
-    width = (width + 3) & ~3;
+    int height = int(baseHeight * scale);
+    int width = int(baseWidth * scale);
 
     // allocate images; one to hold hit object ids, one to hold depth to
     // the first interseciton
@@ -208,19 +208,42 @@ int main(int argc, char *argv[]) {
     float *image = new float[width*height];
 
     //
-    // Run 3 iterations with ispc, record the minimum time
+    // Run 3 iterations with ispc + 1 core, record the minimum time
     //
     double minTimeISPC = 1e30;
     for (int i = 0; i < 3; ++i) {
         reset_and_start_timer();
-        raytrace(width, height, raster2camera, camera2world, 
-                 image, id, nodes, triangles);
+        raytrace_ispc(width, height, baseWidth, baseHeight, raster2camera, 
+                      camera2world, image, id, nodes, triangles);
         double dt = get_elapsed_mcycles();
         minTimeISPC = std::min(dt, minTimeISPC);
     }
-    printf("[rt ispc]:\t\t\t[%.3f] million cycles for %d x %d image\n", minTimeISPC, width, height);
+    printf("[rt ispc, 1 core]:\t\t[%.3f] million cycles for %d x %d image\n", 
+           minTimeISPC, width, height);
 
-    writeImage(id, image, width, height, "rt-ispc.ppm");
+    writeImage(id, image, width, height, "rt-ispc-1core.ppm");
+
+    memset(id, 0, width*height*sizeof(int));
+    memset(image, 0, width*height*sizeof(float));
+
+    //
+    // Run 3 iterations with ispc + 1 core, record the minimum time
+    //
+    double minTimeISPCtasks = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        raytrace_ispc_tasks(width, height, baseWidth, baseHeight, raster2camera,
+                            camera2world, image, id, nodes, triangles);
+        double dt = get_elapsed_mcycles();
+        minTimeISPCtasks = std::min(dt, minTimeISPCtasks);
+    }
+    printf("[rt ispc + tasks]:\t\t[%.3f] million cycles for %d x %d image\n", 
+           minTimeISPCtasks, width, height);
+
+    writeImage(id, image, width, height, "rt-ispc-tasks.ppm");
+
+    memset(id, 0, width*height*sizeof(int));
+    memset(image, 0, width*height*sizeof(float));
 
     //
     // And 3 iterations with the serial implementation, reporting the
@@ -229,14 +252,15 @@ int main(int argc, char *argv[]) {
     double minTimeSerial = 1e30;
     for (int i = 0; i < 3; ++i) {
         reset_and_start_timer();
-        raytrace_serial(width, height, raster2camera, camera2world, 
-                        image, id, nodes, triangles);
+        raytrace_serial(width, height, baseWidth, baseHeight, raster2camera, 
+                        camera2world, image, id, nodes, triangles);
         double dt = get_elapsed_mcycles();
         minTimeSerial = std::min(dt, minTimeSerial);
     }
     printf("[rt serial]:\t\t\t[%.3f] million cycles for %d x %d image\n", 
            minTimeSerial, width, height);
-    printf("\t\t\t\t(%.2fx speedup from ISPC)\n", minTimeSerial / minTimeISPC);
+    printf("\t\t\t\t(%.2fx speedup from ISPC, %.2f from ISPC + tasks)\n", 
+           minTimeSerial / minTimeISPC, minTimeSerial / minTimeISPCtasks);
 
     writeImage(id, image, width, height, "rt-serial.ppm");
 

examples/rt/rt.ispc

@@ -43,28 +43,19 @@ struct Ray {
 };
 
 struct Triangle {
-    uniform float3 p[3];
+    uniform float p[3][4];
     uniform int id;
+    uniform int pad[3];
 };
 
 struct LinearBVHNode {
-    uniform float3 bounds[2];
+    uniform float bounds[2][3];
     uniform unsigned int offset;     // num primitives for leaf, second child for interior
-    uniform unsigned int primsAxis;  // 0:7 nPrimitives, 8:15 split axis, 16:31 padding
+    uniform unsigned int8 nPrimitives;
+    uniform unsigned int8 splitAxis;
+    uniform unsigned int16 pad;
 };
 
-static inline uniform int nPrims(const reference LinearBVHNode node) {
-    return (node.primsAxis & 0xff);
-}
-
-static inline uniform int axis(const reference LinearBVHNode node) {
-    return ((node.primsAxis >> 8) & 0xff);
-}
-
-static inline uniform bool isInterior(const reference LinearBVHNode node) {
-    return nPrims(node) == 0;
-}
-
 static inline float3 Cross(const float3 v1, const float3 v2) {
     float v1x = v1.x, v1y = v1.y, v1z = v1.z;
     float v2x = v2.x, v2y = v2.y, v2z = v2.z;
@@ -82,7 +73,7 @@ static inline float Dot(const float3 a, const float3 b) {
 
 static void generateRay(uniform const float raster2camera[4][4], 
                         uniform const float camera2world[4][4],
-                        float x, float y, reference Ray ray) {
+                        float x, float y, Ray &ray) {
     ray.mint = 0.f;
     ray.maxt = 1e30f;
 
@@ -113,14 +104,16 @@ static void generateRay(uniform const float raster2camera[4][4],
 }
 
 
-static inline bool BBoxIntersect(const reference uniform float3 bounds[2], 
-                                 const reference Ray ray) {
+static inline bool BBoxIntersect(const uniform float bounds[2][3], 
+                                 const Ray &ray) {
+    uniform float3 bounds0 = { bounds[0][0], bounds[0][1], bounds[0][2] };
+    uniform float3 bounds1 = { bounds[1][0], bounds[1][1], bounds[1][2] };
     float t0 = ray.mint, t1 = ray.maxt;
 
     // Check all three axis-aligned slabs.  Don't try to early out; it's
     // not worth the trouble
-    float3 tNear = (bounds[0] - ray.origin) * ray.invDir;
-    float3 tFar  = (bounds[1] - ray.origin) * ray.invDir;
+    float3 tNear = (bounds0 - ray.origin) * ray.invDir;
+    float3 tFar  = (bounds1 - ray.origin) * ray.invDir;
     if (tNear.x > tFar.x) {
         float tmp = tNear.x;
         tNear.x = tFar.x;
@@ -150,9 +143,12 @@ static inline bool BBoxIntersect(const reference uniform float3 bounds[2],
 
 
 
-static inline bool TriIntersect(const reference Triangle tri, reference Ray ray) {
-    uniform float3 e1 = tri.p[1] - tri.p[0];
-    uniform float3 e2 = tri.p[2] - tri.p[0];
+static inline bool TriIntersect(const Triangle &tri, Ray &ray) {
+    uniform float3 p0 = { tri.p[0][0], tri.p[0][1], tri.p[0][2] };
+    uniform float3 p1 = { tri.p[1][0], tri.p[1][1], tri.p[1][2] };
+    uniform float3 p2 = { tri.p[2][0], tri.p[2][1], tri.p[2][2] };
+    uniform float3 e1 = p1 - p0;
+    uniform float3 e2 = p2 - p0;
 
     float3 s1 = Cross(ray.dir, e2);
     float divisor = Dot(s1, e1);
@@ -163,7 +159,7 @@ static inline bool TriIntersect(const reference Triangle tri, reference Ray ray)
     float invDivisor = 1.f / divisor;
 
     // Compute first barycentric coordinate
-    float3 d = ray.origin - tri.p[0];
+    float3 d = ray.origin - p0;
     float b1 = Dot(d, s1) * invDivisor;
     if (b1 < 0. || b1 > 1.)
         hit = false;
@@ -188,7 +184,7 @@ static inline bool TriIntersect(const reference Triangle tri, reference Ray ray)
 
 
 bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[], 
-                  reference Ray r) {
+                  Ray &r) {
     Ray ray = r;
     bool hit = false;
     // Follow ray through BVH nodes to find primitive intersections
@@ -199,7 +195,7 @@ bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[],
         // Check ray against BVH node
         LinearBVHNode node = nodes[nodeNum];
         if (any(BBoxIntersect(node.bounds, ray))) {
-            uniform unsigned int nPrimitives = nPrims(node);
+            uniform unsigned int nPrimitives = node.nPrimitives;
             if (nPrimitives > 0) {
                 // Intersect ray with primitives in leaf BVH node
                 uniform unsigned int primitivesOffset = node.offset;
@@ -213,7 +209,7 @@ bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[],
             }
             else {
                 // Put far BVH node on _todo_ stack, advance to near node
-                if (r.dirIsNeg[axis(node)]) {
+                if (r.dirIsNeg[node.splitAxis]) {
                    todo[todoOffset++] = nodeNum + 1;
                    nodeNum = node.offset;
                 }
@@ -236,38 +232,77 @@ bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[],
 }
 
 
-export void raytrace(uniform int width, uniform int height,
-                     const uniform float raster2camera[4][4], 
-                     const uniform float camera2world[4][4],
-                     uniform float image[], uniform int id[],
-                     const LinearBVHNode nodes[],
-                     const Triangle triangles[]) {
-    static const uniform float udx[16] = { 0, 1, 0, 1, 2, 3, 2, 3, 
-                                           0, 1, 0, 1, 2, 3, 2, 3 };
-    static const uniform float udy[16] = { 0, 0, 1, 1, 0, 0, 1, 1, 
-                                           2, 2, 3, 3, 2, 2, 3, 3 };
+static void raytrace_tile(uniform int x0, uniform int x1,
+                          uniform int y0, uniform int y1, 
+                          uniform int width, uniform int height,
+                          uniform int baseWidth, uniform int baseHeight,
+                          const uniform float raster2camera[4][4], 
+                          const uniform float camera2world[4][4],
+                          uniform float image[], uniform int id[],
+                          const LinearBVHNode nodes[],
+                          const Triangle triangles[]) {
+    uniform float widthScale = (float)(baseWidth) / (float)(width);
+    uniform float heightScale = (float)(baseHeight) / (float)(height);
 
-    // The outer loops are always over blocks of 4x4 pixels
-    for (uniform int y = 0; y < height; y += 4) {
-        for (uniform int x = 0; x < width; x += 4) {
-            // Now we have a block of 4x4=16 pixels to process; it will
-            // take 16/programCount iterations of this loop to process
-            // them.
-            for (uniform int o = 0; o < 16 / programCount; ++o) {
-                // Map program instances to samples in the udx/udy arrays
-                // to figure out which pixel each program instance is
-                // responsible for
-                const float dx = udx[o * programCount + programIndex];
-                const float dy = udy[o * programCount + programIndex];
+    foreach_tiled (y = y0 ... y1, x = x0 ... x1) {
+        Ray ray;
+        generateRay(raster2camera, camera2world, x*widthScale,
+                    y*heightScale, ray);
+        BVHIntersect(nodes, triangles, ray);
 
-                Ray ray;
-                generateRay(raster2camera, camera2world, x+dx, y+dy, ray);
-                BVHIntersect(nodes, triangles, ray);
-
-                int offset = (y + (int)dy) * width + (x + (int)dx);
-                image[offset] = ray.maxt;
-                id[offset] = ray.hitId;
-            }
-        }
+        int offset = y * width + x;
+        image[offset] = ray.maxt;
+        id[offset] = ray.hitId;
     }
 }
+
+
+export void raytrace_ispc(uniform int width, uniform int height,
+                          uniform int baseWidth, uniform int baseHeight,
+                          const uniform float raster2camera[4][4], 
+                          const uniform float camera2world[4][4],
+                          uniform float image[], uniform int id[],
+                          const LinearBVHNode nodes[],
+                          const Triangle triangles[]) {
+    raytrace_tile(0, width, 0, height, width, height, baseWidth, baseHeight,
+                  raster2camera, camera2world, image,
+                  id, nodes, triangles);
+}
+
+
+task void raytrace_tile_task(uniform int width, uniform int height,
+                             uniform int baseWidth, uniform int baseHeight,
+                             const uniform float raster2camera[4][4], 
+                             const uniform float camera2world[4][4],
+                             uniform float image[], uniform int id[],
+                             const LinearBVHNode nodes[],
+                             const Triangle triangles[]) {
+    uniform int dx = 16, dy = 16; // must match dx, dy below
+    uniform int xBuckets = (width + (dx-1)) / dx;
+    uniform int x0 = (taskIndex % xBuckets) * dx;
+    uniform int x1 = min(x0 + dx, width);
+    uniform int y0 = (taskIndex / xBuckets) * dy;
+    uniform int y1 = min(y0 + dy, height);
+                             
+    raytrace_tile(x0, x1, y0, y1, width, height, baseWidth, baseHeight, 
+                  raster2camera, camera2world, image,
+                  id, nodes, triangles);
+}
+
+
+export void raytrace_ispc_tasks(uniform int width, uniform int height,
+                                uniform int baseWidth, uniform int baseHeight,
+                                const uniform float raster2camera[4][4], 
+                                const uniform float camera2world[4][4],
+                                uniform float image[], uniform int id[],
+                                const LinearBVHNode nodes[],
+                                const Triangle triangles[]) {
+    uniform int dx = 16, dy = 16;
+    uniform int xBuckets = (width + (dx-1)) / dx;
+    uniform int yBuckets = (height + (dy-1)) / dy;
+    uniform int nTasks = xBuckets * yBuckets;
+    launch[nTasks] < raytrace_tile_task(width, height, baseWidth, baseHeight, 
+                                        raster2camera, camera2world, 
+                                        image, id, nodes, triangles) >;
+}
+

examples/rt/rt.vcxproj

@@ -64,15 +64,19 @@
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -81,6 +85,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -94,6 +101,9 @@
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <IntrinsicFunctions>true</IntrinsicFunctions>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -109,6 +119,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -126,6 +138,8 @@
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
+      <FloatingPointModel>Fast</FloatingPointModel>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -138,26 +152,27 @@
     <CustomBuild Include="rt.ispc">
       <FileType>Document</FileType>
       <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx
 </Command>
       <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
       <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2,sse4-x2,avx
 </Command>
       <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2,sse4-x2,avx
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_sse2.obj;$(TargetDir)%(Filename)_sse4.obj;$(TargetDir)%(Filename)_avx.obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="rt.cpp" />
     <ClCompile Include="rt_serial.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

examples/rt/rt_serial.cpp

@@ -39,6 +39,7 @@
 #endif
 
 #include <algorithm>
+#include <stdint.h>
 
 // Just enough of a float3 class to do what we need in this file.
 #ifdef _MSC_VER
@@ -74,31 +75,22 @@ struct Ray {
 // Declare these in a namespace so the mangling matches
 namespace ispc {
     struct Triangle {
-        float3 p[3];
-        int id;
+        float p[3][4]; // extra float pad after each vertex
+        int32_t id;
+        int32_t pad[3]; // make 16 x 32-bits
     };
 
     struct LinearBVHNode {
-        float3 bounds[2];
-        unsigned int offset;     // primitives for leaf, second child for interior
-        unsigned int primsAxis;  // 0:7 nPrimitives, 8:15 split axis, 16:31 padding
+        float bounds[2][3];
+        int32_t offset;     // primitives for leaf, second child for interior
+        uint8_t nPrimitives;
+        uint8_t splitAxis;
+        uint16_t pad;
     };
 }
 
 using namespace ispc;
 
-inline int nPrims(const LinearBVHNode &node) {
-    return (node.primsAxis & 0xff);
-}
-
-inline int axis(const LinearBVHNode &node) {
-    return ((node.primsAxis >> 8) & 0xff);
-}
-
-inline bool isInterior(const LinearBVHNode &node) {
-    return nPrims(node) == 0;
-}
-
 inline float3 Cross(const float3 &v1, const float3 &v2) {
     float v1x = v1.x, v1y = v1.y, v1z = v1.z;
     float v2x = v2.x, v2y = v2.y, v2z = v2.z;
@@ -149,12 +141,14 @@ static void generateRay(const float raster2camera[4][4],
 }
 
 
-static inline bool BBoxIntersect(const float3 bounds[2], 
+static inline bool BBoxIntersect(const float bounds[2][3], 
                                  const Ray &ray) {
+    float3 bounds0(bounds[0][0], bounds[0][1], bounds[0][2]);
+    float3 bounds1(bounds[1][0], bounds[1][1], bounds[1][2]);
     float t0 = ray.mint, t1 = ray.maxt;
 
-    float3 tNear = (bounds[0] - ray.origin) * ray.invDir;
-    float3 tFar  = (bounds[1] - ray.origin) * ray.invDir;
+    float3 tNear = (bounds0 - ray.origin) * ray.invDir;
+    float3 tFar  = (bounds1 - ray.origin) * ray.invDir;
     if (tNear.x > tFar.x) {
         float tmp = tNear.x;
         tNear.x = tFar.x;
@@ -185,8 +179,11 @@ static inline bool BBoxIntersect(const float3 bounds[2],
 
 
 inline bool TriIntersect(const Triangle &tri, Ray &ray) {
-    float3 e1 = tri.p[1] - tri.p[0];
-    float3 e2 = tri.p[2] - tri.p[0];
+    float3 p0(tri.p[0][0], tri.p[0][1], tri.p[0][2]);
+    float3 p1(tri.p[1][0], tri.p[1][1], tri.p[1][2]);
+    float3 p2(tri.p[2][0], tri.p[2][1], tri.p[2][2]);
+    float3 e1 = p1 - p0;
+    float3 e2 = p2 - p0;
 
     float3 s1 = Cross(ray.dir, e2);
     float divisor = Dot(s1, e1);
@@ -196,7 +193,7 @@ inline bool TriIntersect(const Triangle &tri, Ray &ray) {
     float invDivisor = 1.f / divisor;
 
     // Compute first barycentric coordinate
-    float3 d = ray.origin - tri.p[0];
+    float3 d = ray.origin - p0;
     float b1 = Dot(d, s1) * invDivisor;
     if (b1 < 0. || b1 > 1.)
         return false;
@@ -230,7 +227,7 @@ bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[],
         // Check ray against BVH node
         const LinearBVHNode &node = nodes[nodeNum];
         if (BBoxIntersect(node.bounds, ray)) {
-            unsigned int nPrimitives = nPrims(node);
+            unsigned int nPrimitives = node.nPrimitives;
             if (nPrimitives > 0) {
                 // Intersect ray with primitives in leaf BVH node
                 unsigned int primitivesOffset = node.offset;
@@ -244,7 +241,7 @@ bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[],
             }
             else {
                 // Put far BVH node on _todo_ stack, advance to near node
-                if (r.dirIsNeg[axis(node)]) {
+                if (r.dirIsNeg[node.splitAxis]) {
                    todo[todoOffset++] = nodeNum + 1;
                    nodeNum = node.offset;
                 }
@@ -267,17 +264,21 @@ bool BVHIntersect(const LinearBVHNode nodes[], const Triangle tris[],
 }
 
 
-void raytrace_serial(int width, int height,
+void raytrace_serial(int width, int height, int baseWidth, int baseHeight,
                      const float raster2camera[4][4], 
                      const float camera2world[4][4],
                      float image[],
                      int id[],
                      const LinearBVHNode nodes[],
                      const Triangle triangles[]) {
+    float widthScale = float(baseWidth) / float(width);
+    float heightScale = float(baseHeight) / float(height);
+
     for (int y = 0; y < height; ++y) {
         for (int x = 0; x < width; ++x) {
                 Ray ray;
-                generateRay(raster2camera, camera2world, x, y, ray);
+                generateRay(raster2camera, camera2world, x * widthScale,
+                            y * heightScale, ray);
                 BVHIntersect(nodes, triangles, ray);
 
                 int offset = y * width + x;

examples/simple/.gitignore

@@ -0,0 +1,2 @@
+simple
+objs

examples/simple/Makefile

@@ -1,8 +1,8 @@
 
-CXX=g++
+CXX=g++ -m64
 CXXFLAGS=-Iobjs/ -O3 -Wall
 ISPC=ispc
-ISPCFLAGS=-O2
+ISPCFLAGS=-O2 --arch=x86-64 --target=sse2
 
 default: simple
 

examples/simple/simple.cpp

@@ -32,21 +32,14 @@
 */
 
 #include <stdio.h>
+#include <stdlib.h>
 
 // Include the header file that the ispc compiler generates
 #include "simple_ispc.h"
 using namespace ispc;
 
 int main() {
-    // Pointers passed to ispc-compiled code are currently required to have
-    // alignment equal to the target's native vector size.  Here we align
-    // to 32 bytes to be safe for both SSE and AVX targets.
-#ifdef _MSC_VER
-    __declspec(align(32)) float vin[16], vout[16];
-#else
-    float vin[16] __attribute__((aligned(32)));
-    float vout[16] __attribute__((aligned(32)));
-#endif
+    float vin[16], vout[16];
 
     // Initialize input buffer
     for (int i = 0; i < 16; ++i)

examples/simple/simple.ispc

@@ -34,9 +34,7 @@
 
 export void simple(uniform float vin[], uniform float vout[], 
                    uniform int count) {
-    // Compute the result for 'programCount' values in parallel
-    for (uniform int i = 0; i < count; i += programCount) {
-        int index = i + programIndex;
+    foreach (index = 0 ... count) {
         // Load the appropriate input value for this program instance.
         float v = vin[index];
 

examples/simple/simple.vcxproj

@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project DefaultTargets="Build" ToolsVersion="4.0" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <ItemGroup Label="ProjectConfigurations">
     <ProjectConfiguration Include="Debug|Win32">
@@ -25,21 +25,21 @@
     <CustomBuild Include="simple.ispc">
       <FileType>Document</FileType>
       <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2
 </Command>
       <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
-cl /E /TP %(Filename).ispc | ispc -O2 -o %(Filename).obj -h %(Filename)_ispc.h
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
       <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --arch=x86 --target=sse2
 </Command>
       <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
-cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
+ispc -O2 %(Filename).ispc -o $(TargetDir)%(Filename).obj -h $(TargetDir)%(Filename)_ispc.h --target=sse2
 </Command>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">$(TargetDir)%(Filename).obj;$(TargetDir)%(Filename)_ispc.h</Outputs>
     </CustomBuild>
   </ItemGroup>
   <PropertyGroup Label="Globals">
@@ -88,15 +88,19 @@ cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
     <LinkIncremental>true</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
     <LinkIncremental>false</LinkIncremental>
+    <ExecutablePath>$(ProjectDir)..\..;$(ExecutablePath)</ExecutablePath>
   </PropertyGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
     <ClCompile>
@@ -105,6 +109,7 @@ cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -118,6 +123,7 @@ cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
       <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -133,6 +139,7 @@ cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -150,6 +157,7 @@ cl /E /TP %(Filename).ispc | ispc -O2 - -o %(Filename).obj -h %(Filename)_ispc.h
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <AdditionalIncludeDirectories>$(TargetDir)</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>

examples/stencil/.gitignore

@@ -0,0 +1,2 @@
+stencil
+objs

examples/stencil/Makefile

@@ -0,0 +1,39 @@
+
+ARCH = $(shell uname)
+
+TASK_CXX=../tasksys.cpp
+TASK_LIB=-lpthread
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
+CXX=g++
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
+ISPC=ispc
+ISPCFLAGS=-O2 --target=sse2,sse4-x2,avx --arch=x86-64
+
+OBJS=objs/stencil.o objs/stencil_serial.o $(TASK_OBJ) objs/stencil_ispc.o \
+	objs/stencil_ispc_sse2.o objs/stencil_ispc_sse4.o \
+	objs/stencil_ispc_avx.o
+
+default: stencil
+
+.PHONY: dirs clean
+
+dirs:
+	/bin/mkdir -p objs/
+
+clean:
+	/bin/rm -rf objs *~ stencil
+
+stencil: dirs $(OBJS)
+	$(CXX) $(CXXFLAGS) -o $@ $(OBJS) -lm $(TASK_LIB)
+
+objs/%.o: %.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/stencil.o: objs/stencil_ispc.h 
+
+objs/%_ispc.h objs/%_ispc.o objs/%_ispc_sse2.o objs/%_ispc_sse4.o objs/%_ispc_avx.o: %.ispc
+	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/stencil/stencil.cpp

@@ -0,0 +1,151 @@
+/*
+  Copyright (c) 2010-2011, Intel Corporation
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are
+  met:
+
+    * Redistributions of source code must retain the above copyright
+      notice, this list of conditions and the following disclaimer.
+
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+
+    * Neither the name of Intel Corporation nor the names of its
+      contributors may be used to endorse or promote products derived from
+      this software without specific prior written permission.
+
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
+   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
+   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  
+*/
+
+#ifdef _MSC_VER
+#define _CRT_SECURE_NO_WARNINGS
+#define NOMINMAX
+#pragma warning (disable: 4244)
+#pragma warning (disable: 4305)
+#endif
+
+#include <stdio.h>
+#include <algorithm>
+#include <math.h>
+#include "../timing.h"
+#include "stencil_ispc.h"
+using namespace ispc;
+
+
+extern void loop_stencil_serial(int t0, int t1, int x0, int x1,
+                                int y0, int y1, int z0, int z1,
+                                int Nx, int Ny, int Nz,
+                                const float coef[5], 
+                                const float vsq[],
+                                float Aeven[], float Aodd[]);
+
+
+void InitData(int Nx, int Ny, int Nz, float *A[2], float *vsq) {
+    int offset = 0;
+    for (int z = 0; z < Nz; ++z)
+        for (int y = 0; y < Ny; ++y)
+            for (int x = 0; x < Nx; ++x, ++offset) {
+                A[0][offset] = (x < Nx / 2) ? x / float(Nx) : y / float(Ny);
+                A[1][offset] = 0;
+                vsq[offset] = x*y*z / float(Nx * Ny * Nz);
+            }
+}
+
+
+int main() {
+    int Nx = 256, Ny = 256, Nz = 256;
+    int width = 4;
+    float *Aserial[2], *Aispc[2];
+    Aserial[0] = new float [Nx * Ny * Nz];
+    Aserial[1] = new float [Nx * Ny * Nz];
+    Aispc[0] = new float [Nx * Ny * Nz];
+    Aispc[1] = new float [Nx * Ny * Nz];
+    float *vsq = new float [Nx * Ny * Nz];
+
+    float coeff[4] = { 0.5, -.25, .125, -.0625 }; 
+
+    InitData(Nx, Ny, Nz, Aispc, vsq);
+
+    //
+    // Compute the image using the ispc implementation on one core; report
+    // the minimum time of three runs.
+    //
+    double minTimeISPC = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        loop_stencil_ispc(0, 6, width, Nx - width, width, Ny - width,
+                          width, Nz - width, Nx, Ny, Nz, coeff, vsq,
+                          Aispc[0], Aispc[1]);
+        double dt = get_elapsed_mcycles();
+        minTimeISPC = std::min(minTimeISPC, dt);
+    }
+
+    printf("[stencil ispc 1 core]:\t\t[%.3f] million cycles\n", minTimeISPC);
+
+    InitData(Nx, Ny, Nz, Aispc, vsq);
+
+    //
+    // Compute the image using the ispc implementation with tasks; report
+    // the minimum time of three runs.
+    //
+    double minTimeISPCTasks = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        loop_stencil_ispc_tasks(0, 6, width, Nx - width, width, Ny - width,
+                                width, Nz - width, Nx, Ny, Nz, coeff, vsq,
+                                Aispc[0], Aispc[1]);
+        double dt = get_elapsed_mcycles();
+        minTimeISPCTasks = std::min(minTimeISPCTasks, dt);
+    }
+
+    printf("[stencil ispc + tasks]:\t\t[%.3f] million cycles\n", minTimeISPCTasks);
+
+    InitData(Nx, Ny, Nz, Aserial, vsq);
+
+    // 
+    // And run the serial implementation 3 times, again reporting the
+    // minimum time.
+    //
+    double minTimeSerial = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        loop_stencil_serial(0, 6, width, Nx-width, width, Ny - width,
+                            width, Nz - width, Nx, Ny, Nz, coeff, vsq,
+                            Aserial[0], Aserial[1]);
+        double dt = get_elapsed_mcycles();
+        minTimeSerial = std::min(minTimeSerial, dt);
+    }
+
+    printf("[stencil serial]:\t\t[%.3f] millon cycles\n", minTimeSerial);
+
+    printf("\t\t\t\t(%.2fx speedup from ISPC, %.2f from ISPC + tasks)\n", 
+           minTimeSerial / minTimeISPC, minTimeSerial / minTimeISPCTasks);
+
+    // Check for agreement
+    int offset = 0;
+    for (int z = 0; z < Nz; ++z)
+        for (int y = 0; y < Ny; ++y)
+            for (int x = 0; x < Nx; ++x, ++offset) {
+                float error = fabsf((Aserial[1][offset] - Aispc[1][offset]) /
+                                    Aserial[1][offset]);
+                if (error > 1e-4)
+                    printf("Error @ (%d,%d,%d): ispc = %f, serial = %f\n",
+                           x, y, z, Aispc[1][offset], Aserial[1][offset]);
+            }
+
+    return 0;
+}