Release notes and doxygen bump for v1.0.10

Within each function that launches tasks, we now can easily track which
tasks that function launched, so that the sync at the end of the function
can just sync on the tasks launched by that function (not all tasks
launched by all functions.)

Implementing this led to a rework of the task system API that ispc generates
code to call; the example task systems in examples/tasksys.cpp have been
updated to conform to this API.  (The updated API is also documented in
the ispc user's guide.)

As part of this, "launch[n]" syntax was added to launch a number of tasks
in a single launch statement, rather than requiring a loop over 'n' to
launch n tasks.

This commit thus fixes issue #84 (enhancement to launch multiple tasks from
a single launch statement) as well as issue #105 (recursive task launches
were broken).

Makefile

@@ -10,9 +10,15 @@ CLANG_LIBS = -lclangFrontend -lclangDriver \
              -lclangSerialization -lclangParse -lclangSema \
              -lclangAnalysis -lclangAST -lclangLex -lclangBasic
 
-LLVM_LIBS=$(shell llvm-config --ldflags --libs) -lpthread -ldl
+ISPC_LIBS=$(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 --abbrev-commit --abbrev=16 | head -1)
@@ -43,12 +49,14 @@ CXX_SRC=builtins.cpp ctx.cpp decl.cpp expr.cpp ispc.cpp \
 	util.cpp
 HEADERS=builtins.h ctx.h decl.h expr.h ispc.h llvmutil.h module.h \
 	opt.h stmt.h sym.h type.h util.h
-BUILTINS_SRC=builtins-avx.ll builtins-sse2.ll builtins-sse4.ll builtins-sse4x2.ll
+BUILTINS_SRC=builtins-avx.ll builtins-avx-x2.ll builtins-sse2.ll \
+	builtins-sse4.ll builtins-sse4x2.ll
 BISON_SRC=parse.yy
 FLEX_SRC=lex.ll
 
-OBJS=$(addprefix objs/, $(CXX_SRC:.cpp=.o) $(BUILTINS_SRC:.ll=.o) builtins-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
 
@@ -77,11 +85,11 @@ doxygen:
 
 ispc: print_llvm_src dirs $(OBJS)
 	@echo Creating ispc executable
-	@$(CXX) $(LDFLAGS) -o $@ $(OBJS) $(CLANG_LIBS) $(LLVM_LIBS)
+	@$(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 $<
@@ -103,19 +111,27 @@ objs/lex.o: objs/lex.cpp $(HEADERS) objs/parse.cc
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
 
-objs/builtins-%.cpp: builtins-%.ll builtins.m4 builtins-sse.ll
+objs/builtins-%.cpp: builtins-%.ll builtins.m4 builtins-sse.ll builtins-avx-common.ll
 	@echo Creating C++ source from builtin definitions file $<
-	@m4 builtins.m4 $< | ./bitcode2cpp.py $< > $@
+	@m4 -DLLVM_VERSION=$(LLVM_VERSION) builtins.m4 $< | ./bitcode2cpp.py $< > $@
 
 objs/builtins-%.o: objs/builtins-%.cpp
 	@echo Compiling $<
 	@$(CXX) $(CXXFLAGS) -o $@ -c $<
 
-objs/builtins-c.cpp: builtins-c.c
+objs/builtins-c-32.cpp: builtins-c.c
 	@echo Creating C++ source from builtins definition file $<
-	@$(CLANG) -I /opt/l1om/usr/include/ -emit-llvm -c $< -o - | llvm-dis - | ./bitcode2cpp.py $< > $@
+	@$(CLANG) -m32 -emit-llvm -c $< -o - | llvm-dis - | ./bitcode2cpp.py builtins-c-32.c > $@
 
-objs/builtins-c.o: objs/builtins-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 $<
 

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.

bitcode2cpp.py

@@ -4,6 +4,8 @@ import sys
 import string
 import re
 import subprocess
+import platform
+import os
 
 length=0
 
@@ -14,8 +16,12 @@ 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)

builtins-avx-common.ll

@@ -0,0 +1,278 @@
+;;  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.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; *** Untested *** AVX target implementation.
+;;
+;; The LLVM AVX code generator is incomplete, so the ispc AVX target
+;; hasn't yet been tested.  There is therefore a higher-than-normal
+;; chance that there are bugs in the code in this file.
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rcp
+
+declare <4 x float> @llvm.x86.sse.rcp.ss(<4 x float>) nounwind readnone
+
+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
+}
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; rounding floats
+
+declare <4 x float> @llvm.x86.sse41.round.ss(<4 x float>, <4 x float>, i32) nounwind readnone
+
+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 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 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 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 internal 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 internal 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 internal 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 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.ss(<4 x float>) nounwind readnone
+
+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
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; fastmath
+
+declare void @llvm.x86.sse.stmxcsr(i8 *) nounwind
+declare void @llvm.x86.sse.ldmxcsr(i8 *) nounwind
+
+define internal 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 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 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 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 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 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 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 internal 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 internal 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 internal 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 internal 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 internal 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 +1,665 @@
+;;  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.  
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; *** Untested *** AVX target implementation.
+;;
+;; The LLVM AVX code generator is incomplete, so the ispc AVX target
+;; hasn't yet been tested.  There is therefore a higher-than-normal
+;; chance that there are bugs in the code in this file.
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; 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 internal <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 internal <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 internal <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 internal <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 internal <16 x double> @__round_varying_double(<16 x double>) nounwind readonly alwaysinline {
+  round4to16double(%0, 8)
+}
+
+define internal <16 x double> @__floor_varying_double(<16 x double>) nounwind readonly alwaysinline {
+  round4to16double(%0, 9)
+}
+
+define internal <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 internal <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 internal <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 internal <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 internal <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 internal <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 internal <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 internal <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 internal <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 internal 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 internal 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 internal float @__reduce_min_float(<16 x float>) nounwind readnone alwaysinline {
+  reduce16(float, @__min_varying_float, @__min_uniform_float)
+}
+
+
+define internal 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 internal <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 internal i32 @__add_uniform_int32(i32, i32) nounwind readnone alwaysinline {
+  %s = add i32 %0, %1
+  ret i32 %s
+}
+
+define internal i32 @__reduce_add_int32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__add_varying_int32, @__add_uniform_int32)
+}
+
+
+define internal i32 @__reduce_min_int32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__min_varying_int32, @__min_uniform_int32)
+}
+
+
+define internal i32 @__reduce_max_int32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__max_varying_int32, @__max_uniform_int32)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;; horizontal uint32 ops
+
+define internal 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 internal i32 @__reduce_min_uint32(<16 x i32>) nounwind readnone alwaysinline {
+  reduce16(i32, @__min_varying_uint32, @__min_uniform_uint32)
+}
+
+
+define internal 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 internal 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 internal double @__reduce_min_double(<16 x double>) nounwind readnone alwaysinline {
+  reduce16(double, @__min_varying_double, @__min_uniform_double)
+}
+
+
+define internal double @__reduce_max_double(<16 x double>) nounwind readnone alwaysinline {
+  reduce16(double, @__max_varying_double, @__max_uniform_double)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; horizontal int64 ops
+
+define internal <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 internal i64 @__add_uniform_int64(i64, i64) nounwind readnone alwaysinline {
+  %s = add i64 %0, %1
+  ret i64 %s
+}
+
+define internal i64 @__reduce_add_int64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__add_varying_int64, @__add_uniform_int64)
+}
+
+
+define internal i64 @__reduce_min_int64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__min_varying_int64, @__min_uniform_int64)
+}
+
+
+define internal i64 @__reduce_max_int64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__max_varying_int64, @__max_uniform_int64)
+}
+
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;; horizontal uint64 ops
+
+define internal 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 internal i64 @__reduce_min_uint64(<16 x i64>) nounwind readnone alwaysinline {
+  reduce16(i64, @__min_varying_uint64, @__min_uniform_uint64)
+}
+
+
+define internal 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 internal <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 internal <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 internal <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

@@ -41,13 +41,15 @@
 
 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
-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 {
   ;  float iv = __rcp_v(v);
@@ -62,25 +64,10 @@ 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
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; rounding floats
 
 declare <8 x float> @llvm.x86.avx.round.ps.256(<8 x float>, i32) nounwind readnone
-declare <4 x float> @llvm.x86.sse.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 {
   ; roundps, round mode nearest 0b00 | don't signal precision exceptions 0b1000 = 8
@@ -88,111 +75,43 @@ define internal <8 x float> @__round_varying_float(<8 x float>) nounwind readonl
   ret <8 x float> %call
 }
 
-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.sse.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
+  ; 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 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.sse.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
+  ; 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
 }
 
-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.sse.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 <4 x double> @llvm.x86.avx.round.pd.256(<4 x double>, i32) nounwind readnone
-declare <2 x double> @llvm.x86.sse41.round.sd(<2 x double>, <2 x double>, i32) nounwind readnone
 
 define internal <8 x double> @__round_varying_double(<8 x double>) nounwind readonly alwaysinline {
   round4to8double(%0, 8)
 }
 
-define internal 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 internal <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 internal 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 0b1000 = 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 internal <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)
 }
 
-define internal 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 0b1000 = 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 <8 x float> @llvm.x86.avx.rsqrt.ps.256(<8 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 {
   ;  float is = __rsqrt_v(v);
@@ -200,64 +119,24 @@ define internal <8 x float> @__rsqrt_varying_float(<8 x float> %v) nounwind read
   ;  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
+  %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
+  %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
 }
 
-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 <8 x float> @llvm.x86.avx.sqrt.ps.256(<8 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 {
   %call = call <8 x float> @llvm.x86.avx.sqrt.ps.256(<8 x float> %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
-}
-
-
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; fastmath
-
-declare void @llvm.x86.sse.stmxcsr(i8 *) nounwind
-declare void @llvm.x86.sse.ldmxcsr(i8 *) nounwind
-
-define internal 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
-}
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; svml
 
@@ -279,9 +158,7 @@ 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 <4 x float> @llvm.x86.sse.max.ss(<4 x float>, <4 x float>) nounwind readnone
 declare <8 x float> @llvm.x86.avx.min.ps.256(<8 x float>, <8 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 {
@@ -289,97 +166,43 @@ define internal <8 x float> @__max_varying_float(<8 x float>,
   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 {
   %call = call <8 x float> @llvm.x86.avx.min.ps.256(<8 x float> %0, <8 x float> %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
 
-declare <8 x i32> @llvm.x86.avx.min.sd.256(<8 x i32>, <8 x i32>) nounwind readnone
-declare <8 x i32> @llvm.x86.avx.max.sd.256(<8 x i32>, <8 x i32>) nounwind readnone
-
 define internal <8 x i32> @__min_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
-  %call = call <8 x i32> @llvm.x86.avx.min.sd.256(<8 x i32> %0, <8 x i32> %1)
-  ret <8 x i32> %call
-}
-
-define internal i32 @__min_uniform_int32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 8, i32, @llvm.x86.avx.min.sd.256, %0, %1)
-  ret i32 %ret
+  binary4to8(ret, i32, @llvm.x86.sse41.pminsd, %0, %1)
+  ret <8 x i32> %ret
 }
 
 define internal <8 x i32> @__max_varying_int32(<8 x i32>, <8 x i32>) nounwind readonly alwaysinline {
-  %call = call <8 x i32> @llvm.x86.avx.max.sd.256(<8 x i32> %0, <8 x i32> %1)
-  ret <8 x i32> %call
-}
-
-define internal i32 @__max_uniform_int32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 8, i32, @llvm.x86.avx.max.sd.256, %0, %1)
-  ret i32 %ret
+  binary4to8(ret, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
+  ret <8 x i32> %ret
 }
 
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; unsigned int min/max
 
-; FIXME: looks like these aren't available in LLVM?
-declare <8 x i32> @llvm.x86.avx.min.ud.256(<8 x i32>, <8 x i32>) nounwind readnone
-declare <8 x i32> @llvm.x86.avx.max.ud.256(<8 x i32>, <8 x i32>) nounwind readnone
-
-define internal <8 x i32> @__min_varying_uint32(<8 x i32>,
-                                                <8 x i32>) nounwind readonly alwaysinline {
-  %call = call <8 x i32> @llvm.x86.avx.min.ud.256(<8 x i32> %0, <8 x i32> %1)
-  ret <8 x i32> %call
+define internal <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 internal i32 @__min_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 8, i32, @llvm.x86.avx.min.ud.256, %0, %1)
-  ret i32 %ret
+define internal <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
 }
 
-define internal <8 x i32> @__max_varying_uint32(<8 x i32>,
-                                                <8 x i32>) nounwind readonly alwaysinline {
-  %call = call <8 x i32> @llvm.x86.avx.max.ud.256(<8 x i32> %0, <8 x i32> %1)
-  ret <8 x i32> %call
-}
-
-define internal i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinline {
-  sse_binary_scalar(ret, 8, i32, @llvm.x86.avx.max.ud.256, %0, %1)
-  ret i32 %ret
-}
-
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; horizontal ops
 
-declare i32 @llvm.ctpop.i32(i32) nounwind readnone
-
-define internal 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 internal i64 @__popcnt_int64(i64) nounwind readonly alwaysinline {
-  %call = call i64 @llvm.ctpop.i64(i64 %0)
-  ret i64 %call
-}
-
 declare i32 @llvm.x86.avx.movmsk.ps.256(<8 x float>) nounwind readnone
 
 define internal i32 @__movmsk(<8 x i32>) nounwind readnone alwaysinline {
@@ -412,6 +235,7 @@ define internal float @__reduce_max_float(<8 x float>) nounwind readnone alwaysi
   reduce8(float, @__max_varying_float, @__max_uniform_float)
 }
 
+reduce_equal(8)
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; horizontal int32 ops
@@ -472,9 +296,10 @@ define internal double @__reduce_add_double(<8 x double>) nounwind readonly alwa
                       <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)
-  %scalar1 = extractelement <4 x double> %sum0, i32 0
-  %scalar2 = extractelement <4 x double> %sum1, i32 1
-  %sum = fadd double %scalar1, %scalar2
+  %final0 = extractelement <4 x double> %sum1, i32 0
+  %final1 = extractelement <4 x double> %sum1, i32 2
+  %sum = fadd double %final0, %final1
+
   ret double %sum
 }
 
@@ -624,13 +449,14 @@ define void @__masked_store_64(<8 x i64>* nocapture, <8 x i64>,
   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>,
+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
@@ -645,7 +471,7 @@ define void @__masked_store_blend_32(<8 x i32>* nocapture, <8 x i32>,
 }
 
 
-define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
+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>
@@ -695,6 +521,7 @@ define void @__masked_store_blend_64(<8 x i64>* nocapture %ptr, <8 x i64> %new,
   ret void
 }
 
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; gather/scatter
 
@@ -712,43 +539,26 @@ gen_scatter(8, i64)
 ;; double precision sqrt
 
 declare <4 x double> @llvm.x86.avx.sqrt.pd.256(<4 x double>) nounwind readnone
-declare <2 x double> @llvm.x86.sse.sqrt.sd(<2 x double>) nounwind readnone
 
 define internal <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
 }
 
-define internal 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 <4 x double> @llvm.x86.avx.max.pd.256(<4 x double>, <4 x double>) nounwind readnone
-declare <2 x double> @llvm.x86.sse.max.sd(<2 x double>, <2 x double>) nounwind readnone
 declare <4 x double> @llvm.x86.avx.min.pd.256(<4 x double>, <4 x double>) nounwind readnone
-declare <2 x double> @llvm.x86.sse.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 {
   binary4to8(ret, double, @llvm.x86.avx.min.pd.256, %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.sse.min.sd, %0, %1)
-  ret double %ret
-}
-
 define internal <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
 }
 
-define internal double @__max_uniform_double(double, double) nounwind readnone alwaysinline {
-  sse_binary_scalar(ret, 2, double, @llvm.x86.sse.max.sd, %0, %1)
-  ret double %ret
-}

builtins-c.c

@@ -51,6 +51,10 @@
   */
 
 
+#ifndef _MSC_VER
+#include <unistd.h>
+#endif // !_MSC_VER
+
 #include <stdint.h>
 #include <stdio.h>
 #include <stdarg.h>
@@ -139,3 +143,28 @@ 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-sse.ll

@@ -376,6 +376,7 @@ define internal i64 @__reduce_max_uint64(<4 x i64>) nounwind readnone {
   reduce4(i64, @__max_varying_uint64, @__max_uniform_uint64)
 }
 
+reduce_equal(4)
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; masked store

builtins-sse2.ll

@@ -35,6 +35,7 @@
 ; Define some basics for a 4-wide target
 stdlib_core(4)
 packed_load_and_store(4)
+scans(4)
 
 ; Include the various definitions of things that only require SSE1 and SSE2
 include(`builtins-sse.ll')
@@ -276,41 +277,17 @@ define internal i32 @__max_uniform_uint32(i32, i32) nounwind readonly alwaysinli
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ; horizontal ops / reductions
 
-; FIXME: this is very inefficient, loops over all 32 bits...
-
-; we could use the LLVM intrinsic declare i32 @llvm.ctpop.i32(i32),
-; although that currently ends up generating a POPCNT instruction even
-; if we give --target=sse2 on the command line.  We probably need to
-; pipe through the 'sse2' request to LLVM via the 'features' string
-; at codegen time...  (If e.g. --cpu=penryn is also passed along, then
-; it does generate non-POPCNT code and in particular better code than
-; the below does.)
+declare i32 @llvm.ctpop.i32(i32)
+declare i64 @llvm.ctpop.i64(i64)
 
 define internal i32 @__popcnt_int32(i32) nounwind readonly alwaysinline {
-entry:
-  br label %loop
-
-loop:
-  %count = phi i32 [ 0, %entry ], [ %newcount, %loop ]
-  %val = phi i32 [ %0, %entry ], [ %newval, %loop ]
-  %delta = and i32 %val, 1
-  %newcount = add i32 %count, %delta
-  %newval = lshr i32 %val, 1
-  %done = icmp eq i32 %newval, 0
-  br i1 %done, label %exit, label %loop
-
-exit:
-  ret i32 %newcount
+  %val = call i32 @llvm.ctpop.i32(i32 %0)
+  ret i32 %val
 }
 
-define internal i32 @__popcnt_int64(i64) nounwind readnone alwaysinline {
-  %vec = bitcast i64 %0 to <2 x i32>
-  %v0 = extractelement <2 x i32> %vec, i32 0
-  %v1 = extractelement <2 x i32> %vec, i32 1
-  %c0 = call i32 @__popcnt_int32(i32 %v0)
-  %c1 = call i32 @__popcnt_int32(i32 %v1)
-  %sum = add i32 %c0, %c1
-  ret i32 %sum
+define internal i64 @__popcnt_int64(i64) nounwind readnone alwaysinline {
+  %val = call i64 @llvm.ctpop.i64(i64 %0)
+  ret i64 %val
 }
 
 

builtins-sse4.ll

@@ -35,6 +35,7 @@
 ; Define common 4-wide stuff
 stdlib_core(4)
 packed_load_and_store(4)
+scans(4)
 
 ; Define the stuff that can be done with base SSE1/SSE2 instructions
 include(`builtins-sse.ll')
@@ -76,7 +77,7 @@ define internal float @__round_uniform_float(float) nounwind readonly alwaysinli
 }
 
 define internal <4 x float> @__floor_varying_float(<4 x float>) nounwind readonly alwaysinline {
-  ; roundps, round down 0b01 | don't signal precision exceptions 0b1000 = 9
+  ; 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
 }
@@ -84,14 +85,14 @@ define internal <4 x float> @__floor_varying_float(<4 x float>) nounwind readonl
 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
+  ; 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 internal <4 x float> @__ceil_varying_float(<4 x float>) nounwind readonly alwaysinline {
-  ; roundps, round up 0b10 | don't signal precision exceptions 0b1000 = 10
+  ; 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
 }
@@ -99,7 +100,7 @@ define internal <4 x float> @__ceil_varying_float(<4 x float>) nounwind readonly
 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
+  ; 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
@@ -123,28 +124,28 @@ define internal double @__round_uniform_double(double) nounwind readonly alwaysi
 }
 
 define internal <4 x double> @__floor_varying_double(<4 x double>) nounwind readonly alwaysinline {
-  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1000 = 9
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1001 = 9
   round2to4double(%0, 9)
 }
 
 define internal 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 0b1000 = 9
+  ; 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 internal <4 x double> @__ceil_varying_double(<4 x double>) nounwind readonly alwaysinline {
-  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1000 = 10
+  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1010 = 10
   round2to4double(%0, 10)
 }
 
 define internal 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 0b1000 = 10
+  ; 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
@@ -229,7 +230,6 @@ define internal float @__reduce_add_float(<4 x float>) nounwind readonly alwaysi
   ret float %scalar
 }
 
-
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; masked store
 

builtins-sse4x2.ll

@@ -38,6 +38,7 @@
 
 stdlib_core(8)
 packed_load_and_store(8)
+scans(8)
 int64minmax(8)
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -434,6 +435,8 @@ define internal i64 @__reduce_max_uint64(<8 x i64>) nounwind readnone {
   reduce8(i64, @__max_varying_uint64, @__max_uniform_uint64)
 }
 
+reduce_equal(8)
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; unaligned loads/loads+broadcasts
 
@@ -495,28 +498,28 @@ define internal float @__round_uniform_float(float) nounwind readonly alwaysinli
 }
 
 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
+  ; 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
+  ; 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 internal <8 x float> @__ceil_varying_float(<8 x float>) nounwind readonly alwaysinline {
-  ; roundps, round up 0b10 | don't signal precision exceptions 0b1000 = 10
+  ; 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
+  ; 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
@@ -540,28 +543,28 @@ define internal double @__round_uniform_double(double) nounwind readonly alwaysi
 }
 
 define internal <8 x double> @__floor_varying_double(<8 x double>) nounwind readonly alwaysinline {
-  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1000 = 9
+  ; roundpd, round down 0b01 | don't signal precision exceptions 0b1001 = 9
   round2to8double(%0, 9)
 }
 
 define internal 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 0b1000 = 9
+  ; 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 internal <8 x double> @__ceil_varying_double(<8 x double>) nounwind readonly alwaysinline {
-  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1000 = 10
+  ; roundpd, round up 0b10 | don't signal precision exceptions 0b1010 = 10
   round2to8double(%0, 10)
 }
 
 define internal 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 0b1000 = 10
+  ; 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

builtins.cpp

@@ -54,6 +54,8 @@
 #include <llvm/Instructions.h>
 #include <llvm/Intrinsics.h>
 #include <llvm/Linker.h>
+#include <llvm/Target/TargetMachine.h>
+#include <llvm/ADT/Triple.h>
 #include <llvm/Support/MemoryBuffer.h>
 #include <llvm/Bitcode/ReaderWriter.h>
 
@@ -170,6 +172,27 @@ lLLVMTypeToISPCType(const llvm::Type *t, bool intAsUnsigned) {
 }
 
 
+static void
+lCreateSymbol(const std::string &name, const Type *returnType, 
+              const std::vector<const Type *> &argTypes, 
+              const llvm::FunctionType *ftype, llvm::Function *func, 
+              SymbolTable *symbolTable) {
+    SourcePos noPos;
+    noPos.name = "__stdlib";
+
+    FunctionType *funcType = new FunctionType(returnType, argTypes, noPos);
+    // set NULL default arguments
+    std::vector<ConstExpr *> defaults;
+    for (unsigned int j = 0; j < ftype->getNumParams(); ++j)
+        defaults.push_back(NULL);
+    funcType->SetArgumentDefaults(defaults);
+
+    Symbol *sym = new Symbol(name, noPos, funcType);
+    sym->function = func;
+    symbolTable->AddFunction(sym);
+}
+
+
 /** Given an LLVM function declaration, synthesize the equivalent ispc
     symbol for the function (if possible).  Returns true on success, false
     on failure.
@@ -221,7 +244,7 @@ lCreateISPCSymbol(llvm::Function *func, SymbolTable *symbolTable) {
 
         // Iterate over the arguments and try to find their equivalent ispc
         // types.  Track if any of the arguments has an integer type.
-        bool anyIntArgs = false;
+        bool anyIntArgs = false, anyReferenceArgs = false;
         std::vector<const Type *> argTypes;
         for (unsigned int j = 0; j < ftype->getNumParams(); ++j) {
             const llvm::Type *llvmArgType = ftype->getParamType(j);
@@ -230,22 +253,26 @@ lCreateISPCSymbol(llvm::Function *func, SymbolTable *symbolTable) {
                 return false;
             anyIntArgs |= 
                 (Type::Equal(type, lLLVMTypeToISPCType(llvmArgType, !intAsUnsigned)) == false);
+            anyReferenceArgs |= (dynamic_cast<const ReferenceType *>(type) != NULL);
             argTypes.push_back(type);
         }
 
         // Always create the symbol the first time through, in particular
         // so that we get symbols for things with no integer types!
-        if (i == 0 || anyIntArgs == true) {
-            FunctionType *funcType = new FunctionType(returnType, argTypes, noPos);
-            // set NULL default arguments
-            std::vector<ConstExpr *> defaults;
-            for (unsigned int j = 0; j < ftype->getNumParams(); ++j)
-                defaults.push_back(NULL);
-            funcType->SetArgumentDefaults(defaults);
+        if (i == 0 || anyIntArgs == true)
+            lCreateSymbol(name, returnType, argTypes, ftype, func, symbolTable);
 
-            Symbol *sym = new Symbol(name, noPos, funcType);
-            sym->function = func;
-            symbolTable->AddFunction(sym);
+        // If there are any reference types, also make a variant of the
+        // symbol that has them as const references.  This obviously
+        // doesn't make sense for many builtins, but we'll give the stdlib
+        // the option to call one if it needs one.
+        if (anyReferenceArgs == true) {
+            for (unsigned int j = 0; j < argTypes.size(); ++j) {
+                if (dynamic_cast<const ReferenceType *>(argTypes[j]) != NULL)
+                    argTypes[j] = argTypes[j]->GetAsConstType();
+                lCreateSymbol(name + "_refsconst", returnType, argTypes, 
+                              ftype, func, symbolTable);
+            }
         }
     }
 
@@ -319,6 +346,22 @@ lAddBitcode(const unsigned char *bitcode, int length,
     if (!bcModule)
         Error(SourcePos(), "Error parsing stdlib bitcode: %s", bcErr.c_str());
     else {
+        // FIXME: this feels like a bad idea, but the issue is that when we
+        // set the llvm::Module's target triple in the ispc Module::Module
+        // constructor, we start by calling llvm::sys::getHostTriple() (and
+        // then change the arch if needed).  Somehow that ends up giving us
+        // strings like 'x86_64-apple-darwin11.0.0', while the stuff we
+        // compile to bitcode with clang has module triples like
+        // 'i386-apple-macosx10.7.0'.  And then LLVM issues a warning about
+        // linking together modules with incompatible target triples..
+        llvm::Triple mTriple(m->module->getTargetTriple());
+        llvm::Triple bcTriple(bcModule->getTargetTriple());
+        assert(bcTriple.getArch() == llvm::Triple::UnknownArch ||
+               mTriple.getArch() == bcTriple.getArch());
+        assert(bcTriple.getVendor() == llvm::Triple::UnknownVendor ||
+               mTriple.getVendor() == bcTriple.getVendor());
+        bcModule->setTargetTriple(mTriple.str());
+
         std::string(linkError);
         if (llvm::Linker::LinkModules(module, bcModule, &linkError))
             Error(SourcePos(), "Error linking stdlib bitcode: %s", linkError.c_str());
@@ -346,6 +389,27 @@ lDefineConstantInt(const char *name, int val, llvm::Module *module,
 }
 
 
+
+static void
+lDefineConstantIntFunc(const char *name, int val, llvm::Module *module,
+                       SymbolTable *symbolTable) {
+    std::vector<const Type *> args;
+    FunctionType *ft = new FunctionType(AtomicType::UniformInt32, args, SourcePos());
+    Symbol *sym = new Symbol(name, SourcePos(), ft);
+    sym->isStatic = true;
+
+    llvm::Function *func = module->getFunction(name);
+    assert(func != NULL); // it should be declared already...
+    func->addFnAttr(llvm::Attribute::AlwaysInline);
+    llvm::BasicBlock *bblock = llvm::BasicBlock::Create(*g->ctx, "entry", func, 0);
+    llvm::ReturnInst::Create(*g->ctx, LLVMInt32(val), bblock);
+
+    sym->function = func;
+    symbolTable->AddVariable(sym);
+}
+
+
+
 static void
 lDefineProgramIndex(llvm::Module *module, SymbolTable *symbolTable) {
     Symbol *pidx = new Symbol("programIndex", SourcePos(), 
@@ -370,9 +434,18 @@ void
 DefineStdlib(SymbolTable *symbolTable, llvm::LLVMContext *ctx, llvm::Module *module,
              bool includeStdlibISPC) {
     // Add the definitions from the compiled builtins-c.c file
-    extern unsigned char builtins_bitcode_c[];
-    extern int builtins_bitcode_c_length;
-    lAddBitcode(builtins_bitcode_c, builtins_bitcode_c_length, module, symbolTable);
+    if (g->target.is32bit) {
+        extern unsigned char builtins_bitcode_c_32[];
+        extern int builtins_bitcode_c_32_length;
+        lAddBitcode(builtins_bitcode_c_32, builtins_bitcode_c_32_length, 
+                    module, symbolTable);
+    }
+    else {
+        extern unsigned char builtins_bitcode_c_64[];
+        extern int builtins_bitcode_c_64_length;
+        lAddBitcode(builtins_bitcode_c_64, builtins_bitcode_c_64_length, 
+                    module, symbolTable);
+    }
 
     // Next, add the target's custom implementations of the various needed
     // builtin functions (e.g. __masked_store_32(), etc).
@@ -402,10 +475,22 @@ DefineStdlib(SymbolTable *symbolTable, llvm::LLVMContext *ctx, llvm::Module *mod
         }
         break;
     case Target::AVX:
-        extern unsigned char builtins_bitcode_avx[];
-        extern int builtins_bitcode_avx_length;
-        lAddBitcode(builtins_bitcode_avx, builtins_bitcode_avx_length, module, 
-                    symbolTable);
+        switch (g->target.vectorWidth) {
+        case 8:
+            extern unsigned char builtins_bitcode_avx[];
+            extern int builtins_bitcode_avx_length;
+            lAddBitcode(builtins_bitcode_avx, builtins_bitcode_avx_length, module, 
+                        symbolTable);
+            break;
+        case 16:
+            extern unsigned char builtins_bitcode_avx_x2[];
+            extern int builtins_bitcode_avx_x2_length;
+            lAddBitcode(builtins_bitcode_avx_x2, builtins_bitcode_avx_x2_length,
+                        module,  symbolTable);
+            break;
+        default:
+            FATAL("logic error in DefineStdlib");
+        }
         break;
     default:
         FATAL("logic error");
@@ -428,6 +513,8 @@ DefineStdlib(SymbolTable *symbolTable, llvm::LLVMContext *ctx, llvm::Module *mod
                        symbolTable);
     lDefineConstantInt("__math_lib_system", (int)Globals::Math_System, module,
                        symbolTable);
+    lDefineConstantIntFunc("__fast_masked_vload", (int)g->opt.fastMaskedVload, module,
+                           symbolTable);
 
     if (includeStdlibISPC) {
         // If the user wants the standard library to be included, parse the

builtins.m4

@@ -111,6 +111,32 @@ define(`reduce8', `
 '
 )
 
+define(`reduce16', `
+  %v1 = shufflevector <16 x $1> %0, <16 x $1> undef,
+        <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15,
+                    i32 undef, i32 undef, i32 undef, i32 undef,
+                    i32 undef, i32 undef, i32 undef, i32 undef>
+  %m1 = call <16 x $1> $2(<16 x $1> %v1, <16 x $1> %0)
+  %v2 = shufflevector <16 x $1> %m1, <16 x $1> undef,
+        <16 x i32> <i32 4, i32 5, i32 6, i32 7,
+                    i32 undef, i32 undef, i32 undef, i32 undef,
+                    i32 undef, i32 undef, i32 undef, i32 undef,
+                    i32 undef, i32 undef, i32 undef, i32 undef>
+  %m2 = call <16 x $1> $2(<16 x $1> %v2, <16 x $1> %m1)
+  %v3 = shufflevector <16 x $1> %m2, <16 x $1> undef,
+        <16 x i32> <i32 2, i32 3, i32 undef, i32 undef,
+                    i32 undef, i32 undef, i32 undef, i32 undef,
+                    i32 undef, i32 undef, i32 undef, i32 undef,
+                    i32 undef, i32 undef, i32 undef, i32 undef>
+  %m3 = call <16 x $1> $2(<16 x $1> %v3, <16 x $1> %m2)
+
+  %m3a = extractelement <16 x $1> %m3, i32 0
+  %m3b = extractelement <16 x $1> %m3, i32 1
+  %m = call $1 $3($1 %m3a, $1 %m3b)
+  ret $1 %m
+'
+)
+
 ;; Do an reduction over an 8-wide vector, using a vector reduction function
 ;; that only takes 4-wide vectors
 ;; $1: type of final scalar result
@@ -211,6 +237,45 @@ define(`unary4to8', `
 '
 )
 
+define(`unary4to16', `
+  %$1_0 = shufflevector <16 x $2> $4, <16 x $2> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %v$1_0 = call <4 x $2> $3(<4 x $2> %$1_0)
+  %$1_1 = shufflevector <16 x $2> $4, <16 x $2> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+  %v$1_1 = call <4 x $2> $3(<4 x $2> %$1_1)
+  %$1_2 = shufflevector <16 x $2> $4, <16 x $2> undef, <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+  %v$1_2 = call <4 x $2> $3(<4 x $2> %$1_2)
+  %$1_3 = shufflevector <16 x $2> $4, <16 x $2> undef, <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+  %v$1_3 = call <4 x $2> $3(<4 x $2> %$1_3)
+
+  %$1a = shufflevector <4 x $2> %v$1_0, <4 x $2> %v$1_1, 
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %$1b = shufflevector <4 x $2> %v$1_2, <4 x $2> %v$1_3, 
+           <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %$1 = shufflevector <8 x $2> %$1a, <8 x $2> %$1b,
+           <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>
+'
+)
+
+;; And so forth...
+;; $1: name of variable into which the final result should go
+;; $2: scalar type of the vector elements
+;; $3: 8-wide unary vector function to apply
+;; $4: 16-wide operand value
+
+define(`unary8to16', `
+  %$1_0 = shufflevector <16 x $2> $4, <16 x $2> undef,
+             <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  %v$1_0 = call <8 x $2> $3(<8 x $2> %$1_0)
+  %$1_1 = shufflevector <16 x $2> $4, <16 x $2> undef,
+             <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+  %v$1_1 = call <8 x $2> $3(<8 x $2> %$1_1)
+  %$1 = shufflevector <8 x $2> %v$1_0, <8 x $2> %v$1_1, 
+           <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>
+'
+)
+
 ;; And along the lines of `binary2to4', this maps a 4-wide binary function to
 ;; two 8-wide vector operands
 ;; $1: name of variable into which the final result should go
@@ -231,6 +296,57 @@ define(`binary4to8', `
 '
 )
 
+define(`binary8to16', `
+%$1_0a = shufflevector <16 x $2> $4, <16 x $2> undef,
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+%$1_0b = shufflevector <16 x $2> $5, <16 x $2> undef,
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+%v$1_0 = call <8 x $2> $3(<8 x $2> %$1_0a, <8 x $2> %$1_0b)
+%$1_1a = shufflevector <16 x $2> $4, <16 x $2> undef,
+          <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+%$1_1b = shufflevector <16 x $2> $5, <16 x $2> undef,
+          <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+%v$1_1 = call <8 x $2> $3(<8 x $2> %$1_1a, <8 x $2> %$1_1b)
+%$1 = shufflevector <8 x $2> %v$1_0, <8 x $2> %v$1_1, 
+         <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>
+'
+)
+
+define(`binary4to16', `
+%$1_0a = shufflevector <16 x $2> $4, <16 x $2> undef,
+          <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+%$1_0b = shufflevector <16 x $2> $5, <16 x $2> undef,
+          <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+%r$1_0 = call <4 x $2> $3(<4 x $2> %$1_0a, <4 x $2> %$1_0b) 
+
+%$1_1a = shufflevector <16 x $2> $4, <16 x $2> undef,
+          <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+%$1_1b = shufflevector <16 x $2> $5, <16 x $2> undef,
+          <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+%r$1_1 = call <4 x $2> $3(<4 x $2> %$1_1a, <4 x $2> %$1_1b) 
+
+%$1_2a = shufflevector <16 x $2> $4, <16 x $2> undef,
+          <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+%$1_2b = shufflevector <16 x $2> $5, <16 x $2> undef,
+          <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+%r$1_2 = call <4 x $2> $3(<4 x $2> %$1_2a, <4 x $2> %$1_2b) 
+
+%$1_3a = shufflevector <16 x $2> $4, <16 x $2> undef,
+          <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+%$1_3b = shufflevector <16 x $2> $5, <16 x $2> undef,
+          <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+%r$1_3 = call <4 x $2> $3(<4 x $2> %$1_3a, <4 x $2> %$1_3b)
+
+%r$1_01 = shufflevector <4 x $2> %r$1_0, <4 x $2> %r$1_1, 
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+%r$1_23 = shufflevector <4 x $2> %r$1_2, <4 x $2> %r$1_3, 
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+
+%$1 = shufflevector <8 x $2> %r$1_01, <8 x $2> %r$1_23, 
+          <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>
+')
 
 ;; Maps a 2-wide unary function to an 8-wide vector operand, returning an 
 ;; 8-wide vector result
@@ -306,6 +422,20 @@ ret <8 x float> %ret
 '
 )
 
+define(`round8to16', `
+%v0 = shufflevector <16 x float> $1, <16 x float> undef,
+        <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+%v1 = shufflevector <16 x float> $1, <16 x float> undef,
+        <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
+%r0 = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %v0, i32 $2)
+%r1 = call <8 x float> @llvm.x86.avx.round.ps.256(<8 x float> %v1, i32 $2)
+%ret = shufflevector <8 x float> %r0, <8 x float> %r1, 
+         <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>
+ret <16 x float> %ret
+'
+)
+
 define(`round4to8double', `
 %v0 = shufflevector <8 x double> $1, <8 x double> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
 %v1 = shufflevector <8 x double> $1, <8 x double> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
@@ -349,6 +479,30 @@ ret <8 x double> %ret
 '
 )
 
+define(`round4to16double', `
+%v0 = shufflevector <16 x double> $1, <16 x double> undef,
+         <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+%v1 = shufflevector <16 x double> $1, <16 x double> undef,
+         <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+%v2 = shufflevector <16 x double> $1, <16 x double> undef,
+         <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+%v3 = shufflevector <16 x double> $1, <16 x double> undef,
+         <4 x i32> <i32 12, i32 13, i32 14, i32 15>
+%r0 = call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %v0, i32 $2)
+%r1 = call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %v1, i32 $2)
+%r2 = call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %v2, i32 $2)
+%r3 = call <4 x double> @llvm.x86.avx.round.pd.256(<4 x double> %v3, i32 $2)
+%ret0 = shufflevector <4 x double> %r0, <4 x double> %r1, 
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+%ret1 = shufflevector <4 x double> %r2, <4 x double> %r3, 
+          <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+%ret = shufflevector <8 x double> %ret0, <8 x double> %ret1,
+          <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>
+ret <16 x double> %ret
+'
+)
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; forloop macro
 
@@ -468,12 +622,91 @@ forloop(i, 1, eval($1-1), `
 }
 ')
 
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; global_atomic
+;; global_atomic_associative
+;; More efficient implementation for atomics that are associative (e.g.,
+;; add, and, ...).  If a basic implementation would do sometihng like:
+;; result0 = atomic_op(ptr, val0)
+;; result1 = atomic_op(ptr, val1)
+;; ..
+;; Then instead we can do:
+;; tmp = (val0 op val1 op ...)
+;; result0 = atomic_op(ptr, tmp)
+;; result1 = (result0 op val0)
+;; ..
+;; And more efficiently compute the same result
+;;
+;; Takes five parameters:
+;; $1: vector width of the target
+;; $2: operation being performed (w.r.t. LLVM atomic intrinsic names)
+;;     (add, sub...)
+;; $3: return type of the LLVM atomic (e.g. i32)
+;; $4: return type of the LLVM atomic type, in ispc naming paralance (e.g. int32)
+;; $5: identity value for the operator (e.g. 0 for add, -1 for AND, ...)
+
+define(`global_atomic_associative', `
+
+define internal <$1 x $3> @__atomic_$2_$4_global($3 * %ptr, <$1 x $3> %val,
+                                                 <$1 x i32> %m) nounwind alwaysinline {
+  ; first, for any lanes where the mask is off, compute a vector where those lanes
+  ; hold the identity value..
+
+  ; for the bit tricks below, we need the mask to be sign extended to be
+  ; the size of the element type.
+  ifelse($3, `i64', `%mask = sext <$1 x i32> %m to <$1 x i64>')
+  ifelse($3, `i32', `
+     ; silly workaround to do %mask = %m, which is not possible directly..
+     %maskmem = alloca <$1 x i32>
+     store <$1 x i32> %m, <$1 x i32> * %maskmem
+     %mask = load <$1 x i32> * %maskmem'
+  )
+  ; zero out any lanes that are off
+  %valoff = and <$1 x $3> %val, %mask
+
+  ; compute an identity vector that is zero in on lanes and has the identiy value
+  ; in the off lanes
+  %idv1 = bitcast $3 $5 to <1 x $3>
+  %idvec = shufflevector <1 x $3> %idv1, <1 x $3> undef,
+     <$1 x i32> < forloop(i, 1, eval($1-1), `i32 0, ') i32 0 >
+  %notmask = xor <$1 x $3> %mask, < forloop(i, 1, eval($1-1), `$3 -1, ') $3 -1 >
+  %idoff = and <$1 x $3> %idvec, %notmask
+
+  ; and comptue the merged vector that holds the identity in the off lanes
+  %valp = or <$1 x $3> %valoff, %idoff
+
+  ; now compute the local reduction (val0 op val1 op ... )--initialize
+  ; %eltvec so that the 0th element is the identity, the first is val0,
+  ; the second is (val0 op val1), ..
+  %red0 = extractelement <$1 x $3> %valp, i32 0
+  %eltvec0 = insertelement <$1 x $3> undef, $3 $5, i32 0
+
+  forloop(i, 1, eval($1-1), `
+  %elt`'i = extractelement <$1 x $3> %valp, i32 i
+  %red`'i = $2 $3 %red`'eval(i-1), %elt`'i
+  %eltvec`'i = insertelement <$1 x $3> %eltvec`'eval(i-1), $3 %red`'eval(i-1), i32 i')
+
+  ; make the atomic call, passing it the final reduced value
+  %final0 = call $3 @llvm.atomic.load.$2.$3.p0$3($3 * %ptr, $3 %red`'eval($1-1))
+
+  ; now go back and compute the values to be returned for each program 
+  ; instance--this just involves smearing the old value returned from the
+  ; actual atomic call across the vector and applying the vector op to the
+  ; %eltvec vector computed above..
+  %finalv1 = bitcast $3 %final0 to <1 x $3>
+  %final_base = shufflevector <1 x $3> %finalv1, <1 x $3> undef,
+     <$1 x i32> < forloop(i, 1, eval($1-1), `i32 0, ') i32 0 >
+  %r = $2 <$1 x $3> %final_base, %eltvec`'eval($1-1)
+
+  ret <$1 x $3> %r
+}
+')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; global_atomic_uniform
 ;; Defines the implementation of a function that handles the mapping from
-;; an ispc atomic function to the underlying LLVM intrinsics.  Specifically,
-;; the function handles loooping over the active lanes, calling the underlying
-;; scalar atomic intrinsic for each one, and assembling the vector result.
+;; an ispc atomic function to the underlying LLVM intrinsics.  This variant
+;; just calls the atomic once, for the given uniform value
 ;;
 ;; Takes four parameters:
 ;; $1: vector width of the target
@@ -482,23 +715,14 @@ forloop(i, 1, eval($1-1), `
 ;; $3: return type of the LLVM atomic (e.g. i32)
 ;; $4: return type of the LLVM atomic type, in ispc naming paralance (e.g. int32)
 
-define(`global_atomic', `
+define(`global_atomic_uniform', `
 
 declare $3 @llvm.atomic.load.$2.$3.p0$3($3 * %ptr, $3 %delta)
 
-define internal <$1 x $3> @__atomic_$2_$4_global($3 * %ptr, <$1 x $3> %val,
-                                                 <$1 x i32> %mask) nounwind alwaysinline {
-  %rptr = alloca <$1 x $3>
-  %rptr32 = bitcast <$1 x $3> * %rptr to $3 *
-
-  per_lane($1, <$1 x i32> %mask, `
-   %v_LANE_ID = extractelement <$1 x $3> %val, i32 LANE
-   %r_LANE_ID = call $3 @llvm.atomic.load.$2.$3.p0$3($3 * %ptr, $3 %v_LANE_ID)
-   %rp_LANE_ID = getelementptr $3 * %rptr32, i32 LANE
-   store $3 %r_LANE_ID, $3 * %rp_LANE_ID')
-
-  %r = load <$1 x $3> * %rptr
-  ret <$1 x $3> %r
+define internal $3 @__atomic_$2_uniform_$4_global($3 * %ptr, $3 %val,
+                                          <$1 x i32> %mask) nounwind alwaysinline {
+  %r = call $3 @llvm.atomic.load.$2.$3.p0$3($3 * %ptr, $3 %val)
+  ret $3 %r
 }
 ')
 
@@ -508,9 +732,10 @@ define internal <$1 x $3> @__atomic_$2_$4_global($3 * %ptr, <$1 x $3> %val,
 ;; $2: llvm type of the vector elements (e.g. i32)
 ;; $3: ispc type of the elements (e.g. int32)
 
-define(`global_swap', `
+declare i32 @llvm.atomic.swap.i32.p0i32(i32 * %ptr, i32 %val)
+declare i64 @llvm.atomic.swap.i64.p0i64(i64 * %ptr, i64 %val)
 
-declare $2 @llvm.atomic.swap.$2.p0$2($2 * %ptr, $2 %val)
+define(`global_swap', `
 
 define internal <$1 x $2> @__atomic_swap_$3_global($2* %ptr, <$1 x $2> %val,
                                                    <$1 x i32> %mask) nounwind alwaysinline {
@@ -526,6 +751,12 @@ define internal <$1 x $2> @__atomic_swap_$3_global($2* %ptr, <$1 x $2> %val,
   %r = load <$1 x $2> * %rptr
   ret <$1 x $2> %r
 }
+
+define internal $2 @__atomic_swap_uniform_$3_global($2* %ptr, $2 %val,
+                                                    <$1 x i32> %mask) nounwind alwaysinline {
+ %r = call $2 @llvm.atomic.swap.$2.p0$2($2 * %ptr, $2 %val)
+ ret $2 %r
+}
 ')
 
 
@@ -555,15 +786,57 @@ define internal <$1 x $2> @__atomic_compare_exchange_$3_global($2* %ptr, <$1 x $
   %r = load <$1 x $2> * %rptr
   ret <$1 x $2> %r
 }
+
+define internal $2 @__atomic_compare_exchange_uniform_$3_global($2* %ptr, $2 %cmp,
+                               $2 %val, <$1 x i32> %mask) nounwind alwaysinline {
+  %r = call $2 @llvm.atomic.cmp.swap.$2.p0$2($2 * %ptr, $2 %cmp, $2 %val)
+  ret $2 %r
+}
 ')
 
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; prefetch definitions
+
+; prefetch has a new parameter in LLVM3.0, to distinguish between instruction
+; and data caches--the declaration is now:
+; declare void @llvm.prefetch(i8* nocapture %ptr, i32 %readwrite, i32 %locality,
+;                             i32 %cachetype)  (cachetype 1 == data cache)
+; however, the version below seems to still work...
+
+declare void @llvm.prefetch(i8* nocapture %ptr, i32 %readwrite, i32 %locality)
+
+define(`prefetch_read', `
+define internal void @__prefetch_read_1_$1($2 *) alwaysinline {
+  %ptr8 = bitcast $2 * %0 to i8 *
+  call void @llvm.prefetch(i8 * %ptr8, i32 0, i32 3)
+  ret void
+}
+define internal void @__prefetch_read_2_$1($2 *) alwaysinline {
+  %ptr8 = bitcast $2 * %0 to i8 *
+  call void @llvm.prefetch(i8 * %ptr8, i32 0, i32 2)
+  ret void
+}
+define internal void @__prefetch_read_3_$1($2 *) alwaysinline {
+  %ptr8 = bitcast $2 * %0 to i8 *
+  call void @llvm.prefetch(i8 * %ptr8, i32 0, i32 1)
+  ret void
+}
+define internal void @__prefetch_read_nt_$1($2 *) alwaysinline {
+  %ptr8 = bitcast $2 * %0 to i8 *
+  call void @llvm.prefetch(i8 * %ptr8, i32 0, i32 0)
+  ret void
+}
+')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 define(`stdlib_core', `
 
-declare i8* @ISPCMalloc(i64, i32) nounwind
-declare i8* @ISPCFree(i8*) nounwind
-declare void @ISPCLaunch(i8*, i8*) nounwind
-declare void @ISPCSync() nounwind
+declare i32 @__fast_masked_vload()
+
+declare i8* @ISPCAlloc(i8**, i64, i32) nounwind
+declare void @ISPCLaunch(i8**, i8*, i8*, i32) nounwind
+declare void @ISPCSync(i8*) nounwind
 declare void @ISPCInstrument(i8*, i8*, i32, i32) nounwind
 
 declare i1 @__is_compile_time_constant_mask(<$1 x i32> %mask)
@@ -779,6 +1052,25 @@ define internal <$1 x i32> @__sext_varying_bool(<$1 x i32>) nounwind readnone al
   ret <$1 x i32> %0
 }
 
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; prefetching
+
+prefetch_read(uniform_bool, i1)
+prefetch_read(uniform_int8, i8)
+prefetch_read(uniform_int16, i16)
+prefetch_read(uniform_int32, i32)
+prefetch_read(uniform_int64, i64)
+prefetch_read(uniform_float, float)
+prefetch_read(uniform_double, double)
+
+prefetch_read(varying_bool, <$1 x i32>)
+prefetch_read(varying_int8, <$1 x i8>)
+prefetch_read(varying_int16, <$1 x i16>)
+prefetch_read(varying_int32, <$1 x i32>)
+prefetch_read(varying_int64, <$1 x i64>)
+prefetch_read(varying_float, <$1 x float>)
+prefetch_read(varying_double, <$1 x double>)
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; stdlib transcendentals
 ;;
@@ -911,25 +1203,35 @@ define internal void @__memory_barrier() nounwind readnone alwaysinline {
   ret void
 }
 
-global_atomic($1, add, i32, int32)
-global_atomic($1, sub, i32, int32)
-global_atomic($1, and, i32, int32)
-global_atomic($1, or, i32, int32)
-global_atomic($1, xor, i32, int32)
-global_atomic($1, min, i32, int32)
-global_atomic($1, max, i32, int32)
-global_atomic($1, umin, i32, uint32)
-global_atomic($1, umax, i32, uint32)
+global_atomic_associative($1, add, i32, int32, 0)
+global_atomic_associative($1, sub, i32, int32, 0)
+global_atomic_associative($1, and, i32, int32, -1)
+global_atomic_associative($1, or, i32, int32, 0)
+global_atomic_associative($1, xor, i32, int32, 0)
+global_atomic_uniform($1, add, i32, int32)
+global_atomic_uniform($1, sub, i32, int32)
+global_atomic_uniform($1, and, i32, int32)
+global_atomic_uniform($1, or, i32, int32)
+global_atomic_uniform($1, xor, i32, int32)
+global_atomic_uniform($1, min, i32, int32)
+global_atomic_uniform($1, max, i32, int32)
+global_atomic_uniform($1, umin, i32, uint32)
+global_atomic_uniform($1, umax, i32, uint32)
 
-global_atomic($1, add, i64, int64)
-global_atomic($1, sub, i64, int64)
-global_atomic($1, and, i64, int64)
-global_atomic($1, or, i64, int64)
-global_atomic($1, xor, i64, int64)
-global_atomic($1, min, i64, int64)
-global_atomic($1, max, i64, int64)
-global_atomic($1, umin, i64, uint64)
-global_atomic($1, umax, i64, uint64)
+global_atomic_associative($1, add, i64, int64, 0)
+global_atomic_associative($1, sub, i64, int64, 0)
+global_atomic_associative($1, and, i64, int64, -1)
+global_atomic_associative($1, or, i64, int64, 0)
+global_atomic_associative($1, xor, i64, int64, 0)
+global_atomic_uniform($1, add, i64, int64)
+global_atomic_uniform($1, sub, i64, int64)
+global_atomic_uniform($1, and, i64, int64)
+global_atomic_uniform($1, or, i64, int64)
+global_atomic_uniform($1, xor, i64, int64)
+global_atomic_uniform($1, min, i64, int64)
+global_atomic_uniform($1, max, i64, int64)
+global_atomic_uniform($1, umin, i64, uint64)
+global_atomic_uniform($1, umax, i64, uint64)
 
 global_swap($1, i32, int32)
 global_swap($1, i64, int64)
@@ -952,6 +1254,24 @@ define internal <$1 x double> @__atomic_swap_double_global(double * %ptr, <$1 x
   ret <$1 x double> %ret
 }
 
+define internal float @__atomic_swap_uniform_float_global(float * %ptr, float %val,
+                                                   <$1 x i32> %mask) nounwind alwaysinline {
+  %iptr = bitcast float * %ptr to i32 *
+  %ival = bitcast float %val to i32
+  %iret = call i32 @__atomic_swap_uniform_int32_global(i32 * %iptr, i32 %ival, <$1 x i32> %mask)
+  %ret = bitcast i32 %iret to float
+  ret float %ret
+}
+
+define internal double @__atomic_swap_uniform_double_global(double * %ptr, double %val,
+                                                   <$1 x i32> %mask) nounwind alwaysinline {
+  %iptr = bitcast double * %ptr to i64 *
+  %ival = bitcast double %val to i64
+  %iret = call i64 @__atomic_swap_uniform_int64_global(i64 * %iptr, i64 %ival, <$1 x i32> %mask)
+  %ret = bitcast i64 %iret to double
+  ret double %ret
+}
+
 global_atomic_exchange($1, i32, int32)
 global_atomic_exchange($1, i64, int64)
 
@@ -976,6 +1296,29 @@ define internal <$1 x double> @__atomic_compare_exchange_double_global(double *
   %ret = bitcast <$1 x i64> %iret to <$1 x double>
   ret <$1 x double> %ret
 }
+
+define internal float @__atomic_compare_exchange_uniform_float_global(float * %ptr, float %cmp, float %val,
+                                                   <$1 x i32> %mask) nounwind alwaysinline {
+  %iptr = bitcast float * %ptr to i32 *
+  %icmp = bitcast float %cmp to i32
+  %ival = bitcast float %val to i32
+  %iret = call i32 @__atomic_compare_exchange_uniform_int32_global(i32 * %iptr, i32 %icmp,
+                                                                   i32 %ival, <$1 x i32> %mask)
+  %ret = bitcast i32 %iret to float
+  ret float %ret
+}
+
+define internal double @__atomic_compare_exchange_uniform_double_global(double * %ptr, double %cmp,
+                                            double %val, <$1 x i32> %mask) nounwind alwaysinline {
+  %iptr = bitcast double * %ptr to i64 *
+  %icmp = bitcast double %cmp to i64
+  %ival = bitcast double %val to i64
+  %iret = call i64 @__atomic_compare_exchange_uniform_int64_global(i64 * %iptr, i64 %icmp,
+                                                                   i64 %ival, <$1 x i32> %mask)
+  %ret = bitcast i64 %iret to double
+  ret double %ret
+}
+
 ')
 
 
@@ -1034,12 +1377,6 @@ i64minmax($1,max,uint64,ugt)
 
 define(`load_and_broadcast', `
 define <$1 x $2> @__load_and_broadcast_$3(i8 *, <$1 x i32> %mask) nounwind alwaysinline {
-  ; must not load if the mask is all off; the address may be invalid
-  %mm = call i32 @__movmsk(<$1 x i32> %mask)
-  %any_on = icmp ne i32 %mm, 0
-  br i1 %any_on, label %load, label %skip
-
-load:
   %ptr = bitcast i8 * %0 to $2 *
   %val = load $2 * %ptr
 
@@ -1047,9 +1384,6 @@ load:
   forloop(i, 1, eval($1-1), `
   %ret`'i = insertelement <$1 x $2> %ret`'eval(i-1), $2 %val, i32 i')
   ret <$1 x $2> %ret`'eval($1-1)
-
-skip:
-  ret <$1 x $2> undef
 }
 ')
 
@@ -1065,14 +1399,20 @@ define(`load_masked', `
 define <$1 x $2> @__load_masked_$3(i8 *, <$1 x i32> %mask) nounwind alwaysinline {
 entry:
   %mm = call i32 @__movmsk(<$1 x i32> %mask)
+  
   ; if the first lane and the last lane are on, then it is safe to do a vector load
   ; of the whole thing--what the lanes in the middle want turns out to not matter...
   %mm_and = and i32 %mm, eval(1 | (1<<($1-1)))
   %can_vload = icmp eq i32 %mm_and, eval(1 | (1<<($1-1)))
+
+  %fast32 = call i32 @__fast_masked_vload()
+  %fast_i1 = trunc i32 %fast32 to i1
+  %can_vload_maybe_fast = or i1 %fast_i1, %can_vload
+
   ; if we are not able to do a singe vload, we will accumulate lanes in this memory..
   %retptr = alloca <$1 x $2>
   %retptr32 = bitcast <$1 x $2> * %retptr to $2 *
-  br i1 %can_vload, label %load, label %loop
+  br i1 %can_vload_maybe_fast, label %load, label %loop
 
 load: 
   %ptr = bitcast i8 * %0 to <$1 x $2> *
@@ -1207,6 +1547,46 @@ define void @__masked_store_blend_16(<8 x i16>* nocapture, <8 x i16>,
 ')
 
 
+define(`masked_store_blend_8_16_by_16', `
+define void @__masked_store_blend_8(<16 x i8>* nocapture, <16 x i8>,
+                                    <16 x i32>) nounwind alwaysinline {
+  %old = load <16 x i8> * %0
+  %old128 = bitcast <16 x i8> %old to i128
+  %new128 = bitcast <16 x i8> %1 to i128
+
+  %mask8 = trunc <16 x i32> %2 to <16 x i8>
+  %mask128 = bitcast <16 x i8> %mask8 to i128
+  %notmask128 = xor i128 %mask128, -1
+
+  %newmasked = and i128 %new128, %mask128
+  %oldmasked = and i128 %old128, %notmask128
+  %result = or i128 %newmasked, %oldmasked
+
+  %resultvec = bitcast i128 %result to <16 x i8>
+  store <16 x i8> %resultvec, <16 x i8> * %0
+  ret void
+}
+
+define void @__masked_store_blend_16(<16 x i16>* nocapture, <16 x i16>,
+                                     <16 x i32>) nounwind alwaysinline {
+  %old = load <16 x i16> * %0
+  %old256 = bitcast <16 x i16> %old to i256
+  %new256 = bitcast <16 x i16> %1 to i256
+
+  %mask16 = trunc <16 x i32> %2 to <16 x i16>
+  %mask256 = bitcast <16 x i16> %mask16 to i256
+  %notmask256 = xor i256 %mask256, -1
+
+  %newmasked = and i256 %new256, %mask256
+  %oldmasked = and i256 %old256, %notmask256
+  %result = or i256 %newmasked, %oldmasked
+
+  %resultvec = bitcast i256 %result to <16 x i16>
+  store <16 x i16> %resultvec, <16 x i16> * %0
+  ret void
+}
+')
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; packed load and store functions
 ;;
@@ -1234,7 +1614,7 @@ entry:
 
 known_mask:
   %allon = icmp eq i32 %mask, eval((1 << $1) -1)
-  br i1 %allon, label %all_on, label %not_all_on
+  br i1 %allon, label %all_on, label %unknown_mask
 
 all_on:
   ;; everyone wants to load, so just load an entire vector width in a single
@@ -1244,14 +1624,6 @@ all_on:
   store <$1 x i32> %vec_load, <$1 x i32> * %val_ptr, align 4
   ret i32 $1
 
-not_all_on:
-  %alloff = icmp eq i32 %mask, 0
-  br i1 %alloff, label %all_off, label %unknown_mask
-
-all_off:
-  ;; no one wants to load
-  ret i32 0
-
 unknown_mask:
   br label %loop
 
@@ -1298,20 +1670,13 @@ entry:
 
 known_mask:
   %allon = icmp eq i32 %mask, eval((1 << $1) -1)
-  br i1 %allon, label %all_on, label %not_all_on
+  br i1 %allon, label %all_on, label %unknown_mask
 
 all_on:
   %vecptr = bitcast i32 *%startptr to <$1 x i32> *
   store <$1 x i32> %vals, <$1 x i32> * %vecptr, align 4
   ret i32 $1
 
-not_all_on:
-  %alloff = icmp eq i32 %mask, 0
-  br i1 %alloff, label %all_off, label %unknown_mask
-
-all_off:
-  ret i32 0
-
 unknown_mask:
   br label %loop
 
@@ -1346,6 +1711,150 @@ done:
 }
 ')
 
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; reduce_equal
+
+; count leading zeros
+declare i32 @llvm.cttz.i32(i32)
+
+define(`reduce_equal_aux', `
+define internal i1 @__reduce_equal_$3(<$1 x $2> %v, $2 * %samevalue,
+                                      <$1 x i32> %mask) nounwind alwaysinline {
+entry:
+   %mm = call i32 @__movmsk(<$1 x i32> %mask)
+   %allon = icmp eq i32 %mm, eval((1<<$1)-1)
+   br i1 %allon, label %check_neighbors, label %domixed
+
+domixed:
+  ; First, figure out which lane is the first active one
+  %first = call i32 @llvm.cttz.i32(i32 %mm)
+  %baseval = extractelement <$1 x $2> %v, i32 %first
+  %basev1 = bitcast $2 %baseval to <1 x $2>
+  ; get a vector that is that value smeared across all elements
+  %basesmear = shufflevector <1 x $2> %basev1, <1 x $2> undef,
+        <$1 x i32> < forloop(i, 0, eval($1-2), `i32 0, ') i32 0 >
+
+  ; now to a blend of that vector with the original vector, such that the
+  ; result will be the original value for the active lanes, and the value
+  ; from the first active lane for the inactive lanes.  Given that, we can
+  ; just unconditionally check if the lanes are all equal in check_neighbors
+  ; below without worrying about inactive lanes...
+  %ptr = alloca <$1 x $2>
+  store <$1 x $2> %basesmear, <$1 x $2> * %ptr
+  %castptr = bitcast <$1 x $2> * %ptr to <$1 x $4> *
+  %castv = bitcast <$1 x $2> %v to <$1 x $4>
+  call void @__masked_store_blend_$6(<$1 x $4> * %castptr, <$1 x $4> %castv, <$1 x i32> %mask)
+  %blendvec = load <$1 x $2> * %ptr
+  br label %check_neighbors
+
+check_neighbors:
+  %vec = phi <$1 x $2> [ %blendvec, %domixed ], [ %v, %entry ]
+  ifelse($6, `32', `
+  ; For 32-bit elements, we rotate once and compare with the vector, which ends 
+  ; up comparing each element to its neighbor on the right.  Then see if
+  ; all of those values are true; if so, then all of the elements are equal..
+  %castvec = bitcast <$1 x $2> %vec to <$1 x $4>
+  %castvr = call <$1 x $4> @__rotate_int$6(<$1 x $4> %castvec, i32 1)
+  %vr = bitcast <$1 x $4> %castvr to <$1 x $2>
+  %eq = $5 eq <$1 x $2> %vec, %vr
+  %eq32 = sext <$1 x i1> %eq to <$1 x i32>
+  %eqmm = call i32 @__movmsk(<$1 x i32> %eq32)
+  %alleq = icmp eq i32 %eqmm, eval((1<<$1)-1)
+  br i1 %alleq, label %all_equal, label %not_all_equal
+  ', `
+  ; But for 64-bit elements, it turns out to be more efficient to just
+  ; scalarize and do a individual pairwise comparisons and AND those
+  ; all together..
+  forloop(i, 0, eval($1-1), `
+  %v`'i = extractelement <$1 x $2> %vec, i32 i')
+
+  forloop(i, 0, eval($1-2), `
+  %eq`'i = $5 eq $2 %v`'i, %v`'eval(i+1)')
+
+  %and0 = and i1 %eq0, %eq1
+  forloop(i, 1, eval($1-3), `
+  %and`'i = and i1 %and`'eval(i-1), %eq`'eval(i+1)')
+
+  br i1 %and`'eval($1-3), label %all_equal, label %not_all_equal
+  ')
+
+all_equal:
+  %the_value = extractelement <$1 x $2> %vec, i32 0
+  store $2 %the_value, $2 * %samevalue
+  ret i1 true
+
+not_all_equal:
+  ret i1 false
+}
+')
+
+define(`reduce_equal', `
+reduce_equal_aux($1, i32, int32, i32, icmp, 32)
+reduce_equal_aux($1, float, float, i32, fcmp, 32)
+reduce_equal_aux($1, i64, int64, i64, icmp, 64)
+reduce_equal_aux($1, double, double, i64, fcmp, 64)
+')
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;; prefix sum stuff
+
+; $1: vector width (e.g. 4)
+; $2: vector element type (e.g. float)
+; $3: bit width of vector element type (e.g. 32)
+; $4: operator to apply (e.g. fadd)
+; $5: identity element value (e.g. 0)
+; $6: suffix for function (e.g. add_float)
+
+define(`exclusive_scan', `
+define internal <$1 x $2> @__exclusive_scan_$6(<$1 x $2> %v,
+                                  <$1 x i32> %mask) nounwind alwaysinline {
+  ; first, set the value of any off lanes to the identity value
+  %ptr = alloca <$1 x $2>
+  %idvec1 = bitcast $2 $5 to <1 x $2>
+  %idvec = shufflevector <1 x $2> %idvec1, <1 x $2> undef,
+      <$1 x i32> < forloop(i, 0, eval($1-2), `i32 0, ') i32 0 >
+  store <$1 x $2> %idvec, <$1 x $2> * %ptr
+  %ptr`'$3 = bitcast <$1 x $2> * %ptr to <$1 x i`'$3> *
+  %vi = bitcast <$1 x $2> %v to <$1 x i`'$3>
+  call void @__masked_store_blend_$3(<$1 x i`'$3> * %ptr`'$3, <$1 x i`'$3> %vi,
+                                     <$1 x i32> %mask)
+  %v_id = load <$1 x $2> * %ptr
+
+  ; extract elements of the vector to use in computing the scan
+  forloop(i, 0, eval($1-1), `
+  %v`'i = extractelement <$1 x $2> %v_id, i32 i')
+
+  ; and just compute the scan directly.
+  ; 0th element is the identity (so nothing to do here),
+  ; 1st element is identity (op) the 0th element of the original vector,
+  ; each successive element is the previous element (op) the previous element
+  ;  of the original vector
+  %s1 = $4 $2 $5, %v0
+  forloop(i, 2, eval($1-1), `
+  %s`'i = $4 $2 %s`'eval(i-1), %v`'eval(i-1)')
+
+  ; and fill in the result vector
+  %r0 = insertelement <$1 x $2> undef, $2 $5, i32 0  ; 0th element gets identity
+  forloop(i, 1, eval($1-1), `
+  %r`'i = insertelement <$1 x $2> %r`'eval(i-1), $2 %s`'i, i32 i')
+
+  ret <$1 x $2> %r`'eval($1-1)
+}
+')
+
+define(`scans', `
+exclusive_scan($1, i32, 32, add, 0, add_i32)
+exclusive_scan($1, float, 32, fadd, zeroinitializer, add_float)
+exclusive_scan($1, i64, 64, add, 0, add_i64)
+exclusive_scan($1, double, 64, fadd, zeroinitializer, add_double)
+
+exclusive_scan($1, i32, 32, and, -1, and_i32)
+exclusive_scan($1, i64, 64, and, -1, and_i64)
+
+exclusive_scan($1, i32, 32, or, 0, or_i32)
+exclusive_scan($1, i64, 64, or, 0, or_i64)
+')
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; per_lane
 ;;
@@ -1371,7 +1880,7 @@ pl_known_mask:
   ;; the mask is known at compile time; see if it is something we can
   ;; handle more efficiently
   %pl_is_allon = icmp eq i32 %pl_mask, eval((1<<$1)-1)
-  br i1 %pl_is_allon, label %pl_all_on, label %pl_not_all_on
+  br i1 %pl_is_allon, label %pl_all_on, label %pl_unknown_mask
 
 pl_all_on:
   ;; the mask is all on--just expand the code for each lane sequentially
@@ -1379,19 +1888,14 @@ pl_all_on:
           `patsubst(`$3', `ID\|LANE', i)')
   br label %pl_done
 
-pl_not_all_on:
-  ;; not all on--see if it is all off or mixed
-  ;; for the mixed case, we just run the general case, though we could
+pl_unknown_mask:
+  ;; we just run the general case, though we could
   ;; try to be smart and just emit the code based on what it actually is,
   ;; for example by emitting the code straight-line without a loop and doing 
   ;; the lane tests explicitly, leaving later optimization passes to eliminate
   ;; the stuff that is definitely not needed.  Not clear if we will frequently 
   ;; encounter a mask that is known at compile-time but is not either all on or
   ;; all off...
-  %pl_alloff = icmp eq i32 %pl_mask, 0
-  br i1 %pl_alloff, label %pl_done, label %pl_unknown_mask
-
-pl_unknown_mask:
   br label %pl_loop
 
 pl_loop:
@@ -1447,20 +1951,6 @@ define internal <$1 x $2> @__gather_elt_$2(i8 * %ptr, <$1 x i32> %offsets, <$1 x
 
 define <$1 x $2> @__gather_base_offsets_$2(i8 * %ptr, <$1 x i32> %offsets,
                                            <$1 x i32> %vecmask) nounwind readonly alwaysinline {
-entry:
-  %mask = call i32 @__movmsk(<$1 x i32> %vecmask)
-
-  %maskKnown = call i1 @__is_compile_time_constant_mask(<$1 x i32> %vecmask)
-  br i1 %maskKnown, label %known_mask, label %unknown_mask
-
-known_mask:
-  %alloff = icmp eq i32 %mask, 0
-  br i1 %alloff, label %gather_all_off, label %unknown_mask
-
-gather_all_off:
-  ret <$1 x $2> undef
-
-unknown_mask:
   ; We can be clever and avoid the per-lane stuff for gathers if we are willing
   ; to require that the 0th element of the array being gathered from is always
   ; legal to read from (and we do indeed require that, given the benefits!) 

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.cpp

@@ -144,6 +144,11 @@ FunctionEmitContext::FunctionEmitContext(const Type *rt, llvm::Function *functio
     returnedLanesPtr = AllocaInst(LLVMTypes::MaskType, "returned_lanes_memory");
     StoreInst(LLVMMaskAllOff, returnedLanesPtr);
 
+    launchedTasks = false;
+    launchGroupHandlePtr = AllocaInst(LLVMTypes::VoidPointerType, "launch_group_handle");
+    StoreInst(llvm::Constant::getNullValue(LLVMTypes::VoidPointerType), 
+              launchGroupHandlePtr);
+
     if (!returnType || returnType == AtomicType::Void)
         returnValuePtr = NULL;
     else {
@@ -153,7 +158,6 @@ FunctionEmitContext::FunctionEmitContext(const Type *rt, llvm::Function *functio
         StoreInst(llvm::Constant::getNullValue(ftype), returnValuePtr);
     }
 
-#ifndef LLVM_2_8
     if (m->diBuilder) {
         /* If debugging is enabled, tell the debug information emission
            code about this new function */
@@ -174,16 +178,12 @@ FunctionEmitContext::FunctionEmitContext(const Type *rt, llvm::Function *functio
         /* And start a scope representing the initial function scope */
         StartScope();
     }
-#endif // LLVM_2_8
-
-    launchedTasks = false;
 
     // connect the funciton's mask memory to the __mask symbol
     Symbol *maskSymbol = m->symbolTable->LookupVariable("__mask");
     assert(maskSymbol != NULL);
     maskSymbol->storagePtr = maskPtr;
 
-#ifndef LLVM_2_8
     // add debugging info for __mask, programIndex, ...
     if (m->diBuilder) {
         maskSymbol->pos = funcStartPos;
@@ -208,15 +208,12 @@ FunctionEmitContext::FunctionEmitContext(const Type *rt, llvm::Function *functio
                                            true /* static */,
                                            programCountSymbol->storagePtr);
     }
-#endif
 }
 
 
 FunctionEmitContext::~FunctionEmitContext() {
     assert(controlFlowInfo.size() == 0);
-#ifndef LLVM_2_8
     assert(debugScopes.size() == (m->diBuilder ? 1 : 0));
-#endif
 }
 
 
@@ -704,6 +701,7 @@ FunctionEmitContext::LaneMask(llvm::Value *v) {
 
 llvm::Value *
 FunctionEmitContext::MasksAllEqual(llvm::Value *v1, llvm::Value *v2) {
+#if 0
     // Compare the two masks to get a vector of i1s
     llvm::Value *cmp = CmpInst(llvm::Instruction::ICmp, llvm::CmpInst::ICMP_EQ,
                                v1, v2, "v1==v2");
@@ -711,6 +709,12 @@ FunctionEmitContext::MasksAllEqual(llvm::Value *v1, llvm::Value *v2) {
     cmp = I1VecToBoolVec(cmp);
     // And see if it's all on
     return All(cmp);
+#else
+    llvm::Value *mm1 = LaneMask(v1);
+    llvm::Value *mm2 = LaneMask(v2);
+    return CmpInst(llvm::Instruction::ICmp, llvm::CmpInst::ICMP_EQ, mm1, mm2,
+                   "v1==v2");
+#endif
 }
 
 
@@ -758,7 +762,7 @@ FunctionEmitContext::I1VecToBoolVec(llvm::Value *b) {
 
 
 llvm::Value *
-FunctionEmitContext::EmitMalloc(LLVM_TYPE_CONST llvm::Type *ty, int align) {
+FunctionEmitContext::SizeOf(LLVM_TYPE_CONST llvm::Type *ty) {
     // Emit code to compute the size of the given type using a GEP with a
     // NULL base pointer, indexing one element of the given type, and
     // casting the resulting 'pointer' to an int giving its size.
@@ -775,24 +779,7 @@ FunctionEmitContext::EmitMalloc(LLVM_TYPE_CONST llvm::Type *ty, int align) {
 #endif
     AddDebugPos(poffset);
     llvm::Value *sizeOf = PtrToIntInst(poffset, LLVMTypes::Int64Type, "offset_int");
-
-    // And given the size, call the malloc function
-    llvm::Function *fmalloc = m->module->getFunction("ISPCMalloc");
-    assert(fmalloc != NULL);
-    llvm::Value *mem = CallInst(fmalloc, sizeOf, LLVMInt32(align), 
-                                "raw_argmem");
-    // Cast the void * back to the result pointer type
-    return BitCastInst(mem, ptrType, "mem_bitcast");
-}
-
-
-void
-FunctionEmitContext::EmitFree(llvm::Value *ptr) {
-    llvm::Value *freeArg = BitCastInst(ptr, LLVMTypes::VoidPointerType,
-                                       "argmemfree");
-    llvm::Function *ffree = m->module->getFunction("ISPCFree");
-    assert(ffree != NULL);
-    CallInst(ffree, freeArg);
+    return sizeOf;
 }
 
 
@@ -850,7 +837,6 @@ FunctionEmitContext::GetDebugPos() const {
 void
 FunctionEmitContext::AddDebugPos(llvm::Value *value, const SourcePos *pos, 
                                  llvm::DIScope *scope) {
-#ifndef LLVM_2_8
     llvm::Instruction *inst = llvm::dyn_cast<llvm::Instruction>(value);
     if (inst != NULL && m->diBuilder) {
         SourcePos p = pos ? *pos : currentPos;
@@ -861,13 +847,11 @@ FunctionEmitContext::AddDebugPos(llvm::Value *value, const SourcePos *pos,
             inst->setDebugLoc(llvm::DebugLoc::get(p.first_line, p.first_column, 
                                                   scope ? *scope : GetDIScope()));
     }
-#endif
 }
 
 
 void
 FunctionEmitContext::StartScope() {
-#ifndef LLVM_2_8
     if (m->diBuilder != NULL) {
         llvm::DIScope parentScope;
         if (debugScopes.size() > 0)
@@ -881,18 +865,15 @@ FunctionEmitContext::StartScope() {
                                              currentPos.first_column);
         debugScopes.push_back(lexicalBlock);
     }
-#endif
 }
 
 
 void
 FunctionEmitContext::EndScope() {
-#ifndef LLVM_2_8
     if (m->diBuilder != NULL) {
         assert(debugScopes.size() > 0);
         debugScopes.pop_back();
     }
-#endif
 }
 
 
@@ -905,7 +886,6 @@ FunctionEmitContext::GetDIScope() const {
 
 void
 FunctionEmitContext::EmitVariableDebugInfo(Symbol *sym) {
-#ifndef LLVM_2_8
     if (m->diBuilder == NULL)
         return;
 
@@ -921,13 +901,11 @@ FunctionEmitContext::EmitVariableDebugInfo(Symbol *sym) {
     llvm::Instruction *declareInst = 
         m->diBuilder->insertDeclare(sym->storagePtr, var, bblock);
     AddDebugPos(declareInst, &sym->pos, &scope);
-#endif
 }
 
 
 void
 FunctionEmitContext::EmitFunctionParameterDebugInfo(Symbol *sym) {
-#ifndef LLVM_2_8
     if (m->diBuilder == NULL)
         return;
 
@@ -943,7 +921,6 @@ FunctionEmitContext::EmitFunctionParameterDebugInfo(Symbol *sym) {
     llvm::Instruction *declareInst = 
         m->diBuilder->insertDeclare(sym->storagePtr, var, bblock);
     AddDebugPos(declareInst, &sym->pos, &scope);
-#endif
 }
 
 
@@ -1501,27 +1478,15 @@ FunctionEmitContext::gather(llvm::Value *lvalue, const Type *type,
 void
 FunctionEmitContext::addGSMetadata(llvm::Instruction *inst, SourcePos pos) {
     llvm::Value *str = llvm::MDString::get(*g->ctx, pos.name);
-#ifdef LLVM_2_8
-    llvm::MDNode *md = llvm::MDNode::get(*g->ctx, &str, 1);
-#else
     llvm::MDNode *md = llvm::MDNode::get(*g->ctx, str);
-#endif
     inst->setMetadata("filename", md);
 
     llvm::Value *line = LLVMInt32(pos.first_line);
-#ifdef LLVM_2_8
-    md = llvm::MDNode::get(*g->ctx, &line, 1);
-#else
     md = llvm::MDNode::get(*g->ctx, line);
-#endif
     inst->setMetadata("line", md);
 
     llvm::Value *column = LLVMInt32(pos.first_column);
-#ifdef LLVM_2_8
-    md = llvm::MDNode::get(*g->ctx, &column, 1);
-#else
     md = llvm::MDNode::get(*g->ctx, column);
-#endif
     inst->setMetadata("column", md);
 }
 
@@ -1838,9 +1803,9 @@ llvm::PHINode *
 FunctionEmitContext::PhiNode(LLVM_TYPE_CONST llvm::Type *type, int count, 
                              const char *name) {
     llvm::PHINode *pn = llvm::PHINode::Create(type, 
-#if !defined(LLVM_2_8) && !defined(LLVM_2_9)
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
                                               count, 
-#endif // !LLVM_2_8 && !LLVM_2_9
+#endif // LLVM_3_0
                                               name ? name : "phi", bblock);
     AddDebugPos(pn);
     return pn;
@@ -1933,15 +1898,9 @@ FunctionEmitContext::CallInst(llvm::Function *func, llvm::Value *arg0,
 
 llvm::Instruction *
 FunctionEmitContext::ReturnInst() {
-    if (launchedTasks) {
-        // Automatically add a sync call at the end of any function that
-        // launched tasks
-        SourcePos noPos;
-        noPos.name = "__auto_sync";
-        ExprStmt *es = new ExprStmt(new SyncExpr(noPos), noPos);
-        es->EmitCode(this); 
-        delete es;
-    }
+    if (launchedTasks)
+        // Add a sync call at the end of any function that launched tasks
+        SyncInst();
 
     llvm::Instruction *rinst = NULL;
     if (returnValuePtr != NULL) {
@@ -1964,7 +1923,8 @@ FunctionEmitContext::ReturnInst() {
 
 llvm::Instruction *
 FunctionEmitContext::LaunchInst(llvm::Function *callee, 
-                                std::vector<llvm::Value *> &argVals) {
+                                std::vector<llvm::Value *> &argVals,
+                                llvm::Value *launchCount) {
     if (callee == NULL) {
         assert(m->errorCount > 0);
         return NULL;
@@ -1981,20 +1941,15 @@ FunctionEmitContext::LaunchInst(llvm::Function *callee,
         static_cast<LLVM_TYPE_CONST llvm::StructType *>(pt->getElementType());
     assert(argStructType->getNumElements() == argVals.size() + 1);
 
+    llvm::Function *falloc = m->module->getFunction("ISPCAlloc");
+    assert(falloc != NULL);
     int align = 4 * RoundUpPow2(g->target.nativeVectorWidth);
-#ifdef ISPC_IS_WINDOWS
-    // Use malloc() to allocate storage on Windows, since the stack is
-    // generally not big enough there to do enough allocations for lots of
-    // tasks and then things crash horribly...
-    llvm::Value *argmem = EmitMalloc(argStructType, align);
-#else
-    // Use alloca for space for the task args on OSX And Linux.  KEY
-    // DETAIL: pass false to the call of FunctionEmitContext::AllocaInst so
-    // that the alloca doesn't happen just once at the top of the function,
-    // but happens each time the enclosing basic block executes.
-    llvm::Value *argmem = AllocaInst(argStructType, "argmem", align, false);
-#endif // ISPC_IS_WINDOWS
-    llvm::Value *voidmem = BitCastInst(argmem, LLVMTypes::VoidPointerType);
+    std::vector<llvm::Value *> allocArgs;
+    allocArgs.push_back(launchGroupHandlePtr);
+    allocArgs.push_back(SizeOf(argStructType));
+    allocArgs.push_back(LLVMInt32(align));
+    llvm::Value *voidmem = CallInst(falloc, allocArgs, "args_ptr");
+    llvm::Value *argmem = BitCastInst(voidmem, pt);
 
     // Copy the values of the parameters into the appropriate place in
     // the argument block
@@ -2016,5 +1971,32 @@ FunctionEmitContext::LaunchInst(llvm::Function *callee,
     llvm::Value *fptr = BitCastInst(callee, LLVMTypes::VoidPointerType);
     llvm::Function *flaunch = m->module->getFunction("ISPCLaunch");
     assert(flaunch != NULL);
-    return CallInst(flaunch, fptr, voidmem, "");
+    std::vector<llvm::Value *> args;
+    args.push_back(launchGroupHandlePtr);
+    args.push_back(fptr);
+    args.push_back(voidmem);
+    args.push_back(launchCount);
+    return CallInst(flaunch, args, "");
+}
+
+
+void
+FunctionEmitContext::SyncInst() {
+    llvm::Value *launchGroupHandle = LoadInst(launchGroupHandlePtr, NULL);
+    llvm::Value *nullPtrValue = llvm::Constant::getNullValue(LLVMTypes::VoidPointerType);
+    llvm::Value *nonNull = CmpInst(llvm::Instruction::ICmp,
+                                   llvm::CmpInst::ICMP_NE,
+                                   launchGroupHandle, nullPtrValue);
+    llvm::BasicBlock *bSync = CreateBasicBlock("call_sync");
+    llvm::BasicBlock *bPostSync = CreateBasicBlock("post_sync");
+    BranchInst(bSync, bPostSync, nonNull);
+
+    SetCurrentBasicBlock(bSync);
+    llvm::Function *fsync = m->module->getFunction("ISPCSync");
+    if (fsync == NULL)
+        FATAL("Couldn't find ISPCSync declaration?!");
+    CallInst(fsync, launchGroupHandle, "");
+    BranchInst(bPostSync);
+
+    SetCurrentBasicBlock(bPostSync);
 }

ctx.h

@@ -210,15 +210,8 @@ 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(LLVM_TYPE_CONST llvm::Type *ty, int align = 0);
-
-    /** 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);
+    /** Returns the size of the given type. */
+    llvm::Value *SizeOf(LLVM_TYPE_CONST llvm::Type *ty);
 
     /** If the user has asked to compile the program with instrumentation,
         this inserts a callback to the user-supplied instrumentation
@@ -399,7 +392,10 @@ public:
     /** 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);
+                                  std::vector<llvm::Value *> &argVals,
+                                  llvm::Value *launchCount);
+
+    void SyncInst();
 
     llvm::Instruction *ReturnInst();
     /** @} */
@@ -489,6 +485,11 @@ 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);
     bool ifsInLoopAllUniform() const;

decl.cpp

@@ -237,7 +237,7 @@ Declarator::GetType(DeclSpecs *ds) const {
                     sprintf(buf, "__anon_parameter_%d", i);
                     sym = new Symbol(buf, pos);
                     Declarator *declarator = new Declarator(sym, sym->pos);
-                    sym->type = declarator->GetType(ds);
+                    sym->type = declarator->GetType(d->declSpecs);
                     d->declarators.push_back(declarator);
                 }
                 else {

docs/ReleaseNotes.txt

@@ -1,3 +1,119 @@
+=== 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

docs/ispc.txt

@@ -33,6 +33,17 @@ The main goals behind ``ispc`` are to:
 number of non-trivial workloads that aren't handled well by other
 compilation approaches (e.g. loop auto-vectorization.)
 
+**We are very interested in your feedback and comments about ispc and
+in hearing your experiences using the system.  We are especially interested
+in hearing if you try using ispc but see results that are not as you
+were expecting or hoping for.** We encourage you to send a note with your
+experiences or comments to the `ispc-users`_ mailing list or to file bug or
+feature requests with the ``ispc`` `bug tracker`_. (Thanks!)
+
+.. _ispc-users: http://groups.google.com/group/ispc-users
+.. _bug tracker: https://github.com/ispc/ispc/issues?state=open
+
+
 Contents:
 
 * `Recent Changes to ISPC`_
@@ -69,7 +80,8 @@ Contents:
   + `Program Instance Convergence`_
   + `Data Races`_
   + `Uniform Variables and Varying Control Flow`_
-  + `Task Parallelism in ISPC`_
+  + `Task Parallelism: Language Syntax`_
+  + `Task Parallelism: Runtime Requirements`_
 
 * `The ISPC Standard Library`_
 
@@ -79,6 +91,8 @@ Contents:
   + `Packed Load and Store Operations`_
   + `Conversions To and From Half-Precision Floats`_
   + `Atomic Operations and Memory Fences`_
+  + `Prefetches`_
+  + `System Information`_
   + `Low-Level Bits`_
 
 * `Interoperability with the Application`_
@@ -100,6 +114,9 @@ Contents:
   + `"Inline" Aggressively`_
   + `Small Performance Tricks`_
   + `Instrumenting Your ISPC Programs`_
+  + `Using Scan Operations For Variable Output`_
+  + `Application-Supplied Execution Masks`_
+  + `Explicit Vector Programming With Uniform Short Vector Types`_
 
 * `Disclaimer and Legal Information`_
 
@@ -822,8 +839,8 @@ by default.  If a function is declared with a ``static`` qualifier, then it
 is only visible in the file in which it was declared.
 
 Any function that can be launched with the ``launch`` construct in ``ispc``
-must have a ``task`` qualifier; see `Task Parallelism in ISPC`_ for more
-discussion of launching tasks in ``ispc``.
+must have a ``task`` qualifier; see `Task Parallelism: Language Syntax`_
+for more discussion of launching tasks in ``ispc``.
 
 Functions that are intended to be called from C/C++ application code must
 have the ``export`` qualifier.  This causes them to have regular C linkage
@@ -924,8 +941,9 @@ execution model is critical for writing efficient and correct programs in
 
 ``ispc`` supports both task parallelism to parallelize across multiple
 cores and SPMD parallelism to parallelize across the SIMD vector lanes on a
-single core.  This section focuses on SPMD parallelism.  See the section
-`Task Parallelism in ISPC`_ for discussion of task parallelism in ``ispc``.
+single core.  This section focuses on SPMD parallelism.  See the sections
+`Task Parallelism: Language Syntax`_ and `Task Parallelism: Runtime
+Requirements`_ for discussion of task parallelism in ``ispc``.
 
 The SPMD-on-SIMD Execution Model
 --------------------------------
@@ -1172,7 +1190,7 @@ This code implicitly assumes that ``programCount`` evenly divides
 ::
 
     for (uniform int i = 0; i < count; i += programCount) {
-        if (i + programIndex < programCount) {
+        if (i + programIndex < count) {
             float d = data[i + programIndex];
             ...
 
@@ -1368,112 +1386,190 @@ be modified in the above code even if *none* of the program instances
 evaluated a true value for the test, given the ``ispc`` execution model.
 
 
-Task Parallelism in ISPC
-------------------------
+Task Parallelism: Language Syntax
+---------------------------------
 
 One option for combining task-parallelism with ``ispc`` is to just use
 regular task parallelism in the C/C++ application code (be it through
-Intel® Cilk(tm), Intel® Thread Building Blocks or another task system,
-etc.), and for tasks to use ``ispc`` for SPMD parallelism across the vector
-lanes as appropriate.  Alternatively, ``ispc`` also has some support for
-launching tasks from ``ispc`` code.  The approach is similar to Intel®
-Cilk's task launch feature.  (See the ``examples/mandelbrot_tasks`` example
-to see it used in a non-trivial example.)
+Intel® Cilk(tm), Intel® Thread Building Blocks or another task system), and
+for tasks to use ``ispc`` for SPMD parallelism across the vector lanes as
+appropriate.  Alternatively, ``ispc`` also has support for launching tasks
+from ``ispc`` code.  The approach is similar to Intel® Cilk's task launch
+feature.  (See the ``examples/mandelbrot_tasks`` example to see it used in
+a small example.)
 
-Any function that is launched as a task must be declared with the ``task``
-qualifier:
+First, any function that is launched as a task must be declared with the
+``task`` qualifier:
 
 ::
 
-    task void func(uniform float a[], uniform int start) {
-        ....
+    task void func(uniform float a[], uniform int index) {
+        ...
+        a[index] = ....
     }
 
 Tasks must return ``void``; a compile time error is issued if a
 non-``void`` task is defined.
 
-Given a task, one can then write code that launches tasks as follows:
+Given a task definitions, there are two ways to write code that launches
+tasks, using the ``launch`` construct.  First, one task can be launched at
+a time, with parameters passed to the task to help it determine what part
+of the overall computation it's responsible for:
 
 ::
 
     for (uniform int i = 0; i < 100; ++i)
-        launch < func(a, i); >
+        launch < func(a, i) >;
 
 Note the ``launch`` keyword and the brackets around the function call.
 This code launches 100 tasks, each of which presumably does some
-computation keyed off of given the value ``i``.  In general, one should
-launch many more tasks than there are processors in the system to
+computation that is keyed off of given the value ``i``.  In general, one
+should launch many more tasks than there are processors in the system to
 ensure good load-balancing, but not so many that the overhead of scheduling
 and running tasks dominates the computation.
 
-Program execution continues asynchronously after task launch; thus, the
-function shouldn't access values being generated by the tasks without
-synchronization.  A function uses a ``sync`` statement to wait for all
-launched tasks to finish:
+Alternatively, a number of tasks may be launched from a single ``launch``
+statement.  We might instead write the above example with a single
+``launch`` like this:
 
 ::
 
-    for (uniform int i = 0; i < 100; ++i)
-        launch < func(a, i); >
+    launch[100] < func2(a) >;
+
+Where an integer value (not necessarily a compile-time constant) is
+provided to the ``launch`` keyword in square brackets; this number of tasks
+will be enqueued to be run asynchronously.  Within each of the tasks, two
+special built-in variables are available--``taskIndex``, and ``taskCount``.
+The first, ``taskIndex``, ranges from zero to one minus the number of tasks
+provided to ``launch``, and ``taskCount`` equals the number of launched
+taks.  Thus, we might use ``taskIndex`` in the implementation of ``func2``
+to determine which array element to process.
+
+::
+
+    task void func2(uniform float a[]) {
+        ...
+        a[taskIndex] = ...
+    }
+
+Program execution continues asynchronously after a ``launch`` statement;
+thus, a function shouldn't access values being generated by the tasks it
+has launched within the function without synchronization.  If results are
+needed before function return, a function can use a ``sync`` statement to
+wait for all launched tasks to finish:
+
+::
+
+    launch[100] < func2(a) >;
     sync;
     // now safe to use computed values in a[]...
 
-Alternatively, any function that launches tasks has an implicit ``sync``
-before it returns, so that functions that call a function that launches
-tasks don't have to worry about outstanding asynchronous computation.
+Alternatively, any function that launches tasks has an automatically-added
+``sync`` statement before it returns, so that functions that call a
+function that launches tasks don't have to worry about outstanding
+asynchronous computation from that function.
 
 Inside functions with the ``task`` qualifier, two additional built-in
-variables are provided: ``threadIndex`` and ``threadCount``.
-``threadCount`` gives the total number of hardware threads that have been
-launched by the task system.  ``threadIndex`` provides an index between
-zero and ``threadCount-1`` that gives a unique index that corresponds to
-the hardware thread that is executing the current task.  The
-``threadIndex`` can be used for accessing data that is private to the
-current thread and thus doesn't require synchronization to access under
-parallel execution.
+variables are provided in addition to ``taskIndex`` and ``taskCount``:
+``threadIndex`` and ``threadCount``.  ``threadCount`` gives the total
+number of hardware threads that have been launched by the task system.
+``threadIndex`` provides an index between zero and ``threadCount-1`` that
+gives a unique index that corresponds to the hardware thread that is
+executing the current task.  The ``threadIndex`` can be used for accessing
+data that is private to the current thread and thus doesn't require
+synchronization to access under parallel execution.
+
+Task Parallelism: Runtime Requirements
+--------------------------------------
 
 If you use the task launch feature in ``ispc``, you must provide C/C++
-implementations of two functions and link them into your final executable
-file.  Although these functions may be implemented in either language, they
-must have "C" linkage (i.e. their prototypes must be declared inside an
-``extern "C"`` block if they are defined in C++.)
+implementations of three specific functions that manage launching and
+synchronizing parallel tasks; these functions must be linked into your
+executable.  Although these functions may be implemented in any
+language, they must have "C" linkage (i.e. their prototypes must be
+declared inside an ``extern "C"`` block if they are defined in C++.)
+
+By using user-supplied versions of these functions, ``ispc`` programs can
+easily interoperate with software systems that have existing task systems
+for managing parallelism.  If you're using ``ispc`` with a system that
+isn't otherwise multi-threaded and don't want to write custom
+implementations of them, you can use the implementations of these functions
+provided in the ``examples/tasksys.cpp`` file in the ``ispc``
+distributions.
+
+If you are implementing your own task system, the remainder of this section
+discusses the requirements for these calls.  You will also likely want to
+review the example task systems in ``examples/tasksys.cpp`` for reference.
+If you are not implmenting your own task system, you can skip reading the
+remainder of this section.
+
+Here are the declarations of the three functions that must be provided to
+manage tasks in ``ispc``:
 
 ::
 
-    void ISPCLaunch(void *funcptr, void *data);
-    void ISPCSync();
+    void *ISPCAlloc(void **handlePtr, int64_t size, int32_t alignment);
+    void ISPCLaunch(void **handlePtr, void *f, void *data, int count);
+    void ISPCSync(void *handle);
 
-On Windows, two additional functions must be provided to dynamically
-allocate and free memory to store the arguments passed to tasks.  (On OSX
-and Linux, the stack provides memory for task arguments; on Windows, the
-stack is generally not large enough to do this for large numbers of tasks.)
+All three of these functions take an opaque handle (or a pointer to an
+opaque handle) as their first parameter.  This handle allows the task
+system runtime to distinguish between calls to these functions from
+different functions in ``ispc`` code.  In this way, the task system
+implementation can efficiently wait for completion on just the tasks
+launched from a single function.
+
+The first time one of ``ISPCLaunch()`` or ``ISPCAlloc()`` is called in an
+``ispc`` functon, the ``void *`` pointed to by the ``handlePtr`` parameter
+will be ``NULL``.  The implementations of these function should then
+initialize ``*handlePtr`` to a unique handle value of some sort.  (For
+example, it might allocate a small structure to record which tasks were
+launched by the current function.)  In subsequent calls to these functions
+in the emitted ``ispc`` code, the same value for ``handlePtr`` will be
+passed in, such that loading from ``*handlePtr`` will retrieve the value
+stored in the first call.
+
+At function exit (or at an explicit ``sync`` statement), a call to
+``ISPCSync()`` will be generated if ``*handlePtr`` is non-``NULL``.
+Therefore, the handle value is passed directly to ``ISPCSync()``, rather
+than a pointer to it, as in the other functions.
+
+The ``ISPCAlloc()`` function is used to allocate small blocks of memory to
+store parameters passed to tasks.  It should return a pointer to memory
+with the given aize and alignment.  Note that there is no explicit
+``ISPCFree()`` call; instead, all memory allocated within an ``ispc``
+function should be freed when ``ISPCSync()`` is called.
+
+``ISPCLaunch()`` is called to launch to launch one or more asynchronous
+tasks.  Each ``launch`` statement in ``ispc`` code causes a call to
+``ISPCLaunch()`` to be emitted in the generated code.  The three parameters
+after the handle pointer to thie function are relatively straightforward;
+the ``void *f`` parameter holds a pointer to a function to call to run the
+work for this task, ``data`` holds a pointer to data to pass to this
+function, and ``count`` is the number of instances of this function to
+enqueue for asynchronous execution.  (In other words, ``count`` corresponds
+to the value ``n`` in a multiple-task launch statement like ``launch[n]``.)
+
+The signature of the provided function pointer ``f`` is
 
 ::
 
-    void *ISPCMalloc(int64_t size, int32_t alignment);
-    void ISPCFree(void *ptr);
+    void (*TaskFuncPtr)(void *data, int threadIndex, int threadCount,
+                        int taskIndex, int taskCount)
 
-These are called by the task launch code generated by the ``ispc``
-compiler; the first is called to launch to launch a task and the second is
-called to wait for, respectively.  (Factoring them out in this way
-allows ``ispc`` to inter-operate with the application's task system, if
-any, rather than having a separate one of its own.)  To run a particular
-task, the task system should cast the function pointer to a ``void (*)(void
-*, int, int)`` function pointer and then call it with the provided ``void
-*`` data and then an index for the current hardware thread and the total
-number of hardware threads the task system has launched--in other words:
-
-::
-
-    typedef void (*TaskFuncType)(void *, int, int);
-    TaskFuncType tft = (TaskFuncType)(funcptr);
-    tft(data, threadIndex, threadCount);
-
-A number of sample task system implementations are provided with ``ispc``; 
-see the files ``tasks_concrt.cpp``, ``tasks_gcd.cpp`` and
-``tasks_pthreads.cpp`` in the ``examples/mandelbrot_tasks`` directory of
-the ``ispc`` distribution.
+When this function pointer is called by one of the hardware threads managed
+bythe task system, the ``data`` pointer passed to ``ISPCLaunch()`` should
+be passed to it for its first parameter; ``threadCount`` gives the total
+number of hardware threads that have been spawned to run tasks and
+``threadIndex`` should be an integer index between zero and ``threadCount``
+uniquely identifying the hardware thread that is running the task.  (These
+values can be used to index into thread-local storage.)
 
+The value of ``taskCount`` should be the number of tasks launched in the
+``launch`` statement that caused the call to ``ISPCLaunch()`` and each of
+the calls to this function should be given a unique value of ``taskIndex``
+between zero and ``taskCount``, to distinguish which of the instances
+of the set of launched tasks is running.
 
 The ISPC Standard Library
 =========================
@@ -1822,6 +1918,71 @@ given value across all of the currently-executing vector lanes.
     uniform int reduce_max(int a, int b)
     uniform unsigned int reduce_max(unsigned int a, unsigned int b)
 
+Finally, you can check to see if a particular value has the same value in
+all of the currently-running program instances:
+
+::
+
+    uniform bool reduce_equal(int32 v)
+    uniform bool reduce_equal(unsigned int32 v)
+    uniform bool reduce_equal(float v)
+    uniform bool reduce_equal(int64 v)
+    uniform bool reduce_equal(unsigned int64 v)
+    uniform bool reduce_equal(double)
+
+There are also variants of these functions that return the value as a
+``uniform`` in the case where the values are all the same.
+
+::
+
+    uniform bool reduce_equal(int32 v, reference uniform int32 sameval)
+    uniform bool reduce_equal(unsigned int32 v,
+                              reference uniform unsigned int32 sameval)
+    uniform bool reduce_equal(float v, reference uniform float sameval)
+    uniform bool reduce_equal(int64 v, reference uniform int64 sameval)
+    uniform bool reduce_equal(unsigned int64 v,
+                              reference uniform unsigned int64 sameval)
+    uniform bool reduce_equal(double, reference uniform double sameval)
+
+If called when none of the program instances are running,
+``reduce_equal()`` will return ``false``.
+
+There are also a number of functions to compute "scan"s of values across
+the program instances.  For example, the ``exclusive_scan_and()`` function
+computes, for each program instance, the sum of the given value over all of
+the preceeding program instances.  (The scans currently available in
+``ispc`` are all so-called "exclusive" scans, meaning that the value
+computed for a given element does not include the value provided for that
+element.)  In C code, an exclusive add scan over an array might be
+implemented as:
+
+::
+
+    void scan_add(int *in_array, int *result_array, int count) {
+        result_array[0] = 0;
+        for (int i = 0; i < count; ++i)
+            result_array[i] = result_array[i-1] + in_array[i-1];
+    }
+
+``ispc`` provides the following scan functions--addition, bitwise-and, and
+bitwise-or are available:
+
+::
+
+    int32 exclusive_scan_add(int32 v) 
+    unsigned int32 exclusive_scan_add(unsigned int32 v) 
+    float exclusive_scan_add(float v) 
+    int64 exclusive_scan_add(int64 v) 
+    unsigned int64 exclusive_scan_add(unsigned int64 v) 
+    double exclusive_scan_add(double v) 
+    int32 exclusive_scan_and(int32 v) 
+    unsigned int32 exclusive_scan_and(unsigned int32 v) 
+    int64 exclusive_scan_and(int64 v) 
+    unsigned int64 exclusive_scan_and(unsigned int64 v) 
+    int32 exclusive_scan_or(int32 v) 
+    unsigned int32 exclusive_scan_or(unsigned int32 v) 
+    int64 exclusive_scan_or(int64 v) 
+    unsigned int64 exclusive_scan_or(unsigned int64 v) 
 
 
 Packed Load and Store Operations
@@ -1921,6 +2082,18 @@ function returns the 16 bits that are the closest match to the given
     int16 float_to_half(float f)
     uniform int16 float_to_half(uniform float f)
 
+There are also faster versions of these functions that don't worry about
+handling floating point infinity, "not a number" and denormalized numbers
+correctly.  These are faster than the above functions, but are less
+precise.
+
+::
+
+    float half_to_float_fast(unsigned int16 h)
+    uniform float half_to_float_fast(uniform unsigned int16 h)
+    int16 float_to_half_fast(float f)
+    uniform int16 float_to_half_fast(uniform float f)
+
 
 Atomic Operations and Memory Fences
 -----------------------------------
@@ -1941,12 +2114,12 @@ end.)
 
 One thing to note is that that the value being added to here is a
 ``uniform`` integer, while the increment amount and the return value are
-``varying``.  In other words, the semantics are that each running program
-instance individually issues the atomic operation with its own ``delta``
-value and gets the previous value of ``val`` back in return.  The atomics
-for the running program instances may be issued in arbitrary order; it's
-not guaranteed that they will be issued in ``programIndex`` order, for
-example.
+``varying``.  In other words, the semantics of this call are that each
+running program instance individually issues the atomic operation with its
+own ``delta`` value and gets the previous value of ``val`` back in return.
+The atomics for the running program instances may be issued in arbitrary
+order; it's not guaranteed that they will be issued in ``programIndex``
+order, for example.
 
 Here are the declarations of the ``int32`` variants of these functions.
 There are also ``int64`` equivalents as well as variants that take
@@ -1964,17 +2137,44 @@ function can be used with ``float`` and ``double`` types as well.)
   int32 atomic_xor_global(reference uniform int32 val, int32 value)
   int32 atomic_swap_global(reference uniform int32 val, int32 newval)
 
-There is also an atomic "compare and exchange" function; it atomically
-compares the value in "val" to "compare"--if they match, it assigns
-"newval" to "val".  In either case, the old value of "val" is returned.
-(As with the other atomic operations, there are also ``unsigned`` and
-64-bit variants of this function.  Furthermore, there are ``float`` and
-``double`` variants as well.)
+There are also variants of these functions that take ``uniform`` values for
+the operand and return a ``uniform`` result:
 
 ::
 
+  uniform int32 atomic_add_global(reference uniform int32 val,
+                                  uniform int32 value)
+  uniform int32 atomic_subtract_global(reference uniform int32 val,
+                                       uniform int32 value)
+  uniform int32 atomic_min_global(reference uniform int32 val,
+                                  uniform int32 value)
+  uniform int32 atomic_max_global(reference uniform int32 val,
+                                  uniform int32 value)
+  uniform int32 atomic_and_global(reference uniform int32 val,
+                                  uniform int32 value)
+  uniform int32 atomic_or_global(reference uniform int32 val,
+                                  uniform int32 value)
+  uniform int32 atomic_xor_global(reference uniform int32 val,
+                                  uniform int32 value)
+  uniform int32 atomic_swap_global(reference uniform int32 val,
+                                   uniform int32 newval)
+
+There are also an atomic swap and "compare and exchange" functions.
+Compare and exchange atomically compares the value in "val" to
+"compare"--if they match, it assigns "newval" to "val".  In either case,
+the old value of "val" is returned.  (As with the other atomic operations,
+there are also ``unsigned`` and 64-bit variants of this function.
+Furthermore, there are ``float`` and ``double`` variants as well.)
+
+::
+
+  int32 atomic_swap_global(reference uniform int32 val, int32 new)
+  uniform int32 atomic_swap_global(reference uniform int32 val,
+                                   uniform int32 new)
   int32 atomic_compare_exchange_global(reference uniform int32 val,
                                        int32 compare, int32 newval)
+  uniform int32 atomic_compare_exchange_global(reference uniform int32 val,
+                                  uniform int32 compare, uniform int32 newval)
 
 ``ispc`` also has a standard library routine that inserts a memory barrier
 into the code; it ensures that all memory reads and writes prior to be
@@ -1990,6 +2190,53 @@ code.
     void memory_barrier();
 
 
+Prefetches
+----------
+
+The standard library has a variety of functions to prefetch data into the
+processor's cache.  While modern CPUs have automatic prefetchers that do a
+reasonable job of prefetching data to the cache before its needed, high
+performance applications may find it helpful to prefetch data before it's
+needed.
+
+For example, this code shows how to prefetch data to the processor's L1
+cache while iterating over the items in an array.  
+
+::
+
+   uniform int32 array[...];
+   for (uniform int i = 0; i < count; ++i) {
+       // do computation with array[i]
+       prefetch_l1(array[i+32]);
+   }
+
+The standard library has routines to prefetch to the L1, L2, and L3
+caches.  It also has a variant, ``prefetch_nt()``, that indicates that the
+value being prefetched isn't expected to be used more than once (so should
+be high priority to be evicted from the cache).
+
+::
+
+    void prefetch_{l1,l2,l3,nt}(reference TYPE)
+
+These functions are available for all of the basic types in the
+language--``int8``, ``int16``, ``int32``, ``float``, and so forth.
+
+
+System Information
+------------------
+
+A routine is available to find the number of CPU cores available in the
+system:
+
+::
+
+    int num_cores()
+
+This value can be useful for adapting the granularity of parallel task
+decomposition depending on the number of processors in the system.
+
+
 Low-Level Bits
 --------------
 
@@ -2097,14 +2344,14 @@ Both the ``foo`` and ``bar`` global variables can be accessed on each
 side.
 
 ``ispc`` code can also call back to C/C++.  On the ``ispc`` side, any
-application functions to be called must be declared with the ``export "C"``
+application functions to be called must be declared with the ``extern "C"``
 qualifier.
 
 ::
 
    extern "C" void foo(uniform float f, uniform float g);
 
-Unlike in C++, ``export "C"`` doesn't take braces to delineate
+Unlike in C++, ``extern "C"`` doesn't take braces to delineate
 multiple functions to be declared; thus, multiple C functions to be called
 from ``ispc`` must be declared as follows:
 
@@ -2699,6 +2946,123 @@ active upon function entry.
     ao.ispc(0088) - function entry: 36928 calls (0 / 0.00% all off!), 97.40% active lanes
     ...
 
+
+Using Scan Operations For Variable Output
+-----------------------------------------
+
+One important application of the ``exclusive_scan_add()`` function in the
+standard library is when program instances want to generate a variable amount
+of output and when one would like that output to be densely packed in a
+single array.  For example, consider the code fragment below:
+
+::
+
+    uniform int func(uniform float outArray[], ...) {
+       int numOut = ...;  // figure out how many to be output
+       float outLocal[MAX_OUT]; // staging area
+       // put 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 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 the
+program instances have written to the array.
+
+Application-Supplied Execution Masks
+------------------------------------
+
+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) {
+        for (uniform int i = 0; i < count; i += programCount) {
+            cif (update[i+programIndex] == 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.)
+
+Explicit Vector Programming With Uniform Short Vector Types
+-----------------------------------------------------------
+
+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 
+(as were introduced in the `Short Vector Types`_ section).  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;
+    }
+
+
 Disclaimer and Legal Information
 ================================
 

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         = 1.0.5
+PROJECT_NUMBER         = 1.0.10
 
 # The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute)
 # base path where the generated documentation will be put.

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
 ================
 
@@ -58,6 +95,7 @@ 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
 =====
 
@@ -71,6 +109,7 @@ Options
 This program implements both the Black-Scholes and Binomial options pricing
 models in both ispc and regular serial C++ code.
 
+
 RT
 ==
 
@@ -87,9 +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/Makefile

@@ -1,8 +1,14 @@
 
-CXX=g++ -m64
-CXXFLAGS=-Iobjs/ -O3 -Wall
+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 --fast-math --arch=x86-64
+ISPCFLAGS=-O2 --target=sse4 --arch=x86-64
 
 default: ao
 
@@ -14,12 +20,15 @@ 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/ao.o objs/ao_serial.o objs/ao_ispc.o $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ objs/ao.o objs/ao_ispc.o objs/ao_serial.o $(TASK_OBJ) -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

examples/aobench/ao.cpp

@@ -101,6 +101,7 @@ 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);
 }
 
 
@@ -172,10 +173,30 @@ 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.
@@ -192,7 +213,8 @@ 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

@@ -203,8 +203,9 @@ 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, 
+                         reference 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 },
@@ -231,6 +232,9 @@ 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 +243,21 @@ 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 +299,7 @@ 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 +321,15 @@ 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

@@ -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">
@@ -21,6 +21,7 @@
   <ItemGroup>
     <ClCompile Include="ao.cpp" />
     <ClCompile Include="ao_serial.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="ao.ispc">

examples/aobench_instrumented/Makefile

@@ -2,7 +2,7 @@
 CXX=g++ -m64
 CXXFLAGS=-Iobjs/ -g3 -Wall
 ISPC=ispc
-ISPCFLAGS=-O2 --fast-math --instrument --arch=x86-64
+ISPCFLAGS=-O2 --instrument --arch=x86-64
 
 default: ao
 

examples/aobench_instrumented/ao.cpp

@@ -100,6 +100,7 @@ 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);
 }
 
 

examples/deferred/Makefile

@@ -0,0 +1,42 @@
+
+ARCH = $(shell uname)
+
+TASK_CXX=../tasks_pthreads.cpp
+TASK_LIB=-lpthread
+
+ifeq ($(ARCH), Darwin)
+  TASK_CXX=../tasks_gcd.cpp
+  TASK_LIB=
+endif
+
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
+
+CXX=g++
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
+ISPC=ispc
+ISPCFLAGS=-O2 --target=sse4x2 --arch=x86-64 --math-lib=fast
+
+OBJS=objs/main.o objs/common.o objs/kernels_ispc.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: %.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(int64_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 +1,170 @@
+<?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>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <LinkIncremental>true</LinkIncremental>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <LinkIncremental>false</LinkIncremental>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <LinkIncremental>false</LinkIncremental>
+  </PropertyGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <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>
+      <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>
+      <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>
+      <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="../tasks_concrt.cpp" />
+  </ItemGroup>
+  <ItemGroup>
+    <CustomBuild Include="kernels.ispc">
+      <FileType>Document</FileType>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+</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'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+</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>
+    </CustomBuild>
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+  </ImportGroup>
+</Project>

examples/deferred/dynamic_c.cpp

@@ -0,0 +1,871 @@
+/*
+  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(int64_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);
+}
+
+
+// numLights need not be a multiple of programCount here, but the input and output arrays
+// should be able to handle programCount-sized load/stores.
+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 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 __cilkplusplus
+
+#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(int64_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 // __cilkplusplus

examples/deferred/kernels.ispc

@@ -0,0 +1,717 @@
+/*
+  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 zBuffer[];
+    uniform unsigned int16 normalEncoded_x[]; // half float
+    uniform unsigned int16 normalEncoded_y[]; // half float
+    uniform unsigned int16 specularAmount[]; // half float
+    uniform unsigned int16 specularPower[]; // half float
+    uniform unsigned int8 albedo_x[]; // unorm8
+    uniform unsigned int8 albedo_y[]; // unorm8
+    uniform unsigned int8 albedo_z[]; // unorm8
+    uniform float lightPositionView_x[];
+    uniform float lightPositionView_y[];
+    uniform float lightPositionView_z[];
+    uniform float lightAttenuationBegin[];
+    uniform float lightColor_x[];
+    uniform float lightColor_y[];
+    uniform float lightColor_z[];
+    uniform float 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];
+};
+
+export void foo(reference InputHeader h) { }
+
+
+///////////////////////////////////////////////////////////////////////////
+// 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, reference float ox, 
+           reference float oy, reference 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);
+}
+
+
+// tile width must be a multiple of programCount (SIMD size)
+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
+    reference uniform float minZ,
+    reference uniform float maxZ
+    )
+{
+    // Find Z bounds
+    float laneMinZ = cameraFar;
+    float laneMaxZ = cameraNear;
+    for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
+        for (uniform int32 x = tileStartX; x < tileEndX; x += programCount) {
+            // Unproject depth buffer Z value into view space
+            float z = zBuffer[(y * gBufferWidth + x) + programIndex];
+            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);
+}
+
+
+// tile width must be a multiple of programCount (SIMD size)
+// numLights must currently be a multiple of programCount (SIMD size)
+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
+    reference 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;
+
+    for (uniform int32 baseLightIndex = 0; baseLightIndex < numLights; 
+         baseLightIndex += programCount) {
+        int32 lightIndex = baseLightIndex + programIndex;
+        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)) 
+            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
+        cif (inFrustum) {
+            tileNumLights += packed_store_active(tileLightIndices, tileNumLights, 
+                                                 lightIndex);
+        }
+    }
+
+    return tileNumLights;
+}
+
+
+// tile width must be a multiple of programCount (SIMD size)
+// numLights must currently be a multiple of programCount (SIMD size)
+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
+    reference 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;
+}
+
+
+// tile width must be a multiple of programCount (SIMD size)
+export void
+ShadeTile(
+    uniform int32 tileStartX, uniform int32 tileEndX,
+    uniform int32 tileStartY, uniform int32 tileEndY,
+    uniform int32 gBufferWidth, uniform int32 gBufferHeight,
+    reference uniform InputDataArrays inputData,
+    // Camera data
+    uniform float cameraProj_11, uniform float cameraProj_22,
+    uniform float cameraProj_33, uniform float cameraProj_43,
+    // Light list
+    reference uniform int32 tileLightIndices[],
+    uniform int32 tileNumLights,
+    // UI
+    uniform bool visualizeLightCount,
+    // Output
+    reference uniform unsigned int8 framebuffer_r[],
+    reference uniform unsigned int8 framebuffer_g[],
+    reference 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) {
+            for (uniform int32 x = tileStartX; x < tileEndX; x += programCount) {
+                int32 framebufferIndex = (y * gBufferWidth + x) + programIndex;
+                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);
+
+            for (uniform int32 x = tileStartX; x < tileEndX; x += programCount) {
+                uniform int32 gBufferOffsetBase = y * gBufferWidth + x;
+                int32 gBufferOffset = gBufferOffsetBase + programIndex;
+                
+                // 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 + programIndex)) * 
+                    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 g, uniform int num_groups_x, uniform int num_groups_y,
+           reference uniform InputHeader inputHeader,
+           reference uniform InputDataArrays inputData,
+           uniform int visualizeLightCount,
+           // Output
+           reference uniform unsigned int8 framebuffer_r[],
+           reference uniform unsigned int8 framebuffer_g[],
+           reference uniform unsigned int8 framebuffer_b[]) {
+    uniform int32 group_y = g / num_groups_x;
+    uniform int32 group_x = g % 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 sTileNumLights = 0;
+    uniform int sTileLightIndices[MAX_LIGHTS];  // Light list for the tile
+
+    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
+    sTileNumLights = 
+        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,
+                                sTileLightIndices);
+
+    ShadeTile(tile_start_x, tile_end_x, tile_start_y, tile_end_y,
+              framebufferWidth, framebufferHeight, inputData,
+              cameraProj_00, cameraProj_11, cameraProj_22, cameraProj_32,
+              sTileLightIndices, sTileNumLights, visualizeLightCount, 
+              framebuffer_r, framebuffer_g, framebuffer_b);
+}
+
+
+export void
+RenderStatic(reference uniform InputHeader inputHeader,
+             reference uniform InputDataArrays inputData,
+             uniform int visualizeLightCount,
+             // Output
+             reference uniform unsigned int8 framebuffer_r[],
+             reference uniform unsigned int8 framebuffer_g[],
+             reference 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;
+
+    for (uniform int g = 0; g < num_groups; ++g)
+        launch < RenderTile(g, num_groups_x, num_groups_y,
+                            inputHeader, inputData, visualizeLightCount,
+                            framebuffer_r, framebuffer_g, framebuffer_b) >;
+}
+
+
+///////////////////////////////////////////////////////////////////////////
+// Routines for dynamic decomposition path
+
+// tile width must be a multiple of programCount (SIMD size)
+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
+    reference uniform float minZArray[],
+    reference 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;
+    }
+}
+
+
+// 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
+    reference 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
+    reference uniform int32 subtileIndices[],
+    uniform int32 subtileIndicesPitch,
+    reference 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;
+
+    for (int32 i = programIndex; i < numLights; i += programCount) {
+        // TODO: ISPC says gather required here when it actually
+        // isn't... this could be fixed this by nesting an if() within a
+        // uniform loop, but I'm not totally sure if that's a win
+        // overall. For now we'll just eat the perf cost for cleanliness
+        // since the below are real gathers anyways.
+        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 +1,137 @@
+/*
+  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>\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 __cilkplusplus
+    InitDynamicCilk(input);
+#endif // __cilkplusplus
+
+    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);
+
+    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\n", 
+           serialCycles);
+    WriteFrame("deferred-serial-dynamic.ppm", input, framebuffer);
+
+#ifdef __cilkplusplus
+    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\n", 
+           dynamicCilkCycles);
+    WriteFrame("deferred-ispc-dynamic.ppm", input, framebuffer);
+
+    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 // __cilkplusplus
+
+    DeleteInputData(input);
+
+    return 0;
+}

examples/examples.sln

@@ -15,8 +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}") = "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
@@ -81,14 +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
+		{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/mandelbrot.cpp

@@ -64,6 +64,7 @@ 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_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/Makefile

@@ -1,18 +1,12 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
+TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=tasks_gcd.cpp
-  TASK_LIB=
-endif
-
-TASK_OBJ=$(addprefix objs/, $(TASK_CXX:.cpp=.o))
-
-CXX=g++ -m64
-CXXFLAGS=-Iobjs/ -O3 -Wall
+CXX=g++
+CXXFLAGS=-Iobjs/ -O3 -Wall -m64
 ISPC=ispc
 ISPCFLAGS=-O2 --target=sse4x2 --arch=x86-64
 
@@ -32,6 +26,9 @@ mandelbrot: dirs objs/mandelbrot.o objs/mandelbrot_serial.o objs/mandelbrot_ispc
 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

examples/mandelbrot_tasks/mandelbrot.cpp

@@ -40,6 +40,7 @@
 
 #include <stdio.h>
 #include <algorithm>
+#include <string.h>
 #include "../timing.h"
 #include "../cpuid.h"
 #include "mandelbrot_ispc.h"
@@ -64,6 +65,7 @@ writePPM(int *buf, int width, int height, const char *fn) {
             fputc(c, fp);
     }
     fclose(fp);
+    printf("Wrote image file %s\n", fn);
 }
 
 
@@ -98,8 +100,12 @@ ensureTargetISAIsSupported() {
     }
 }
 
+static void usage() {
+    fprintf(stderr, "usage: mandelbrot [--scale=<factor>]\n");
+    exit(1);
+}
 
-int main() {
+int main(int argc, char *argv[]) {
     unsigned int width = 1536;
     unsigned int height = 1024;
     float x0 = -2;
@@ -107,10 +113,26 @@ int main() {
     float y0 = -1;
     float y1 = 1;
 
-    ensureTargetISAIsSupported();
+    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();
 
-    extern void TasksInit();
-    TasksInit();
+    ensureTargetISAIsSupported();
 
     int maxIterations = 512;
     int *buf = new int[width*height];
@@ -121,6 +143,9 @@ 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();
@@ -130,9 +155,6 @@ 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
@@ -140,6 +162,9 @@ 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

@@ -53,11 +53,14 @@ 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[]) {
+    uniform int ystart = ybase + taskIndex * span;
+    uniform int yend = ystart + span;
+
     for (uniform int j = ystart; j < yend; ++j) {
         for (uniform int i = 0; i < width; i += programCount) {
             float x = x0 + (programIndex + i) * dx;
@@ -70,6 +73,20 @@ mandelbrot_scanlines(uniform int ystart, uniform int yend,
 }
                                
 
+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, reference 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,
@@ -78,9 +95,6 @@ mandelbrot_ispc(uniform float x0, uniform float y0,
     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">
@@ -143,7 +143,7 @@
   <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">

examples/mandelbrot_tasks/tasks_concrt.cpp

@@ -1,141 +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 <stdint.h>
-#include <assert.h>
-#include <stdio.h>
-#include <stdlib.h>
-
-// ispc expects these functions to have C linkage / not be mangled
-extern "C" { 
-    void ISPCLaunch(void *f, void *data);
-    void ISPCSync();
-    void *ISPCMalloc(int64_t size, int32_t alignment);
-    void ISPCFree(void *ptr);
-}
-
-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;
-static bool initialized = false;
-
-void
-TasksInit() {
-    InitializeCriticalSection(&criticalSection);
-    for (int i = 0; i < MAX_TASKS; ++i)
-        events[i] = new event;
-    initialized = true;
-}
-
-
-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) {
-    if (!initialized) {
-        fprintf(stderr, "You must call TasksInit() before launching tasks.\n");
-        exit(1);
-    }
-
-    // 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() {
-    if (!initialized) {
-        fprintf(stderr, "You must call TasksInit() before launching tasks.\n");
-        exit(1);
-    }
-
-    event::wait_for_multiple(&events[0], taskOffset, true, 
-                             COOPERATIVE_TIMEOUT_INFINITE);
-
-    for (int i = 0; i < taskOffset; ++i)
-        events[i]->reset();
-
-    taskOffset = 0;
-}
-
-
-void *ISPCMalloc(int64_t size, int32_t alignment) {
-    return _aligned_malloc(size, alignment);
-}
-
-
-void ISPCFree(void *ptr) {
-    _aligned_free(ptr);
-}

examples/mandelbrot_tasks/tasks_gcd.cpp

@@ -1,103 +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.  
-*/
-
-/* A simple task system for ispc programs based on Apple's Grand Central
-   Dispatch. */
-
-#include <dispatch/dispatch.h>
-#include <stdio.h>
-#include <stdlib.h>
-
-static bool initialized = false;
-static dispatch_queue_t gcdQueue;
-static dispatch_group_t gcdGroup;
-
-// ispc expects these functions to have C linkage / not be mangled
-extern "C" {
-    void ISPCLaunch(void *f, void *data);
-    void ISPCSync();
-}
-
-struct TaskInfo {
-    void *func;
-    void *data;
-};
-
-
-void
-TasksInit() {
-    gcdQueue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
-    gcdGroup = dispatch_group_create();
-    initialized = true;
-}
-
-
-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) {
-    if (!initialized) {
-        fprintf(stderr, "You must call TasksInit() before launching tasks.\n");
-        exit(1);
-    }
-    TaskInfo *ti = new TaskInfo;
-    ti->func = func;
-    ti->data = data;
-    dispatch_group_async_f(gcdGroup, gcdQueue, ti, lRunTask);
-}
-
-
-void ISPCSync() {
-    if (!initialized) {
-        fprintf(stderr, "You must call TasksInit() before launching tasks.\n");
-        exit(1);
-    }
-
-    // Wait for all of the tasks in the group to complete before returning
-    dispatch_group_wait(gcdGroup, DISPATCH_TIME_FOREVER);
-}

examples/mandelbrot_tasks/tasks_pthreads.cpp

@@ -1,295 +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) {
-    if (threads == NULL) {
-        fprintf(stderr, "You must call TasksInit() before launching tasks.\n");
-        exit(1);
-    }
-
-    //
-    // 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() {
-    if (threads == NULL) {
-        fprintf(stderr, "You must call TasksInit() before launching tasks.\n");
-        exit(1);
-    }
-
-    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/noise.ispc

@@ -131,11 +131,11 @@ static float Noise(float x, float y, float z) {
 }
 
 
-static float Turbulence(float x, float y, float z, int octaves) {
+static float Turbulence(float x, float y, float z, uniform int octaves) {
     float omega = 0.6;
 
     float sum = 0., lambda = 1., o = 1.;
-    for (int i = 0; i < octaves; ++i) {
+    for (uniform int i = 0; i < octaves; ++i) {
         sum += abs(o * Noise(lambda * x, lambda * y, lambda * z));
         lambda *= 1.99f;
         o *= omega;

examples/noise/noise.vcxproj

@@ -164,4 +164,4 @@
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

examples/noise/noise_serial.cpp

@@ -104,7 +104,7 @@ inline float NoiseWeight(float t) {
 
 
 inline float Lerp(float t, float low, float high) {
-    return (1. - t) * low + t * high;
+    return (1.f - t) * low + t * high;
 }
 
 
@@ -147,7 +147,7 @@ static float Turbulence(float x, float y, float z, int octaves) {
         lambda *= 1.99f;
         o *= omega;
     }
-    return sum * 0.5;
+    return sum * 0.5f;
 }
 
 
@@ -163,7 +163,7 @@ void noise_serial(float x0, float y0, float x1, float y1,
             float y = y0 + j * dy;
 
             int index = (j * width + i);
-            output[index] = Turbulence(x, y, 0.6, 8);
+            output[index] = Turbulence(x, y, 0.6f, 8);
         }
     }
 }

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/Makefile

@@ -1,6 +1,12 @@
 
-CXX=g++ -m64
-CXXFLAGS=-Iobjs/ -O3 -Wall
+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=sse4x2 --arch=x86-64
 
@@ -14,11 +20,16 @@ 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/rt.o objs/rt_serial.o objs/rt_ispc.o $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ objs/rt.o objs/rt_ispc.o objs/rt_serial.o $(TASK_OBJ) -lm $(TASK_LIB)
 
-objs/%.o: %.cpp objs/rt_ispc.h
+objs/%.o: %.cpp
 	$(CXX) $< $(CXXFLAGS) -c -o $@
 
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/rt.o: objs/rt_ispc.h 
+
 objs/%_ispc.h objs/%_ispc.o: %.ispc
 	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/rt/rt.cpp

@@ -42,6 +42,7 @@
 #include <math.h>
 #include <algorithm>
 #include <assert.h>
+#include <string.h>
 #include <sys/types.h>
 #include "../timing.h"
 #include "../cpuid.h"
@@ -51,7 +52,8 @@ using namespace ispc;
 
 typedef unsigned int uint;
 
-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[]);
@@ -90,6 +92,7 @@ static void writeImage(int *idImage, float *depthImage, int width, int height,
         }
     }            
     fclose(f);
+    printf("Wrote image file %s\n", filename);
 }
 
 
@@ -125,11 +128,28 @@ ensureTargetISAIsSupported() {
 }
 
 
+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();
 
     ensureTargetISAIsSupported();
 
@@ -143,10 +163,10 @@ 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;
     }
 
@@ -154,20 +174,20 @@ 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;
     }
 
@@ -214,10 +234,10 @@ int main(int argc, char *argv[]) {
     }
     fclose(f);
 
-    // round image resolution up to multiple of 4 to makethings easy for
+    // round image resolution up to multiple of 16 to make things easy for
     // the code that assigns pixels to ispc program instances
-    height = (height + 3) & ~3;
-    width = (width + 3) & ~3;
+    int height = (int(baseHeight * scale) + 0xf) & ~0xf;
+    int width = (int(baseWidth * scale) + 0xf) & ~0xf;
 
     // allocate images; one to hold hit object ids, one to hold depth to
     // the first interseciton
@@ -225,19 +245,42 @@ 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
@@ -246,14 +289,15 @@ 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

@@ -226,20 +226,26 @@ 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 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);
+
     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 };
 
     // 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) {
+    for (uniform int y = y0; y < y1; y += 4) {
+        for (uniform int x = x0; x < x1; 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.
@@ -251,7 +257,8 @@ export void raytrace(uniform int width, uniform int height,
                 const float dy = udy[o * programCount + programIndex];
 
                 Ray ray;
-                generateRay(raster2camera, camera2world, x+dx, y+dy, ray);
+                generateRay(raster2camera, camera2world, (x+dx)*widthScale,
+                            (y+dy)*heightScale, ray);
                 BVHIntersect(nodes, triangles, ray);
 
                 int offset = (y + (int)dy) * width + (x + (int)dx);
@@ -261,3 +268,54 @@ export void raytrace(uniform int width, uniform int height,
         }
     }
 }
+
+
+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 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 int dx = 16; // must match dx below
+    uniform int xTasks = (width + (dx-1)) / dx;
+    uniform int x0 = (taskIndex % xTasks) * dx;
+    uniform int x1 = x0 + dx;
+    x1 = min(x1, width);
+                             
+    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 nTasks = (width + (dx-1)) / dx;
+    for (uniform int y = 0; y < height; y += dy) {
+        uniform int y1 = min(y + dy, height);
+        launch[nTasks] < raytrace_tile_task(y, y1, width, height, baseWidth,
+                                            baseHeight, raster2camera, camera2world, 
+                                            image, id, nodes, triangles) >;
+    }
+}

examples/rt/rt.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">
@@ -164,6 +164,7 @@ ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h
   <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

@@ -258,17 +258,21 @@ 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/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">
@@ -28,7 +28,7 @@
 ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
 </Command>
       <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
-ispc -O2 %(Filename).ispco %(Filename).obj -h %(Filename)_ispc.h
+ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h
 </Command>
       <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%(Filename).obj</Outputs>
       <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">%(Filename).obj</Outputs>
@@ -161,4 +161,4 @@ ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

examples/stencil/.gitignore

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

examples/stencil/Makefile

@@ -0,0 +1,35 @@
+
+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=sse4x2 --arch=x86-64
+
+default: stencil
+
+.PHONY: dirs clean
+
+dirs:
+	/bin/mkdir -p objs/
+
+clean:
+	/bin/rm -rf objs *~ stencil
+
+stencil: dirs objs/stencil.o objs/stencil_serial.o objs/stencil_ispc.o $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ objs/stencil.o objs/stencil_ispc.o objs/stencil_serial.o $(TASK_OBJ) -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: %.ispc
+	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/stencil/stencil.cpp

@@ -0,0 +1,186 @@
+/*
+  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 "../cpuid.h"
+#include "stencil_ispc.h"
+using namespace ispc;
+
+
+// Make sure that the vector ISA used during compilation is supported by
+// the processor.  The ISPC_TARGET_* macro is set in the ispc-generated
+// header file that we include above.
+static void
+ensureTargetISAIsSupported() {
+#if defined(ISPC_TARGET_SSE2)
+    bool isaSupported = CPUSupportsSSE2();
+    const char *target = "SSE2";
+#elif defined(ISPC_TARGET_SSE4)
+    bool isaSupported = CPUSupportsSSE4();
+    const char *target = "SSE4";
+#elif defined(ISPC_TARGET_AVX)
+    bool isaSupported = CPUSupportsAVX();
+    const char *target = "AVX";
+#else
+#error "Unknown ISPC_TARGET_* value"
+#endif
+    if (!isaSupported) {
+        fprintf(stderr, "***\n*** Error: the ispc-compiled code uses the %s instruction "
+                "set, which isn't\n***        supported by this computer's CPU!\n", target);
+        fprintf(stderr, "***\n***        Please modify the "
+#ifdef _MSC_VER
+                "MSVC project file "
+#else
+                "Makefile "
+#endif
+                "to select another target (e.g. sse2)\n***\n");
+        exit(1);
+    }
+}
+
+
+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() {
+    ensureTargetISAIsSupported();
+
+    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;
+}

examples/stencil/stencil.ispc

@@ -0,0 +1,129 @@
+/*
+  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.  
+*/
+
+
+static void
+stencil_step(uniform int x0, uniform int x1,
+             uniform int y0, uniform int y1,
+             uniform int z0, uniform int z1,
+             uniform int Nx, uniform int Ny, uniform int Nz,
+             uniform const float coef[4], uniform const float vsq[],
+             uniform const float Ain[], uniform float Aout[]) {
+    const uniform int Nxy = Nx * Ny;
+
+    for (uniform int z = z0; z < z1; ++z) {
+        for (uniform int y = y0; y < y1; ++y) {
+            // Assumes that (x1-x0) % programCount == 0
+            for (uniform int x = x0; x < x1; x += programCount) {
+                int index = (z * Nxy) + (y * Nx) + x + programIndex;
+#define A_cur(x, y, z) Ain[index + (x) + ((y) * Nx) + ((z) * Nxy)]
+#define A_next(x, y, z) Aout[index + (x) + ((y) * Nx) + ((z) * Nxy)]
+                float div = coef[0] * A_cur(0, 0, 0) +
+                            coef[1] * (A_cur(+1, 0, 0) + A_cur(-1, 0, 0) +
+                                       A_cur(0, +1, 0) + A_cur(0, -1, 0) +
+                                       A_cur(0, 0, +1) + A_cur(0, 0, -1)) +
+                            coef[2] * (A_cur(+2, 0, 0) + A_cur(-2, 0, 0) +
+                                       A_cur(0, +2, 0) + A_cur(0, -2, 0) +
+                                       A_cur(0, 0, +2) + A_cur(0, 0, -2)) +
+                            coef[3] * (A_cur(+3, 0, 0) + A_cur(-3, 0, 0) +
+                                       A_cur(0, +3, 0) + A_cur(0, -3, 0) +
+                                       A_cur(0, 0, +3) + A_cur(0, 0, -3));
+
+                A_next(0, 0, 0) = 2 * A_cur(0, 0, 0) - A_next(0, 0, 0) + 
+                    vsq[index] * div;
+            }
+        }
+    }
+}
+
+
+static task void
+stencil_step_task(uniform int x0, uniform int x1,
+                  uniform int y0, uniform int y1,
+                  uniform int z0, uniform int z1,
+                  uniform int Nx, uniform int Ny, uniform int Nz,
+                  uniform const float coef[4], uniform const float vsq[],
+                  uniform const float Ain[], uniform float Aout[]) {
+    stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq, Ain, Aout);
+}
+
+
+export void
+loop_stencil_ispc_tasks(uniform int t0, uniform int t1, 
+                        uniform int x0, uniform int x1,
+                        uniform int y0, uniform int y1,
+                        uniform int z0, uniform int z1,
+                        uniform int Nx, uniform int Ny, uniform int Nz,
+                        uniform const float coef[4], 
+                        uniform const float vsq[],
+                        uniform float Aeven[], uniform float Aodd[])
+{
+    for (uniform int t = t0; t < t1; ++t) {
+        // Parallelize across cores as well: each task will work on a slice
+        // of "dz" in the z extent of the volume.  (dz=1 seems to work
+        // better than any larger values.)
+        uniform int dz = 1;
+        for (uniform int z = z0; z < z1; z += dz) {
+            if ((t & 1) == 0)
+                launch < stencil_step_task(x0, x1, y0, y1, z, z+dz, Nx, Ny, Nz, 
+                                           coef, vsq, Aeven, Aodd) >;
+            else
+                launch < stencil_step_task(x0, x1, y0, y1, z, z+dz, Nx, Ny, Nz, 
+                                           coef, vsq, Aodd, Aeven) >;
+        }
+        // We need to wait for all of the launched tasks to finish before
+        // starting the next iteration.
+        sync;
+    }
+}
+
+
+export void
+loop_stencil_ispc(uniform int t0, uniform int t1, 
+                  uniform int x0, uniform int x1,
+                  uniform int y0, uniform int y1,
+                  uniform int z0, uniform int z1,
+                  uniform int Nx, uniform int Ny, uniform int Nz,
+                  uniform const float coef[4], 
+                  uniform const float vsq[],
+                  uniform float Aeven[], uniform float Aodd[])
+{
+    for (uniform int t = t0; t < t1; ++t) {
+        if ((t & 1) == 0)
+            stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq, 
+                         Aeven, Aodd);
+        else
+            stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq, 
+                         Aodd, Aeven);
+    }
+}

examples/stencil/stencil.vcxproj

@@ -0,0 +1,172 @@
+<?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>{2ef070a1-f62f-4e6a-944b-88d140945c3c}</ProjectGuid>
+    <Keyword>Win32Proj</Keyword>
+    <RootNamespace>rt</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>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <LinkIncremental>true</LinkIncremental>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <LinkIncremental>false</LinkIncremental>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <LinkIncremental>false</LinkIncremental>
+  </PropertyGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <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>
+      <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>
+      <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>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <OptimizeReferences>true</OptimizeReferences>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <CustomBuild Include="stencil.ispc">
+      <FileType>Document</FileType>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h
+</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'">
+ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h
+</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%(Filename).obj</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|x64'">%(Filename).obj</Outputs>
+    </CustomBuild>
+  </ItemGroup>
+  <ItemGroup>
+    <ClCompile Include="stencil.cpp" />
+    <ClCompile Include="stencil_serial.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+  </ImportGroup>
+</Project>

examples/stencil/stencil_serial.cpp

@@ -0,0 +1,86 @@
+/*
+  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.  
+*/
+
+
+static void
+stencil_step(int x0, int x1,
+             int y0, int y1,
+             int z0, int z1,
+             int Nx, int Ny, int Nz,
+             const float coef[4], const float vsq[],
+             const float Ain[], float Aout[]) {
+    int Nxy = Nx * Ny;
+
+    for (int z = z0; z < z1; ++z) {
+        for (int y = y0; y < y1; ++y) {
+            for (int x = x0; x < x1; ++x) {
+                int index = (z * Nxy) + (y * Nx) + x;
+#define A_cur(x, y, z) Ain[index + (x) + ((y) * Nx) + ((z) * Nxy)]
+#define A_next(x, y, z) Aout[index + (x) + ((y) * Nx) + ((z) * Nxy)]
+                float div = coef[0] * A_cur(0, 0, 0) +
+                            coef[1] * (A_cur(+1, 0, 0) + A_cur(-1, 0, 0) +
+                                       A_cur(0, +1, 0) + A_cur(0, -1, 0) +
+                                       A_cur(0, 0, +1) + A_cur(0, 0, -1)) +
+                            coef[2] * (A_cur(+2, 0, 0) + A_cur(-2, 0, 0) +
+                                       A_cur(0, +2, 0) + A_cur(0, -2, 0) +
+                                       A_cur(0, 0, +2) + A_cur(0, 0, -2)) +
+                            coef[3] * (A_cur(+3, 0, 0) + A_cur(-3, 0, 0) +
+                                       A_cur(0, +3, 0) + A_cur(0, -3, 0) +
+                                       A_cur(0, 0, +3) + A_cur(0, 0, -3));
+
+                A_next(0, 0, 0) = 2 * A_cur(0, 0, 0) - A_next(0, 0, 0) + 
+                    vsq[index] * div;
+            }
+        }
+    }
+}
+
+
+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[4], 
+                         const float vsq[],
+                         float Aeven[], float Aodd[])
+{
+    for (int t = t0; t < t1; ++t) {
+        if ((t & 1) == 0)
+            stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq, 
+                         Aeven, Aodd);
+        else
+            stencil_step(x0, x1, y0, y1, z0, z1, Nx, Ny, Nz, coef, vsq, 
+                         Aodd, Aeven);
+    }
+}

examples/tasksys.cpp

@@ -0,0 +1,868 @@
+/*
+  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 implements simple task systems that provide the three
+  entrypoints used by ispc-generated to code to handle 'launch' and 'sync'
+  statements in ispc programs.  See the section "Task Parallelism: Language
+  Syntax" in the ispc documentation for information about using task
+  parallelism in ispc programs, and see the section "Task Parallelism:
+  Runtime Requirements" for information about the task-related entrypoints
+  that are implemented here.
+
+  There are three task systems in this file: one built using Microsoft's
+  Concurrency Runtime, one built with Apple's Grand Central Dispatch, and
+  one built on top of bare pthreads.
+*/
+
+#if defined(_WIN32) || defined(_WIN64)
+  #define ISPC_IS_WINDOWS
+  #define ISPC_USE_CONCRT
+#elif defined(__linux__)
+  #define ISPC_IS_LINUX
+  #define ISPC_USE_PTHREADS
+#elif defined(__APPLE__)
+  #define ISPC_IS_APPLE
+  // pthreads is noticably more efficient than GCD on OSX
+  #define ISPC_USE_PTHREADS
+  //#define ISPC_USE_GCD
+#endif
+
+#define DBG(x) 
+
+#ifdef ISPC_IS_WINDOWS
+  #define NOMINMAX
+  #include <windows.h>
+#endif // ISPC_IS_WINDOWS
+#ifdef ISPC_USE_CONCRT
+  #include <concrt.h>
+  using namespace Concurrency;
+#endif // ISPC_USE_CONCRT
+#ifdef ISPC_USE_GCD
+  #include <dispatch/dispatch.h>
+  #include <pthread.h>
+#endif // ISPC_USE_GCD
+#ifdef ISPC_USE_PTHREADS
+  #include <pthread.h>
+  #include <semaphore.h>
+  #include <unistd.h>
+  #include <fcntl.h>
+  #include <errno.h>
+  #include <sys/types.h>
+  #include <sys/stat.h>
+  #include <sys/param.h>
+  #include <sys/sysctl.h>
+  #include <vector>
+  #include <algorithm>
+#endif // ISPC_USE_PTHREADS
+#ifdef ISPC_IS_LINUX
+  #include <malloc.h>
+#endif // ISPC_IS_LINUX
+#include <stdio.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <assert.h>
+#include <string.h>
+#include <algorithm>
+
+// Signature of ispc-generated 'task' functions
+typedef void (*TaskFuncType)(void *data, int threadIndex, int threadCount,
+                             int taskIndex, int taskCount);
+
+// Small structure used to hold the data for each task
+struct TaskInfo {
+    TaskFuncType func;
+    void *data;
+    int taskIndex, taskCount;
+#if defined(ISPC_IS_WINDOWS)
+    event taskEvent;
+#endif
+};
+
+///////////////////////////////////////////////////////////////////////////
+// TaskGroupBase
+
+#define LOG_TASK_QUEUE_CHUNK_SIZE 12
+#define MAX_TASK_QUEUE_CHUNKS 8
+#define TASK_QUEUE_CHUNK_SIZE (1<<LOG_TASK_QUEUE_CHUNK_SIZE)
+
+#define MAX_LAUNCHED_TASKS (MAX_TASK_QUEUE_CHUNKS * TASK_QUEUE_CHUNK_SIZE)
+
+#define NUM_MEM_BUFFERS 16
+
+class TaskGroup;
+
+/** The TaskGroupBase structure provides common functionality for "task
+    groups"; a task group is the set of tasks launched from within a single
+    ispc function.  When the function is ready to return, it waits for all
+    of the tasks in its task group to finish before it actually returns.
+ */
+class TaskGroupBase {
+public:
+    void Reset();
+
+    int AllocTaskInfo(int count);
+    TaskInfo *GetTaskInfo(int index);
+
+    void *AllocMemory(int64_t size, int32_t alignment);
+
+protected:
+    TaskGroupBase();
+    ~TaskGroupBase();
+
+    int nextTaskInfoIndex;
+
+private:
+    /* We allocate blocks of TASK_QUEUE_CHUNK_SIZE TaskInfo structures as
+       needed by the calling function.  We hold up to MAX_TASK_QUEUE_CHUNKS
+       of these (and then exit at runtime if more than this many tasks are
+       launched.)
+     */
+    TaskInfo *taskInfo[MAX_TASK_QUEUE_CHUNKS];
+
+    /* We also allocate chunks of memory to service ISPCAlloc() calls.  The
+       memBuffers[] array holds pointers to this memory.  The first element
+       of this array is initialized to point to mem and then any subsequent
+       elements required are initialized with dynamic allocation.
+     */
+    int curMemBuffer, curMemBufferOffset;
+    int memBufferSize[NUM_MEM_BUFFERS];
+    char *memBuffers[NUM_MEM_BUFFERS];
+    char mem[256];
+
+};
+
+
+inline TaskGroupBase::TaskGroupBase() { 
+    nextTaskInfoIndex = 0; 
+
+    curMemBuffer = 0; 
+    curMemBufferOffset = 0;
+    memBuffers[0] = mem;
+    memBufferSize[0] = sizeof(mem) / sizeof(mem[0]);
+    for (int i = 1; i < NUM_MEM_BUFFERS; ++i) {
+        memBuffers[i] = NULL;
+        memBufferSize[i] = 0;
+    }
+
+    for (int i = 0; i < MAX_TASK_QUEUE_CHUNKS; ++i)
+        taskInfo[i] = NULL;
+}
+
+
+inline TaskGroupBase::~TaskGroupBase() {
+    // Note: don't delete memBuffers[0], since it points to the start of
+    // the "mem" member!
+    for (int i = 1; i < NUM_MEM_BUFFERS; ++i)
+        delete[] memBuffers[i];
+}
+
+
+inline void
+TaskGroupBase::Reset() {
+    nextTaskInfoIndex = 0; 
+    curMemBuffer = 0; 
+    curMemBufferOffset = 0;
+}
+
+
+inline int
+TaskGroupBase::AllocTaskInfo(int count) {
+    int ret = nextTaskInfoIndex;
+    nextTaskInfoIndex += count;
+    return ret;
+}
+
+
+inline TaskInfo *
+TaskGroupBase::GetTaskInfo(int index) {
+    int chunk = (index >> LOG_TASK_QUEUE_CHUNK_SIZE);
+    int offset = index & (TASK_QUEUE_CHUNK_SIZE-1);
+
+    if (chunk == MAX_TASK_QUEUE_CHUNKS) {
+        fprintf(stderr, "A total of %d tasks have been launched from the "
+                "current function--the simple built-in task system can handle "
+                "no more. You can increase the values of TASK_QUEUE_CHUNK_SIZE "
+                "and LOG_TASK_QUEUE_CHUNK_SIZE to work around this limitation.  "
+                "Sorry!  Exiting.\n", index);
+        exit(1);
+    }
+
+    if (taskInfo[chunk] == NULL)
+        taskInfo[chunk] = new TaskInfo[TASK_QUEUE_CHUNK_SIZE];
+    return &taskInfo[chunk][offset];
+}
+
+
+inline void *
+TaskGroupBase::AllocMemory(int64_t size, int32_t alignment) {
+    char *basePtr = memBuffers[curMemBuffer];
+    int64_t iptr = (int64_t)(basePtr + curMemBufferOffset);
+    iptr = (iptr + (alignment-1)) & ~(alignment-1);
+
+    int newOffset = int(iptr + size - (int64_t)basePtr);
+    if (newOffset < memBufferSize[curMemBuffer]) {
+        curMemBufferOffset = newOffset;
+        return (char *)iptr;
+    }
+
+    ++curMemBuffer;
+    curMemBufferOffset = 0;
+    assert(curMemBuffer < NUM_MEM_BUFFERS);
+
+    int allocSize = 1 << (12 + curMemBuffer);
+    allocSize = std::max(int(size+alignment), allocSize);
+    char *newBuf = new char[allocSize];
+    memBufferSize[curMemBuffer] = allocSize;
+    memBuffers[curMemBuffer] = newBuf;
+    return AllocMemory(size, alignment);
+}
+
+
+///////////////////////////////////////////////////////////////////////////
+// Atomics and the like
+
+#ifndef ISPC_IS_WINDOWS
+static inline void
+lMemFence() {
+    __asm__ __volatile__("mfence":::"memory");
+}
+#endif // !ISPC_IS_WINDOWS
+
+
+#if (__SIZEOF_POINTER__ == 4) || defined(__i386__) || defined(_WIN32)
+#define ISPC_POINTER_BYTES 4
+#elif (__SIZEOF_POINTER__ == 8) || defined(__x86_64__) || defined(__amd64__) || defined(_WIN64)
+#define ISPC_POINTER_BYTES 8
+#else
+#error "Pointer size unknown!"
+#endif // __SIZEOF_POINTER__
+
+
+static void *
+lAtomicCompareAndSwapPointer(void **v, void *newValue, void *oldValue) {
+#ifdef ISPC_IS_WINDOWS
+    return InterlockedCompareExchangePointer(v, newValue, oldValue);
+#else
+    void *result;
+#if (ISPC_POINTER_BYTES == 4)
+    __asm__ __volatile__("lock\ncmpxchgd %2,%1"
+                          : "=a"(result), "=m"(*v)
+                          : "q"(newValue), "0"(oldValue)
+                          : "memory");
+#else
+    __asm__ __volatile__("lock\ncmpxchgq %2,%1"
+                          : "=a"(result), "=m"(*v)
+                          : "q"(newValue), "0"(oldValue)
+                          : "memory");
+#endif // ISPC_POINTER_BYTES
+    lMemFence();
+    return result;
+#endif // ISPC_IS_WINDOWS
+}
+
+
+
+#ifndef ISPC_IS_WINDOWS
+static int32_t 
+lAtomicCompareAndSwap32(volatile int32_t *v, int32_t newValue, int32_t oldValue) {
+    int32_t result;
+    __asm__ __volatile__("lock\ncmpxchgl %2,%1"
+                          : "=a"(result), "=m"(*v)
+                          : "q"(newValue), "0"(oldValue)
+                          : "memory");
+    lMemFence();
+    return result;
+}
+#endif // !ISPC_IS_WINDOWS
+
+
+///////////////////////////////////////////////////////////////////////////
+
+#ifdef ISPC_USE_CONCRT
+// With ConcRT, we don't need to extend TaskGroupBase at all.
+class TaskGroup : public TaskGroupBase {
+public:
+    void Launch(int baseIndex, int count);
+    void Sync();
+};
+#endif // ISPC_USE_CONCRT
+
+#ifdef ISPC_USE_GCD
+/* With Grand Central Dispatch, we associate a GCD dispatch group with each
+   task group.  (We'll later wait on this dispatch group when we need to
+   wait on all of the tasks in the group to finish.)
+ */
+class TaskGroup : public TaskGroupBase {
+public:
+    TaskGroup() {
+        gcdGroup = dispatch_group_create();
+    }
+
+    void Launch(int baseIndex, int count);
+    void Sync();
+
+private:
+    dispatch_group_t gcdGroup;
+};
+#endif // ISPC_USE_GCD
+
+#ifdef ISPC_USE_PTHREADS
+static void *lTaskEntry(void *arg);
+
+class TaskGroup : public TaskGroupBase {
+public:
+    TaskGroup() {
+        numUnfinishedTasks = 0;
+        waitingTasks.reserve(128);
+        inActiveList = false;
+    }
+
+    void Reset() {
+        TaskGroupBase::Reset();
+        numUnfinishedTasks = 0;
+        assert(inActiveList == false);
+        lMemFence();
+    }
+
+    void Launch(int baseIndex, int count);
+    void Sync();
+
+private:
+    friend void *lTaskEntry(void *arg);
+
+    int32_t numUnfinishedTasks;
+    int32_t pad[3];
+    std::vector<int> waitingTasks;
+    bool inActiveList;
+};
+
+#endif // ISPC_USE_PTHREADS
+
+
+///////////////////////////////////////////////////////////////////////////
+// Grand Central Dispatch
+
+#ifdef ISPC_USE_GCD
+
+/* A simple task system for ispc programs based on Apple's Grand Central
+   Dispatch. */
+
+static dispatch_queue_t gcdQueue;
+static volatile int32_t lock = 0;
+
+static void
+InitTaskSystem() {
+    if (gcdQueue != NULL)
+        return;
+
+    while (1) {
+        if (lAtomicCompareAndSwap32(&lock, 1, 0) == 0) {
+            if (gcdQueue == NULL) {
+                gcdQueue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
+                assert(gcdQueue != NULL);
+                lMemFence();
+            }
+            lock = 0;
+            break;
+        }
+    }
+}
+
+
+static void
+lRunTask(void *ti) {
+    TaskInfo *taskInfo = (TaskInfo *)ti;
+    // 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
+    taskInfo->func(taskInfo->data, threadIndex, threadCount, 
+                   taskInfo->taskIndex, taskInfo->taskCount);
+}
+
+
+inline void
+TaskGroup::Launch(int baseIndex, int count) {
+    for (int i = 0; i < count; ++i) {
+        TaskInfo *ti = GetTaskInfo(baseIndex + i);
+        dispatch_group_async_f(gcdGroup, gcdQueue, ti, lRunTask);
+    }
+}
+
+
+inline void
+TaskGroup::Sync() {
+    dispatch_group_wait(gcdGroup, DISPATCH_TIME_FOREVER);
+}
+
+#endif // ISPC_USE_GCD
+
+///////////////////////////////////////////////////////////////////////////
+// Concurrency Runtime
+
+#ifdef ISPC_USE_CONCRT
+
+static void
+InitTaskSystem() {
+    // No initialization needed
+}
+
+
+static void __cdecl
+lRunTask(LPVOID param) {
+    TaskInfo *ti = (TaskInfo *)param;
+    
+    // Actually run the task. 
+    // FIXME: like the GCD implementation for OS X, this is passing bogus
+    // values for the threadIndex and threadCount builtins, which in turn
+    // will cause bugs in code that uses those.
+    int threadIndex = 0;
+    int threadCount = 1;
+    ti->func(ti->data, threadIndex, threadCount, ti->taskIndex, ti->taskCount);
+
+    // Signal the event that this task is done
+    ti->taskEvent.set();
+}
+
+
+inline void
+TaskGroup::Launch(int baseIndex, int count) {
+    for (int i = 0; i < count; ++i)
+        CurrentScheduler::ScheduleTask(lRunTask, GetTaskInfo(baseIndex + i));
+}
+
+
+inline void
+TaskGroup::Sync() {
+    for (int i = 0; i < nextTaskInfoIndex; ++i) {
+        TaskInfo *ti = GetTaskInfo(i);
+        ti->taskEvent.wait();
+        ti->taskEvent.reset();
+    }
+}
+
+#endif // ISPC_USE_CONCRT
+
+///////////////////////////////////////////////////////////////////////////
+// pthreads
+
+#ifdef ISPC_USE_PTHREADS
+
+static volatile int32_t lock = 0;
+
+static int nThreads;
+static pthread_t *threads = NULL;
+
+static pthread_mutex_t taskSysMutex;
+static std::vector<TaskGroup *> activeTaskGroups;
+static sem_t *workerSemaphore;
+
+
+static inline int32_t 
+lAtomicAdd(int32_t *v, int32_t delta) {
+    int32_t origValue;
+    __asm__ __volatile__("lock\n"
+                         "xaddl %0,%1"
+                         : "=r"(origValue), "=m"(*v) : "0"(delta)
+                         : "memory");
+    return origValue;
+}
+
+
+static void *
+lTaskEntry(void *arg) {
+    int threadIndex = (int)((int64_t)arg);
+    int threadCount = nThreads;
+
+    while (1) {
+        int err;
+        //
+        // Wait on the semaphore until we're woken up due to the arrival of
+        // more work.
+        //
+        if ((err = sem_wait(workerSemaphore)) != 0) {
+            fprintf(stderr, "Error from sem_wait: %s\n", strerror(err));
+            exit(1);
+        }
+
+        //
+        // Acquire the mutex
+        //
+        if ((err = pthread_mutex_lock(&taskSysMutex)) != 0) {
+            fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
+            exit(1);
+        }
+
+        if (activeTaskGroups.size() == 0) {
+            //
+            // Task queue is empty, go back and wait on the semaphore
+            //
+            if ((err = pthread_mutex_unlock(&taskSysMutex)) != 0) {
+                fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
+                exit(1);
+            }
+            continue;
+        }
+
+        //
+        // Get the last task group on the active list and the last task
+        // from its waiting tasks list.
+        //
+        TaskGroup *tg = activeTaskGroups.back();
+        assert(tg->waitingTasks.size() > 0);
+        int taskNumber = tg->waitingTasks.back();
+        tg->waitingTasks.pop_back();
+
+        if (tg->waitingTasks.size() == 0) {
+            // We just took the last task from this task group, so remove
+            // it from the active list.
+            activeTaskGroups.pop_back();
+            tg->inActiveList = false;
+        }
+    
+        if ((err = pthread_mutex_unlock(&taskSysMutex)) != 0) {
+            fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
+            exit(1);
+        }
+
+        //
+        // And now actually run the task
+        //
+        DBG(fprintf(stderr, "running task %d from group %p\n", taskNumber, tg));
+        TaskInfo *myTask = tg->GetTaskInfo(taskNumber);
+        myTask->func(myTask->data, threadIndex, threadCount, myTask->taskIndex,
+                     myTask->taskCount);
+
+        //
+        // Decrement the "number of unfinished tasks" counter in the task
+        // group.
+        //
+        lMemFence();
+        lAtomicAdd(&tg->numUnfinishedTasks, -1);
+    }
+
+    pthread_exit(NULL);
+    return 0;
+}
+
+
+static void
+InitTaskSystem() {
+    if (threads == NULL) {
+        while (1) {
+            if (lAtomicCompareAndSwap32(&lock, 1, 0) == 0) {
+                if (threads == NULL) {
+                    // We launch one fewer thread than there are cores,
+                    // since the main thread here will also grab jobs from
+                    // the task queue itself.
+                    nThreads = sysconf(_SC_NPROCESSORS_ONLN) - 1;
+
+                    int err;
+                    if ((err = pthread_mutex_init(&taskSysMutex, NULL)) != 0) {
+                        fprintf(stderr, "Error creating mutex: %s\n", strerror(err));
+                        exit(1);
+                    }
+
+                    char name[32];
+                    sprintf(name, "ispc_task.%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);
+                    }
+
+                    threads = (pthread_t *)malloc(nThreads * sizeof(pthread_t));
+                    for (int i = 0; i < nThreads; ++i) {
+                        err = pthread_create(&threads[i], NULL, &lTaskEntry, (void *)(i));
+                        if (err != 0) {
+                            fprintf(stderr, "Error creating pthread %d: %s\n", i, strerror(err));
+                            exit(1);
+                        }
+                    }
+
+                    activeTaskGroups.reserve(64);
+                }
+
+                // Make sure all of the above goes to memory before we
+                // clear the lock.
+                lMemFence();
+                lock = 0;
+                break;
+            }
+        }
+    }
+}
+
+
+inline void
+TaskGroup::Launch(int baseCoord, int count) {
+    //
+    // Acquire mutex, add task
+    //
+    int err;
+    if ((err = pthread_mutex_lock(&taskSysMutex)) != 0) {
+        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
+        exit(1);
+    }
+
+    // Add the corresponding set of tasks to the waiting-to-be-run list for
+    // this task group.
+    //
+    // FIXME: it's a little ugly to hold a global mutex for this when we
+    // only need to make sure no one else is accessing this task group's
+    // waitingTasks list.  (But a small experiment in switching to a
+    // per-TaskGroup mutex showed worse performance!)
+    for (int i = 0; i < count; ++i)
+        waitingTasks.push_back(baseCoord + i);
+
+    // Add the task group to the global active list if it isn't there
+    // already.
+    if (inActiveList == false) {
+        activeTaskGroups.push_back(this);
+        inActiveList = true;
+    }
+
+    if ((err = pthread_mutex_unlock(&taskSysMutex)) != 0) {
+        fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
+        exit(1);
+    }
+
+    //
+    // Update the count of the number of tasks left to run in this task
+    // group.
+    //
+    lMemFence();
+    lAtomicAdd(&numUnfinishedTasks, count);
+
+    //
+    // Post to the worker semaphore to wake up worker threads that are
+    // sleeping waiting for tasks to show up
+    //
+    for (int i = 0; i < count; ++i)
+        if ((err = sem_post(workerSemaphore)) != 0) {
+            fprintf(stderr, "Error from sem_post: %s\n", strerror(err));
+            exit(1);
+        }
+}
+
+
+inline void
+TaskGroup::Sync() {
+    DBG(fprintf(stderr, "syncing %p - %d unfinished\n", tg, numUnfinishedTasks));
+
+    while (numUnfinishedTasks > 0) {
+        // All of the tasks in this group aren't finished yet.  We'll try
+        // to help out here since we don't have anything else to do...
+
+        DBG(fprintf(stderr, "while syncing %p - %d unfinished\n", tg, 
+                    numUnfinishedTasks));
+
+        //
+        // Acquire the global task system mutex to grab a task to work on
+        //
+        int err;
+        if ((err = pthread_mutex_lock(&taskSysMutex)) != 0) {
+            fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
+            exit(1);
+        }
+
+        TaskInfo *myTask = NULL;
+        TaskGroup *runtg = this;
+        if (waitingTasks.size() > 0) {
+            int taskNumber = waitingTasks.back();
+            waitingTasks.pop_back();
+
+            if (waitingTasks.size() == 0) {
+                // There's nothing left to start running from this group,
+                // so remove it from the active task list.
+                activeTaskGroups.erase(std::find(activeTaskGroups.begin(),
+                                                 activeTaskGroups.end(), this));
+                inActiveList = false;
+            }
+            myTask = GetTaskInfo(taskNumber);
+            DBG(fprintf(stderr, "running task %d from group %p in sync\n", taskNumber, tg));
+        }
+        else {
+            // Other threads are already working on all of the tasks in
+            // this group, so we can't help out by running one ourself.
+            // We'll try to run one from another group to make ourselves
+            // useful here.
+            if (activeTaskGroups.size() == 0) {
+                // No active task groups left--there's nothing for us to do.
+                if ((err = pthread_mutex_unlock(&taskSysMutex)) != 0) {
+                    fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
+                    exit(1);
+                }
+                // FIXME: We basically end up busy-waiting here, which is
+                // extra wasteful in a world with hyperthreading.  It would
+                // be much better to put this thread to sleep on a
+                // condition variable that was signaled when the last task
+                // in this group was finished.
+                sleep(0);
+                continue;
+            }
+
+            // Get a task to run from another task group.
+            runtg = activeTaskGroups.back();
+            assert(runtg->waitingTasks.size() > 0);
+
+            int taskNumber = runtg->waitingTasks.back();
+            runtg->waitingTasks.pop_back();
+            if (runtg->waitingTasks.size() == 0) {
+                // There's left to start running from this group, so remove
+                // it from the active task list.
+                activeTaskGroups.pop_back();
+                runtg->inActiveList = false;
+            }
+            myTask = runtg->GetTaskInfo(taskNumber);
+            DBG(fprintf(stderr, "running task %d from other group %p in sync\n", 
+                        taskNumber, runtg));
+        }
+
+        if ((err = pthread_mutex_unlock(&taskSysMutex)) != 0) {
+            fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
+            exit(1);
+        }
+    
+        //
+        // Do work for _myTask_
+        //
+        // FIXME: bogus values for thread index/thread count here as well..
+        myTask->func(myTask->data, 0, 1, myTask->taskIndex, myTask->taskCount);
+
+        //
+        // Decrement the number of unfinished tasks counter
+        //
+        lMemFence();
+        lAtomicAdd(&runtg->numUnfinishedTasks, -1);
+    }
+    DBG(fprintf(stderr, "sync for %p done!n", tg));
+}
+
+#endif // ISPC_USE_PTHREADS
+
+///////////////////////////////////////////////////////////////////////////
+
+#define MAX_FREE_TASK_GROUPS 64
+static TaskGroup *freeTaskGroups[MAX_FREE_TASK_GROUPS];
+
+static inline TaskGroup *
+AllocTaskGroup() {
+    for (int i = 0; i < MAX_FREE_TASK_GROUPS; ++i) {
+        TaskGroup *tg = freeTaskGroups[i];
+        if (tg != NULL) {
+            void *ptr = lAtomicCompareAndSwapPointer((void **)(&freeTaskGroups[i]), NULL, tg);
+            if (ptr != NULL) {
+                assert(ptr == tg);
+                return (TaskGroup *)ptr;
+            }
+        }
+    }
+
+    return new TaskGroup;
+}
+
+
+static inline void
+FreeTaskGroup(TaskGroup *tg) {
+    tg->Reset();
+
+    for (int i = 0; i < MAX_FREE_TASK_GROUPS; ++i) {
+        if (freeTaskGroups[i] == NULL) {
+            void *ptr = lAtomicCompareAndSwapPointer((void **)&freeTaskGroups[i], tg, NULL);
+            if (ptr == NULL)
+                return;
+        }
+    }
+
+    delete tg;
+}
+
+///////////////////////////////////////////////////////////////////////////
+
+// ispc expects these functions to have C linkage / not be mangled
+extern "C" { 
+    void ISPCLaunch(void **handlePtr, void *f, void *data, int count);
+    void *ISPCAlloc(void **handlePtr, int64_t size, int32_t alignment);
+    void ISPCSync(void *handle);
+}
+
+void
+ISPCLaunch(void **taskGroupPtr, void *func, void *data, int count) {
+    TaskGroup *taskGroup;
+    if (*taskGroupPtr == NULL) {
+        InitTaskSystem();
+        taskGroup = AllocTaskGroup();
+        *taskGroupPtr = taskGroup;
+    }
+    else
+        taskGroup = (TaskGroup *)(*taskGroupPtr);
+
+    int baseIndex = taskGroup->AllocTaskInfo(count);
+    for (int i = 0; i < count; ++i) {
+        TaskInfo *ti = taskGroup->GetTaskInfo(baseIndex+i);
+        ti->func = (TaskFuncType)func;
+        ti->data = data;
+        ti->taskIndex = i;
+        ti->taskCount = count;
+    }
+    taskGroup->Launch(baseIndex, count);
+}
+
+
+void
+ISPCSync(void *h) {
+    TaskGroup *taskGroup = (TaskGroup *)h;
+    if (taskGroup != NULL) {
+        taskGroup->Sync();
+        FreeTaskGroup(taskGroup);
+    }
+}
+
+
+void *
+ISPCAlloc(void **taskGroupPtr, int64_t size, int32_t alignment) {
+    TaskGroup *taskGroup;
+    if (*taskGroupPtr == NULL) {
+        InitTaskSystem();
+        taskGroup = AllocTaskGroup();
+        *taskGroupPtr = taskGroup;
+    }
+    else
+        taskGroup = (TaskGroup *)(*taskGroupPtr);
+
+    return taskGroup->AllocMemory(size, alignment);
+}

examples/timing.h

@@ -38,7 +38,9 @@
 #include <windows.h>
 #define rdtsc __rdtsc
 #else
+#ifdef __cplusplus
 extern "C" {
+#endif /* __cplusplus */
     __inline__ uint64_t rdtsc() {
         uint32_t low, high;
         __asm__ __volatile__ (
@@ -48,7 +50,9 @@ extern "C" {
                               "rdtsc" : "=a" (low), "=d" (high));
         return (uint64_t)high << 32 | low;
     }
+#ifdef __cplusplus
 }
+#endif /* __cplusplus */
 #endif            
             
 static uint64_t start, end;

examples/volume_rendering/.gitignore

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

examples/volume_rendering/Makefile

@@ -0,0 +1,35 @@
+
+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=sse4x2 --arch=x86-64
+
+default: volume
+
+.PHONY: dirs clean
+
+dirs:
+	/bin/mkdir -p objs/
+
+clean:
+	/bin/rm -rf objs *~ volume
+
+volume: dirs objs/volume.o objs/volume_serial.o objs/volume_ispc.o $(TASK_OBJ)
+	$(CXX) $(CXXFLAGS) -o $@ objs/volume.o objs/volume_ispc.o objs/volume_serial.o $(TASK_OBJ) -lm $(TASK_LIB)
+
+objs/%.o: %.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/%.o: ../%.cpp
+	$(CXX) $< $(CXXFLAGS) -c -o $@
+
+objs/volume.o: objs/volume_ispc.h 
+
+objs/%_ispc.h objs/%_ispc.o: %.ispc
+	$(ISPC) $(ISPCFLAGS) $< -o objs/$*_ispc.o -h objs/$*_ispc.h

examples/volume_rendering/camera.dat

@@ -0,0 +1,11 @@
+896 1184
+
+0.000155 0.000000 0.000000 -0.069927
+0.000000 -0.000155 0.000000 0.093236
+0.000000 0.000000 0.000000 1.000000
+0.000000 0.000000 -99.999001 100.000000
+
+1.000000 0.000000 0.000000 1.000000
+0.000000 0.980129 -0.198360 2.900000
+0.000000 0.198360 0.980129 -10.500000
+0.000000 0.000000 0.000000 1.000000

examples/volume_rendering/volume.cpp

@@ -0,0 +1,248 @@
+/*
+  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 NOMINMAX
+#pragma warning (disable: 4244)
+#pragma warning (disable: 4305)
+#endif
+
+#include <stdio.h>
+#include <algorithm>
+#include "../timing.h"
+#include "../cpuid.h"
+#include "volume_ispc.h"
+using namespace ispc;
+
+extern void volume_serial(float density[], int nVoxels[3], 
+                          const float raster2camera[4][4],
+                          const float camera2world[4][4], 
+                          int width, int height, float image[]);
+
+/* 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.f) v = 0.f;
+        else if (v > 255.f) v = 255.f;
+        unsigned char c = (unsigned char)v;
+        for (int j = 0; j < 3; ++j)
+            fputc(c, fp);
+    }
+    fclose(fp);
+    printf("Wrote image file %s\n", fn);
+}
+
+
+// Make sure that the vector ISA used during compilation is supported by
+// the processor.  The ISPC_TARGET_* macro is set in the ispc-generated
+// header file that we include above.
+static void
+ensureTargetISAIsSupported() {
+#if defined(ISPC_TARGET_SSE2)
+    bool isaSupported = CPUSupportsSSE2();
+    const char *target = "SSE2";
+#elif defined(ISPC_TARGET_SSE4)
+    bool isaSupported = CPUSupportsSSE4();
+    const char *target = "SSE4";
+#elif defined(ISPC_TARGET_AVX)
+    bool isaSupported = CPUSupportsAVX();
+    const char *target = "AVX";
+#else
+#error "Unknown ISPC_TARGET_* value"
+#endif
+    if (!isaSupported) {
+        fprintf(stderr, "***\n*** Error: the ispc-compiled code uses the %s instruction "
+                "set, which isn't\n***        supported by this computer's CPU!\n", target);
+        fprintf(stderr, "***\n***        Please modify the "
+#ifdef _MSC_VER
+                "MSVC project file "
+#else
+                "Makefile "
+#endif
+                "to select another target (e.g. sse2)\n***\n");
+        exit(1);
+    }
+}
+
+/* Load image and viewing parameters from a camera data file.
+   FIXME: we should add support to be able to specify viewing parameters
+   in the program here directly. */
+static void
+loadCamera(const char *fn, int *width, int *height, float raster2camera[4][4],
+           float camera2world[4][4]) {
+    FILE *f = fopen(fn, "r");
+    if (!f) {
+        perror(fn);
+        exit(1);
+    }
+    if (fscanf(f, "%d %d", width, height) != 2) {
+        fprintf(stderr, "Unexpected end of file in camera file\n");
+        exit(1);
+    }
+
+    for (int i = 0; i < 4; ++i) {
+        for (int j = 0; j < 4; ++j) {
+            if (fscanf(f, "%f", &raster2camera[i][j]) != 1) {
+                fprintf(stderr, "Unexpected end of file in camera file\n");
+                exit(1);
+            }
+        }
+    }
+    for (int i = 0; i < 4; ++i) {
+        for (int j = 0; j < 4; ++j) {
+            if (fscanf(f, "%f", &camera2world[i][j]) != 1) {
+                fprintf(stderr, "Unexpected end of file in camera file\n");
+                exit(1);
+            }
+        }
+    }
+    fclose(f);
+}
+
+
+/* Load a volume density file.  Expects the number of x, y, and z samples
+   as the first three values (as integer strings), then x*y*z
+   floating-point values (also as strings) to give the densities.  */
+static float *
+loadVolume(const char *fn, int n[3]) {
+    FILE *f = fopen(fn, "r");
+    if (!f) {
+        perror(fn);
+        exit(1);
+    }
+
+    if (fscanf(f, "%d %d %d", &n[0], &n[1], &n[2]) != 3) {
+        fprintf(stderr, "Couldn't find resolution at start of density file\n");
+        exit(1);
+    }
+
+    int count = n[0] * n[1] * n[2];
+    float *v = new float[count];
+    for (int i = 0; i < count; ++i) {
+        if (fscanf(f, "%f", &v[i]) != 1) {
+            fprintf(stderr, "Unexpected end of file at %d'th density value\n", i);
+            exit(1);
+        }
+    }
+
+    return v;
+}
+
+
+int main(int argc, char *argv[]) {
+    if (argc != 3) {
+        fprintf(stderr, "usage: volume <camera.dat> <volume_density.vol>\n");
+        return 1;
+    }
+
+    ensureTargetISAIsSupported();
+
+    //
+    // Load viewing data and the volume density data
+    //
+    int width, height;
+    float raster2camera[4][4], camera2world[4][4];
+    loadCamera(argv[1], &width, &height, raster2camera, camera2world);
+    float *image = new float[width*height];
+
+    int n[3];
+    float *density = loadVolume(argv[2], n);
+
+    //
+    // 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();
+        volume_ispc(density, n, raster2camera, camera2world,
+                    width, height, image);
+        double dt = get_elapsed_mcycles();
+        minISPC = std::min(minISPC, dt);
+    }
+
+    printf("[volume ispc 1 core]:\t\t[%.3f] million cycles\n", minISPC);
+    writePPM(image, width, height, "volume-ispc-1core.ppm");
+
+    // Clear out the buffer
+    for (int i = 0; i < width * height; ++i)
+        image[i] = 0.;
+
+    //
+    // Compute the image using the ispc implementation that also uses
+    // tasks; report the minimum time of three runs.
+    //
+    double minISPCtasks = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        reset_and_start_timer();
+        volume_ispc_tasks(density, n, raster2camera, camera2world,
+                          width, height, image);
+        double dt = get_elapsed_mcycles();
+        minISPCtasks = std::min(minISPCtasks, dt);
+    }
+
+    printf("[volume ispc + tasks]:\t\t[%.3f] million cycles\n", minISPCtasks);
+    writePPM(image, width, height, "volume-ispc-tasks.ppm");
+
+    // Clear out the buffer
+    for (int i = 0; i < width * height; ++i)
+        image[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();
+        volume_serial(density, n, raster2camera, camera2world,
+                      width, height, image);
+        double dt = get_elapsed_mcycles();
+        minSerial = std::min(minSerial, dt);
+    }
+
+    printf("[volume serial]:\t\t[%.3f] millon cycles\n", minSerial);
+    writePPM(image, width, height, "volume-serial.ppm");
+
+    printf("\t\t\t\t(%.2fx speedup from ISPC serial, %.2fx from ISPC+tasks)\n", 
+           minSerial/minISPC, minSerial / minISPCtasks);
+
+    return 0;
+}

examples/volume_rendering/volume.ispc

@@ -0,0 +1,385 @@
+/*
+  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.  
+*/
+
+typedef float<3> float3;
+
+struct Ray {
+    float3 origin, dir;
+};
+
+
+static void
+generateRay(const uniform float raster2camera[4][4], 
+            const uniform float camera2world[4][4],
+            float x, float y, reference Ray ray) {
+    // transform raster coordinate (x, y, 0) to camera space
+    float camx = raster2camera[0][0] * x + raster2camera[0][1] * y + raster2camera[0][3];
+    float camy = raster2camera[1][0] * x + raster2camera[1][1] * y + raster2camera[1][3];
+    float camz = raster2camera[2][3];
+    float camw = raster2camera[3][3];
+    camx /= camw;
+    camy /= camw;
+    camz /= camw;
+
+    ray.dir.x = camera2world[0][0] * camx + camera2world[0][1] * camy + camera2world[0][2] * camz;
+    ray.dir.y = camera2world[1][0] * camx + camera2world[1][1] * camy + camera2world[1][2] * camz;
+    ray.dir.z = camera2world[2][0] * camx + camera2world[2][1] * camy + camera2world[2][2] * camz;
+
+    ray.origin.x = camera2world[0][3] / camera2world[3][3];
+    ray.origin.y = camera2world[1][3] / camera2world[3][3];
+    ray.origin.z = camera2world[2][3] / camera2world[3][3];
+}
+
+
+static inline bool
+Inside(float3 p, float3 pMin, float3 pMax) {
+    return (p.x >= pMin.x && p.x <= pMax.x &&
+            p.y >= pMin.y && p.y <= pMax.y &&
+            p.z >= pMin.z && p.z <= pMax.z);
+}
+
+
+static bool
+IntersectP(Ray ray, float3 pMin, float3 pMax, reference float hit0, reference float hit1) {
+    float t0 = -1e30, t1 = 1e30;
+
+    float3 tNear = (pMin - ray.origin) / ray.dir;
+    float3 tFar  = (pMax - ray.origin) / ray.dir;
+    if (tNear.x > tFar.x) {
+        float tmp = tNear.x;
+        tNear.x = tFar.x;
+        tFar.x = tmp;
+    }
+    t0 = max(tNear.x, t0);
+    t1 = min(tFar.x, t1);
+
+    if (tNear.y > tFar.y) {
+        float tmp = tNear.y;
+        tNear.y = tFar.y;
+        tFar.y = tmp;
+    }
+    t0 = max(tNear.y, t0);
+    t1 = min(tFar.y, t1);
+
+    if (tNear.z > tFar.z) {
+        float tmp = tNear.z;
+        tNear.z = tFar.z;
+        tFar.z = tmp;
+    }
+    t0 = max(tNear.z, t0);
+    t1 = min(tFar.z, t1);
+    
+    if (t0 <= t1) {
+        hit0 = t0;
+        hit1 = t1;
+        return true;
+    }
+    else
+        return false;
+}
+
+
+static inline float Lerp(float t, float a, float b) {
+    return (1.f - t) * a + t * b;
+}
+
+
+static inline float D(int x, int y, int z, uniform int nVoxels[3], 
+                      uniform float density[]) {
+    x = clamp(x, 0, nVoxels[0]-1);
+    y = clamp(y, 0, nVoxels[1]-1);
+    z = clamp(z, 0, nVoxels[2]-1);
+
+    return density[z*nVoxels[0]*nVoxels[1] + y*nVoxels[0] + x];
+}
+
+
+static inline float Du(uniform int x, uniform int y, uniform int z, 
+                       uniform int nVoxels[3], uniform float density[]) {
+    x = clamp(x, 0, nVoxels[0]-1);
+    y = clamp(y, 0, nVoxels[1]-1);
+    z = clamp(z, 0, nVoxels[2]-1);
+
+    return density[z*nVoxels[0]*nVoxels[1] + y*nVoxels[0] + x];
+}
+
+
+static inline float3 Offset(float3 p, float3 pMin, float3 pMax) {
+    return (p - pMin) / (pMax - pMin);
+}
+
+
+static inline float Density(float3 Pobj, float3 pMin, float3 pMax, 
+                            uniform float density[], uniform int nVoxels[3],
+                            reference uniform bool checkForSameVoxel) {
+    if (!Inside(Pobj, pMin, pMax)) 
+        return 0;
+    // Compute voxel coordinates and offsets for _Pobj_
+    float3 vox = Offset(Pobj, pMin, pMax);
+    vox.x = vox.x * nVoxels[0] - .5f;
+    vox.y = vox.y * nVoxels[1] - .5f;
+    vox.z = vox.z * nVoxels[2] - .5f;
+    int vx = (int)(vox.x), vy = (int)(vox.y), vz = (int)(vox.z);
+    float dx = vox.x - vx, dy = vox.y - vy, dz = vox.z - vz;
+
+    // Trilinearly interpolate density values to compute local density
+    float d00, d10, d01, d11;
+    uniform int uvx, uvy, uvz;
+    if (checkForSameVoxel && reduce_equal(vx, uvx) && reduce_equal(vy, uvy) &&
+        reduce_equal(vz, uvz)) {
+        // If all of the program instances are inside the same voxel, then
+        // we'll call the 'uniform' variant of the voxel density lookup
+        // function, thus doing a single load for each value rather than a
+        // gather.
+        d00 = Lerp(dx, Du(uvx, uvy, uvz, nVoxels, density),     
+                       Du(uvx+1, uvy, uvz, nVoxels, density));
+        d10 = Lerp(dx, Du(uvx, uvy+1, uvz, nVoxels, density),   
+                       Du(uvx+1, uvy+1, uvz, nVoxels, density));
+        d01 = Lerp(dx, Du(uvx, uvy, uvz+1, nVoxels, density),   
+                       Du(uvx+1, uvy, uvz+1, nVoxels, density));
+        d11 = Lerp(dx, Du(uvx, uvy+1, uvz+1, nVoxels, density), 
+                       Du(uvx+1, uvy+1, uvz+1, nVoxels, density));
+    }
+    else {
+        // Otherwise, we have to do an actual gather in the more general
+        // D() function.  Once the reduce_equal tests above fail, we stop
+        // checking in subsequent steps, since it's unlikely that this will
+        // be true in the future once they've diverged into different
+        // voxels.
+        checkForSameVoxel = false;
+        d00 = Lerp(dx, D(vx, vy, vz, nVoxels, density),     
+                       D(vx+1, vy, vz, nVoxels, density));
+        d10 = Lerp(dx, D(vx, vy+1, vz, nVoxels, density),   
+                       D(vx+1, vy+1, vz, nVoxels, density));
+        d01 = Lerp(dx, D(vx, vy, vz+1, nVoxels, density),   
+                       D(vx+1, vy, vz+1, nVoxels, density));
+        d11 = Lerp(dx, D(vx, vy+1, vz+1, nVoxels, density), 
+                       D(vx+1, vy+1, vz+1, nVoxels, density));
+    }
+    float d0 = Lerp(dy, d00, d10);
+    float d1 = Lerp(dy, d01, d11);
+    return Lerp(dz, d0, d1);
+}
+
+
+/* Returns the transmittance between two points p0 and p1, in a volume
+   with extent (pMin,pMax) with transmittance coefficient sigma_t,
+   defined by nVoxels[3] voxels in each dimension in the given density
+   array. */
+static float
+transmittance(uniform float3 p0, float3 p1, uniform float3 pMin,
+              uniform float3 pMax, uniform float sigma_t, 
+              uniform float density[], uniform int nVoxels[3]) {
+    float rayT0, rayT1;
+    Ray ray;
+    ray.origin = p1;
+    ray.dir = p0 - p1;
+
+    // Find the parametric t range along the ray that is inside the volume.
+    if (!IntersectP(ray, pMin, pMax, rayT0, rayT1))
+        return 1.;
+
+    rayT0 = max(rayT0, 0.f);
+
+    // Accumulate beam transmittance in tau
+    float tau = 0;
+    float rayLength = sqrt(ray.dir.x * ray.dir.x + ray.dir.y * ray.dir.y +
+                           ray.dir.z * ray.dir.z);
+    uniform float stepDist = 0.2;
+    float stepT = stepDist / rayLength;
+
+    float t = rayT0;
+    float3 pos = ray.origin + ray.dir * rayT0;
+    float3 dirStep = ray.dir * stepT;
+    uniform bool checkForSameVoxel = true;
+    while (t < rayT1) {
+        tau += stepDist * sigma_t * Density(pos, pMin, pMax, density, nVoxels,
+                                            checkForSameVoxel);
+        pos = pos + dirStep;
+        t += stepT;
+    }
+
+    return exp(-tau);
+}
+
+
+static inline float
+distanceSquared(float3 a, float3 b) {
+    float3 d = a-b;
+    return d.x*d.x + d.y*d.y + d.z*d.z;
+}
+
+
+static float 
+raymarch(uniform float density[], uniform int nVoxels[3], Ray ray) {
+    float rayT0, rayT1;
+    uniform float3 pMin = {.3, -.2, .3}, pMax = {1.8, 2.3, 1.8};
+    uniform float3 lightPos = { -1, 4, 1.5 };
+
+    cif (!IntersectP(ray, pMin, pMax, rayT0, rayT1))
+        return 0.;
+
+    rayT0 = max(rayT0, 0.f);
+
+    // Parameters that define the volume scattering characteristics and
+    // sampling rate for raymarching
+    uniform float Le = .25;            // Emission coefficient
+    uniform float sigma_a = 10;        // Absorption coefficient
+    uniform float sigma_s = 10;        // Scattering coefficient
+    uniform float stepDist = 0.025;    // Ray step amount
+    uniform float lightIntensity = 40; // Light source intensity
+
+    float tau = 0.f;  // accumulated beam transmittance
+    float L = 0;      // radiance along the ray
+    float rayLength = sqrt(ray.dir.x * ray.dir.x + ray.dir.y * ray.dir.y +
+                           ray.dir.z * ray.dir.z);
+    float stepT = stepDist / rayLength;
+
+    float t = rayT0;
+    float3 pos = ray.origin + ray.dir * rayT0;
+    float3 dirStep = ray.dir * stepT;
+    uniform bool checkForSameVoxel = true;
+    cwhile (t < rayT1) {
+        float d = Density(pos, pMin, pMax, density, nVoxels, checkForSameVoxel);
+
+        // terminate once attenuation is high
+        float atten = exp(-tau);
+        if (atten < .005)
+            cbreak;
+
+        // direct lighting
+        float Li = lightIntensity / distanceSquared(lightPos, pos) * 
+            transmittance(lightPos, pos, pMin, pMax, sigma_a + sigma_s,
+                          density, nVoxels);
+        L += stepDist * atten * d * sigma_s * (Li + Le);
+
+        // update beam transmittance
+        tau += stepDist * (sigma_a + sigma_s) * d;
+
+        pos = pos + dirStep;
+        t += stepT;
+    }
+
+    // Gamma correction
+    return pow(L, 1.f / 2.2f);
+}
+
+
+/* Utility routine used by both the task-based and the single-core entrypoints.
+   Renders a tile of the image, covering [x0,x0) * [y0, y1), storing the
+   result into the image[] array.
+ */
+static void
+volume_tile(uniform int x0, uniform int y0, uniform int x1,
+            uniform int y1, uniform float density[], uniform int nVoxels[3], 
+            const uniform float raster2camera[4][4],
+            const uniform float camera2world[4][4], 
+            uniform int width, uniform int height, uniform float image[]) {
+    // Work on 4x4=16 pixel big tiles of the image.  This function thus
+    // implicitly assumes that both (x1-x0) and (y1-y0) are evenly divisble
+    // by 4.
+    for (uniform int y = y0; y < y1; y += 4) {
+        for (uniform int x = x0; x < x1; x += 4) {
+            // For each such tile, process programCount pixels at a time,
+            // until we've done all 16 of them.  Thus, we're also assuming
+            // that programCount <= 16 and that 16 is evenly dividible by
+            // programCount.
+            for (uniform int o = 0; o < 16; o += programCount) {
+                // These two arrays encode the mapping from [0,15] to
+                // offsets within the 4x4 pixel block so that we render
+                // each pixel inside the block
+                const uniform int xoffsets[16] = { 0, 1, 0, 1, 2, 3, 2, 3,
+                                                   0, 1, 0, 1, 2, 3, 2, 3 };
+                const uniform int yoffsets[16] = { 0, 0, 1, 1, 0, 0, 1, 1,
+                                                   2, 2, 3, 3, 2, 2, 3, 3 };
+
+                // Figure out the pixel to render for this program instance
+                int xo = x + xoffsets[o + programIndex];
+                int yo = y + yoffsets[o + programIndex];
+
+                // Use viewing parameters to compute the corresponding ray
+                // for the pixel
+                Ray ray;
+                generateRay(raster2camera, camera2world, xo, yo, ray);
+
+                // And raymarch through the volume to compute the pixel's
+                // value
+                int offset = yo * width + xo;
+                image[offset] = raymarch(density, nVoxels, ray);
+            }
+        }
+    }
+}
+
+
+task void
+volume_task(uniform float density[], uniform int nVoxels[3], 
+            const uniform float raster2camera[4][4],
+            const uniform float camera2world[4][4], 
+            uniform int width, uniform int height, uniform float image[]) {
+    uniform int dx = 8, dy = 8; // must match value in volume_ispc_tasks
+    uniform int xbuckets = (width + (dx-1)) / dx;
+    uniform int ybuckets = (height + (dy-1)) / dy;
+
+    uniform int x0 = (taskIndex % xbuckets) * dx;
+    uniform int y0 = (taskIndex / ybuckets) * dy;
+    uniform int x1 = x0 + dx, y1 = y0 + dy;
+    x1 = min(x1, width);
+    y1 = min(y1, height);
+
+    volume_tile(x0, y0, x1, y1, density, nVoxels, raster2camera,
+                 camera2world, width, height, image);
+}
+
+
+export void
+volume_ispc(uniform float density[], uniform int nVoxels[3], 
+            const uniform float raster2camera[4][4],
+            const uniform float camera2world[4][4], 
+            uniform int width, uniform int height, uniform float image[]) {
+    volume_tile(0, 0, width, height, density, nVoxels, raster2camera, 
+                camera2world, width, height,  image);
+}
+
+
+export void
+volume_ispc_tasks(uniform float density[], uniform int nVoxels[3], 
+                  const uniform float raster2camera[4][4],
+                  const uniform float camera2world[4][4], 
+                  uniform int width, uniform int height, uniform float image[]) {
+    // Launch tasks to work on (dx,dy)-sized tiles of the image
+    uniform int dx = 8, dy = 8;
+    uniform int nTasks = ((width+(dx-1))/dx) * ((height+(dy-1))/dy);
+    launch[nTasks] < volume_task(density, nVoxels, raster2camera, camera2world, 
+                                 width, height, image) >;
+}

examples/volume_rendering/volume.vcxproj

@@ -0,0 +1,168 @@
+<?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>{dee5733a-e93e-449d-9114-9bffcaeb4df9}</ProjectGuid>
+    <Keyword>Win32Proj</Keyword>
+    <RootNamespace>volume</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>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">
+    <LinkIncremental>true</LinkIncremental>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
+    <LinkIncremental>false</LinkIncremental>
+  </PropertyGroup>
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|x64'">
+    <LinkIncremental>false</LinkIncremental>
+  </PropertyGroup>
+  <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">
+    <ClCompile>
+      <PrecompiledHeader>
+      </PrecompiledHeader>
+      <WarningLevel>Level3</WarningLevel>
+      <Optimization>Disabled</Optimization>
+      <PreprocessorDefinitions>WIN32;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <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>
+      <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>
+      <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>
+      <FloatingPointModel>Fast</FloatingPointModel>
+    </ClCompile>
+    <Link>
+      <SubSystem>Console</SubSystem>
+      <GenerateDebugInformation>true</GenerateDebugInformation>
+      <EnableCOMDATFolding>true</EnableCOMDATFolding>
+      <OptimizeReferences>true</OptimizeReferences>
+    </Link>
+  </ItemDefinitionGroup>
+  <ItemGroup>
+    <ClCompile Include="volume.cpp" />
+    <ClCompile Include="volume_serial.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
+  </ItemGroup>
+  <ItemGroup>
+    <CustomBuild Include="volume.ispc">
+      <FileType>Document</FileType>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+</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'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --arch=x86 --target=sse4x2
+</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|x64'">ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h --target=sse4x2
+</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>
+    </CustomBuild>
+  </ItemGroup>
+  <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <ImportGroup Label="ExtensionTargets">
+  </ImportGroup>
+</Project>

examples/volume_rendering/volume_serial.cpp

@@ -0,0 +1,305 @@
+/*
+  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 <assert.h>
+#include <math.h>
+#include <algorithm>
+
+// Just enough of a float3 class to do what we need in this file.
+#ifdef _MSC_VER
+__declspec(align(16)) 
+#endif
+struct float3 {
+    float3() { }
+    float3(float xx, float yy, float zz) { x = xx; y = yy; z = zz; }
+
+    float3 operator*(float f) const { return float3(x*f, y*f, z*f); }
+    float3 operator-(const float3 &f2) const { 
+        return float3(x-f2.x, y-f2.y, z-f2.z); 
+    }
+    float3 operator*(const float3 &f2) const { 
+        return float3(x*f2.x, y*f2.y, z*f2.z); 
+    }
+    float3 operator+(const float3 &f2) const { 
+        return float3(x+f2.x, y+f2.y, z+f2.z); 
+    }
+    float3 operator/(const float3 &f2) const { 
+        return float3(x/f2.x, y/f2.y, z/f2.z); 
+    }
+    float operator[](int i) const { return (&x)[i]; }
+    float &operator[](int i) { return (&x)[i]; }
+
+    float x, y, z;
+    float pad;  // match padding/alignment of ispc version 
+}
+#ifndef _MSC_VER
+__attribute__ ((aligned(16)))
+#endif
+;
+
+struct Ray {
+    float3 origin, dir;
+};
+
+
+static void
+generateRay(const float raster2camera[4][4], const float camera2world[4][4],
+            float x, float y, Ray &ray) {
+    // transform raster coordinate (x, y, 0) to camera space
+    float camx = raster2camera[0][0] * x + raster2camera[0][1] * y + raster2camera[0][3];
+    float camy = raster2camera[1][0] * x + raster2camera[1][1] * y + raster2camera[1][3];
+    float camz = raster2camera[2][3];
+    float camw = raster2camera[3][3];
+    camx /= camw;
+    camy /= camw;
+    camz /= camw;
+
+    ray.dir.x = camera2world[0][0] * camx + camera2world[0][1] * camy + camera2world[0][2] * camz;
+    ray.dir.y = camera2world[1][0] * camx + camera2world[1][1] * camy + camera2world[1][2] * camz;
+    ray.dir.z = camera2world[2][0] * camx + camera2world[2][1] * camy + camera2world[2][2] * camz;
+
+    ray.origin.x = camera2world[0][3] / camera2world[3][3];
+    ray.origin.y = camera2world[1][3] / camera2world[3][3];
+    ray.origin.z = camera2world[2][3] / camera2world[3][3];
+}
+
+
+static bool
+Inside(float3 p, float3 pMin, float3 pMax) {
+    return (p.x >= pMin.x && p.x <= pMax.x &&
+            p.y >= pMin.y && p.y <= pMax.y &&
+            p.z >= pMin.z && p.z <= pMax.z);
+}
+
+
+static bool
+IntersectP(const Ray &ray, float3 pMin, float3 pMax, float *hit0, float *hit1) {
+    float t0 = -1e30f, t1 = 1e30f;
+
+    float3 tNear = (pMin - ray.origin) / ray.dir;
+    float3 tFar  = (pMax - ray.origin) / ray.dir;
+    if (tNear.x > tFar.x) {
+        float tmp = tNear.x;
+        tNear.x = tFar.x;
+        tFar.x = tmp;
+    }
+    t0 = std::max(tNear.x, t0);
+    t1 = std::min(tFar.x, t1);
+
+    if (tNear.y > tFar.y) {
+        float tmp = tNear.y;
+        tNear.y = tFar.y;
+        tFar.y = tmp;
+    }
+    t0 = std::max(tNear.y, t0);
+    t1 = std::min(tFar.y, t1);
+
+    if (tNear.z > tFar.z) {
+        float tmp = tNear.z;
+        tNear.z = tFar.z;
+        tFar.z = tmp;
+    }
+    t0 = std::max(tNear.z, t0);
+    t1 = std::min(tFar.z, t1);
+    
+    if (t0 <= t1) {
+        *hit0 = t0;
+        *hit1 = t1;
+        return true;
+    }
+    else
+        return false;
+}
+
+
+static inline float Lerp(float t, float a, float b) {
+    return (1.f - t) * a + t * b;
+}
+
+
+static inline int Clamp(int v, int low, int high) {
+    return std::min(std::max(v, low), high);
+}
+
+
+static inline float D(int x, int y, int z, int nVoxels[3], float density[]) {
+    x = Clamp(x, 0, nVoxels[0]-1);
+    y = Clamp(y, 0, nVoxels[1]-1);
+    z = Clamp(z, 0, nVoxels[2]-1);
+    return density[z*nVoxels[0]*nVoxels[1] + y*nVoxels[0] + x];
+}
+
+
+static inline float3 Offset(float3 p, float3 pMin, float3 pMax) {
+    return float3((p.x - pMin.x) / (pMax.x - pMin.x),
+                  (p.y - pMin.y) / (pMax.y - pMin.y),
+                  (p.z - pMin.z) / (pMax.z - pMin.z));
+}
+
+
+static inline float Density(float3 Pobj, float3 pMin, float3 pMax, 
+                            float density[], int nVoxels[3]) {
+    if (!Inside(Pobj, pMin, pMax)) 
+        return 0;
+    // Compute voxel coordinates and offsets for _Pobj_
+    float3 vox = Offset(Pobj, pMin, pMax);
+    vox.x = vox.x * nVoxels[0] - .5f;
+    vox.y = vox.y * nVoxels[1] - .5f;
+    vox.z = vox.z * nVoxels[2] - .5f;
+    int vx = (int)(vox.x), vy = (int)(vox.y), vz = (int)(vox.z);
+    float dx = vox.x - vx, dy = vox.y - vy, dz = vox.z - vz;
+
+    // Trilinearly interpolate density values to compute local density
+    float d00 = Lerp(dx, D(vx, vy, vz, nVoxels, density),     
+                         D(vx+1, vy, vz, nVoxels, density));
+    float d10 = Lerp(dx, D(vx, vy+1, vz, nVoxels, density),   
+                         D(vx+1, vy+1, vz, nVoxels, density));
+    float d01 = Lerp(dx, D(vx, vy, vz+1, nVoxels, density),   
+                         D(vx+1, vy, vz+1, nVoxels, density));
+    float d11 = Lerp(dx, D(vx, vy+1, vz+1, nVoxels, density), 
+                         D(vx+1, vy+1, vz+1, nVoxels, density));
+    float d0 = Lerp(dy, d00, d10);
+    float d1 = Lerp(dy, d01, d11);
+    return Lerp(dz, d0, d1);
+}
+
+
+
+static float
+transmittance(float3 p0, float3 p1, float3 pMin,
+              float3 pMax, float sigma_t, float density[], int nVoxels[3]) {
+    float rayT0, rayT1;
+    Ray ray;
+    ray.origin = p1;
+    ray.dir = p0 - p1;
+
+    // Find the parametric t range along the ray that is inside the volume.
+    if (!IntersectP(ray, pMin, pMax, &rayT0, &rayT1))
+        return 1.;
+
+    rayT0 = std::max(rayT0, 0.f);
+
+    // Accumulate beam transmittance in tau
+    float tau = 0;
+    float rayLength = sqrtf(ray.dir.x * ray.dir.x + ray.dir.y * ray.dir.y +
+                            ray.dir.z * ray.dir.z);
+    float stepDist = 0.2f;
+    float stepT = stepDist / rayLength;
+
+    float t = rayT0;
+    float3 pos = ray.origin + ray.dir * rayT0;
+    float3 dirStep = ray.dir * stepT;
+    while (t < rayT1) {
+        tau += stepDist * sigma_t * Density(pos, pMin, pMax, density, nVoxels);
+        pos = pos + dirStep;
+        t += stepT;
+    }
+
+    return expf(-tau);
+}
+
+
+static float
+distanceSquared(float3 a, float3 b) {
+    float3 d = a-b;
+    return d.x*d.x + d.y*d.y + d.z*d.z;
+}
+
+
+static float 
+raymarch(float density[], int nVoxels[3], const Ray &ray) {
+    float rayT0, rayT1;
+    float3 pMin(.3f, -.2f, .3f), pMax(1.8f, 2.3f, 1.8f);
+    float3 lightPos(-1.f, 4.f, 1.5f);
+
+    if (!IntersectP(ray, pMin, pMax, &rayT0, &rayT1))
+        return 0.;
+
+    rayT0 = std::max(rayT0, 0.f);
+
+    // Parameters that define the volume scattering characteristics and
+    // sampling rate for raymarching
+    float Le = .25f;           // Emission coefficient
+    float sigma_a = 10;        // Absorption coefficient
+    float sigma_s = 10;        // Scattering coefficient
+    float stepDist = 0.025f;   // Ray step amount
+    float lightIntensity = 40; // Light source intensity
+
+    float tau = 0.f;  // accumulated beam transmittance
+    float L = 0;      // radiance along the ray
+    float rayLength = sqrtf(ray.dir.x * ray.dir.x + ray.dir.y * ray.dir.y +
+                            ray.dir.z * ray.dir.z);
+    float stepT = stepDist / rayLength;
+
+    float t = rayT0;
+    float3 pos = ray.origin + ray.dir * rayT0;
+    float3 dirStep = ray.dir * stepT;
+    while (t < rayT1) {
+        float d = Density(pos, pMin, pMax, density, nVoxels);
+
+        // terminate once attenuation is high
+        float atten = expf(-tau);
+        if (atten < .005f)
+            break;
+
+        // direct lighting
+        float Li = lightIntensity / distanceSquared(lightPos, pos) * 
+            transmittance(lightPos, pos, pMin, pMax, sigma_a + sigma_s,
+                          density, nVoxels);
+        L += stepDist * atten * d * sigma_s * (Li + Le);
+
+        // update beam transmittance
+        tau += stepDist * (sigma_a + sigma_s) * d;
+
+        pos = pos + dirStep;
+        t += stepT;
+    }
+
+    // Gamma correction
+    return powf(L, 1.f / 2.2f);
+}
+
+
+void
+volume_serial(float density[], int nVoxels[3], const float raster2camera[4][4],
+              const float camera2world[4][4], 
+              int width, int height, float image[]) {
+    int offset = 0;
+    for (int y = 0; y < height; ++y) {
+        for (int x = 0; x < width; ++x, ++offset) {
+            Ray ray;
+            generateRay(raster2camera, camera2world, x, y, ray);
+            image[offset] = raymarch(density, nVoxels, ray);
+        }
+    }
+}

expr.cpp

@@ -741,6 +741,12 @@ UnaryExpr::TypeCheck() {
 }
 
 
+int
+UnaryExpr::EstimateCost() const {
+    return (expr ? expr->EstimateCost() : 0) + COST_SIMPLE_ARITH_LOGIC_OP;
+}
+
+
 void
 UnaryExpr::Print() const {
     if (!expr || !GetType())
@@ -799,11 +805,17 @@ lOpString(BinaryExpr::Op op) {
 */
 static llvm::Value *
 lEmitBinaryBitOp(BinaryExpr::Op op, llvm::Value *arg0Val,
-                 llvm::Value *arg1Val, FunctionEmitContext *ctx) {
+                 llvm::Value *arg1Val, bool isUnsigned,
+                 FunctionEmitContext *ctx) {
     llvm::Instruction::BinaryOps inst;
     switch (op) {
     case BinaryExpr::Shl:    inst = llvm::Instruction::Shl;  break;
-    case BinaryExpr::Shr:    inst = llvm::Instruction::AShr; break; 
+    case BinaryExpr::Shr:
+        if (isUnsigned)
+            inst = llvm::Instruction::LShr; 
+        else
+            inst = llvm::Instruction::AShr; 
+        break; 
     case BinaryExpr::BitAnd: inst = llvm::Instruction::And;  break;
     case BinaryExpr::BitXor: inst = llvm::Instruction::Xor;  break;
     case BinaryExpr::BitOr:  inst = llvm::Instruction::Or;   break;
@@ -949,7 +961,8 @@ BinaryExpr::GetValue(FunctionEmitContext *ctx) const {
             dynamic_cast<ConstExpr *>(arg1) == NULL)
             PerformanceWarning(pos, "Shift right is extremely inefficient for "
                                "varying shift amounts.");
-        return lEmitBinaryBitOp(op, e0Val, e1Val, ctx);
+        return lEmitBinaryBitOp(op, e0Val, e1Val, 
+                                arg0->GetType()->IsUnsignedType(), ctx);
     }
     case LogicalAnd:
         return ctx->BinaryOperator(llvm::Instruction::And, e0Val, e1Val,
@@ -1176,10 +1189,10 @@ BinaryExpr::Optimize() {
                     m->symbolTable->LookupFunction("rcp");
                 if (rcpFuns != NULL) {
                     assert(rcpFuns->size() == 2);
-                    Expr *rcpSymExpr = new FunctionSymbolExpr(rcpFuns, pos);
+                    Expr *rcpSymExpr = new FunctionSymbolExpr("rcp", rcpFuns, pos);
                     ExprList *args = new ExprList(arg1, arg1->pos);
                     Expr *rcpCall = new FunctionCallExpr(rcpSymExpr, args, 
-                                                         arg1->pos, false);
+                                                         arg1->pos);
                     rcpCall = rcpCall->TypeCheck();
                     if (rcpCall == NULL)
                         return NULL;
@@ -1292,6 +1305,17 @@ BinaryExpr::TypeCheck() {
     if (type0 == NULL || type1 == NULL)
         return NULL;
 
+    if (dynamic_cast<const ReferenceType *>(type0) != NULL) {
+        arg0 = new DereferenceExpr(arg0, arg0->pos);
+        type0 = arg0->GetType();
+        assert(type0 != NULL);
+    }
+    if (dynamic_cast<const ReferenceType *>(type1) != NULL) {
+        arg1 = new DereferenceExpr(arg1, arg1->pos);
+        type1 = arg1->GetType();
+        assert(type1 != NULL);
+    }
+
     switch (op) {
     case Shl:
     case Shr:
@@ -1438,6 +1462,15 @@ BinaryExpr::TypeCheck() {
 }
 
 
+int
+BinaryExpr::EstimateCost() const {
+    return ((arg0 ? arg0->EstimateCost() : 0) +
+            (arg1 ? arg1->EstimateCost() : 0) +
+            ((op == Div || op == Mod) ? COST_COMPLEX_ARITH_OP : 
+                                        COST_SIMPLE_ARITH_LOGIC_OP));
+}
+
+
 void
 BinaryExpr::Print() const {
     if (!arg0 || !arg1 || !GetType())
@@ -1533,7 +1566,8 @@ lEmitOpAssign(AssignExpr::Op op, Expr *arg0, Expr *arg1, const Type *type,
     case AssignExpr::AndAssign:
     case AssignExpr::XorAssign:
     case AssignExpr::OrAssign:
-        newValue = lEmitBinaryBitOp(basicop, oldLHS, rvalue, ctx);
+        newValue = lEmitBinaryBitOp(basicop, oldLHS, rvalue, 
+                                    arg0->GetType()->IsUnsignedType(), ctx);
         break;
     default:
         FATAL("logic error in lEmitOpAssign");
@@ -1688,6 +1722,20 @@ AssignExpr::TypeCheck() {
 }
 
 
+int
+AssignExpr::EstimateCost() const {
+    int cost = ((lvalue ? lvalue->EstimateCost() : 0) +
+                (rvalue ? rvalue->EstimateCost() : 0));
+    cost += COST_ASSIGN;
+    if (op == Assign)
+        return cost;
+    if (op == DivAssign || op == ModAssign)
+        return cost + COST_COMPLEX_ARITH_OP;
+    else
+        return cost + COST_SIMPLE_ARITH_LOGIC_OP;
+}
+
+
 void
 AssignExpr::Print() const {
     if (!lvalue || !rvalue || !GetType())
@@ -1936,6 +1984,12 @@ SelectExpr::TypeCheck() {
 }
 
 
+int
+SelectExpr::EstimateCost() const {
+    return COST_SELECT;
+}
+
+
 void
 SelectExpr::Print() const {
     if (!test || !expr1 || !expr2 || !GetType())
@@ -2100,7 +2154,8 @@ FunctionCallExpr::tryResolve(bool (*matchFunc)(Expr *, const Type *)) {
             // It's kind of a silly to redundantly discover this for each
             // potential match versus detecting this earlier in the
             // matching process and just giving up.
-            if (!callArgs[i] || !callArgs[i]->GetType() || !candArgTypes[i])
+            if (!callArgs[i] || !callArgs[i]->GetType() || !candArgTypes[i] ||
+                dynamic_cast<const FunctionType *>(callArgs[i]->GetType()) != NULL)
                 return false;
             
             // See if this caller argument matches the type of the
@@ -2158,7 +2213,7 @@ FunctionCallExpr::tryResolve(bool (*matchFunc)(Expr *, const Type *)) {
 
 
 void
-FunctionCallExpr::resolveFunctionOverloads() {
+FunctionCallExpr::resolveFunctionOverloads(bool exactMatchOnly) {
     FunctionSymbolExpr *fse = dynamic_cast<FunctionSymbolExpr *>(func);
     if (!fse) 
         // error will be issued later if not calling an actual function
@@ -2172,93 +2227,55 @@ FunctionCallExpr::resolveFunctionOverloads() {
     if (tryResolve(lExactMatch))
         return;
 
-    // Try to find a single match ignoring references
-    if (tryResolve(lMatchIgnoringReferences))
-        return;
+    if (!exactMatchOnly) {
+        // Try to find a single match ignoring references
+        if (tryResolve(lMatchIgnoringReferences))
+            return;
 
-    // TODO: next, try to find an exact match via type promotion--i.e. char
-    // -> int, etc--things that don't lose data
+        // TODO: next, try to find an exact match via type promotion--i.e. char
+        // -> int, etc--things that don't lose data
 
-    // Next try to see if there's a match via just uniform -> varying
-    // promotions.  TODO: look for one with a minimal number of them?
-    if (tryResolve(lMatchIgnoringUniform))
-        return;
+        // Next try to see if there's a match via just uniform -> varying
+        // promotions.  TODO: look for one with a minimal number of them?
+        if (tryResolve(lMatchIgnoringUniform))
+            return;
 
-    // Try to find a match via type conversion, but don't change
-    // unif->varying
-    if (tryResolve(lMatchWithTypeConvSameVariability))
-        return;
+        // Try to find a match via type conversion, but don't change
+        // unif->varying
+        if (tryResolve(lMatchWithTypeConvSameVariability))
+            return;
     
-    // Last chance: try to find a match via arbitrary type conversion.
-    if (tryResolve(lMatchWithTypeConv))
-        return;
+        // Last chance: try to find a match via arbitrary type conversion.
+        if (tryResolve(lMatchWithTypeConv))
+            return;
+    }
 
     // failure :-(
     const char *funName = fse->candidateFunctions->front()->name.c_str();
-    Error(pos, "Unable to find matching overload for call to function \"%s\".",
-          funName);
+    Error(pos, "Unable to find matching overload for call to function \"%s\"%s.",
+          funName, exactMatchOnly ? " only considering exact matches" : "");
     fprintf(stderr, "Candidates are:\n");
     lPrintFunctionOverloads(*fse->candidateFunctions);
     lPrintPassedTypes(funName, args->exprs);
 }
 
 
-FunctionCallExpr::FunctionCallExpr(Expr *f, ExprList *a, SourcePos p, bool il) 
-    : Expr(p) {
+FunctionCallExpr::FunctionCallExpr(Expr *f, ExprList *a, SourcePos p, 
+                                   bool il, Expr *lce) 
+    : Expr(p), isLaunch(il) {
     func = f;
     args = a;
-    isLaunch = il;
+    launchCountExpr = lce;
 
-    resolveFunctionOverloads();
-}
-
-
-/** Starting from the function initialFunction, we're calling into
-    calledFunc.  The question is: is this a recursive call back to
-    initialFunc?  If it definitely is or if it may be, then return true.
-    Return false if it definitely is not.
- */
-static bool
-lMayBeRecursiveCall(llvm::Function *calledFunc, 
-                    llvm::Function *initialFunc,
-                    std::set<llvm::Function *> &seenFuncs) {
-    // Easy case: intrinsics aren't going to call functions themselves
-    if (calledFunc->isIntrinsic())
-        return false;
-
-    std::string name = calledFunc->getName();
-    if (name.size() > 2 && name[0] == '_' && name[1] == '_')
-        // builtin stdlib function; none of these are recursive...
-        return false;
-
-    if (calledFunc->isDeclaration())
-        // There's visibility into what the called function does without a
-        // definition, so we have to be conservative
-        return true;
-
-    if (calledFunc == initialFunc)
-        // hello recursive call
-        return true;
-
-    // Otherwise iterate over all of the instructions in the function.  If
-    // any of them is a function call then check recursively..
-    llvm::inst_iterator iter;
-    for (iter = llvm::inst_begin(calledFunc); 
-         iter != llvm::inst_end(calledFunc); ++iter) {
-        llvm::Instruction *inst = &*iter;
-        llvm::CallInst *ci = llvm::dyn_cast<llvm::CallInst>(inst);
-        if (ci != NULL) {
-            llvm::Function *nextCalledFunc = ci->getCalledFunction();
-            // Don't repeatedly test functions we've seen before 
-            if (seenFuncs.find(nextCalledFunc) == seenFuncs.end()) {
-                seenFuncs.insert(nextCalledFunc);
-                if (lMayBeRecursiveCall(nextCalledFunc, initialFunc, 
-                                        seenFuncs))
-                    return true;
-            }
-        }
-    }
-    return false;
+    FunctionSymbolExpr *fse = dynamic_cast<FunctionSymbolExpr *>(func);
+    // Functions with names that start with "__" should only be various
+    // builtins.  For those, we'll demand an exact match, since we'll
+    // expect whichever function in stdlib.ispc is calling out to one of
+    // those to be matching the argument types exactly; this is to be a bit
+    // extra safe to be sure that the expected builtin is in fact being
+    // called.
+    bool exactMatchOnly = (fse != NULL) && (fse->name.substr(0,2) == "__");
+    resolveFunctionOverloads(exactMatchOnly);
 }
 
 
@@ -2382,47 +2399,18 @@ FunctionCallExpr::GetValue(FunctionEmitContext *ctx) const {
         }
     }
 
-    // We sometimes need to check to see if the mask is all off here;
-    // specifically, if the mask is all off and we call a recursive
-    // function, then we will probably have an unsesirable infinite loop.
-    ctx->SetDebugPos(pos);
-    llvm::BasicBlock *bDoCall = ctx->CreateBasicBlock("funcall_mask_ok");
-    llvm::BasicBlock *bSkip = ctx->CreateBasicBlock("funcall_mask_off");
-    llvm::BasicBlock *bAfter = ctx->CreateBasicBlock("after_funcall");
-    llvm::Function *currentFunc = ctx->GetCurrentBasicBlock()->getParent();
-
-    // If we need to check the mask (it may be a recursive call, possibly
-    // transitively), or we're launching a task, which is expensive and
-    // thus probably always worth checking, then use the mask to choose
-    // whether to go to the bDoCallBlock or the bSkip block
-    std::set<llvm::Function *> seenFuncs;
-    seenFuncs.insert(currentFunc);
-    if (ft->isTask || lMayBeRecursiveCall(callee, currentFunc, seenFuncs)) {
-        Debug(pos, "Checking mask before function call \"%s\".", funSym->name.c_str());
-        ctx->BranchIfMaskAny(bDoCall, bSkip);
-    }
-    else
-        // If we don't need to check the mask, then always to the call;
-        // just jump to bDoCall
-        ctx->BranchInst(bDoCall);
-    
-    // And the bSkip block just jumps immediately to bAfter.  So why do we
-    // need it?  So the phi node below can easily tell what paths are
-    // going into it
-    ctx->SetCurrentBasicBlock(bSkip);
-    ctx->BranchInst(bAfter);
-
-    // Emit the code to do the function call
-    ctx->SetCurrentBasicBlock(bDoCall);
-
     llvm::Value *retVal = NULL;
     ctx->SetDebugPos(pos);
-    if (ft->isTask)
-        ctx->LaunchInst(callee, argVals);
+    if (ft->isTask) {
+        assert(launchCountExpr != NULL);
+        llvm::Value *launchCount = launchCountExpr->GetValue(ctx);
+        if (launchCount != NULL)
+            ctx->LaunchInst(callee, argVals, launchCount);
+    }
     else {
         // Most of the time, the mask is passed as the last argument.  this
-        // isn't the case for things like SSE intrinsics and extern "C"
-        // functions from the application.
+        // isn't the case for things like intrinsics, builtins, and extern
+        // "C" functions from the application.
         assert(callargs.size() + 1 == callee->arg_size() ||
                callargs.size() == callee->arg_size());
 
@@ -2449,22 +2437,10 @@ FunctionCallExpr::GetValue(FunctionEmitContext *ctx) const {
         }
     }
 
-    // And jump out to the 'after funciton call' basic block
-    ctx->BranchInst(bAfter);
-    ctx->SetCurrentBasicBlock(bAfter);
-
     if (isVoidFunc)
         return NULL;
-
-    // The return value for the non-void case is either undefined or the
-    // function return value, depending on whether we actually ran the code
-    // path that called the function or not.
-    LLVM_TYPE_CONST llvm::Type *lrType = ft->GetReturnType()->LLVMType(g->ctx);
-    llvm::PHINode *ret = ctx->PhiNode(lrType, 2, "fun_ret");
-    assert(retVal != NULL);
-    ret->addIncoming(llvm::UndefValue::get(lrType), bSkip);
-    ret->addIncoming(retVal, bDoCall);
-    return ret;
+    else
+        return retVal;
 }
 
 
@@ -2506,10 +2482,21 @@ FunctionCallExpr::TypeCheck() {
                     if (!isLaunch)
                         Error(pos, "\"launch\" expression needed to call function "
                               "with \"task\" qualifier.");
+                    if (!launchCountExpr)
+                        return NULL;
+
+                    launchCountExpr = 
+                        launchCountExpr->TypeConv(AtomicType::UniformInt32,
+                                                  "task launch count");
+                    if (!launchCountExpr)
+                        return NULL;
+                }
+                else {
+                    if (isLaunch)
+                        Error(pos, "\"launch\" expression illegal with non-\"task\"-"
+                              "qualified function.");
+                    assert(launchCountExpr == NULL);
                 }
-                else if (isLaunch)
-                    Error(pos, "\"launch\" expression illegal with non-\"task\"-"
-                          "qualified function.");
             }
             else
                 Error(pos, "Valid function name must be used for function call.");
@@ -2525,6 +2512,13 @@ FunctionCallExpr::TypeCheck() {
 }
 
 
+int
+FunctionCallExpr::EstimateCost() const {
+    return ((args ? args->EstimateCost() : 0) +
+            (isLaunch ? COST_TASK_LAUNCH : COST_FUNCALL));
+}
+
+
 void
 FunctionCallExpr::Print() const {
     if (!func || !args || !GetType())
@@ -2613,7 +2607,7 @@ ExprList::GetConstant(const Type *type) const {
     }
 
     if (dynamic_cast<const StructType *>(type) != NULL) {
-#if defined(LLVM_2_8) || defined(LLVM_2_9)
+#if defined(LLVM_2_9)
         return llvm::ConstantStruct::get(*g->ctx, cv, false);
 #else
         LLVM_TYPE_CONST llvm::StructType *llvmStructType =
@@ -2636,6 +2630,17 @@ ExprList::GetConstant(const Type *type) const {
 }
 
 
+int
+ExprList::EstimateCost() const {
+    int cost = 0;
+    for (unsigned int i = 0; i < exprs.size(); ++i) {
+        if (exprs[i] != NULL)
+            cost += exprs[i]->EstimateCost();
+    }
+    return cost;
+}
+
+
 void
 ExprList::Print() const {
     printf("expr list (");
@@ -2766,6 +2771,22 @@ IndexExpr::GetLValue(FunctionEmitContext *ctx) const {
     if (!basePtr)
         return NULL;
 
+    // If the array index is a compile time constant, check to see if it
+    // may lead to an out-of-bounds access.
+    ConstExpr *ce = dynamic_cast<ConstExpr *>(index);
+    const SequentialType *seqType = dynamic_cast<const SequentialType *>(type);
+    assert(seqType != NULL);
+    int nElements = seqType->GetElementCount();
+    if (ce != NULL && nElements > 0) {
+        int32_t indices[ISPC_MAX_NVEC];
+        int count = ce->AsInt32(indices);
+        for (int i = 0; i < count; ++i) {
+            if (indices[i] < 0 || indices[i] >= nElements)
+                Warning(index->pos, "Array index \"%d\" may be out of bounds for "
+                        "\"%d\" element array.", indices[i], nElements);
+        }
+    }
+
     basePtr = lCastUniformVectorBasePtr(basePtr, ctx);
 
     ctx->SetDebugPos(pos);
@@ -2818,6 +2839,16 @@ IndexExpr::TypeCheck() {
 }
 
 
+int
+IndexExpr::EstimateCost() const {
+    // be pessimistic
+    if (index && index->GetType()->IsVaryingType())
+        return COST_GATHER;
+    else
+        return COST_LOAD;
+}
+
+
 void
 IndexExpr::Print() const {
     if (!arrayOrVector || !index || !GetType())
@@ -3117,6 +3148,7 @@ MemberExpr::create(Expr *e, const char *id, SourcePos p, SourcePos idpos) {
     return new MemberExpr(e, id, p, idpos);
 }
 
+
 MemberExpr::MemberExpr(Expr *e, const char *id, SourcePos p, SourcePos idpos) 
     : Expr(p), identifierPos(idpos) {
     expr = e;
@@ -3213,6 +3245,14 @@ MemberExpr::Optimize() {
 }
 
 
+int
+MemberExpr::EstimateCost() const {
+    // FIXME: return gather cost when we can tell a gather is going to be
+    // needed
+    return COST_SIMPLE_ARITH_LOGIC_OP;
+}
+
+
 void
 MemberExpr::Print() const {
     if (!expr || !GetType())
@@ -3280,7 +3320,7 @@ ConstExpr::ConstExpr(const Type *t, uint8_t u, SourcePos p)
   : Expr(p) {
     type = t;
     type = type->GetAsConstType();
-    assert(type == AtomicType::UniformUInt8);
+    assert(type == AtomicType::UniformConstUInt8);
     uint8Val[0] = u;
 }
 
@@ -3320,7 +3360,7 @@ ConstExpr::ConstExpr(const Type *t, uint16_t u, SourcePos p)
   : Expr(p) {
     type = t;
     type = type->GetAsConstType();
-    assert(type == AtomicType::UniformUInt16);
+    assert(type == AtomicType::UniformConstUInt16);
     uint16Val[0] = u;
 }
 
@@ -3423,7 +3463,7 @@ ConstExpr::ConstExpr(const Type *t, uint64_t u, SourcePos p)
   : Expr(p) {
     type = t;
     type = type->GetAsConstType();
-    assert(type == AtomicType::UniformUInt64);
+    assert(type == AtomicType::UniformConstUInt64);
     uint64Val[0] = u;
 }
 
@@ -4008,6 +4048,12 @@ ConstExpr::TypeCheck() {
 }
 
 
+int
+ConstExpr::EstimateCost() const {
+    return 0;
+}
+
+
 void
 ConstExpr::Print() const {
     printf("[%s] (", GetType()->GetString().c_str());
@@ -4094,7 +4140,7 @@ lTypeConvAtomic(FunctionEmitContext *ctx, llvm::Value *exprVal,
         case AtomicType::TYPE_BOOL:
             if (fromType->IsVaryingType() && 
                 LLVMTypes::BoolVectorType == LLVMTypes::Int32VectorType)
-                // If we have a bool vector of i32 element,s first truncate
+                // If we have a bool vector of i32 elements, first truncate
                 // down to a single bit
                 exprVal = ctx->TruncInst(exprVal, LLVMTypes::Int1VectorType, "bool_to_i1");
             // And then do an unisgned int->float cast
@@ -4154,9 +4200,6 @@ lTypeConvAtomic(FunctionEmitContext *ctx, llvm::Value *exprVal,
         case AtomicType::TYPE_UINT16:
         case AtomicType::TYPE_UINT32:
         case AtomicType::TYPE_UINT64:
-            if (fromType->IsVaryingType())
-                PerformanceWarning(pos, "Conversion from unsigned int64 to float is slow. "
-                                   "Use \"int64\" if possible");
             cast = ctx->CastInst(llvm::Instruction::UIToFP, // unsigned int
                                  exprVal, targetType, "uint2double");
             break;
@@ -4928,6 +4971,13 @@ TypeCastExpr::Optimize() {
 }
 
 
+int
+TypeCastExpr::EstimateCost() const {
+    // FIXME: return COST_TYPECAST_COMPLEX when appropriate
+    return COST_TYPECAST_SIMPLE;
+}
+
+
 void
 TypeCastExpr::Print() const {
     printf("[%s] type cast (", GetType()->GetString().c_str());
@@ -4993,6 +5043,12 @@ ReferenceExpr::TypeCheck() {
 }
 
 
+int
+ReferenceExpr::EstimateCost() const {
+    return 0;
+}
+
+
 void
 ReferenceExpr::Print() const {
     if (expr == NULL || GetType() == NULL)
@@ -5071,6 +5127,12 @@ DereferenceExpr::Optimize() {
 }
 
 
+int
+DereferenceExpr::EstimateCost() const {
+    return COST_DEREF;
+}
+
+
 void
 DereferenceExpr::Print() const {
     if (expr == NULL || GetType() == NULL)
@@ -5142,6 +5204,15 @@ SymbolExpr::Optimize() {
 }
 
 
+int
+SymbolExpr::EstimateCost() const {
+    if (symbol->constValue != NULL)
+        return 0;
+    else
+        return COST_LOAD;
+}
+
+
 void
 SymbolExpr::Print() const {
     if (symbol == NULL || GetType() == NULL)
@@ -5156,9 +5227,11 @@ SymbolExpr::Print() const {
 ///////////////////////////////////////////////////////////////////////////
 // FunctionSymbolExpr
 
-FunctionSymbolExpr::FunctionSymbolExpr(std::vector<Symbol *> *candidates,
+FunctionSymbolExpr::FunctionSymbolExpr(const char *n,
+                                       std::vector<Symbol *> *candidates,
                                        SourcePos p) 
   : Expr(p) {
+    name = n;
     matchingFunc = NULL;
     candidateFunctions = candidates;
 }
@@ -5195,6 +5268,12 @@ FunctionSymbolExpr::Optimize() {
 }
 
 
+int
+FunctionSymbolExpr::EstimateCost() const {
+    return 0;
+}
+
+
 void
 FunctionSymbolExpr::Print() const {
     if (!matchingFunc || !GetType())
@@ -5218,14 +5297,14 @@ SyncExpr::GetType() const {
 llvm::Value *
 SyncExpr::GetValue(FunctionEmitContext *ctx) const {
     ctx->SetDebugPos(pos);
-    std::vector<llvm::Value *> noArg;
-    llvm::Function *fsync = m->module->getFunction("ISPCSync");
-    if (fsync == NULL) {
-        FATAL("Couldn't find ISPCSync declaration?!");
-        return NULL;
-    }
+    ctx->SyncInst();
+    return NULL;
+}
 
-    return ctx->CallInst(fsync, noArg, "");
+
+int
+SyncExpr::EstimateCost() const {
+    return COST_SYNC;
 }
 
 

expr.h

@@ -121,8 +121,8 @@ public:
     void Print() const;
     Expr *Optimize();
     Expr *TypeCheck();
+    int EstimateCost() const;
 
-private:
     const Op op;
     Expr *expr;
 };
@@ -164,8 +164,8 @@ public:
 
     Expr *Optimize();
     Expr *TypeCheck();
+    int EstimateCost() const;
 
-private:
     const Op op;
     Expr *arg0, *arg1;
 };
@@ -196,8 +196,8 @@ public:
 
     Expr *Optimize();
     Expr *TypeCheck();
+    int EstimateCost() const;
 
-private:
     const Op op;
     Expr *lvalue, *rvalue;
 };
@@ -217,8 +217,8 @@ public:
 
     Expr *Optimize();
     Expr *TypeCheck();
+    int EstimateCost() const;
 
-private:
     Expr *test, *expr1, *expr2;
 };
 
@@ -240,6 +240,7 @@ public:
     llvm::Constant *GetConstant(const Type *type) const;
     ExprList *Optimize();
     ExprList *TypeCheck();
+    int EstimateCost() const;
 
     std::vector<Expr *> exprs;
 };
@@ -249,7 +250,8 @@ public:
  */
 class FunctionCallExpr : public Expr {
 public:
-    FunctionCallExpr(Expr *func, ExprList *args, SourcePos p, bool isLaunch);
+    FunctionCallExpr(Expr *func, ExprList *args, SourcePos p, 
+                     bool isLaunch = false, Expr *launchCountExpr = NULL);
 
     llvm::Value *GetValue(FunctionEmitContext *ctx) const;
     const Type *GetType() const;
@@ -257,13 +259,15 @@ public:
 
     Expr *Optimize();
     Expr *TypeCheck();
+    int EstimateCost() const;
 
-private:
     Expr *func;
     ExprList *args;
     bool isLaunch;
+    Expr *launchCountExpr;
 
-    void resolveFunctionOverloads();
+private:
+    void resolveFunctionOverloads(bool exactMatchOnly);
     bool tryResolve(bool (*matchFunc)(Expr *, const Type *));
 };
 
@@ -285,8 +289,8 @@ public:
 
     Expr *Optimize();
     Expr *TypeCheck();
+    int EstimateCost() const;
 
-private:
     Expr *arrayOrVector, *index;
 };
 
@@ -303,16 +307,17 @@ public:
     MemberExpr(Expr *expr, const char *identifier, SourcePos pos, 
                SourcePos identifierPos);
 
-    virtual llvm::Value *GetValue(FunctionEmitContext *ctx) const;
-    virtual llvm::Value *GetLValue(FunctionEmitContext *ctx) const;
-    virtual const Type *GetType() const;
-    virtual Symbol *GetBaseSymbol() const;
-    virtual void Print() const;
-    virtual Expr *Optimize();
-    virtual Expr *TypeCheck();
+    llvm::Value *GetValue(FunctionEmitContext *ctx) const;
+    llvm::Value *GetLValue(FunctionEmitContext *ctx) const;
+    const Type *GetType() const;
+    Symbol *GetBaseSymbol() const;
+    void Print() const;
+    Expr *Optimize();
+    Expr *TypeCheck();
+    int EstimateCost() const;
+
     virtual int getElementNumber() const;
 
-protected:
     std::string getCandidateNearMatches() const;
 
     Expr *expr;
@@ -392,6 +397,7 @@ public:
 
     Expr *TypeCheck();
     Expr *Optimize();
+    int EstimateCost() const;
 
     /** Return the ConstExpr's values as booleans, doing type conversion
         from the actual type if needed.  If forceVarying is true, then type
@@ -495,8 +501,8 @@ public:
     void Print() const;
     Expr *TypeCheck();
     Expr *Optimize();
+    int EstimateCost() const;
 
-private:
     const Type *type;
     Expr *expr;
 };
@@ -514,8 +520,8 @@ public:
     void Print() const;
     Expr *TypeCheck();
     Expr *Optimize();
+    int EstimateCost() const;
 
-private:
     Expr *expr;
 };
 
@@ -533,8 +539,8 @@ public:
     void Print() const;
     Expr *TypeCheck();
     Expr *Optimize();
+    int EstimateCost() const;
 
-private:
     Expr *expr;
 };
 
@@ -551,6 +557,7 @@ public:
     Expr *TypeCheck();
     Expr *Optimize();
     void Print() const;
+    int EstimateCost() const;
 
 private:
     Symbol *symbol;
@@ -562,7 +569,7 @@ private:
  */    
 class FunctionSymbolExpr : public Expr {
 public:
-    FunctionSymbolExpr(std::vector<Symbol *> *candidateFunctions, 
+    FunctionSymbolExpr(const char *name, std::vector<Symbol *> *candidateFunctions,
                        SourcePos pos);
 
     llvm::Value *GetValue(FunctionEmitContext *ctx) const;
@@ -571,10 +578,14 @@ public:
     Expr *TypeCheck();
     Expr *Optimize();
     void Print() const;
+    int EstimateCost() const;
 
 private:
     friend class FunctionCallExpr;
 
+    /** Name of the function that is being called. */
+    std::string name;
+
     /** All of the functions with the name given in the function call;
         there may be more then one, in which case we need to resolve which
         overload is the best match. */
@@ -597,6 +608,7 @@ public:
     Expr *TypeCheck();
     Expr *Optimize();
     void Print() const;
+    int EstimateCost() const;
 };
 
 #endif // ISPC_EXPR_H

failing_tests/masked-scatter-vector.ispc

@@ -14,7 +14,7 @@ export void f_fu(uniform float ret[], uniform float aa[], uniform float b) {
     varying int3 vv = array[a];
     ++vv.y;
     array[a] = vv;
-    print("fin %\n", array[programIndex].y);
+//CO    print("fin %\n", array[programIndex].y);
     ret[programIndex] = array[programIndex].y;
 }
 

failing_tests/max-uint-1.ispc

@@ -1,19 +1,14 @@
-static float float4(uniform float a, uniform float b, uniform float c, 
-                    uniform float d) {
-    float ret = 0;
-    for (uniform int i = 0; i < programCount; i += 4) {
-        ret = insert(ret, i + 0, a);
-        ret = insert(ret, i + 1, b);
-        ret = insert(ret, i + 2, c);
-        ret = insert(ret, i + 3, d);
-    }
-    return ret;
+
+export uniform int width() { return programCount; }
+
+export void f_f(uniform float r[], uniform float a[]) {
+    unsigned int i = (unsigned int)a[programIndex];
+    r[programIndex] = max((unsigned int)2, i);
 }
 
-export float f_f(float a) {
-    unsigned int i = (unsigned int)a;
-    return max((unsigned int)2, i);
+export void result(uniform float r[]) { 
+    r[programIndex] = 1+programIndex;
+    r[0] = 2;
 }
 
-export float result() { return float4(2,2,3,4); }
 

failing_tests/max-uint.ispc

@@ -1,8 +1,10 @@
 
-export float f_f(float a) {
-    unsigned int i = (unsigned int)a;
-    return max((unsigned int)10, i);
+export uniform int width() { return programCount; }
+
+export void f_f(uniform float result[], uniform float aa[]) {
+    unsigned int i = (unsigned int)aa[programIndex];
+    result[programIndex] = max((unsigned int)100, i);
 }
 
-export float result() { return 10; }
+export void result(uniform float r[]) { r[programIndex] = 100; }
 

failing_tests/min-uint-1.ispc

@@ -1,19 +1,14 @@
-static float float4(uniform float a, uniform float b, uniform float c, 
-                    uniform float d) {
-    float ret = 0;
-    for (uniform int i = 0; i < programCount; i += 4) {
-        ret = insert(ret, i + 0, a);
-        ret = insert(ret, i + 1, b);
-        ret = insert(ret, i + 2, c);
-        ret = insert(ret, i + 3, d);
-    }
-    return ret;
+
+export uniform int width() { return programCount; }
+
+export void f_f(uniform float result[], uniform float aa[]) {
+    unsigned int i = (unsigned int)aa[programIndex];
+    result[programIndex] = min((unsigned int)2, i);
 }
 
-export float f_f(float a) {
-    unsigned int i = (unsigned int)a;
-    return min((unsigned int)2, i);
+export void result(uniform float r[]) { 
+    r[programIndex] = 2;
+    r[0] = 1;
 }
 
-export float result() { return float4(1,2,2,2); }
 

failing_tests/min-uint-2.ispc

@@ -1,19 +1,13 @@
-static float float4(uniform float a, uniform float b, uniform float c, 
-                    uniform float d) {
-    float ret = 0;
-    for (uniform int i = 0; i < programCount; i += 4) {
-        ret = insert(ret, i + 0, a);
-        ret = insert(ret, i + 1, b);
-        ret = insert(ret, i + 2, c);
-        ret = insert(ret, i + 3, d);
-    }
-    return ret;
+
+export uniform int width() { return programCount; }
+
+export void f_f(uniform float r[], uniform float a[]) {
+    unsigned int i = (unsigned int)a[programIndex];
+    r[programIndex] =  min((unsigned int)20, i);
 }
 
-export float f_f(float a) {
-    unsigned int i = (unsigned int)a;
-    return min((unsigned int)20, i);
+export void result(uniform float r[]) { 
+    r[programIndex] = 1+programIndex;
 }
 
-export float result() { return float4(1,2,3,4); }
 

failing_tests/struct-array-assign.ispc

@@ -1,11 +0,0 @@
-
-struct Foo {
-    float f;
-};
-
-
-export float foo(Foo f[], int i, uniform int j) {
-    Foo x = f[i];
-    return x.f;
-}
-

ispc.cpp

@@ -42,14 +42,25 @@
 #ifdef ISPC_IS_WINDOWS
 #include <windows.h>
 #include <direct.h>
+#define strcasecmp stricmp
 #endif
 #include <llvm/LLVMContext.h>
 #include <llvm/Module.h>
-#ifndef LLVM_2_8
 #include <llvm/Analysis/DIBuilder.h>
-#endif
 #include <llvm/Analysis/DebugInfo.h>
 #include <llvm/Support/Dwarf.h>
+#include <llvm/Target/TargetMachine.h>
+#include <llvm/Target/TargetOptions.h>
+#include <llvm/Target/TargetData.h>
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+  #include <llvm/Support/TargetRegistry.h>
+  #include <llvm/Support/TargetSelect.h>
+#else
+  #include <llvm/Target/TargetRegistry.h>
+  #include <llvm/Target/TargetSelect.h>
+  #include <llvm/Target/SubtargetFeature.h>
+#endif
+#include <llvm/Support/Host.h>
 
 Globals *g;
 Module *m;
@@ -57,20 +68,198 @@ Module *m;
 ///////////////////////////////////////////////////////////////////////////
 // Target
 
-Target::Target() {
-    arch = "x86-64";
-    cpu = "nehalem";
-    isa = SSE4;
-    nativeVectorWidth = 4;
-    vectorWidth = 4;
+bool
+Target::GetTarget(const char *arch, const char *cpu, const char *isa,
+                  bool pic, Target *t) {
+    if (cpu == NULL) {
+        std::string hostCPU = llvm::sys::getHostCPUName();
+        if (hostCPU.size() > 0)
+            cpu = hostCPU.c_str();
+        else {
+            fprintf(stderr, "Warning: unable to determine host CPU!\n");
+            cpu = "generic";
+        }
+    }
+    t->cpu = cpu;
+
+    if (isa == NULL) {
+        if (!strcasecmp(cpu, "atom"))
+            isa = "sse2";
+#if defined(LLVM_3_0) || defined(LLVM_3_0_svn)
+        else if (!strcasecmp(cpu, "sandybridge") ||
+                 !strcasecmp(cpu, "corei7-avx"))
+            isa = "avx";
+#endif // LLVM_3_0
+        else
+            isa = "sse4";
+    }
+    if (arch == NULL)
+        arch = "x86-64";
+
+    bool error = false;
+
+    t->generatePIC = pic;
+
+    // Make sure the target architecture is a known one; print an error
+    // with the valid ones otherwise.
+    t->target = NULL;
+    for (llvm::TargetRegistry::iterator iter = llvm::TargetRegistry::begin();
+         iter != llvm::TargetRegistry::end(); ++iter) {
+        if (std::string(arch) == iter->getName()) {
+            t->target = &*iter;
+            break;
+        }
+    }
+    if (t->target == NULL) {
+        fprintf(stderr, "Invalid architecture \"%s\"\nOptions: ", arch);
+        llvm::TargetRegistry::iterator iter;
+        for (iter = llvm::TargetRegistry::begin();
+             iter != llvm::TargetRegistry::end(); ++iter)
+            fprintf(stderr, "%s ", iter->getName());
+        fprintf(stderr, "\n");
+        error = true;
+    }
+    else {
+        t->arch = arch;
+    }
+
+    if (!strcasecmp(isa, "sse2")) {
+        t->isa = Target::SSE2;
+        t->nativeVectorWidth = 4;
+        t->vectorWidth = 4;
+        t->attributes = "+sse,+sse2,-sse3,-sse41,-sse42,-sse4a,-ssse3,-popcnt";
+    }
+    else if (!strcasecmp(isa, "sse4")) {
+        t->isa = Target::SSE4;
+        t->nativeVectorWidth = 4;
+        t->vectorWidth = 4;
+        t->attributes = "+sse,+sse2,+sse3,+sse41,-sse42,-sse4a,+ssse3,-popcnt,+cmov";
+    }
+    else if (!strcasecmp(isa, "sse4x2")) {
+        t->isa = Target::SSE4;
+        t->nativeVectorWidth = 4;
+        t->vectorWidth = 8;
+        t->attributes = "+sse,+sse2,+sse3,+sse41,-sse42,-sse4a,+ssse3,-popcnt,+cmov";
+    }
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+    else if (!strcasecmp(isa, "avx")) {
+        t->isa = Target::AVX;
+        t->nativeVectorWidth = 8;
+        t->vectorWidth = 8;
+        t->attributes = "+avx,+popcnt,+cmov";
+    }
+    else if (!strcasecmp(isa, "avx-x2")) {
+        t->isa = Target::AVX;
+        t->nativeVectorWidth = 8;
+        t->vectorWidth = 16;
+        t->attributes = "+avx,+popcnt,+cmov";
+    }
+#endif // LLVM 3.0
+    else {
+        fprintf(stderr, "Target ISA \"%s\" is unknown.  Choices are: %s\n", 
+                isa, SupportedTargetISAs());
+        error = true;
+    }
+
+    if (!error) {
+        llvm::TargetMachine *targetMachine = t->GetTargetMachine();
+        const llvm::TargetData *targetData = targetMachine->getTargetData();
+        t->is32bit = (targetData->getPointerSize() == 4);
+    }
+
+    return !error;
 }
 
+
+const char *
+Target::SupportedTargetCPUs() {
+    return "atom, barcelona, core2, corei7, "
+#if defined(LLVM_3_0) || defined(LLVM_3_0_svn)
+        "corei7-avx, "
+#endif
+        "istanbul, nocona, penryn, "
+#ifdef LLVM_2_9
+        "sandybridge, "
+#endif
+        "westmere";
+}
+
+
+const char *
+Target::SupportedTargetArchs() {
+    return "x86, x86-64";
+}
+
+
+const char *
+Target::SupportedTargetISAs() {
+    return "sse2, sse4, sse4x2"
+#if defined(LLVM_3_0) || defined(LLVM_3_0_svn)
+        ", avx, avx-x2"
+#endif
+        ;
+}
+
+
+std::string
+Target::GetTripleString() const {
+    llvm::Triple triple;
+    // Start with the host triple as the default
+    triple.setTriple(llvm::sys::getHostTriple());
+
+    // And override the arch in the host triple based on what the user
+    // specified.  Here we need to deal with the fact that LLVM uses one
+    // naming convention for targets TargetRegistry, but wants some
+    // slightly different ones for the triple.  TODO: is there a way to
+    // have it do this remapping, which would presumably be a bit less
+    // error prone?
+    if (arch == "x86")
+        triple.setArchName("i386");
+    else if (arch == "x86-64")
+        triple.setArchName("x86_64");
+    else
+        triple.setArchName(arch);
+
+    return triple.str();
+}
+
+
+llvm::TargetMachine *
+Target::GetTargetMachine() const {
+    std::string triple = GetTripleString();
+
+    llvm::Reloc::Model relocModel = generatePIC ? llvm::Reloc::PIC_ : 
+                                                  llvm::Reloc::Default;
+#if defined(LLVM_3_0svn) || defined(LLVM_3_0)
+    std::string featuresString = attributes;
+    llvm::TargetMachine *targetMachine = 
+        target->createTargetMachine(triple, cpu, featuresString, relocModel);
+#else
+#ifdef ISPC_IS_APPLE
+    relocModel = llvm::Reloc::PIC_;
+#endif // ISPC_IS_APPLE
+    std::string featuresString = cpu + std::string(",") + attributes;
+    llvm::TargetMachine *targetMachine = 
+        target->createTargetMachine(triple, featuresString);
+#ifndef ISPC_IS_WINDOWS
+    targetMachine->setRelocationModel(relocModel);
+#endif // !ISPC_IS_WINDOWS
+#endif
+    assert(targetMachine != NULL);
+
+    targetMachine->setAsmVerbosityDefault(true);
+    return targetMachine;
+}
+
+
 ///////////////////////////////////////////////////////////////////////////
 // Opt
 
 Opt::Opt() {
     level = 1;
     fastMath = false;
+    fastMaskedVload = false;
+    unrollLoops = true;
     disableBlendedMaskedStores = false;
     disableCoherentControlFlow = false;
     disableUniformControlFlow = false;
@@ -120,13 +309,9 @@ SourcePos::SourcePos(const char *n, int l, int c) {
 }
 
 llvm::DIFile SourcePos::GetDIFile() const {
-#ifdef LLVM_2_8
-    return llvm::DIFile();
-#else
     std::string directory, filename;
     GetDirectoryAndFileName(g->currentDirectory, name, &directory, &filename);
     return m->diBuilder->createFile(filename, directory);
-#endif // LLVM_2_8
 }
 
 

ispc.h

@@ -69,6 +69,8 @@ namespace llvm {
     class FunctionType;
     class LLVMContext;
     class Module;
+    class Target;
+    class TargetMachine;
     class Type;
     class Value;
 }
@@ -146,6 +148,8 @@ public:
         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. */
     const SourcePos pos;
@@ -156,7 +160,34 @@ public:
     This structure defines a compilation target for the ispc compiler.
 */
 struct Target {
-    Target();
+    /** Initializes the given Target pointer for a target of the given
+        name, if the name is a known target.  Returns true if the
+        target was initialized and false if the name is unknown. */
+    static bool GetTarget(const char *arch, const char *cpu, const char *isa,
+                          bool pic, Target *);
+
+    /** Returns a comma-delimited string giving the names of the currently
+        supported target ISAs. */
+    static const char *SupportedTargetISAs();
+
+    /** Returns a comma-delimited string giving the names of the currently
+        supported target CPUs. */
+    static const char *SupportedTargetCPUs();
+
+    /** Returns a comma-delimited string giving the names of the currently
+        supported target architectures. */
+    static const char *SupportedTargetArchs();
+
+    /** Returns a triple string specifying the target architecture, vendor,
+        and environment. */
+    std::string GetTripleString() const;
+
+    /** Returns the LLVM TargetMachine object corresponding to this
+        target. */
+    llvm::TargetMachine *GetTargetMachine() const;
+
+    /** llvm Target object representing this target. */
+    const llvm::Target *target;
 
     /** Enumerator giving the instruction sets that the compiler can
         target. */
@@ -168,9 +199,15 @@ struct Target {
     /** Target system architecture.  (e.g. "x86-64", "x86"). */
     std::string arch;
 
+    /** Is the target architecture 32 or 64 bit */
+    bool is32bit;
+
     /** Target CPU. (e.g. "corei7", "corei7-avx", ..) */
     std::string cpu;
 
+    /** Target-specific attributes to pass along to the LLVM backend */
+    std::string attributes;
+
     /** Native vector width of the vector instruction set.  Note that this
         value is directly derived from the ISA Being used (e.g. it's 4 for
         SSE, 8 for AVX, etc.) */
@@ -180,8 +217,12 @@ struct Target {
         integer multiple of the native vector width, for example if we're
         "doubling up" and compiling 8-wide on a 4-wide SSE system. */
     int vectorWidth;
+
+    /** Indicates whether position independent code should be generated. */
+    bool generatePIC;
 };
 
+
 /** @brief Structure that collects optimization options
 
     This structure collects all of the options related to optimization of
@@ -199,6 +240,16 @@ struct Opt {
         should be performed.  This is false by default. */
     bool fastMath;
 
+    /** Indicates whether an vector load should be issued for masked loads
+        on platforms that don't have a native masked vector load.  (This may
+        lead to accessing memory up to programCount-1 elements past the end of
+        arrays, so is unsafe in general.) */
+    bool fastMaskedVload;
+
+    /** Indicates when loops should be unrolled (when doing so seems like
+        it will make sense. */
+    bool unrollLoops;
+
     /** On targets that don't have a masked store instruction but do have a
         blending instruction, by default, we simulate masked stores by
         loading the old value, blending, and storing the result.  This can
@@ -316,6 +367,29 @@ struct Globals {
     std::vector<std::string> cppArgs;
 };
 
+enum {
+    COST_ASSIGN = 1,
+    COST_COHERENT_BREAK_CONTINE = 4,
+    COST_COMPLEX_ARITH_OP = 4,
+    COST_DEREF = 4,
+    COST_FUNCALL = 4,
+    COST_GATHER = 8,
+    COST_LOAD = 2,
+    COST_REGULAR_BREAK_CONTINUE = 2,
+    COST_RETURN = 4,
+    COST_SELECT = 4,
+    COST_SIMPLE_ARITH_LOGIC_OP = 1,
+    COST_SYNC = 32,
+    COST_TASK_LAUNCH = 16,
+    COST_TYPECAST_COMPLEX = 4,
+    COST_TYPECAST_SIMPLE = 1,
+    COST_UNIFORM_LOOP = 4,
+    COST_VARYING_LOOP = 6,
+
+    CHECK_MASK_AT_FUNCTION_START_COST = 16,
+    PREDICATE_SAFE_IF_STATEMENT_COST = 6,
+};
+
 extern Globals *g;
 extern Module *m;
 

ispc.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">
@@ -16,7 +16,9 @@
     <ClCompile Include="decl.cpp" />
     <ClCompile Include="expr.cpp" />
     <ClCompile Include="gen-bitcode-avx.cpp" />
-    <ClCompile Include="gen-bitcode-c.cpp" />
+    <ClCompile Include="gen-bitcode-avx-x2.cpp" />
+    <ClCompile Include="gen-bitcode-c-32.cpp" />
+    <ClCompile Include="gen-bitcode-c-64.cpp" />
     <ClCompile Include="gen-bitcode-sse2.cpp" />
     <ClCompile Include="gen-bitcode-sse4.cpp" />
     <ClCompile Include="gen-bitcode-sse4x2.cpp" />
@@ -29,12 +31,14 @@
     <ClCompile Include="opt.cpp" />
     <ClCompile Include="parse.cc" />
     <CustomBuild Include="builtins-c.c">
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%LLVM_INSTALL_DIR%\bin\clang -emit-llvm builtins-c.c -c -o - | %LLVM_INSTALL_DIR%\bin\llvm-dis - | python bitcode2cpp.py builtins-c.c &gt; gen-bitcode-c.cpp</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%LLVM_INSTALL_DIR%\bin\clang -m32 -emit-llvm builtins-c.c -c -o - | %LLVM_INSTALL_DIR%\bin\llvm-dis - | python bitcode2cpp.py builtins-c-32.c &gt; gen-bitcode-c-32.cpp;
+%LLVM_INSTALL_DIR%\bin\clang -m64 -emit-llvm builtins-c.c -c -o - | %LLVM_INSTALL_DIR%\bin\llvm-dis - | python bitcode2cpp.py builtins-c-64.c &gt; gen-bitcode-c-64.cpp</Command>
       <Message Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">clang builtins-c.c</Message>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%LLVM_INSTALL_DIR%\bin\clang -emit-llvm builtins-c.c -c -o - | %LLVM_INSTALL_DIR%\bin\llvm-dis - | python bitcode2cpp.py builtins-c.c &gt; gen-bitcode-c.cpp</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%LLVM_INSTALL_DIR%\bin\clang -m32 -emit-llvm builtins-c.c -c -o - | %LLVM_INSTALL_DIR%\bin\llvm-dis - | python bitcode2cpp.py builtins-c-32.c &gt; gen-bitcode-c-32.cpp;
+%LLVM_INSTALL_DIR%\bin\clang -m64 -emit-llvm builtins-c.c -c -o - | %LLVM_INSTALL_DIR%\bin\llvm-dis - | python bitcode2cpp.py builtins-c-64.c &gt; gen-bitcode-c-64.cpp</Command>
       <Message Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">clang builtins-c.c</Message>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">gen-bitcode-c.cpp</Outputs>
-      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">gen-bitcode-c.cpp</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">gen-bitcode-c-32.cpp;gen-bitcore-c-64.cpp</Outputs>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">gen-bitcode-c-32.cpp;gen-bitcore-c-64.cpp</Outputs>
     </CustomBuild>
     <ClCompile Include="stmt.cpp" />
     <ClCompile Include="sym.cpp" />
@@ -59,9 +63,9 @@
   <ItemGroup>
     <CustomBuild Include="stdlib.ispc">
       <FileType>Document</FileType>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">clang -E -x c %(Filename).ispc -DISPC=1 -DPI=3.1415926535 | python stdlib2cpp.py &gt; gen-stdlib.cpp</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">%LLVM_INSTALL_DIR%\bin\clang -E -x c %(Filename).ispc -DISPC=1 -DPI=3.1415926535 | python stdlib2cpp.py &gt; gen-stdlib.cpp</Command>
       <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">gen-stdlib.cpp</Outputs>
-      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">clang -E -x c %(Filename).ispc -DISPC=1 -DPI=3.1415926535 | python stdlib2cpp.py &gt; gen-stdlib.cpp</Command>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">%LLVM_INSTALL_DIR%\bin\clang -E -x c %(Filename).ispc -DISPC=1 -DPI=3.1415926535 | python stdlib2cpp.py &gt; gen-stdlib.cpp</Command>
       <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">gen-stdlib.cpp</Outputs>
       <Message Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">Building gen-stdlib.cpp</Message>
       <Message Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">Building gen-stdlib.cpp</Message>
@@ -119,6 +123,19 @@
       <Message Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">Building gen-bitcode-avx.cpp</Message>
     </CustomBuild>
   </ItemGroup>
+  <ItemGroup>
+    <CustomBuild Include="builtins-avx-x2.ll">
+      <FileType>Document</FileType>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">m4 builtins.m4 builtins-avx-x2.ll | python bitcode2cpp.py builtins-avx-x2.ll &gt; gen-bitcode-avx-x2.cpp</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">gen-bitcode-avx-x2.cpp</Outputs>
+      <AdditionalInputs Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">builtins.m4;builtins-sse.ll</AdditionalInputs>
+      <Command Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">m4 builtins.m4 builtins-avx-x2.ll | python bitcode2cpp.py builtins-avx-x2.ll &gt; gen-bitcode-avx-x2.cpp</Command>
+      <Outputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">gen-bitcode-avx-x2.cpp</Outputs>
+      <AdditionalInputs Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">builtins.m4;builtins-sse.ll</AdditionalInputs>
+      <Message Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">Building gen-bitcode-avx-x2.cpp</Message>
+      <Message Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">Building gen-bitcode-avx-x2.cpp</Message>
+    </CustomBuild>
+  </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="lex.ll">
       <FileType>Document</FileType>
@@ -179,7 +196,7 @@
       <PrecompiledHeader>NotUsing</PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <PreprocessorDefinitions>NOMINMAX;LLVM_2_9</PreprocessorDefinitions>
+      <PreprocessorDefinitions>NOMINMAX;LLVM_3_0</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>$(LLVM_INSTALL_DIR)\include;.;.\winstuff;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4146;4800;4996;4355;4624</DisableSpecificWarnings>
     </ClCompile>
@@ -187,7 +204,7 @@
       <SubSystem>Console</SubSystem>
       <GenerateDebugInformation>true</GenerateDebugInformation>
       <AdditionalLibraryDirectories>$(LLVM_INSTALL_DIR)\lib;%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
-      <AdditionalDependencies>clangFrontend.lib;clangDriver.lib;clangSerialization.lib;clangParse.lib;clangSema.lib;clangAnalysis.lib;clangAST.lib;clangLex.lib;clangBasic.lib;LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMipa.lib;LLVMipo.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Disassembler.lib;LLVMX86Info.lib;shlwapi.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>clangFrontend.lib;clangDriver.lib;clangSerialization.lib;clangParse.lib;clangSema.lib;clangAnalysis.lib;clangAST.lib;clangLex.lib;clangBasic.lib;LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmParser.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMDebugInfo.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCDisassembler.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Desc.lib;LLVMX86Disassembler.lib;LLVMX86Info.lib;LLVMipa.lib;LLVMipo.lib;shlwapi.lib;%(AdditionalDependencies)</AdditionalDependencies>
     </Link>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
@@ -197,7 +214,7 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <PreprocessorDefinitions>NOMINMAX;LLVM_2_9</PreprocessorDefinitions>
+      <PreprocessorDefinitions>NOMINMAX;LLVM_3_0</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>$(LLVM_INSTALL_DIR)\include;.;.\winstuff;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4146;4800;4996;4355;4624</DisableSpecificWarnings>
     </ClCompile>
@@ -207,10 +224,10 @@
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
       <OptimizeReferences>true</OptimizeReferences>
       <AdditionalLibraryDirectories>$(LLVM_INSTALL_DIR)\lib;%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
-      <AdditionalDependencies>clangFrontend.lib;clangDriver.lib;clangSerialization.lib;clangParse.lib;clangSema.lib;clangAnalysis.lib;clangAST.lib;clangLex.lib;clangBasic.lib;LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMipa.lib;LLVMipo.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Disassembler.lib;LLVMX86Info.lib;shlwapi.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>clangFrontend.lib;clangDriver.lib;clangSerialization.lib;clangParse.lib;clangSema.lib;clangAnalysis.lib;clangAST.lib;clangLex.lib;clangBasic.lib;LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmParser.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMDebugInfo.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCDisassembler.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Desc.lib;LLVMX86Disassembler.lib;LLVMX86Info.lib;LLVMipa.lib;LLVMipo.lib;shlwapi.lib;%(AdditionalDependencies)</AdditionalDependencies>
     </Link>
   </ItemDefinitionGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

ispc_test.cpp

@@ -33,12 +33,25 @@
 
 #define _CRT_SECURE_NO_WARNINGS
 
+#if defined(_WIN32) || defined(_WIN64)
+#define ISPC_IS_WINDOWS
+#elif defined(__linux__)
+#define ISPC_IS_LINUX
+#elif defined(__APPLE__)
+#define ISPC_IS_APPLE
+#endif
+
 #ifdef ISPC_IS_WINDOWS
 #define NOMINMAX
 #include <windows.h>
 #endif
 #include <stdio.h>
 #include <stdint.h>
+#include <stdlib.h>
+#include <memory.h>
+#ifdef ISPC_IS_LINUX
+#include <malloc.h>
+#endif
 
 #ifdef ISPC_HAVE_SVML
 #include <xmmintrin.h>
@@ -61,8 +74,14 @@ extern "C" {
 #include <llvm/DerivedTypes.h>
 #include <llvm/Instructions.h>
 #include <llvm/ExecutionEngine/ExecutionEngine.h>
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+  #include <llvm/Support/TargetRegistry.h>
+  #include <llvm/Support/TargetSelect.h>
+#else
+  #include <llvm/Target/TargetRegistry.h>
+  #include <llvm/Target/TargetSelect.h>
+#endif
 #include <llvm/ExecutionEngine/JIT.h>
-#include <llvm/Target/TargetSelect.h>
 #include <llvm/Target/TargetOptions.h>
 #include <llvm/Target/TargetData.h>
 #include <llvm/Transforms/Scalar.h>
@@ -74,42 +93,53 @@ extern "C" {
 #include <llvm/Support/raw_ostream.h>
 #include <llvm/Bitcode/ReaderWriter.h>
 #include <llvm/Support/MemoryBuffer.h>
-#ifndef LLVM_2_8
 #include <llvm/Support/system_error.h>
-#endif
+
+bool shouldFail = false;
 
 extern "C" { 
-    void ISPCLaunch(void *, void *);
-    void ISPCSync();
-    void *ISPCMalloc(int64_t size, int32_t alignment);
-    void ISPCFree(void *ptr);
+    void ISPCLaunch(void **, void *, void *, int32_t);
+    void ISPCSync(void *);
+    void *ISPCAlloc(void **, int64_t size, int32_t alignment);
 }
 
-void ISPCLaunch(void *func, void *data) {
-    typedef void (*TaskFuncType)(void *, int, int);
+void ISPCLaunch(void **handle, void *func, void *data, int32_t count) {
+    *handle = (void *)0xdeadbeef;
+    typedef void (*TaskFuncType)(void *, int, int, int, int);
     TaskFuncType tft = (TaskFuncType)(func);
-    tft(data, 0, 1);
+    for (int i = 0; i < count; ++i)
+        tft(data, 0, 1, i, count);
 }
 
 
-void ISPCSync() {
+void ISPCSync(void *) {
 }
 
 
+void *ISPCAlloc(void **handle, int64_t size, int32_t alignment) {
+    *handle = (void *)0xdeadbeef;
+    // leak time!
 #ifdef ISPC_IS_WINDOWS
-void *ISPCMalloc(int64_t size, int32_t alignment) {
     return _aligned_malloc(size, alignment);
-}
-
-
-void ISPCFree(void *ptr) {
-    _aligned_free(ptr);
-}
 #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 usage(int ret) {
     fprintf(stderr, "usage: ispc_test\n");
     fprintf(stderr, "\t[-h/--help]\tprint help\n");
+    fprintf(stderr, "\t[-f]\t\tindicates that test is expected to fail\n");
     fprintf(stderr, "\t<files>\n");
     exit(ret);
 }
@@ -119,20 +149,22 @@ static void svml_missing() {
     exit(1);
 }
 
+// On Windows, sin() is an overloaded function, so we need an unambiguous
+// function we can take the address of when wiring up the external references
+// below.
+
+double Sin(double x) { return sin(x); }
+double Cos(double x) { return cos(x); }
+double Tan(double x) { return tan(x); }
+double Atan(double x) { return atan(x); }
+double Atan2(double y, double x) { return atan2(y, x); }
+double Pow(double a, double b) { return pow(a, b); }
+double Exp(double x) { return exp(x); }
+double Log(double x) { return log(x); }
+
 static bool lRunTest(const char *fn) {
     llvm::LLVMContext *ctx = new llvm::LLVMContext;
 
-#ifdef LLVM_2_8
-    std::string err;
-    llvm::MemoryBuffer *buf = llvm::MemoryBuffer::getFileOrSTDIN(fn, &err);
-    if (!buf) {
-        fprintf(stderr, "Unable to open file \"%s\": %s\n", fn, err.c_str());
-        delete ctx;
-        return false;
-    }
-    std::string bcErr;
-    llvm::Module *module = llvm::ParseBitcodeFile(buf, *ctx, &bcErr);
-#else
     llvm::OwningPtr<llvm::MemoryBuffer> buf;
     llvm::error_code err = llvm::MemoryBuffer::getFileOrSTDIN(fn, buf);
     if (err) {
@@ -142,7 +174,6 @@ static bool lRunTest(const char *fn) {
     }
     std::string bcErr;
     llvm::Module *module = llvm::ParseBitcodeFile(buf.get(), *ctx, &bcErr);
-#endif
 
     if (!module) {
         fprintf(stderr, "Bitcode reader failed for \"%s\": %s\n", fn, bcErr.c_str());
@@ -151,7 +182,21 @@ static bool lRunTest(const char *fn) {
     }
 
     std::string eeError;
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+    llvm::EngineBuilder engineBuilder(module);
+    engineBuilder.setErrorStr(&eeError);
+    engineBuilder.setEngineKind(llvm::EngineKind::JIT);
+#if 0
+    std::vector<std::string> attributes;
+    if (target != NULL && !strcmp(target, "avx"))
+        attributes.push_back("+avx");
+    engineBuilder.setMAttrs(attributes);
+    engineBuilder.setUseMCJIT(true);
+#endif
+    llvm::ExecutionEngine *ee = engineBuilder.create();
+#else
     llvm::ExecutionEngine *ee = llvm::ExecutionEngine::createJIT(module, &eeError);
+#endif
     if (!ee) {
         fprintf(stderr, "Unable to create ExecutionEngine: %s\n", eeError.c_str());
         return false;
@@ -163,10 +208,7 @@ static bool lRunTest(const char *fn) {
         ee->addGlobalMapping(func, (void *)FUNC)
     DO_FUNC(ISPCLaunch, "ISPCLaunch");
     DO_FUNC(ISPCSync, "ISPCSync");
-#ifdef ISPC_IS_WINDOWS
-    DO_FUNC(ISPCMalloc, "ISPCMalloc");
-    DO_FUNC(ISPCFree, "ISPCFree");
-#endif // ISPC_IS_WINDOWS
+    DO_FUNC(ISPCAlloc, "ISPCAlloc");
     DO_FUNC(putchar, "putchar");
     DO_FUNC(printf, "printf");
     DO_FUNC(fflush, "fflush");
@@ -178,14 +220,14 @@ static bool lRunTest(const char *fn) {
     DO_FUNC(powf, "powf");
     DO_FUNC(expf, "expf");
     DO_FUNC(logf, "logf");
-    DO_FUNC(sin, "sin");
-    DO_FUNC(cos, "cos");
-    DO_FUNC(tan, "tan");
-    DO_FUNC(atan, "atan");
-    DO_FUNC(atan2, "atan2");
-    DO_FUNC(pow, "pow");
-    DO_FUNC(exp, "exp");
-    DO_FUNC(log, "log");
+    DO_FUNC(Sin, "sin");
+    DO_FUNC(Cos, "cos");
+    DO_FUNC(Tan, "tan");
+    DO_FUNC(Atan, "atan");
+    DO_FUNC(Atan2, "atan2");
+    DO_FUNC(Pow, "pow");
+    DO_FUNC(Exp, "exp");
+    DO_FUNC(Log, "log");
     DO_FUNC(memset, "memset");
 #ifdef ISPC_IS_APPLE
     DO_FUNC(memset_pattern4, "memset_pattern4");
@@ -233,7 +275,6 @@ static bool lRunTest(const char *fn) {
     float result[16];
     for (int i = 0; i < 16; ++i)
         result[i] = 0;
-    bool ok = true;
     if (foundResult) {
         typedef void (*PFN)(float *);
         PFN pfn = reinterpret_cast<PFN>(ee->getPointerToFunction(func));
@@ -290,50 +331,49 @@ static bool lRunTest(const char *fn) {
     }
     else {
         fprintf(stderr, "Unable to find runnable function in file \"%s\"\n", fn);
-        ok = false;
+        return false;
     }
 
     // see if we got the right result
-    if (ok) {
-        if (foundResult) {
-            for (int i = 0; i < width; ++i)
-                if (returned[i] != result[i]) {
-                    ok = false;
-                    fprintf(stderr, "Test \"%s\" RETURNED %d: %g / %a EXPECTED %g / %a\n",
-                            fn, i, returned[i], returned[i], result[i], result[i]);
-                }
-        }
-        else {
-            for (int i = 0; i < width; ++i)
-                fprintf(stderr, "Test \"%s\" returned %d: %g / %a\n",
-                        fn, i, returned[i], returned[i]);
-        }
+    bool resultsMatch = true;
+    if (foundResult) {
+        for (int i = 0; i < width; ++i)
+            if (returned[i] != result[i]) {
+                resultsMatch = false;
+                fprintf(stderr, "Test \"%s\" RETURNED %d: %g / %a EXPECTED %g / %a\n",
+                        fn, i, returned[i], returned[i], result[i], result[i]);
+            }
     }
+    else {
+        for (int i = 0; i < width; ++i)
+            fprintf(stderr, "Test \"%s\" returned %d: %g / %a\n",
+                    fn, i, returned[i], returned[i]);
+    }
+    if (foundResult && shouldFail && resultsMatch)
+        fprintf(stderr, "Test %s unexpectedly passed\n", fn);
 
     delete ee;
     delete ctx;
 
-    return ok && foundResult;
+    return foundResult && resultsMatch;
 }
 
+
 int main(int argc, char *argv[]) {
     llvm::InitializeNativeTarget();
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+    LLVMLinkInJIT();
+#endif
 
-    std::vector<const char *> files;
+    const char *filename = NULL;
     for (int i = 1; i < argc; ++i) {
         if (!strcmp(argv[i], "--help") || !strcmp(argv[i], "-h"))
             usage(0);
+        if (!strcmp(argv[i], "-f"))
+            shouldFail = true;
         else
-            files.push_back(argv[i]);
+            filename = argv[i];
     }
 
-    int passes = 0, fails = 0;
-    for (unsigned int i = 0; i < files.size(); ++i) {
-        if (lRunTest(files[i])) ++passes;
-        else ++fails;
-    }
-
-    if (fails > 0)
-        fprintf(stderr, "%d/%d tests passed\n", passes, passes+fails);
-    return fails > 0;
+    return (lRunTest(filename) == true) ? 0 : 1;
 }

ispc_test.vcxproj

@@ -52,14 +52,14 @@
       </PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <PreprocessorDefinitions>ISPC_IS_WINDOWS;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>LLVM_3_0;_DEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>$(LLVM_INSTALL_DIR)/include</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
       <GenerateDebugInformation>true</GenerateDebugInformation>
       <AdditionalLibraryDirectories>$(LLVM_INSTALL_DIR)/lib</AdditionalLibraryDirectories>
-      <AdditionalDependencies>LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMipa.lib;LLVMipo.lib;LLVMJIT.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Disassembler.lib;LLVMX86Info.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMipa.lib;LLVMipo.lib;LLVMJIT.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Disassembler.lib;LLVMX86Desc.lib;LLVMX86Info.lib;%(AdditionalDependencies)</AdditionalDependencies>
     </Link>
   </ItemDefinitionGroup>
   <ItemDefinitionGroup Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">
@@ -70,7 +70,7 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <PreprocessorDefinitions>ISPC_IS_WINDOWS;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>LLVM_3_0;NDEBUG;_CONSOLE;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>$(LLVM_INSTALL_DIR)/include</AdditionalIncludeDirectories>
     </ClCompile>
     <Link>
@@ -79,10 +79,10 @@
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
       <OptimizeReferences>true</OptimizeReferences>
       <AdditionalLibraryDirectories>$(LLVM_INSTALL_DIR)/lib</AdditionalLibraryDirectories>
-      <AdditionalDependencies>LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMipa.lib;LLVMipo.lib;LLVMJIT.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Disassembler.lib;LLVMX86Info.lib;%(AdditionalDependencies)</AdditionalDependencies>
+      <AdditionalDependencies>LLVMAnalysis.lib;LLVMArchive.lib;LLVMAsmPrinter.lib;LLVMBitReader.lib;LLVMBitWriter.lib;LLVMCodeGen.lib;LLVMCore.lib;LLVMExecutionEngine.lib;LLVMInstCombine.lib;LLVMInstrumentation.lib;LLVMipa.lib;LLVMipo.lib;LLVMJIT.lib;LLVMLinker.lib;LLVMMC.lib;LLVMMCParser.lib;LLVMObject.lib;LLVMScalarOpts.lib;LLVMSelectionDAG.lib;LLVMSupport.lib;LLVMTarget.lib;LLVMTransformUtils.lib;LLVMX86ASMPrinter.lib;LLVMX86ASMParser.lib;LLVMX86Utils.lib;LLVMX86CodeGen.lib;LLVMX86Disassembler.lib;LLVMX86Desc.lib;LLVMX86Info.lib;%(AdditionalDependencies)</AdditionalDependencies>
     </Link>
   </ItemDefinitionGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

main.cpp

@@ -40,10 +40,14 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <llvm/Support/PrettyStackTrace.h>
-#ifdef LLVM_2_8
-#include <llvm/System/Signals.h>
-#else
 #include <llvm/Support/Signals.h>
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+  #include <llvm/Support/TargetRegistry.h>
+  #include <llvm/Support/TargetSelect.h>
+#else
+  #include <llvm/Target/TargetRegistry.h>
+  #include <llvm/Target/TargetSelect.h>
+  #include <llvm/Target/SubtargetFeature.h>
 #endif
 
 #ifdef ISPC_IS_WINDOWS
@@ -53,36 +57,36 @@
 #endif // ISPC_IS_WINDOWS
 
 static void usage(int ret) {
-    printf("This is the Intel(r) SPMD Program Compiler (ispc), build %s (%s)\n\n", BUILD_DATE, BUILD_VERSION);
+    printf("This is the Intel(r) SPMD Program Compiler (ispc), build %s (%s)\n\n", 
+           BUILD_DATE, BUILD_VERSION);
     printf("usage: ispc\n");
-    printf("    [--arch={x86,x86-64}]\t\tSelect target architecture\n");
+    printf("    [--arch={%s}]\t\tSelect target architecture\n", 
+           Target::SupportedTargetArchs());
     printf("    [--cpu=<cpu>]\t\t\tSelect target CPU type\n");
-    printf("         (atom, barcelona, core2, corei7, corei7-avx, istanbul, nocona,\n");
-    printf("          penryn, westmere)\n");
-#ifndef ISPC_IS_WINDOWS
-    printf("    [-D<foo>]\t\t\t\t#define value when running preprocessor\n");
-#endif
+    printf("         <cpu>={%s}\n", Target::SupportedTargetCPUs());
+    printf("    [-D<foo>]\t\t\t\t#define given value when running preprocessor\n");
     printf("    [--debug]\t\t\t\tPrint information useful for debugging ispc\n");
     printf("    [--emit-asm]\t\t\tGenerate assembly language file as output\n");
     printf("    [--emit-llvm]\t\t\tEmit LLVM bitode file as output\n");
-    printf("    [--emit-obj]\t\t\tGenerate object file file as output\n");
-    printf("    [--fast-math]\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
+    printf("    [--emit-obj]\t\t\tGenerate object file file as output (default)\n");
     printf("    [-g]\t\t\t\tGenerate debugging information\n");
     printf("    [--help]\t\t\t\tPrint help\n");
-    printf("    [-h] <name>\t\t\t\tOutput filename for header\n");
+    printf("    [-h <name>/--header-outfile=<name>]\tOutput filename for header\n");
     printf("    [--instrument]\t\t\tEmit instrumentation to gather performance data\n");
     printf("    [--math-lib=<option>]\t\tSelect math library\n");
     printf("        default\t\t\t\tUse ispc's built-in math functions\n");
     printf("        fast\t\t\t\tUse high-performance but lower-accuracy math functions\n");
-    printf("        svml\t\t\t\tUse the Intel SVML math libraries\n");
+    printf("        svml\t\t\t\tUse the Intel(r) SVML math libraries\n");
     printf("        system\t\t\t\tUse the system's math library (*may be quite slow*)\n");
     printf("    [--nostdlib]\t\t\tDon't make the ispc standard library available\n");
-#ifndef ISPC_IS_WINDOWS
     printf("    [--nocpp]\t\t\t\tDon't run the C preprocessor\n");
-#endif
-    printf("    [-o/--outfile] <name>\t\tOutput filename for bitcode (may be \"-\" for standard output)\n");
-    printf("    [-O0/-O1]\t\t\t\tSet optimization level\n");
+    printf("    [-o <name>/--outfile=<name>]\tOutput filename (may be \"-\" for standard output)\n");
+    printf("    [-O0/-O1]\t\t\t\tSet optimization level (-O1 is default)\n");
     printf("    [--opt=<option>]\t\t\tSet optimization option\n");
+    printf("        disable-loop-unroll\t\tDisable loop unrolling.\n");
+    printf("        fast-masked-vload\t\tFaster masked vector loads on SSE (may go past end of array)\n");
+    printf("        fast-math\t\t\tPerform non-IEEE-compliant optimizations of numeric expressions\n");
+#if 0
     printf("        disable-blended-masked-stores\t\tScalarize masked stores on SSE (vs. using vblendps)\n");
     printf("        disable-coherent-control-flow\t\tDisable coherent control flow optimizations\n");
     printf("        disable-uniform-control-flow\t\tDisable uniform control flow optimizations\n");
@@ -91,11 +95,11 @@ static void usage(int ret) {
     printf("        disable-gather-scatter-flattening\tDisable flattening when all lanes are on\n");
     printf("        disable-uniform-memory-optimizations\tDisable uniform-based coherent memory access\n");
     printf("        disable-masked-store-optimizations\tDisable lowering to regular stores when possible\n");
-#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
-    printf("    [--target={sse2,sse4,sse4x2,avx}] Select target ISA (SSE4 is default unless compiling for atom; then SSE2 is.)\n");
-#else
-    printf("    [--target={sse2,sse4,sse4x2}] Select target ISA (SSE4 is default unless compiling for atom; then SSE2 is.)\n");
-#endif // LLVM 3.0
+#endif
+#ifndef ISPC_IS_WINDOWS
+    printf("    [--pic]\t\t\t\tGenerate position-independent code\n");
+#endif // !ISPC_IS_WINDOWS
+    printf("    [--target=<isa>]\t\t\tSelect target ISA. <isa>={%s}\n", Target::SupportedTargetISAs());
     printf("    [--version]\t\t\t\tPrint ispc version\n");
     printf("    [--woff]\t\t\t\tDisable warnings\n");
     printf("    [--wno-perf]\t\t\tDon't issue warnings related to performance-related issues\n");
@@ -103,35 +107,6 @@ static void usage(int ret) {
     exit(ret);
 }
 
-/** Given a target name string, set initialize the global g->target
-    structure appropriately. 
-*/
-static void lDoTarget(const char *target) {
-    if (!strcasecmp(target, "sse2")) {
-        g->target.isa = Target::SSE2;
-        g->target.nativeVectorWidth = 4;
-        g->target.vectorWidth = 4;
-    }
-    else if (!strcasecmp(target, "sse4")) {
-        g->target.isa = Target::SSE4;
-        g->target.nativeVectorWidth = 4;
-        g->target.vectorWidth = 4;
-    }
-    else if (!strcasecmp(target, "sse4x2")) {
-        g->target.isa = Target::SSE4;
-        g->target.nativeVectorWidth = 4;
-        g->target.vectorWidth = 8;
-    }
-#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
-    else if (!strcasecmp(target, "avx")) {
-        g->target.isa = Target::AVX;
-        g->target.nativeVectorWidth = 8;
-        g->target.vectorWidth = 8;
-    }
-#endif // LLVM 3.0
-    else
-        usage(1);
-}
 
 
 /** We take arguments from both the command line as well as from the
@@ -190,6 +165,16 @@ int main(int Argc, char *Argv[]) {
     llvm::sys::PrintStackTraceOnErrorSignal();
     llvm::PrettyStackTraceProgram X(argc, argv);
 
+    // initialize available LLVM targets
+    LLVMInitializeX86TargetInfo();
+    LLVMInitializeX86Target();
+    LLVMInitializeX86AsmPrinter();
+    LLVMInitializeX86AsmParser();
+    LLVMInitializeX86Disassembler();
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+    LLVMInitializeX86TargetMC();
+#endif
+
     char *file = NULL;
     const char *headerFileName = NULL;
     const char *outFileName = NULL;
@@ -198,23 +183,29 @@ int main(int Argc, char *Argv[]) {
     // as we're parsing below
     g = new Globals;
 
-    bool debugSet = false, optSet = false, targetSet = false;
+    bool debugSet = false, optSet = false;
     Module::OutputType ot = Module::Object;
+    bool generatePIC = false;
+    const char *arch = NULL, *cpu = NULL, *target = NULL;
 
     for (int i = 1; i < argc; ++i) {
         if (!strcmp(argv[i], "--help"))
             usage(0);
-#ifndef ISPC_IS_WINDOWS
-        else if (!strncmp(argv[i], "-D", 2)) {
+        else if (!strncmp(argv[i], "-D", 2))
             g->cppArgs.push_back(argv[i]);
-        }
-#endif // !ISPC_IS_WINDOWS
         else if (!strncmp(argv[i], "--arch=", 7))
-            g->target.arch = argv[i] + 7;
+            arch = argv[i] + 7;
         else if (!strncmp(argv[i], "--cpu=", 6))
-            g->target.cpu = argv[i] + 6;
-        else if (!strcmp(argv[i], "--fast-math"))
-            g->opt.fastMath = true;
+            cpu = argv[i] + 6;
+        else if (!strcmp(argv[i], "--fast-math")) {
+            fprintf(stderr, "--fast-math option has been renamed to --opt=fast-math!\n");
+            usage(1);
+        }
+        else if (!strcmp(argv[i], "--fast-masked-vload")) {
+            fprintf(stderr, "--fast-masked-vload option has been renamed to "
+                    "--opt=fast-masked-vload!\n");
+            usage(1);
+        }
         else if (!strcmp(argv[i], "--debug"))
             g->debugPrint = true;
         else if (!strcmp(argv[i], "--instrument"))
@@ -230,14 +221,12 @@ int main(int Argc, char *Argv[]) {
         else if (!strcmp(argv[i], "--emit-obj"))
             ot = Module::Object;
         else if (!strcmp(argv[i], "--target")) {
+            // FIXME: should remove this way of specifying the target...
             if (++i == argc) usage(1);
-            lDoTarget(argv[i]);
-            targetSet = true;
-        }
-        else if (!strncmp(argv[i], "--target=", 9)) {
-            const char *target = argv[i] + 9;
-            lDoTarget(target);
+            target = argv[i];
         }
+        else if (!strncmp(argv[i], "--target=", 9))
+            target = argv[i] + 9;
         else if (!strncmp(argv[i], "--math-lib=", 11)) {
             const char *lib = argv[i] + 11;
             if (!strcmp(lib, "default"))
@@ -253,7 +242,16 @@ int main(int Argc, char *Argv[]) {
         }
         else if (!strncmp(argv[i], "--opt=", 6)) {
             const char *opt = argv[i] + 6;
-            if (!strcmp(opt, "disable-blended-masked-stores"))
+            if (!strcmp(opt, "fast-math"))
+                g->opt.fastMath = true;
+            else if (!strcmp(opt, "fast-masked-vload"))
+                g->opt.fastMaskedVload = true;
+            else if (!strcmp(opt, "disable-loop-unroll"))
+                g->opt.unrollLoops = false;
+
+            // These are only used for performance tests of specific
+            // optimizations
+            else if (!strcmp(opt, "disable-blended-masked-stores"))
                 g->opt.disableBlendedMaskedStores = true;
             else if (!strcmp(opt, "disable-coherent-control-flow"))
                 g->opt.disableCoherentControlFlow = true;
@@ -278,14 +276,19 @@ int main(int Argc, char *Argv[]) {
         }
         else if (!strcmp(argv[i], "--wno-perf") || !strcmp(argv[i], "-wno-perf"))
             g->emitPerfWarnings = false;
-        else if (!strcmp(argv[i], "-o") || !strcmp(argv[i], "--outfile")) {
+        else if (!strcmp(argv[i], "-o")) {
             if (++i == argc) usage(1);
             outFileName = argv[i];
         }
-        else if (!strcmp(argv[i], "-h") || !strcmp(argv[i], "--header-outfile")) {
+        else if (!strcmp(argv[i], "--outfile="))
+            outFileName = argv[i] + strlen("--outfile=");
+        else if (!strcmp(argv[i], "-h")) {
             if (++i == argc) usage(1);
             headerFileName = argv[i];
         }
+        else if (!strcmp(argv[i], "--header-outfile=")) {
+            headerFileName = argv[i] + strlen("--header-outfile=");
+        }
         else if (!strcmp(argv[i], "-O0")) {
             g->opt.level = 0;
             optSet = true;
@@ -301,6 +304,10 @@ int main(int Argc, char *Argv[]) {
             g->includeStdlib = false;
         else if (!strcmp(argv[i], "--nocpp"))
             g->runCPP = false;
+#ifndef ISPC_IS_WINDOWS
+        else if (!strcmp(argv[i], "--pic"))
+            generatePIC = true;
+#endif // !ISPC_IS_WINDOWS
         else if (!strcmp(argv[i], "-v") || !strcmp(argv[i], "--version")) {
             printf("Intel(r) SPMD Program Compiler (ispc) build %s (%s)\n", 
                    BUILD_DATE, BUILD_VERSION);
@@ -322,10 +329,8 @@ int main(int Argc, char *Argv[]) {
     if (debugSet && !optSet)
         g->opt.level = 0;
 
-    // Make SSE2 the default target on atom unless the target has been set
-    // explicitly.
-    if (!targetSet && (g->target.cpu == "atom"))
-        lDoTarget("sse2");
+    if (!Target::GetTarget(arch, cpu, target, generatePIC, &g->target))
+        usage(1);
 
     m = new Module(file);
     if (m->CompileFile() == 0) {

module.cpp

@@ -72,23 +72,17 @@
 #include <llvm/Support/FormattedStream.h>
 #include <llvm/Support/FileUtilities.h>
 #include <llvm/Target/TargetMachine.h>
-#include <llvm/Target/TargetRegistry.h>
-#include <llvm/Target/TargetSelect.h>
 #include <llvm/Target/TargetOptions.h>
 #include <llvm/Target/TargetData.h>
-#include <llvm/Target/SubtargetFeature.h>
 #include <llvm/PassManager.h>
 #include <llvm/Analysis/Verifier.h>
 #include <llvm/Support/CFG.h>
 #include <clang/Frontend/CompilerInstance.h>
+#include <clang/Frontend/TextDiagnosticPrinter.h>
 #include <clang/Frontend/Utils.h>
 #include <clang/Basic/TargetInfo.h>
-#ifndef LLVM_2_8
 #include <llvm/Support/ToolOutputFile.h>
 #include <llvm/Support/Host.h>
-#else // !LLVM_2_8
-#include <llvm/System/Host.h>
-#endif // LLVM_2_8
 #include <llvm/Assembly/PrintModulePass.h>
 #include <llvm/Support/raw_ostream.h>
 #include <llvm/Bitcode/ReaderWriter.h>
@@ -107,14 +101,13 @@ Module::Module(const char *fn) {
     symbolTable = new SymbolTable;
     module = new llvm::Module(filename ? filename : "<stdin>", *g->ctx);
 
-#ifndef LLVM_2_8
+    module->setTargetTriple(g->target.GetTripleString());
+
     if (g->generateDebuggingSymbols)
         diBuilder = new llvm::DIBuilder(*module);
     else
         diBuilder = NULL;
-#endif // LLVM_2_8
 
-#ifndef LLVM_2_8
     // If we're generating debugging symbols, let the DIBuilder know that
     // we're starting a new compilation unit.
     if (diBuilder != NULL) {
@@ -140,7 +133,6 @@ Module::Module(const char *fn) {
                                          0 /* run time version */);
         }
     }
-#endif // LLVM_2_8
 }
 
 
@@ -154,6 +146,9 @@ extern void yy_delete_buffer(YY_BUFFER_STATE);
 
 int
 Module::CompileFile() {
+    if (g->opt.fastMath == true)
+        llvm::UnsafeFPMath = true;
+
     // FIXME: it'd be nice to do this in the Module constructor, but this
     // function ends up calling into routines that expect the global
     // variable 'm' to be initialized and available (which it isn't until
@@ -458,6 +453,10 @@ Module::AddGlobal(DeclSpecs *ds, Declarator *decl) {
     // declarations, typedefs, and global variables declarations /
     // definitions.  Figure out what we've got and take care of it.
 
+    if (ds == NULL || decl == NULL)
+        // Error happened earlier during parsing
+        return;
+
     if (decl->isFunction) {
         // function declaration
         const Type *t = decl->GetType(ds);
@@ -558,7 +557,6 @@ Module::AddGlobal(DeclSpecs *ds, Declarator *decl) {
                                                          decl->sym->name.c_str());
         m->symbolTable->AddVariable(decl->sym);
 
-#ifndef LLVM_2_8
         if (diBuilder && (ds->storageClass != SC_EXTERN)) {
             llvm::DIFile file = decl->pos.GetDIFile();
             diBuilder->createGlobalVariable(decl->sym->name, 
@@ -568,7 +566,6 @@ Module::AddGlobal(DeclSpecs *ds, Declarator *decl) {
                                             (ds->storageClass == SC_STATIC),
                                             decl->sym->storagePtr);
         }
-#endif // LLVM_2_8
     }
 }
 
@@ -606,6 +603,7 @@ lCopyInTaskParameter(int i, llvm::Value *structArgPtr, Declarator *decl,
     // memory
     llvm::Value *ptrval = ctx->LoadInst(ptr, NULL, sym->name.c_str());
     ctx->StoreInst(ptrval, sym->storagePtr);
+    ctx->EmitFunctionParameterDebugInfo(sym);
 }
 
 
@@ -629,6 +627,8 @@ lEmitFunctionCode(FunctionEmitContext *ctx, llvm::Function *function,
         llvm::Value *structParamPtr = argIter++;
         llvm::Value *threadIndex = argIter++;
         llvm::Value *threadCount = argIter++;
+        llvm::Value *taskIndex = argIter++;
+        llvm::Value *taskCount = argIter++;
 
         // Copy the function parameter values from the structure into local
         // storage
@@ -656,13 +656,17 @@ lEmitFunctionCode(FunctionEmitContext *ctx, llvm::Function *function,
         threadCountSym->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "threadCount");
         ctx->StoreInst(threadCount, threadCountSym->storagePtr);
 
-#ifdef ISPC_IS_WINDOWS
-        // On Windows, we dynamically-allocate space for the task arguments
-        // (see FunctionEmitContext::LaunchInst().)  Here is where we emit
-        // the code to free that memory, now that we've copied the
-        // parameter values out of the structure.
-        ctx->EmitFree(structParamPtr);
-#endif // ISPC_IS_WINDOWS
+        // Copy taskIndex and taskCount into stack-allocated storage so
+        // that their symbols point to something reasonable.
+        Symbol *taskIndexSym = m->symbolTable->LookupVariable("taskIndex");
+        assert(taskIndexSym);
+        taskIndexSym->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "taskIndex");
+        ctx->StoreInst(taskIndex, taskIndexSym->storagePtr);
+
+        Symbol *taskCountSym = m->symbolTable->LookupVariable("taskCount");
+        assert(taskCountSym);
+        taskCountSym->storagePtr = ctx->AllocaInst(LLVMTypes::Int32Type, "taskCount");
+        ctx->StoreInst(taskCount, taskCountSym->storagePtr);
     }
     else {
         // Regular, non-task function
@@ -700,8 +704,18 @@ lEmitFunctionCode(FunctionEmitContext *ctx, llvm::Function *function,
 
     // Finally, we can generate code for the function
     if (code != NULL) {
+        int costEstimate = code->EstimateCost();
         bool checkMask = (ft->isTask == true) || 
-            (function->hasFnAttr(llvm::Attribute::AlwaysInline) == false);
+            ((function->hasFnAttr(llvm::Attribute::AlwaysInline) == false) &&
+             costEstimate > CHECK_MASK_AT_FUNCTION_START_COST);
+        Debug(code->pos, "Estimated cost for function \"%s\" = %d\n", 
+              funSym->name.c_str(), costEstimate);
+        // If the body of the function is non-trivial, then we wrap the
+        // entire thing around a varying "cif (true)" test in order to reap
+        // the side-effect benefit of checking to see if the execution mask
+        // is all on and thence having a specialized code path for that
+        // case.  If this is a simple function, then this isn't worth the
+        // code bloat / overhead.
         if (checkMask) {
             bool allTrue[ISPC_MAX_NVEC];
             for (int i = 0; i < g->target.vectorWidth; ++i)
@@ -849,6 +863,11 @@ Module::AddFunction(DeclSpecs *ds, Declarator *decl, Stmt *code) {
 
 bool
 Module::WriteOutput(OutputType outputType, const char *outFileName) {
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+    if (diBuilder != NULL && outputType != Header)
+        diBuilder->finalize();
+#endif // LLVM_3_0
+
     // First, issue a warning if the output file suffix and the type of
     // file being created seem to mismatch.  This can help catch missing
     // command-line arguments specifying the output file type.
@@ -909,12 +928,7 @@ Module::WriteOutput(OutputType outputType, const char *outFileName) {
             return true;
         }
         else {
-#ifdef LLVM_2_8
-            fprintf(stderr, "Direct object file emission not supported in this build.\n");
-            return false;
-#else
             return writeObjectFileOrAssembly(outputType, outFileName);
-#endif // LLVM_2_8
         }
     }
 }
@@ -922,79 +936,7 @@ Module::WriteOutput(OutputType outputType, const char *outFileName) {
 
 bool
 Module::writeObjectFileOrAssembly(OutputType outputType, const char *outFileName) {
-    llvm::InitializeAllTargets();
-#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
-    llvm::InitializeAllTargetMCs();
-#endif
-    llvm::InitializeAllAsmPrinters();
-    llvm::InitializeAllAsmParsers();
-
-    llvm::Triple triple(module->getTargetTriple());
-    if (triple.getTriple().empty())
-        triple.setTriple(llvm::sys::getHostTriple());
-
-    const llvm::Target *target = NULL;
-    if (g->target.arch != "") {
-        // If the user specified a target architecture, see if it's a known
-        // one; print an error with the valid ones otherwise.
-        for (llvm::TargetRegistry::iterator iter = llvm::TargetRegistry::begin();
-             iter != llvm::TargetRegistry::end(); ++iter) {
-            if (g->target.arch == iter->getName()) {
-                target = &*iter;
-                break;
-            }
-        }
-        if (!target) {
-            fprintf(stderr, "Invalid target \"%s\"\nOptions: ", 
-                    g->target.arch.c_str());
-            llvm::TargetRegistry::iterator iter;
-            for (iter = llvm::TargetRegistry::begin();
-                 iter != llvm::TargetRegistry::end(); ++iter)
-                fprintf(stderr, "%s ", iter->getName());
-            fprintf(stderr, "\n");
-            return false;
-        }
-
-        llvm::Triple::ArchType archType = 
-            llvm::Triple::getArchTypeForLLVMName(g->target.arch);
-        if (archType != llvm::Triple::UnknownArch)
-            triple.setArch(archType);
-    }
-    else {
-        // Otherwise get the target either based on the host or the
-        // module's target, if it has been set there.
-        std::string error;
-        target = llvm::TargetRegistry::lookupTarget(triple.getTriple(), error);
-        if (!target) {
-            fprintf(stderr, "Unable to select target for module: %s\n", 
-                    error.c_str());
-            return false;
-        }
-    }
-
-    std::string featuresString;
-    llvm::TargetMachine *targetMachine = NULL;
-#if defined LLVM_3_0svn || defined LLVM_3_0
-    if (g->target.isa == Target::AVX)
-        featuresString = "+avx";
-    targetMachine = target->createTargetMachine(triple.getTriple(), g->target.cpu,
-                                                featuresString);
-#else
-    if (g->target.cpu.size()) {
-        llvm::SubtargetFeatures features;
-        features.setCPU(g->target.cpu);
-        featuresString = features.getString();
-    }
-
-    targetMachine = target->createTargetMachine(triple.getTriple(), 
-                                                featuresString);
-#endif
-    if (targetMachine == NULL) {
-        fprintf(stderr, "Unable to create target machine for target \"%s\"!",
-                triple.str().c_str());
-        return false;
-    }
-    targetMachine->setAsmVerbosityDefault(true);
+    llvm::TargetMachine *targetMachine = g->target.GetTargetMachine();
 
     // Figure out if we're generating object file or assembly output, and
     // set binary output for object files
@@ -1021,9 +963,8 @@ Module::writeObjectFileOrAssembly(OutputType outputType, const char *outFileName
         (g->opt.level > 0) ? llvm::CodeGenOpt::Aggressive : llvm::CodeGenOpt::None;
 
     if (targetMachine->addPassesToEmitFile(pm, fos, fileType, optLevel)) {
-        fprintf(stderr, "Fatal error adding passes to emit object file for "
-                "target %s!\n", triple.str().c_str());
-        return false;
+        fprintf(stderr, "Fatal error adding passes to emit object file!");
+        exit(1);
     }
 
     // Finally, run the passes to emit the object file/assembly
@@ -1189,6 +1130,12 @@ lEmitVectorTypedefs(const std::vector<const VectorType *> &types, FILE *file) {
     for (unsigned int i = 0; i < types.size(); ++i) {
         std::string baseDecl;
         const VectorType *vt = types[i]->GetAsNonConstType();
+        if (!vt->IsUniformType())
+            // Varying stuff shouldn't be visibile to / used by the
+            // application, so at least make it not simple to access it by
+            // not declaring the type here...
+            continue;
+
         int size = vt->GetElementCount();
 
         baseDecl = vt->GetBaseType()->GetCDeclaration("");
@@ -1361,6 +1308,7 @@ Module::writeHeader(const char *fn) {
     default:
         FATAL("Unhandled target in header emission");
     }
+    fprintf(f, "#define ISPC_TARGET_VECTOR_WIDTH %d\n", g->target.vectorWidth);
 
     fprintf(f, "#ifdef __cplusplus\nnamespace ispc {\n#endif // __cplusplus\n\n");
 
@@ -1397,14 +1345,6 @@ Module::writeHeader(const char *fn) {
     lEmitEnumDecls(exportedEnumTypes, f);
     lEmitStructDecls(exportedStructTypes, f);
 
-    // emit externs for globals
-    if (externGlobals.size() > 0) {
-        fprintf(f, "///////////////////////////////////////////////////////////////////////////\n");
-        fprintf(f, "// Globals declared \"extern\" from ispc code\n");
-        fprintf(f, "///////////////////////////////////////////////////////////////////////////\n");
-        lPrintExternGlobals(f, externGlobals);
-    }
-
     // emit function declarations for exported stuff...
     if (exportedFuncs.size() > 0) {
         fprintf(f, "\n");
@@ -1426,6 +1366,15 @@ Module::writeHeader(const char *fn) {
     // end namespace
     fprintf(f, "\n#ifdef __cplusplus\n}\n#endif // __cplusplus\n");
 
+    // and only now emit externs for globals, outside of the ispc namespace
+    if (externGlobals.size() > 0) {
+        fprintf(f, "\n");
+        fprintf(f, "///////////////////////////////////////////////////////////////////////////\n");
+        fprintf(f, "// Globals declared \"extern\" from ispc code\n");
+        fprintf(f, "///////////////////////////////////////////////////////////////////////////\n");
+        lPrintExternGlobals(f, externGlobals);
+    }
+
     // end guard
     fprintf(f, "\n#endif // %s\n", guard.c_str());
 
@@ -1441,23 +1390,26 @@ Module::execPreprocessor(const char* infilename, llvm::raw_string_ostream* ostre
     std::string error;
 
     inst.createFileManager();
-    inst.createDiagnostics(0, NULL);
-    clang::TargetOptions& options = inst.getTargetOpts();
 
+    llvm::raw_fd_ostream stderrRaw(2, false);
+    clang::TextDiagnosticPrinter *diagPrinter = 
+        new clang::TextDiagnosticPrinter(stderrRaw, clang::DiagnosticOptions());
+    inst.createDiagnostics(0, NULL, diagPrinter);
+
+    clang::TargetOptions &options = inst.getTargetOpts();
     llvm::Triple triple(module->getTargetTriple());
     if (triple.getTriple().empty())
         triple.setTriple(llvm::sys::getHostTriple());
-    
     options.Triple = triple.getTriple();
 
-    clang::TargetInfo* target 
-        = clang::TargetInfo::CreateTargetInfo(inst.getDiagnostics(), options);
+    clang::TargetInfo *target =
+        clang::TargetInfo::CreateTargetInfo(inst.getDiagnostics(), options);
 
     inst.setTarget(target);
     inst.createSourceManager(inst.getFileManager());
     inst.InitializeSourceManager(infilename);
 
-    clang::PreprocessorOptions& opts = inst.getPreprocessorOpts();
+    clang::PreprocessorOptions &opts = inst.getPreprocessorOpts();
 
     //Add defs for ISPC and PI
     opts.addMacroDef("ISPC");
@@ -1470,7 +1422,10 @@ Module::execPreprocessor(const char* infilename, llvm::raw_string_ostream* ostre
         }
     }    
     inst.createPreprocessor();
+
+    clang::LangOptions langOptions;
+    diagPrinter->BeginSourceFile(langOptions, &inst.getPreprocessor());
     clang::DoPrintPreprocessedInput(inst.getPreprocessor(),
                                     ostream, inst.getPreprocessorOutputOpts());
+    diagPrinter->EndSourceFile();
 }
-

module.h

@@ -91,11 +91,8 @@ public:
     /** llvm Module object into which globals and functions are added. */
     llvm::Module *module; 
 
-#ifndef LLVM_2_8
-    /** The diBuilder manages generating debugging information (only
-        supported in LLVM 2.9 and beyond...) */
+    /** The diBuilder manages generating debugging information */
     llvm::DIBuilder *diBuilder;
-#endif
 
     GatherBuffer *gatherBuffer;
 

opt.cpp

@@ -55,13 +55,12 @@
 #include <llvm/Instructions.h>
 #include <llvm/Intrinsics.h>
 #include <llvm/Constants.h>
-#ifndef LLVM_2_8
-    #include <llvm/Target/TargetLibraryInfo.h>
-    #ifdef LLVM_2_9
-        #include <llvm/Support/StandardPasses.h>
-    #else
-        #include <llvm/Support/PassManagerBuilder.h>
-    #endif // LLVM_2_9
+#include <llvm/Analysis/ConstantFolding.h>
+#include <llvm/Target/TargetLibraryInfo.h>
+#ifdef LLVM_2_9
+    #include <llvm/Support/StandardPasses.h>
+#else
+    #include <llvm/Transforms/IPO/PassManagerBuilder.h>
 #endif // LLVM_2_8
 #include <llvm/ADT/Triple.h>
 #include <llvm/Transforms/Scalar.h>
@@ -69,13 +68,18 @@
 #include <llvm/Transforms/Utils/BasicBlockUtils.h>
 #include <llvm/Target/TargetOptions.h>
 #include <llvm/Target/TargetData.h>
+#include <llvm/Target/TargetMachine.h>
 #include <llvm/Analysis/Verifier.h>
 #include <llvm/Support/raw_ostream.h>
-#ifndef LLVM_2_8
 #include <llvm/Analysis/DIBuilder.h>
-#endif
 #include <llvm/Analysis/DebugInfo.h>
 #include <llvm/Support/Dwarf.h>
+#ifdef ISPC_IS_LINUX
+  #include <alloca.h>
+#elif defined(ISPC_IS_WINDOWS)
+  #include <malloc.h>
+  #define alloca _alloca
+#endif // ISPC_IS_WINDOWS
 
 static llvm::Pass *CreateIntrinsicsOptPass();
 static llvm::Pass *CreateGatherScatterFlattenPass();
@@ -178,19 +182,22 @@ Optimize(llvm::Module *module, int optLevel) {
     llvm::PassManager optPM;
     llvm::FunctionPassManager funcPM(module);
 
-#ifndef LLVM_2_8
     llvm::TargetLibraryInfo *targetLibraryInfo =
         new llvm::TargetLibraryInfo(llvm::Triple(module->getTargetTriple()));
     optPM.add(targetLibraryInfo);
-#endif
     optPM.add(new llvm::TargetData(module));
 
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
+    optPM.add(llvm::createIndVarSimplifyPass());
+#endif
+
     if (optLevel == 0) {
         // This is more or less the minimum set of optimizations that we
         // need to do to generate code that will actually run.  (We can't
         // run absolutely no optimizations, since the front-end needs us to
         // take the various __pseudo_* functions it has emitted and turn
         // them into something that can actually execute.
+        optPM.add(llvm::createPromoteMemoryToRegisterPass());
         optPM.add(CreateGatherScatterFlattenPass());
         optPM.add(CreateLowerGatherScatterPass());
         optPM.add(CreateLowerMaskedStorePass());
@@ -211,7 +218,6 @@ Optimize(llvm::Module *module, int optLevel) {
         // only later in the optimization process as things like constant
         // propagation have done their thing, and then when they do kick
         // in, they can often open up new opportunities for optimization...
-#ifndef LLVM_2_8
         llvm::PassRegistry *registry = llvm::PassRegistry::getPassRegistry();
         llvm::initializeCore(*registry);
         llvm::initializeScalarOpts(*registry);
@@ -222,7 +228,7 @@ Optimize(llvm::Module *module, int optLevel) {
         llvm::initializeInstCombine(*registry);
         llvm::initializeInstrumentation(*registry);
         llvm::initializeTarget(*registry);
-#endif
+
         // Early optimizations to try to reduce the total amount of code to
         // work with if we can
         optPM.add(CreateGatherScatterFlattenPass());
@@ -279,13 +285,11 @@ Optimize(llvm::Module *module, int optLevel) {
         optPM.add(llvm::createConstantPropagationPass());
         optPM.add(CreateIntrinsicsOptPass());
 
-#if defined(LLVM_2_8)
-        optPM.add(CreateIsCompileTimeConstantPass(true));
-#elif defined(LLVM_2_9)
+#if defined(LLVM_2_9)
         llvm::createStandardModulePasses(&optPM, 3, 
                                          false /* opt size */,
                                          true /* unit at a time */, 
-                                         false /* unroll loops */,
+                                         g->opt.unrollLoops,
                                          true /* simplify lib calls */,
                                          false /* may have exceptions */,
                                          llvm::createFunctionInliningPass());
@@ -300,7 +304,7 @@ Optimize(llvm::Module *module, int optLevel) {
         llvm::createStandardModulePasses(&optPM, 3, 
                                          false /* opt size */,
                                          true /* unit at a time */, 
-                                         false /* unroll loops */,
+                                         g->opt.unrollLoops,
                                          true /* simplify lib calls */,
                                          false /* may have exceptions */,
                                          llvm::createFunctionInliningPass());
@@ -309,6 +313,8 @@ Optimize(llvm::Module *module, int optLevel) {
         llvm::PassManagerBuilder builder;
         builder.OptLevel = 3;
         builder.Inliner = llvm::createFunctionInliningPass();
+        if (g->opt.unrollLoops == false)
+            builder.DisableUnrollLoops = true;
         builder.populateFunctionPassManager(funcPM);
         builder.populateModulePassManager(optPM);
         optPM.add(CreateIsCompileTimeConstantPass(true));
@@ -421,8 +427,11 @@ IntrinsicsOpt::IntrinsicsOpt()
     blendInstructions.push_back(BlendInstruction(
         llvm::Intrinsic::getDeclaration(m->module, llvm::Intrinsic::x86_sse41_blendvps),
         0xf, 0, 1, 2));
+#if defined(LLVM_3_0) || defined(LLVM_3_0svn)
     blendInstructions.push_back(BlendInstruction(
-        m->module->getFunction("llvm.x86.avx.blendvps"), 0xff, 0, 1, 2));
+        llvm::Intrinsic::getDeclaration(m->module, llvm::Intrinsic::x86_avx_blendv_ps_256),
+        0xff, 0, 1, 2));
+#endif
 }
 
 
@@ -469,8 +478,18 @@ lGetMask(llvm::Value *factor) {
     else if (llvm::isa<llvm::ConstantAggregateZero>(factor))
         return 0;
     else {
+#if 0
+        llvm::ConstantExpr *ce = llvm::dyn_cast<llvm::ConstantExpr>(factor);
+        if (ce != NULL) {
+            llvm::TargetMachine *targetMachine = g->target.GetTargetMachine();
+            const llvm::TargetData *td = targetMachine->getTargetData();
+            llvm::Constant *c = llvm::ConstantFoldConstantExpression(ce, td);
+            c->dump();
+            factor = c;
+        }
         // else we should be able to handle it above...
         assert(!llvm::isa<llvm::Constant>(factor));
+#endif
         return -1;
     }
 }
@@ -608,9 +627,10 @@ IntrinsicsOpt::runOnBasicBlock(llvm::BasicBlock &bb) {
                                           llvm::PointerType::get(returnType, 0), 
                                           "ptr2vec", callInst);
                 lCopyMetadata(castPtr, callInst);
+                int align = callInst->getCalledFunction() == avxMaskedLoad32 ? 4 : 8;
                 llvm::Instruction *loadInst = 
                     new llvm::LoadInst(castPtr, "load", false /* not volatile */,
-                                       0 /* align */, (llvm::Instruction *)NULL);
+                                       align, (llvm::Instruction *)NULL);
                 lCopyMetadata(loadInst, callInst);
                 llvm::ReplaceInstWithInst(callInst, loadInst);
                 modifiedAny = true;
@@ -630,17 +650,21 @@ IntrinsicsOpt::runOnBasicBlock(llvm::BasicBlock &bb) {
             }
             else if (mask == 0xff) {
                 // all lanes storing, so replace with a regular store
-                llvm::Value *rvalue = callInst->getArgOperand(1);
+                llvm::Value *rvalue = callInst->getArgOperand(2);
                 llvm::Type *storeType = rvalue->getType();
                 llvm::Value *castPtr = 
                     new llvm::BitCastInst(callInst->getArgOperand(0),
                                           llvm::PointerType::get(storeType, 0), 
                                           "ptr2vec", callInst);
                 lCopyMetadata(castPtr, callInst);
-                llvm::Instruction *storeInst = 
+
+                llvm::StoreInst *storeInst = 
                     new llvm::StoreInst(rvalue, castPtr, (llvm::Instruction *)NULL);
+                int align = callInst->getCalledFunction() == avxMaskedStore32 ? 4 : 8;
+                storeInst->setAlignment(align);
                 lCopyMetadata(storeInst, callInst);
                 llvm::ReplaceInstWithInst(callInst, storeInst);
+
                 modifiedAny = true;
                 goto restart;
             }
@@ -1416,15 +1440,12 @@ LowerMaskedStorePass::runOnBasicBlock(llvm::BasicBlock &bb) {
         llvm::Value *rvalue  = callInst->getArgOperand(1);
         llvm::Value *mask = callInst->getArgOperand(2);
 
-        // On SSE, we need to choose between doing the load + blend + store
-        // trick, or serializing the masked store.  On targets with a
-        // native masked store instruction, the implementations of
-        // __masked_store_blend_* should be the same as __masked_store_*,
-        // so this doesn't matter.  On SSE, blending is generally more
-        // efficient and is always safe to do on stack-allocated values.(?)
-        bool doBlend = lIsStackVariablePointer(lvalue);
-        if (g->target.isa == Target::SSE4 || g->target.isa == Target::SSE2)
-            doBlend |= !g->opt.disableBlendedMaskedStores;
+        // We need to choose between doing the load + blend + store trick,
+        // or serializing the masked store.  Even on targets with a native
+        // masked store instruction, this is preferable since it lets us
+        // keep values in registers rather than going out to the stack.
+        bool doBlend = (!g->opt.disableBlendedMaskedStores ||
+                        lIsStackVariablePointer(lvalue));
 
         // Generate the call to the appropriate masked store function and
         // replace the __pseudo_* one with it.
@@ -1502,8 +1523,8 @@ static void lPrintVector(const char *info, llvm::Value *elements[ISPC_MAX_NVEC])
 
 
 /** Given an LLVM vector in vec, return a 'scalarized' version of the
-    vector in the provided offsets[] array.  For example, if the vector
-    value passed in is:  
+    vector in the provided scalarizedVector[] array.  For example, if the
+    vector value passed in is:
 
     add <4 x i32> %a_smear, <4 x i32> <4, 8, 12, 16>,
 
@@ -1524,28 +1545,39 @@ static void lPrintVector(const char *info, llvm::Value *elements[ISPC_MAX_NVEC])
     @param vec               Vector to be scalarized
     @param scalarizedVector  Array in which to store the individual vector 
                              elements
+    @param vectorLength      Number of elements in the given vector. (The
+                             passed scalarizedVector array must also be at least
+                             this length as well.)
     @returns                 True if the vector was successfully scalarized and
                              the values in offsets[] are valid; false otherwise
  */
 static bool
-lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC]) {
+lScalarizeVector(llvm::Value *vec, llvm::Value **scalarizedVector,
+                 int vectorLength) {
     // First initialize the values of scalarizedVector[] to NULL.
-    for (int i = 0; i < g->target.vectorWidth; ++i)
+    for (int i = 0; i < vectorLength; ++i)
         scalarizedVector[i] = NULL;
+    
+    // It may be ok for the vector to be an undef vector; these come up for
+    // example in shufflevector instructions.  As long as elements of the
+    // undef vector aren't referenced by the shuffle indices, this is fine.
+    if (llvm::isa<llvm::UndefValue>(vec))
+        return true;
 
     // ConstantVectors are easy; just pull out the individual constant
     // element values
     llvm::ConstantVector *cv = llvm::dyn_cast<llvm::ConstantVector>(vec);
     if (cv != NULL) {
-        for (int i = 0; i < g->target.vectorWidth; ++i)
+        for (int i = 0; i < vectorLength; ++i)
             scalarizedVector[i] = cv->getOperand(i);
         return true;
     }
 
     // It's also easy if it's just a vector of all zeros
-    llvm::ConstantAggregateZero *caz = llvm::dyn_cast<llvm::ConstantAggregateZero>(vec);
-    if (caz) {
-        for (int i = 0; i < g->target.vectorWidth; ++i)
+    llvm::ConstantAggregateZero *caz = 
+        llvm::dyn_cast<llvm::ConstantAggregateZero>(vec);
+    if (caz != NULL) {
+        for (int i = 0; i < vectorLength; ++i)
             scalarizedVector[i] = LLVMInt32(0);
         return true;
     }
@@ -1557,13 +1589,16 @@ lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC])
         // scalar values we return from here are synthesized with scalar
         // versions of the original vector binary operator
         llvm::Instruction::BinaryOps opcode = bo->getOpcode();
-        llvm::Value *v0[ISPC_MAX_NVEC], *v1[ISPC_MAX_NVEC];
+        llvm::Value **v0 = 
+            (llvm::Value **)alloca(vectorLength * sizeof(llvm::Value *));
+        llvm::Value **v1 = 
+            (llvm::Value **)alloca(vectorLength * sizeof(llvm::Value *));
 
-        if (!lScalarizeVector(bo->getOperand(0), v0) || 
-            !lScalarizeVector(bo->getOperand(1), v1))
+        if (!lScalarizeVector(bo->getOperand(0), v0, vectorLength) || 
+            !lScalarizeVector(bo->getOperand(1), v1, vectorLength))
             return false;
 
-        for (int i = 0; i < g->target.vectorWidth; ++i) {
+        for (int i = 0; i < vectorLength; ++i) {
             scalarizedVector[i] = 
                 llvm::BinaryOperator::Create(opcode, v0[i], v1[i], "flat_bop", bo);
             lCopyMetadata(scalarizedVector[i], bo);
@@ -1588,7 +1623,7 @@ lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC])
         // vaue in scalarizedVector[] based on the value being inserted.
         while (ie != NULL) {
             uint64_t iOffset = lGetIntValue(ie->getOperand(2));
-            assert((int)iOffset < g->target.vectorWidth);
+            assert((int)iOffset < vectorLength);
             assert(scalarizedVector[iOffset] == NULL);
 
             scalarizedVector[iOffset] = ie->getOperand(1);
@@ -1602,15 +1637,17 @@ lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC])
     }
 
     llvm::CastInst *ci = llvm::dyn_cast<llvm::CastInst>(vec);
-    if (ci) {
+    if (ci != NULL) {
         // Casts are similar to BinaryOperators in that we attempt to
         // scalarize the vector being cast and if successful, we apply
         // equivalent scalar cast operators to each of the values in the
         // scalarized vector.
         llvm::Instruction::CastOps op = ci->getOpcode();
 
-        llvm::Value *scalarizedTarget[ISPC_MAX_NVEC];
-        if (!lScalarizeVector(ci->getOperand(0), scalarizedTarget))
+        llvm::Value **scalarizedTarget = 
+            (llvm::Value **)alloca(vectorLength * sizeof(llvm::Value *));
+        if (!lScalarizeVector(ci->getOperand(0), scalarizedTarget,
+                              vectorLength))
             return false;
 
         LLVM_TYPE_CONST llvm::Type *destType = ci->getDestTy();
@@ -1619,7 +1656,7 @@ lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC])
         assert(vectorDestType != NULL);
         LLVM_TYPE_CONST llvm::Type *elementType = vectorDestType->getElementType();
 
-        for (int i = 0; i < g->target.vectorWidth; ++i) {
+        for (int i = 0; i < vectorLength; ++i) {
             scalarizedVector[i] = 
                 llvm::CastInst::Create(op, scalarizedTarget[i], elementType,
                                        "cast", ci);
@@ -1629,16 +1666,11 @@ lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC])
     }
 
     llvm::ShuffleVectorInst *svi = llvm::dyn_cast<llvm::ShuffleVectorInst>(vec);
-    if (svi) {
-        // Note that the code for shufflevector instructions is untested.
-        // (We haven't yet had a case where it needs to run).  Therefore,
-        // an assert at the bottom of this routien will hit the first time
-        // it runs as a reminder that this needs to be tested further.
-
+    if (svi != NULL) {
         LLVM_TYPE_CONST llvm::VectorType *svInstType = 
             llvm::dyn_cast<LLVM_TYPE_CONST llvm::VectorType>(svi->getType());
         assert(svInstType != NULL);
-        assert((int)svInstType->getNumElements() == g->target.vectorWidth);
+        assert((int)svInstType->getNumElements() == vectorLength);
 
         // Scalarize the two vectors being shuffled.  First figure out how
         // big they are.
@@ -1653,58 +1685,90 @@ lScalarizeVector(llvm::Value *vec, llvm::Value *scalarizedVector[ISPC_MAX_NVEC])
         int n0 = vectorType0->getNumElements();
         int n1 = vectorType1->getNumElements();
 
-        // FIXME: It's actually totally legitimate for these two to have
-        // different sizes; the final result just needs to have the native
-        // vector width.  To handle this, not only do we need to
-        // potentially dynamically allocate space for the arrays passed
-        // into lScalarizeVector, but we need to change the rest of its
-        // implementation to not key off g->target.vectorWidth everywhere
-        // to get the sizes of the arrays to iterate over, etc.
-        assert(n0 == g->target.vectorWidth && n1 == g->target.vectorWidth);
-
         // Go ahead and scalarize the two input vectors now.
-        // FIXME: it's ok if some or all of the values of these two vectors
-        // have undef values, so long as we don't try to access undef
-        // values with the vector indices provided to the instruction.
-        // Should fix lScalarizeVector so that it doesn't return false in
-        // this case and just leaves the elements of the arrays with undef
-        // values as NULL.
-        llvm::Value *v0[ISPC_MAX_NVEC], *v1[ISPC_MAX_NVEC];
-        if (!lScalarizeVector(svi->getOperand(0), v0) ||
-            !lScalarizeVector(svi->getOperand(1), v1))
+        llvm::Value **v0 = (llvm::Value **)alloca(n0 * sizeof(llvm::Value *));
+        llvm::Value **v1 = (llvm::Value **)alloca(n1 * sizeof(llvm::Value *));
+
+        if (!lScalarizeVector(svi->getOperand(0), v0, n0) ||
+            !lScalarizeVector(svi->getOperand(1), v1, n1))
             return false;
 
-        llvm::ConstantVector *shuffleIndicesVector = 
-            llvm::dyn_cast<llvm::ConstantVector>(svi->getOperand(2));
-        // I think this has to be a ConstantVector.  If this ever hits,
-        // we'll dig into what we got instead and figure out how to handle
-        // that...
-        assert(shuffleIndicesVector != NULL);
-
-        // Get the integer indices for each element of the returned vector
-        llvm::SmallVector<llvm::Constant *, ISPC_MAX_NVEC> shuffleIndices;
-        shuffleIndicesVector->getVectorElements(shuffleIndices);
-        assert((int)shuffleIndices.size() == g->target.vectorWidth);
-
-        // And loop over the indices, setting the i'th element of the
-        // result vector with the source vector element that corresponds to
-        // the i'th shuffle index value.
-        for (unsigned int i = 0; i < shuffleIndices.size(); ++i) {
-            if (!llvm::isa<llvm::ConstantInt>(shuffleIndices[i]))
-                // I'm not sure when this case would ever happen, though..
-                return false;
-            int offset = (int)lGetIntValue(shuffleIndices[i]);
-            assert(offset >= 0 && offset < n0+n1);
-
-            if (offset < n0)
-                // Offsets from 0 to n0-1 index into the first vector
-                scalarizedVector[i] = v0[offset];
-            else
-                // And offsets from n0 to (n0+n1-1) index into the second
-                // vector
-                scalarizedVector[i] = v1[offset - n0];
+        llvm::ConstantAggregateZero *caz = 
+            llvm::dyn_cast<llvm::ConstantAggregateZero>(svi->getOperand(2));
+        if (caz != NULL) {
+            for (int i = 0; i < vectorLength; ++i)
+                scalarizedVector[i] = v0[0];
+        }
+        else {
+            llvm::ConstantVector *shuffleIndicesVector = 
+                llvm::dyn_cast<llvm::ConstantVector>(svi->getOperand(2));
+            // I think this has to be a ConstantVector.  If this ever hits,
+            // we'll dig into what we got instead and figure out how to handle
+            // that...
+            assert(shuffleIndicesVector != NULL);
+
+            // Get the integer indices for each element of the returned vector
+            llvm::SmallVector<llvm::Constant *, ISPC_MAX_NVEC> shuffleIndices;
+            shuffleIndicesVector->getVectorElements(shuffleIndices);
+            assert((int)shuffleIndices.size() == vectorLength);
+
+            // And loop over the indices, setting the i'th element of the
+            // result vector with the source vector element that corresponds to
+            // the i'th shuffle index value.
+            for (unsigned int i = 0; i < shuffleIndices.size(); ++i) {
+                // I'm not sure when this case would ever happen, though..
+                assert(llvm::isa<llvm::ConstantInt>(shuffleIndices[i]));
+
+                int offset = (int)lGetIntValue(shuffleIndices[i]);
+                assert(offset >= 0 && offset < n0+n1);
+
+                if (offset < n0)
+                    // Offsets from 0 to n0-1 index into the first vector
+                    scalarizedVector[i] = v0[offset];
+                else
+                    // And offsets from n0 to (n0+n1-1) index into the second
+                    // vector
+                    scalarizedVector[i] = v1[offset - n0];
+            }
+        }
+        return true;
+    }
+
+    llvm::LoadInst *li = llvm::dyn_cast<llvm::LoadInst>(vec);
+    if (li != NULL) {
+        llvm::Value *baseAddr = li->getOperand(0);
+        llvm::Value *baseInt = new llvm::PtrToIntInst(baseAddr, LLVMTypes::Int64Type,
+                                                      "base2int", li);
+        lCopyMetadata(baseInt, li);
+
+        LLVM_TYPE_CONST llvm::PointerType *ptrType = 
+            llvm::dyn_cast<llvm::PointerType>(baseAddr->getType());
+        assert(ptrType != NULL);
+        LLVM_TYPE_CONST llvm::VectorType *vecType = 
+            llvm::dyn_cast<llvm::VectorType>(ptrType->getElementType());
+        assert(vecType != NULL);
+        LLVM_TYPE_CONST llvm::Type *elementType = vecType->getElementType();
+        uint64_t elementSize;
+        bool sizeKnown = lSizeOfIfKnown(elementType, &elementSize);
+        assert(sizeKnown == true);
+
+        LLVM_TYPE_CONST llvm::Type *eltPtrType = llvm::PointerType::get(elementType, 0);
+
+        for (int i = 0; i < vectorLength; ++i) {
+            llvm::Value *intPtrOffset = 
+                llvm::BinaryOperator::Create(llvm::Instruction::Add, baseInt,
+                                             LLVMInt64(i * elementSize), "baseoffset",
+                                             li);
+            lCopyMetadata(intPtrOffset, li);
+            llvm::Value *scalarLoadPtr = 
+                new llvm::IntToPtrInst(intPtrOffset, eltPtrType, "int2ptr", li);
+            lCopyMetadata(scalarLoadPtr, li);
+
+            llvm::Instruction *scalarLoad = 
+                new llvm::LoadInst(scalarLoadPtr, "loadelt", li);
+            lCopyMetadata(scalarLoad, li);
+            scalarizedVector[i] = scalarLoad;
         }
-        FATAL("the above code is untested so far; check now that it's actually running");
         return true;
     }
 
@@ -2116,11 +2180,18 @@ GSImprovementsPass::runOnBasicBlock(llvm::BasicBlock &bb) {
         if (ce && ce->getOpcode() == llvm::Instruction::BitCast)
             base = ce->getOperand(0);
 
-        // Try to out the offsets; the i'th element of the offsetElements
-        // array should be an i32 with the value of the offset for the i'th
-        // vector lane.  This may fail; if so, just give up.
+        // Try to find out the offsets; the i'th element of the
+        // offsetElements array should be an i32 with the value of the
+        // offset for the i'th vector lane.  This may fail; if so, just
+        // give up.
+        llvm::Value *vecValue = callInst->getArgOperand(1);
+        LLVM_TYPE_CONST llvm::VectorType *vt = 
+            llvm::dyn_cast<llvm::VectorType>(vecValue->getType());
+        assert(vt != NULL);
+        int vecLength = vt->getNumElements();
+        assert(vecLength == g->target.vectorWidth);
         llvm::Value *offsetElements[ISPC_MAX_NVEC];
-        if (!lScalarizeVector(callInst->getArgOperand(1), offsetElements))
+        if (!lScalarizeVector(vecValue, offsetElements, vecLength))
             continue;
 
         llvm::Value *mask = callInst->getArgOperand((gatherInfo != NULL) ? 2 : 3);
@@ -2497,7 +2568,7 @@ llvm::RegisterPass<MakeInternalFuncsStaticPass>
 bool
 MakeInternalFuncsStaticPass::runOnModule(llvm::Module &module) {
     const char *names[] = {
-        "__do_print",
+        "__do_print", "__fast_masked_vload", "__num_cores",
         "__gather_base_offsets_i8", "__gather_base_offsets_i16",
         "__gather_base_offsets_i32", "__gather_base_offsets_i64",
         "__gather_elt_8", "__gather_elt_16", 

parse.yy

@@ -165,7 +165,7 @@ static const char *lParamListTokens[] = {
 %token TOKEN_CBREAK TOKEN_CCONTINUE TOKEN_CRETURN TOKEN_SYNC TOKEN_PRINT
 
 %type <expr> primary_expression postfix_expression
-%type <expr> unary_expression cast_expression
+%type <expr> unary_expression cast_expression launch_expression
 %type <expr> multiplicative_expression additive_expression shift_expression
 %type <expr> relational_expression equality_expression and_expression
 %type <expr> exclusive_or_expression inclusive_or_expression
@@ -177,6 +177,7 @@ static const char *lParamListTokens[] = {
 %type <stmt> statement labeled_statement compound_statement for_init_statement
 %type <stmt> expression_statement selection_statement iteration_statement
 %type <stmt> jump_statement statement_list declaration_statement print_statement
+%type <stmt> sync_statement
 
 %type <declaration> declaration parameter_declaration
 %type <declarators> init_declarator_list 
@@ -221,7 +222,7 @@ primary_expression
         else {
             std::vector<Symbol *> *funs = m->symbolTable->LookupFunction(name);
             if (funs)
-                $$ = new FunctionSymbolExpr(funs, @1);
+                $$ = new FunctionSymbolExpr(name, funs, @1);
         }
         if ($$ == NULL) {
             std::vector<std::string> alternates = 
@@ -256,18 +257,32 @@ primary_expression
     | '(' expression ')' { $$ = $2; }
     ;
 
+launch_expression
+    : TOKEN_LAUNCH '<' postfix_expression '(' argument_expression_list ')' '>'
+      { 
+          ConstExpr *oneExpr = new ConstExpr(AtomicType::UniformInt32, (int32_t)1, @3);
+          $$ = new FunctionCallExpr($3, $5, @3, true, oneExpr);
+      }
+    | TOKEN_LAUNCH '<' postfix_expression '(' ')' '>'
+      {
+          ConstExpr *oneExpr = new ConstExpr(AtomicType::UniformInt32, (int32_t)1, @3);
+          $$ = new FunctionCallExpr($3, new ExprList(@3), @3, true, oneExpr);
+       }
+    | TOKEN_LAUNCH '[' expression ']' '<' postfix_expression '(' argument_expression_list ')' '>'
+      { $$ = new FunctionCallExpr($6, $8, @6, true, $3); }
+    | TOKEN_LAUNCH '[' expression ']' '<' postfix_expression '(' ')' '>'
+      { $$ = new FunctionCallExpr($6, new ExprList(@6), @6, true, $3); }
+    ;
+
 postfix_expression
     : primary_expression
     | postfix_expression '[' expression ']'
       { $$ = new IndexExpr($1, $3, @1); }
     | postfix_expression '(' ')'
-      { $$ = new FunctionCallExpr($1, new ExprList(@1), @1, false); }
+      { $$ = new FunctionCallExpr($1, new ExprList(@1), @1); }
     | postfix_expression '(' argument_expression_list ')'
-      { $$ = new FunctionCallExpr($1, $3, @1, false); }
-    | TOKEN_LAUNCH '<' postfix_expression '(' argument_expression_list ')' '>'
-      { $$ = new FunctionCallExpr($3, $5, @3, true); }
-    | TOKEN_LAUNCH '<' postfix_expression '(' ')' '>'
-      { $$ = new FunctionCallExpr($3, new ExprList(@3), @3, true); }
+      { $$ = new FunctionCallExpr($1, $3, @1); }
+    | launch_expression
     | postfix_expression '.' TOKEN_IDENTIFIER
       { $$ = MemberExpr::create($1, yytext, @1, @3); }
 /*    | postfix_expression TOKEN_PTR_OP TOKEN_IDENTIFIER
@@ -436,8 +451,6 @@ assignment_expression
 
 expression
     : assignment_expression
-    | TOKEN_SYNC 
-      { $$ = new SyncExpr(@1); }
     | expression ',' assignment_expression
       { $$ = new BinaryExpr(BinaryExpr::Comma, $1, $3, @2); }
     ;
@@ -928,9 +941,13 @@ parameter_list
             builtinTokens.push_back(*token);
             ++token;
         }
-        std::vector<std::string> alternates = MatchStrings(yytext, builtinTokens);
-        std::string alts = lGetAlternates(alternates);
-        Error(@1, "Syntax error--token \"%s\" unknown.%s", yytext, alts.c_str());
+        if (strlen(yytext) == 0)
+            Error(@1, "Syntax error--premature end of file.");
+        else {
+            std::vector<std::string> alternates = MatchStrings(yytext, builtinTokens);
+            std::string alts = lGetAlternates(alternates);
+            Error(@1, "Syntax error--token \"%s\" unknown.%s", yytext, alts.c_str());
+        }
         $$ = NULL;
     }
     ;
@@ -1019,6 +1036,7 @@ statement
     | jump_statement
     | declaration_statement
     | print_statement
+    | sync_statement
     | error
     {
         std::vector<std::string> builtinTokens;
@@ -1027,9 +1045,13 @@ statement
             builtinTokens.push_back(*token);
             ++token;
         }
-        std::vector<std::string> alternates = MatchStrings(yytext, builtinTokens);
-        std::string alts = lGetAlternates(alternates);
-        Error(@1, "Syntax error--token \"%s\" unknown.%s", yytext, alts.c_str());
+        if (strlen(yytext) == 0)
+            Error(@1, "Syntax error--premature end of file.");
+        else {
+            std::vector<std::string> alternates = MatchStrings(yytext, builtinTokens);
+            std::string alts = lGetAlternates(alternates);
+            Error(@1, "Syntax error--token \"%s\" unknown.%s", yytext, alts.c_str());
+        }
         $$ = NULL;
     }
     ;
@@ -1155,6 +1177,11 @@ jump_statement
       { $$ = new ReturnStmt($2, true, @1); }
     ;
 
+sync_statement
+    : TOKEN_SYNC 
+      { $$ = new ExprStmt(new SyncExpr(@1), @1); }
+    ;
+
 print_statement
     : TOKEN_PRINT '(' string_constant ')'
       {
@@ -1177,9 +1204,13 @@ translation_unit
             builtinTokens.push_back(*token);
             ++token;
         }
-        std::vector<std::string> alternates = MatchStrings(yytext, builtinTokens);
-        std::string alts = lGetAlternates(alternates);
-        Error(@1, "Syntax error--token \"%s\" unknown.%s", yytext, alts.c_str());
+        if (strlen(yytext) == 0)
+            Error(@1, "Syntax error--premature end of file.");
+        else {
+            std::vector<std::string> alternates = MatchStrings(yytext, builtinTokens);
+            std::string alts = lGetAlternates(alternates);
+            Error(@1, "Syntax error--token \"%s\" unknown.%s", yytext, alts.c_str());
+        }
     }
     ;
 
@@ -1266,6 +1297,12 @@ static void lAddThreadIndexCountToSymbolTable(SourcePos pos) {
 
     Symbol *threadCountSym = new Symbol("threadCount", pos, AtomicType::UniformConstUInt32);
     m->symbolTable->AddVariable(threadCountSym);
+
+    Symbol *taskIndexSym = new Symbol("taskIndex", pos, AtomicType::UniformConstUInt32);
+    m->symbolTable->AddVariable(taskIndexSym);
+
+    Symbol *taskCountSym = new Symbol("taskCount", pos, AtomicType::UniformConstUInt32);
+    m->symbolTable->AddVariable(taskCountSym);
 }
 
 

run_tests.py

@@ -0,0 +1,218 @@
+#!/usr/bin/python
+
+# test-running driver for ispc
+
+# TODO: windows support (mostly should be calling CL.exe rather than gcc
+#   for static linking?)
+
+from optparse import OptionParser
+import multiprocessing
+from ctypes import c_int
+import os
+import sys
+import glob
+import re
+import signal
+import random
+import string
+import mutex
+import subprocess
+import platform
+
+parser = OptionParser()
+parser.add_option("-r", "--random-shuffle", dest="random", help="Randomly order tests",
+                  default=False, action="store_true")
+parser.add_option("-s", "--static-exe", dest="static_exe", 
+                  help="Create and run a regular executable for each test (rather than using the LLVM JIT).",
+                  default=False, action="store_true")
+parser.add_option('-t', '--target', dest='target',
+                  help='Set compilation target (sse2, sse4, sse4x2, avx, avx-x2)',
+                  default="sse4")
+parser.add_option('-a', '--arch', dest='arch',
+                  help='Set architecture (x86, x86-64)',
+                  default="x86-64")
+parser.add_option('-o', '--no-opt', dest='no_opt', help='Disable optimization',
+                  default=False, action="store_true")
+
+(options, args) = parser.parse_args()
+
+# if no specific test files are specified, run all of the tests in tests/
+# and failing_tests/
+if len(args) == 0:
+    files = glob.glob("tests/*ispc") + glob.glob("failing_tests/*ispc")
+else:
+    files = args
+
+# randomly shuffle the tests if asked to do so
+if (options.random):
+    random.seed()
+    random.shuffle(files)
+
+# counter
+total_tests = 0
+finished_tests_counter = multiprocessing.Value(c_int)
+
+# We'd like to use the Lock class from the multiprocessing package to
+# serialize accesses to finished_tests_counter.  Unfortunately, the version of
+# python that ships with OSX 10.5 has this bug:
+# http://bugs.python.org/issue5261.  Therefore, we use the (deprecated but
+# still available) mutex class.
+#finished_tests_counter_lock = multiprocessing.Lock()
+finished_tests_mutex = mutex.mutex()
+
+# utility routine to print an update on the number of tests that have been
+# finished.  Should be called with the mutex (or lock) held..
+def update_progress(fn):
+    finished_tests_counter.value = finished_tests_counter.value + 1
+    progress_str = " Done %d / %d [%s]" % (finished_tests_counter.value, total_tests, fn)
+    # spaces to clear out detrius from previous printing...
+    for x in range(30):
+        progress_str += ' '
+    progress_str += '\r'
+    sys.stdout.write(progress_str)
+    sys.stdout.flush()
+    finished_tests_mutex.unlock()
+
+fnull = open(os.devnull, 'w')
+
+# run the commands in cmd_list
+def run_cmds(cmd_list, filename, expect_failure):
+    for cmd in cmd_list:
+        if expect_failure:
+            failed = (subprocess.call(cmd, shell = True, stdout = fnull, stderr = fnull) != 0)
+        else:
+            failed = (os.system(cmd) != 0)
+        if failed:
+            break
+
+    surprise = ((expect_failure and not failed) or (not expect_failure and failed))
+    if surprise == True:
+        print "Test %s %s                 " % \
+            (filename, "unexpectedly passed" if expect_failure else "failed")
+    return surprise
+
+
+# pull tests to run from the given queue and run them.  Multiple copies of
+# this function will be running in parallel across all of the CPU cores of
+# the system.
+def run_tasks_from_queue(queue):
+    error_count = 0
+    while True:
+        filename = queue.get()
+        if (filename == 'STOP'):
+            sys.exit(error_count)
+
+        # do we expect this test to fail?
+        should_fail = (filename.find("failing_") != -1)
+
+        if options.static_exe == True:
+            # if the user wants us to build a static executable to run for
+            # this test, we need to figure out the signature of the test
+            # function that this test has.
+            sig2def = { "f_v(" : 0, "f_f(" : 1, "f_fu(" : 2, "f_fi(" : 3, 
+                        "f_du(" : 4, "f_duf(" : 5, "f_di(" : 6 }
+            file = open(filename, 'r')
+            match = -1
+            for line in file:
+                # look for lines with 'export'...
+                if line.find("export") == -1:
+                    continue
+                # one of them should have a function with one of the
+                # declarations in sig2def
+                for pattern, ident in sig2def.items():
+                    if line.find(pattern) != -1:
+                        match = ident
+                        break
+            file.close()
+            if match == -1:
+                print "Fatal error: unable to find function signature in test %s" % filename
+                error_count += 1
+            else:
+                obj_name = "%s.o" % filename
+                exe_name = "%s.run" % filename
+                ispc_cmd = "ispc --woff %s -o %s --arch=%s --target=%s" % \
+                    (filename, obj_name, options.arch, options.target)
+                if options.no_opt:
+                    ispc_cmd += " -O0" 
+                if options.arch == 'x86':
+                    gcc_arch = '-m32'
+                else:
+                    gcc_arch = '-m64'
+                gcc_cmd = "g++ %s test_static.cpp -DTEST_SIG=%d %s.o -o %s" % \
+                    (gcc_arch, match, filename, exe_name)
+                if platform.system() == 'Darwin':
+                    gcc_cmd += ' -Wl,-no_pie'
+                if should_fail:
+                    gcc_cmd += " -DEXPECT_FAILURE"
+                    
+                # compile the ispc code, make the executable, and run it...
+                error_count += run_cmds([ispc_cmd, gcc_cmd, exe_name], filename, should_fail)
+
+                # clean up after running the test
+                try:
+                    os.unlink(exe_name)
+                    os.unlink(obj_name)
+                except:
+                    None
+        else:
+            # otherwise we'll use ispc_test + the LLVM JIT to run the test
+            bitcode_file = "%s.bc" % filename
+            compile_cmd = "ispc --woff --emit-llvm %s --target=%s -o %s" % \
+                (filename, options.target, bitcode_file)
+            if options.no_opt:
+                compile_cmd += " -O0"
+            test_cmd = "ispc_test %s" % bitcode_file
+
+            error_count += run_cmds([compile_cmd, test_cmd], filename, should_fail)
+
+            try:
+                os.unlink(bitcode_file)
+            except:
+                None
+
+        # If not for http://bugs.python.org/issue5261 on OSX, we'd like to do this:
+        #with finished_tests_counter_lock:
+            #update_progress(filename)
+        # but instead we do this...
+        finished_tests_mutex.lock(update_progress, filename)
+
+
+task_threads = []
+
+def sigint(signum, frame):
+    for t in task_threads:
+        t.terminate()
+    sys.exit(1)
+
+if __name__ == '__main__':
+    nthreads = multiprocessing.cpu_count()
+    total_tests = len(files)
+    print "Found %d CPUs. Running %d tests." % (nthreads, total_tests)
+
+    # put each of the test filenames into a queue
+    q = multiprocessing.Queue()
+    for fn in files:
+        q.put(fn)
+    for x in range(nthreads):
+        q.put('STOP')
+
+    # need to catch sigint so that we can terminate all of the tasks if
+    # we're interrupted
+    signal.signal(signal.SIGINT, sigint)
+
+    # launch jobs to run tests
+    for x in range(nthreads):
+        t = multiprocessing.Process(target=run_tasks_from_queue, args=(q,))
+        task_threads.append(t)
+        t.start()
+
+    # wait for them to all finish and then return the number that failed
+    # (i.e. return 0 if all is ok)
+    error_count = 0
+    for t in task_threads:
+        t.join()
+        error_count += t.exitcode
+    print
+    if error_count > 0:
+        print "%d / %d tests FAILED!" % (error_count, total_tests)
+    sys.exit(error_count)

run_tests.sh

@@ -23,6 +23,12 @@ EOF
     esac
 done
 
+ISPC_ARCH=x86-64
+if [[ $OS == "Windows_NT" ]]; then
+  ISPC_ARCH=x86
+fi
+ISPC_ARGS="--target=$target --arch=$ISPC_ARCH -O2 --woff"
+
 shift $(( $OPTIND - 1 ))
 if [[ "$1" > 0 ]]; then
     while [[ "$1" > 0 ]]; do
@@ -31,7 +37,7 @@ if [[ "$1" > 0 ]]; then
         echo Running test $i
 
         bc=${i%%ispc}bc
-        ispc -O2 $i -woff -o $bc --emit-llvm --target=$target
+        ispc $ISPC_ARGS $i -o $bc --emit-llvm
         if [[ $? != 0 ]]; then
             surprises=1
             echo Test $i FAILED ispc compile
@@ -55,7 +61,7 @@ else
         fi
         (( counter++ ))
         bc=${i%%ispc}bc
-        ispc -O2 $i -woff -o $bc --emit-llvm --target=$target
+        ispc $ISPC_ARGS $i -o $bc --emit-llvm 
         if [[ $? != 0 ]]; then
             surprises=1
             echo Test $i FAILED ispc compile

stdlib.ispc

@@ -315,6 +315,39 @@ static inline uniform int lanemask() {
     return __movmsk(__mask);
 }
 
+///////////////////////////////////////////////////////////////////////////
+// Prefetching
+
+#define PREFETCHES(NAME, TYPE)                                  \
+static inline void prefetch_l1(const reference TYPE ptr) {      \
+    __prefetch_read_1_##NAME##_refsconst(ptr);                  \
+}                                                               \
+static inline void prefetch_l2(const reference TYPE ptr) {      \
+    __prefetch_read_2_##NAME##_refsconst(ptr);                  \
+}                                                               \
+static inline void prefetch_l3(const reference TYPE ptr) {      \
+    __prefetch_read_3_##NAME##_refsconst(ptr);                  \
+}                                                               \
+ static inline void prefetch_nt(const reference TYPE ptr) {     \
+     __prefetch_read_nt_##NAME##_refsconst(ptr);                \
+}
+
+PREFETCHES(uniform_int8, uniform int8)
+PREFETCHES(uniform_int16, uniform int16)
+PREFETCHES(uniform_int32, uniform int32)
+PREFETCHES(uniform_int64, uniform int64)
+PREFETCHES(uniform_float, uniform float)
+PREFETCHES(uniform_double, uniform double)
+
+PREFETCHES(varying_int8, int8)
+PREFETCHES(varying_int16, int16)
+PREFETCHES(varying_int32, int32)
+PREFETCHES(varying_int64, int64)
+PREFETCHES(varying_float, float)
+PREFETCHES(varying_double, double)
+
+#undef PREFETCHES
+
 ///////////////////////////////////////////////////////////////////////////
 // Horizontal ops / reductions
 
@@ -336,7 +369,7 @@ static inline uniform float reduce_min(float v) {
 static inline uniform float reduce_max(float v) {
     // For the lanes where the mask is off, replace the given value with
     // negative infinity, so that it doesn't affect the result.
-    const uniform int iflt_neg_max = 0xff800000; // -infinity
+    const int iflt_neg_max = 0xff800000; // -infinity
     // Must use __floatbits_varying_int32, not floatbits(), since with the
     // latter the current mask enters into the returned result...
     return __reduce_max_float(__mask ? v : __floatbits_varying_int32(iflt_neg_max));
@@ -394,7 +427,7 @@ static inline uniform double reduce_min(double v) {
 }
 
 static inline uniform double reduce_max(double v) {
-    const uniform int64 iflt_neg_max = 0xfff0000000000000; // -infinity
+    const int64 iflt_neg_max = 0xfff0000000000000; // -infinity
     // Must use __doublebits_varying_int64, not doublebits(), since with the
     // latter the current mask enters into the returned result...
     return __reduce_max_double(__mask ? v : __doublebits_varying_int64(iflt_neg_max));
@@ -438,29 +471,110 @@ static inline uniform unsigned int64 reduce_max(unsigned int64 v) {
     return __reduce_max_uint64(__mask ? v : 0);
 }
 
+#define REDUCE_EQUAL(TYPE, FUNCTYPE, MASKTYPE)                     \
+static inline uniform bool reduce_equal(TYPE v) {                  \
+    uniform TYPE unusedValue;                                      \
+    return __reduce_equal_##FUNCTYPE(v, unusedValue, (MASKTYPE)__mask); \
+}                                                                  \
+static inline uniform bool reduce_equal(TYPE v, reference uniform TYPE value) { \
+    return __reduce_equal_##FUNCTYPE(v, value, (MASKTYPE)__mask);       \
+}
+
+REDUCE_EQUAL(int32, int32, int32)
+REDUCE_EQUAL(unsigned int32, int32, unsigned int32)
+REDUCE_EQUAL(float, float, int32)
+REDUCE_EQUAL(int64, int64, int32)
+REDUCE_EQUAL(unsigned int64, int64, unsigned int32)
+REDUCE_EQUAL(double, double, int32)
+
+static int32 exclusive_scan_add(int32 v) {
+    return __exclusive_scan_add_i32(v, (int32)__mask);
+}
+
+static unsigned int32 exclusive_scan_add(unsigned int32 v) {
+    return __exclusive_scan_add_i32(v, __mask);
+}
+
+static float exclusive_scan_add(float v) {
+    return __exclusive_scan_add_float(v, __mask);
+}
+
+static int64 exclusive_scan_add(int64 v) {
+    return __exclusive_scan_add_i64(v, (int32)__mask);
+}
+
+static unsigned int64 exclusive_scan_add(unsigned int64 v) {
+    return __exclusive_scan_add_i64(v, __mask);
+}
+
+static double exclusive_scan_add(double v) {
+    return __exclusive_scan_add_double(v, __mask);
+}
+
+static int32 exclusive_scan_and(int32 v) {
+    return __exclusive_scan_and_i32(v, (int32)__mask);
+}
+
+static unsigned int32 exclusive_scan_and(unsigned int32 v) {
+    return __exclusive_scan_and_i32(v, __mask);
+}
+
+static int64 exclusive_scan_and(int64 v) {
+    return __exclusive_scan_and_i64(v, (int32)__mask);
+}
+
+static unsigned int64 exclusive_scan_and(unsigned int64 v) {
+    return __exclusive_scan_and_i64(v, __mask);
+}
+
+static int32 exclusive_scan_or(int32 v) {
+    return __exclusive_scan_or_i32(v, (int32)__mask);
+}
+
+static unsigned int32 exclusive_scan_or(unsigned int32 v) {
+    return __exclusive_scan_or_i32(v, __mask);
+}
+
+static int64 exclusive_scan_or(int64 v) {
+    return __exclusive_scan_or_i64(v, (int32)__mask);
+}
+
+static unsigned int64 exclusive_scan_or(unsigned int64 v) {
+    return __exclusive_scan_or_i64(v, __mask);
+}
+
 ///////////////////////////////////////////////////////////////////////////
 // packed load, store
 
 static inline uniform int 
 packed_load_active(uniform unsigned int a[], uniform int start,
                    reference unsigned int vals) {
-    return __packed_load_active(a, start, vals, __mask);
+    return __packed_load_active(a, (unsigned int)start, vals,
+                                (unsigned int32)__mask);
 }
 
 static inline uniform int
 packed_store_active(uniform unsigned int a[], uniform int start,
                     unsigned int vals) {
-    return __packed_store_active(a, start, vals, __mask);
+    return __packed_store_active(a, (unsigned int)start, vals,
+                                 (unsigned int32)__mask);
 }
 
 static inline uniform int packed_load_active(uniform int a[], uniform int start,
                                              reference int vals) {
-    return __packed_load_active(a, start, vals, __mask);
+    return __packed_load_active(a, start, vals, (int32)__mask);
 }
 
 static inline uniform int packed_store_active(uniform int a[], uniform int start,
                                               int vals) {
-    return __packed_store_active(a, start, vals, __mask);
+    return __packed_store_active(a, start, vals, (int32)__mask);
+}
+
+///////////////////////////////////////////////////////////////////////////
+// System information
+
+static inline int num_cores() {
+    return __num_cores();
 }
 
 ///////////////////////////////////////////////////////////////////////////
@@ -470,73 +584,110 @@ static inline void memory_barrier() {
     __memory_barrier();
 }
 
-#define DEFINE_ATOMIC_OP(TA,TB,OPA,OPB)                                 \
+#define DEFINE_ATOMIC_OP(TA,TB,OPA,OPB,MASKTYPE)                        \
 static inline TA atomic_##OPA##_global(uniform reference TA ref, TA value) { \
     memory_barrier();                                                   \
-    TA ret = __atomic_##OPB##_##TB##_global(ref, value, __mask);  \
+    TA ret = __atomic_##OPB##_##TB##_global(ref, value, (MASKTYPE)__mask); \
+    memory_barrier();                                                   \
+    return ret;                                                         \
+}                                                                       \
+static inline uniform TA atomic_##OPA##_global(uniform reference TA ref, \
+                                               uniform TA value) {      \
+    memory_barrier();                                                   \
+    uniform TA ret = __atomic_##OPB##_uniform_##TB##_global(ref, value, (MASKTYPE)__mask); \
     memory_barrier();                                                   \
     return ret;                                                         \
 }
 
-DEFINE_ATOMIC_OP(int32,int32,add,add)
-DEFINE_ATOMIC_OP(int32,int32,subtract,sub)
-DEFINE_ATOMIC_OP(int32,int32,min,min)
-DEFINE_ATOMIC_OP(int32,int32,max,max)
-DEFINE_ATOMIC_OP(int32,int32,and,and)
-DEFINE_ATOMIC_OP(int32,int32,or,or)
-DEFINE_ATOMIC_OP(int32,int32,xor,xor)
-DEFINE_ATOMIC_OP(int32,int32,swap,swap)
+#define DEFINE_ATOMIC_MINMAX_OP(TA,TB,OPA,OPB, MASKTYPE)                \
+static inline TA atomic_##OPA##_global(uniform reference TA ref, TA value) { \
+    uniform TA oneval = reduce_##OPA(value);                            \
+    TA ret;                                                             \
+    if (lanemask() != 0) {                                              \
+        memory_barrier();                                               \
+        ret = __atomic_##OPB##_uniform_##TB##_global(ref, oneval, (MASKTYPE)__mask); \
+        memory_barrier();                                               \
+    }                                                                   \
+    return ret;                                                         \
+}                                                                       \
+static inline uniform TA atomic_##OPA##_global(uniform reference TA ref, \
+                                               uniform TA value) {      \
+    memory_barrier();                                                   \
+    uniform TA ret = __atomic_##OPB##_uniform_##TB##_global(ref, value, (MASKTYPE)__mask); \
+    memory_barrier();                                                   \
+    return ret;                                                         \
+}
+
+DEFINE_ATOMIC_OP(int32,int32,add,add,int32)
+DEFINE_ATOMIC_OP(int32,int32,subtract,sub,int32)
+DEFINE_ATOMIC_MINMAX_OP(int32,int32,min,min,int32)
+DEFINE_ATOMIC_MINMAX_OP(int32,int32,max,max,int32)
+DEFINE_ATOMIC_OP(int32,int32,and,and,int32)
+DEFINE_ATOMIC_OP(int32,int32,or,or,int32)
+DEFINE_ATOMIC_OP(int32,int32,xor,xor,int32)
+DEFINE_ATOMIC_OP(int32,int32,swap,swap,int32)
 
 // For everything but atomic min and max, we can use the same
 // implementations for unsigned as for signed.
-DEFINE_ATOMIC_OP(unsigned int32,int32,add,add)
-DEFINE_ATOMIC_OP(unsigned int32,int32,subtract,sub)
-DEFINE_ATOMIC_OP(unsigned int32,uint32,min,umin)
-DEFINE_ATOMIC_OP(unsigned int32,uint32,max,umax)
-DEFINE_ATOMIC_OP(unsigned int32,int32,and,and)
-DEFINE_ATOMIC_OP(unsigned int32,int32,or,or)
-DEFINE_ATOMIC_OP(unsigned int32,int32,xor,xor)
-DEFINE_ATOMIC_OP(unsigned int32,int32,swap,swap)
+DEFINE_ATOMIC_OP(unsigned int32,int32,add,add,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int32,int32,subtract,sub,unsigned int32)
+DEFINE_ATOMIC_MINMAX_OP(unsigned int32,uint32,min,umin,unsigned int32)
+DEFINE_ATOMIC_MINMAX_OP(unsigned int32,uint32,max,umax,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int32,int32,and,and,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int32,int32,or,or,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int32,int32,xor,xor,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int32,int32,swap,swap,unsigned int32)
 
-DEFINE_ATOMIC_OP(float,float,swap,swap)
+DEFINE_ATOMIC_OP(float,float,swap,swap,int32)
 
-DEFINE_ATOMIC_OP(int64,int64,add,add)
-DEFINE_ATOMIC_OP(int64,int64,subtract,sub)
-DEFINE_ATOMIC_OP(int64,int64,min,min)
-DEFINE_ATOMIC_OP(int64,int64,max,max)
-DEFINE_ATOMIC_OP(int64,int64,and,and)
-DEFINE_ATOMIC_OP(int64,int64,or,or)
-DEFINE_ATOMIC_OP(int64,int64,xor,xor)
-DEFINE_ATOMIC_OP(int64,int64,swap,swap)
+DEFINE_ATOMIC_OP(int64,int64,add,add,int32)
+DEFINE_ATOMIC_OP(int64,int64,subtract,sub,int32)
+DEFINE_ATOMIC_MINMAX_OP(int64,int64,min,min,int32)
+DEFINE_ATOMIC_MINMAX_OP(int64,int64,max,max,int32)
+DEFINE_ATOMIC_OP(int64,int64,and,and,int32)
+DEFINE_ATOMIC_OP(int64,int64,or,or,int32)
+DEFINE_ATOMIC_OP(int64,int64,xor,xor,int32)
+DEFINE_ATOMIC_OP(int64,int64,swap,swap,int32)
 
 // For everything but atomic min and max, we can use the same
 // implementations for unsigned as for signed.
-DEFINE_ATOMIC_OP(unsigned int64,int64,add,add)
-DEFINE_ATOMIC_OP(unsigned int64,int64,subtract,sub)
-DEFINE_ATOMIC_OP(unsigned int64,uint64,min,umin)
-DEFINE_ATOMIC_OP(unsigned int64,uint64,max,umax)
-DEFINE_ATOMIC_OP(unsigned int64,int64,and,and)
-DEFINE_ATOMIC_OP(unsigned int64,int64,or,or)
-DEFINE_ATOMIC_OP(unsigned int64,int64,xor,xor)
-DEFINE_ATOMIC_OP(unsigned int64,int64,swap,swap)
+DEFINE_ATOMIC_OP(unsigned int64,int64,add,add,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int64,int64,subtract,sub,unsigned int32)
+DEFINE_ATOMIC_MINMAX_OP(unsigned int64,uint64,min,umin,unsigned int32)
+DEFINE_ATOMIC_MINMAX_OP(unsigned int64,uint64,max,umax,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int64,int64,and,and,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int64,int64,or,or,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int64,int64,xor,xor,unsigned int32)
+DEFINE_ATOMIC_OP(unsigned int64,int64,swap,swap,unsigned int32)
 
-DEFINE_ATOMIC_OP(double,double,swap,swap)
+DEFINE_ATOMIC_OP(double,double,swap,swap,int32)
 
-#define ATOMIC_DECL_CMPXCHG(TA, TB)                                        \
+#undef DEFINE_ATOMIC_OP
+
+#define ATOMIC_DECL_CMPXCHG(TA, TB, MASKTYPE)                           \
 static inline TA atomic_compare_exchange_global(                           \
          uniform reference TA ref, TA oldval, TA newval) {                 \
     memory_barrier();                                                      \
-    TA ret = __atomic_compare_exchange_##TB##_global(ref, oldval, newval, __mask); \
+    TA ret = __atomic_compare_exchange_##TB##_global(ref, oldval, newval, (MASKTYPE)__mask); \
     memory_barrier();                                                      \
     return ret;                                                            \
+} \
+static inline uniform TA atomic_compare_exchange_global(               \
+         uniform reference TA ref, uniform TA oldval, uniform TA newval) {                 \
+    memory_barrier();                                                   \
+    uniform TA ret = __atomic_compare_exchange_uniform_##TB##_global(ref, oldval, newval, (MASKTYPE)__mask); \
+    memory_barrier();                                                   \
+    return ret;                                                         \
 }
 
-ATOMIC_DECL_CMPXCHG(int32, int32)
-ATOMIC_DECL_CMPXCHG(unsigned int32, int32)
-ATOMIC_DECL_CMPXCHG(float, float)
-ATOMIC_DECL_CMPXCHG(int64, int64)
-ATOMIC_DECL_CMPXCHG(unsigned int64, int64)
-ATOMIC_DECL_CMPXCHG(double, double)
+ATOMIC_DECL_CMPXCHG(int32, int32, int32)
+ATOMIC_DECL_CMPXCHG(unsigned int32, int32, unsigned int32)
+ATOMIC_DECL_CMPXCHG(float, float, int32)
+ATOMIC_DECL_CMPXCHG(int64, int64, int32)
+ATOMIC_DECL_CMPXCHG(unsigned int64, int64, unsigned int32)
+ATOMIC_DECL_CMPXCHG(double, double, int32)
+
+#undef ATOMIC_DECL_CMPXCHG
 
 ///////////////////////////////////////////////////////////////////////////
 // Floating-Point Math
@@ -2600,6 +2751,80 @@ static inline int16 float_to_half(float f) {
 }
 
 
+static inline uniform float half_to_float_fast(uniform unsigned int16 h) {
+    uniform unsigned int32 hs = h & (int32)0x8000u;  // Pick off sign bit
+    uniform unsigned int32 he = h & (int32)0x7C00u;  // Pick off exponent bits
+    uniform unsigned int32 hm = h & (int32)0x03FFu;  // Pick off mantissa bits
+
+    // sign
+    uniform unsigned int32 xs = ((unsigned int32) hs) << 16; 
+    // Exponent: unbias the halfp, then bias the single
+    uniform int32 xes = ((int32) (he >> 10)) - 15 + 127; 
+    // Exponent
+    uniform unsigned int32 xe = (unsigned int32) (xes << 23);
+    // Mantissa
+    uniform unsigned int32 xm = ((unsigned int32) hm) << 13; 
+    return floatbits(xs | xe | xm);
+
+}
+
+static inline float half_to_float_fast(unsigned int16 h) {
+    unsigned int32 hs = h & (int32)0x8000u;  // Pick off sign bit
+    unsigned int32 he = h & (int32)0x7C00u;  // Pick off exponent bits
+    unsigned int32 hm = h & (int32)0x03FFu;  // Pick off mantissa bits
+
+    // sign
+    unsigned int32 xs = ((unsigned int32) hs) << 16; 
+    // Exponent: unbias the halfp, then bias the single
+    int32 xes = ((int32) (he >> 10)) - 15 + 127; 
+    // Exponent
+    unsigned int32 xe = (unsigned int32) (xes << 23);
+    // Mantissa
+    unsigned int32 xm = ((unsigned int32) hm) << 13; 
+    return floatbits(xs | xe | xm);
+
+}
+
+static inline uniform int16 float_to_half_fast(uniform float f) {
+    uniform int32 x = intbits(f);
+    uniform unsigned int32 xs = x & 0x80000000u;  // Pick off sign bit
+    uniform unsigned int32 xe = x & 0x7F800000u;  // Pick off exponent bits
+    uniform unsigned int32 xm = x & 0x007FFFFFu;  // Pick off mantissa bits
+
+    uniform unsigned int32 hs = (xs >> 16); // Sign bit
+    // Exponent unbias the single, then bias the halfp
+    uniform int32 hes = ((int)(xe >> 23)) - 127 + 15; 
+    uniform unsigned int32 he = (hes << 10); // Exponent
+    uniform int32 hm = (xm >> 13); // Mantissa
+    uniform int32 ret = (hs | he | hm);
+
+    if (xm & 0x00001000u) // Check for rounding
+        // Round, might overflow to inf, this is OK
+        ret += 1u; 
+
+    return (int16)ret;
+}
+
+static inline int16 float_to_half_fast(float f) {
+    int32 x = intbits(f);
+    unsigned int32 xs = x & 0x80000000u;  // Pick off sign bit
+    unsigned int32 xe = x & 0x7F800000u;  // Pick off exponent bits
+    unsigned int32 xm = x & 0x007FFFFFu;  // Pick off mantissa bits
+
+    unsigned int32 hs = (xs >> 16); // Sign bit
+    // Exponent unbias the single, then bias the halfp
+    int32 hes = ((int)(xe >> 23)) - 127 + 15; 
+    unsigned int32 he = (hes << 10); // Exponent
+    int32 hm = (xm >> 13); // Mantissa
+    int32 ret = (hs | he | hm);
+
+    if (xm & 0x00001000u) // Check for rounding
+        // Round, might overflow to inf, this is OK
+        ret += 1u; 
+
+    return (int16)ret;
+}
+
 ///////////////////////////////////////////////////////////////////////////
 // RNG stuff
 
@@ -2624,7 +2849,9 @@ static inline unsigned int random(reference RNGState state)
 
 static inline float frandom(reference RNGState state)
 {
-    return ((int)(random(state) & ((1<<24)-1))) / (float)(1 << 24);
+    unsigned int irand = random(state);
+    irand &= (1<<23)-1;
+    return floatbits(0x3F800000 | irand)-1.0f;
 }
 
 static inline uniform unsigned int __seed4(reference RNGState state, 
@@ -2665,6 +2892,12 @@ static inline void seed_rng(reference uniform RNGState state, uniform unsigned i
     seed = __seed4(state, 0, seed);
     if (programCount == 8)
         __seed4(state, 4, seed ^ 0xbeeff00d);
+    if (programCount == 16) {
+        __seed4(state, 4,  seed ^ 0xbeeff00d);
+        __seed4(state, 8,  ((seed & 0xffff) << 16) | (seed >> 16));
+        __seed4(state, 12, (((seed & 0xff) << 24) | ((seed & 0xff00)  << 8) |
+                            ((seed & 0xff0000) >> 8) | (seed & 0xff000000) >> 24));
+    }
 }
 
 static inline void fastmath() {