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).

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-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.m4

@@ -622,40 +622,6 @@ forloop(i, 1, eval($1-1), `
 }
 ')
 
-;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
-;; global_atomic
-;; 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.
-;;
-;; Takes four 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)
-
-define(`global_atomic', `
-
-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
-}
-')
-
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; global_atomic_associative
@@ -681,18 +647,20 @@ define internal <$1 x $3> @__atomic_$2_$4_global($3 * %ptr, <$1 x $3> %val,
 
 define(`global_atomic_associative', `
 
-declare $3 @llvm.atomic.load.$2.$3.p0$3($3 * %ptr, $3 %delta)
-
-;; note that the mask is expected to be of type $3, so the caller must ensure
-;; that for 64-bit types, the mask is cast to a signed int before being passed
-;; to this so that it is properly sign extended...  (The code in stdlib.ispc
-;; does do this..)
-
 define internal <$1 x $3> @__atomic_$2_$4_global($3 * %ptr, <$1 x $3> %val,
-                                                 <$1 x $3> %mask) nounwind alwaysinline {
+                                                 <$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
 
@@ -751,7 +719,7 @@ define(`global_atomic_uniform', `
 
 declare $3 @llvm.atomic.load.$2.$3.p0$3($3 * %ptr, $3 %delta)
 
-define internal $3 @__atomic_$2_$4_global($3 * %ptr, $3 %val,
+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
@@ -764,9 +732,10 @@ define internal $3 @__atomic_$2_$4_global($3 * %ptr, $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 {
@@ -782,6 +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
+}
 ')
 
 
@@ -811,6 +786,12 @@ 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
+}
 ')
 
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
@@ -853,10 +834,9 @@ define(`stdlib_core', `
 
 declare i32 @__fast_masked_vload()
 
-declare i8* @ISPCMalloc(i64, i32) nounwind
-declare i8* @ISPCFree(i8*) nounwind
-declare void @ISPCLaunch(i8*, i8*) nounwind
-declare void @ISPCSync() nounwind
+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)
@@ -1228,6 +1208,11 @@ 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)
@@ -1238,6 +1223,11 @@ 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)
@@ -1264,6 +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)
 
@@ -1288,6 +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
+}
+
 ')
 
 

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 {
@@ -174,8 +179,6 @@ FunctionEmitContext::FunctionEmitContext(const Type *rt, llvm::Function *functio
         StartScope();
     }
 
-    launchedTasks = false;
-
     // connect the funciton's mask memory to the __mask symbol
     Symbol *maskSymbol = m->symbolTable->LookupVariable("__mask");
     assert(maskSymbol != NULL);
@@ -759,7 +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.
@@ -776,24 +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;
 }
 
 
@@ -1912,15 +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) {
@@ -1943,7 +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;
@@ -1960,29 +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);
-    llvm::Value *argmem;
-#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...
-    argmem = EmitMalloc(argStructType, align);
-#else
-    // Otherwise, use alloca for space for the task args, ** unless we're 
-    // compiling to AVX, in which case we use malloc after all **. (See
-    // http://llvm.org/bugs/show_bug.cgi?id=10841 for details.  There are
-    // limitations in LLVM with respect to dynamic allocas of this sort
-    // when the stack also has to be 32-byte aligned...).
-    if (g->target.isa == Target::AVX)
-        argmem = EmitMalloc(argStructType, align);
-    else
-        // 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.
-        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
@@ -2004,5 +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 +1,38 @@
+=== 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

docs/ispc.txt

@@ -80,7 +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`_
 
@@ -91,6 +92,7 @@ Contents:
   + `Conversions To and From Half-Precision Floats`_
   + `Atomic Operations and Memory Fences`_
   + `Prefetches`_
+  + `System Information`_
   + `Low-Level Bits`_
 
 * `Interoperability with the Application`_
@@ -837,8 +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
@@ -939,8 +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
 --------------------------------
@@ -1383,112 +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
 =========================
@@ -2033,12 +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
@@ -2056,17 +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
@@ -2115,6 +2223,20 @@ 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
 --------------
 

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         = 1.0.8
+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
 ==
 
@@ -87,6 +126,7 @@ 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
 ======
 
@@ -94,6 +134,7 @@ 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
 ======
 

examples/aobench/Makefile

@@ -1,14 +1,8 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=../tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
-
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=../tasks_gcd.cpp
-  TASK_LIB=
-endif
-
 TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
 CXX=g++

examples/aobench/ao.ispc

@@ -323,16 +323,13 @@ export void ao_ispc(uniform int w, uniform int h, uniform int nsubsamples,
 }
 
 
-static void task ao_task(uniform int y0, uniform int y1, uniform int width,
-                         uniform int height, uniform int nsubsamples, 
-                         uniform float image[]) {
-    ao_scanlines(y0, y1, width, height, 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[]) {
-    uniform int dy = 1;
-    for (uniform int y = 0; y < h; y += dy)
-        launch < ao_task(y, y+dy, w, h, nsubsamples, image) >;
+    launch[h] < ao_task(w, h, nsubsamples, image) >;
 }

examples/aobench/aobench.vcxproj

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

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

@@ -18,8 +18,11 @@ 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
@@ -108,6 +111,14 @@ Global
 		{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_tasks/Makefile

@@ -1,14 +1,8 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=../tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
-
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=../tasks_gcd.cpp
-  TASK_LIB=
-endif
-
 TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
 CXX=g++

examples/mandelbrot_tasks/mandelbrot.cpp

@@ -101,7 +101,7 @@ ensureTargetISAIsSupported() {
 }
 
 static void usage() {
-    fprintf(stderr, "usage: mandelbrot [--scale=<factor]\n");
+    fprintf(stderr, "usage: mandelbrot [--scale=<factor>]\n");
     exit(1);
 }
 
@@ -143,6 +143,9 @@ int main(int argc, char *argv[]) {
     //
     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();
@@ -152,9 +155,6 @@ int main(int argc, char *argv[]) {
     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
@@ -162,6 +162,9 @@ int main(int argc, char *argv[]) {
     //
     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_tasks.vcxproj

@@ -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/noise/noise.vcxproj

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

examples/rt/Makefile

@@ -1,14 +1,8 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=../tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
-
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=../tasks_gcd.cpp
-  TASK_LIB=
-endif
-
 TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
 CXX=g++

examples/rt/rt.ispc

@@ -283,8 +283,7 @@ export void raytrace_ispc(uniform int width, uniform int height,
 }
 
 
-task void raytrace_tile_task(uniform int x0, uniform int x1,
-                             uniform int y0, uniform int y1, 
+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], 
@@ -292,6 +291,12 @@ task void raytrace_tile_task(uniform int x0, uniform int x1,
                              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);
@@ -306,13 +311,11 @@ export void raytrace_ispc_tasks(uniform int width, uniform int height,
                                 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);
-        for (uniform int x = 0; x < width; x += dx) {
-            uniform int x1 = min(x + dx, width);
-            launch < raytrace_tile_task(x, x1, y, y1, width, height, baseWidth,
-                                        baseHeight, raster2camera, camera2world, 
-                                        image, id, nodes, triangles) >;
-         }
+        launch[nTasks] < raytrace_tile_task(y, y1, width, height, baseWidth,
+                                            baseHeight, raster2camera, camera2world, 
+                                            image, id, nodes, triangles) >;
     }
 }

examples/rt/rt.vcxproj

@@ -164,7 +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="../tasks_concrt.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

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/Makefile

@@ -1,14 +1,8 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=../tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
-
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=../tasks_gcd.cpp
-  TASK_LIB=
-endif
-
 TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
 CXX=g++

examples/stencil/stencil.vcxproj

@@ -164,7 +164,7 @@ ispc -O2 %(Filename).ispc -o %(Filename).obj -h %(Filename)_ispc.h
   <ItemGroup>
     <ClCompile Include="stencil.cpp" />
     <ClCompile Include="stencil_serial.cpp" />
-    <ClCompile Include="../tasks_concrt.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

examples/taskinfo.h

@@ -1,180 +0,0 @@
-/*
-  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 TASKINFO_H
-#define TASKINFO_H 1
-
-#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_WINDOWS
-#define NOMINMAX
-#include <windows.h>
-#include <concrt.h>
-using namespace Concurrency;
-#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__
-
-#include <stdint.h>
-#include <stdlib.h>
-#include <stdio.h>
-#include <assert.h>
-
-typedef struct TaskInfo {
-    void *func;
-    void *data;
-#if defined(ISPC_IS_WINDOWS)
-    event taskEvent;
-#endif
-} TaskInfo;
-
-
-#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");
-    __asm__ __volatile__("mfence":::"memory");
-    return result;
-}
-#endif // !ISPC_IS_WINDOWS
-
-
-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
-    __asm__ __volatile__("mfence":::"memory");
-    return result;
-#endif // ISPC_IS_WINDOWS
-}
-
-
-#ifndef ISPC_IS_WINDOWS
-static int32_t 
-lAtomicAdd32(volatile int32_t *v, int32_t delta) {
-    // Do atomic add with gcc x86 inline assembly
-    int32_t origValue;
-    __asm__ __volatile__("lock\n"
-                         "xaddl %0,%1"
-                         : "=r"(origValue), "=m"(*v) : "0"(delta)
-                         : "memory");
-    return origValue;
-}
-#endif
-
-#define LOG_TASK_QUEUE_CHUNK_SIZE 13
-#define MAX_TASK_QUEUE_CHUNKS 1024
-#define TASK_QUEUE_CHUNK_SIZE (1<<LOG_TASK_QUEUE_CHUNK_SIZE)
-
-#define MAX_LAUNCHED_TASKS (MAX_TASK_QUEUE_CHUNKS * TASK_QUEUE_CHUNK_SIZE)
-
-typedef void (*TaskFuncType)(void *, int, int);
-
-#ifdef ISPC_IS_WINDOWS
-static volatile LONG nextTaskInfoCoordinate;
-#else
-static volatile int nextTaskInfoCoordinate;
-#endif
-
-static TaskInfo *taskInfo[MAX_TASK_QUEUE_CHUNKS];
-
-static inline void
-lInitTaskInfo() {
-    taskInfo[0] = new TaskInfo[TASK_QUEUE_CHUNK_SIZE];
-}
-
-
-static inline TaskInfo *
-lGetTaskInfo() {
-#ifdef ISPC_IS_WINDOWS
-    int myCoord = InterlockedAdd(&nextTaskInfoCoordinate, 1)-1;
-#else
-    int myCoord = lAtomicAdd32(&nextTaskInfoCoordinate, 1);
-#endif
-	int index = (myCoord >> LOG_TASK_QUEUE_CHUNK_SIZE);
-    int offset = myCoord & (TASK_QUEUE_CHUNK_SIZE-1);
-    if (index == MAX_TASK_QUEUE_CHUNKS) {
-        fprintf(stderr, "A total of %d tasks have been launched--the simple "
-                "built-in task system can handle no more. Exiting.", myCoord);
-        exit(1);
-    }
-
-    if (taskInfo[index] == NULL) {
-        TaskInfo *newChunk = new TaskInfo[TASK_QUEUE_CHUNK_SIZE];
-        if (lAtomicCompareAndSwapPointer((void **)&taskInfo[index], newChunk, 
-                                         NULL) != NULL) {
-            // failure--someone else got it, but that's cool
-            assert(taskInfo[index] != NULL);
-            free(newChunk);
-        }
-    }
-
-    return &taskInfo[index][offset];
-}
-
-
-static inline void
-lResetTaskInfo() {
-    nextTaskInfoCoordinate = 0;
-}
-
-#endif // TASKINFO_H

examples/tasks_concrt.cpp

@@ -1,104 +0,0 @@
-/*
-  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 "taskinfo.h"
-
-/* 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>
-#include <algorithm>
-
-// 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);
-}
-
-
-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;
-    TaskFuncType func = (TaskFuncType)ti->func;
-    func(ti->data, threadIndex, threadCount);
-
-    // Signal the event that this task is done
-    ti->taskEvent.set();
-}
-
-
-void
-ISPCLaunch(void *func, void *data) {
-    TaskInfo *ti = lGetTaskInfo();
-    ti->func = (TaskFuncType)func;
-    ti->data = data;
-	ti->taskEvent.reset();
-    CurrentScheduler::ScheduleTask(lRunTask, ti);
-}
-
-
-void ISPCSync() {
-    for (int i = 0; i < nextTaskInfoCoordinate; ++i) {
-		int index = (i >> LOG_TASK_QUEUE_CHUNK_SIZE);
-		int offset = i & (TASK_QUEUE_CHUNK_SIZE-1);
-		taskInfo[index][offset].taskEvent.wait();
-		taskInfo[index][offset].taskEvent.reset();
-    }
-
-    lResetTaskInfo();
-}
-
-
-void *ISPCMalloc(int64_t size, int32_t alignment) {
-    return _aligned_malloc(size, alignment);
-}
-
-
-void ISPCFree(void *ptr) {
-    _aligned_free(ptr);
-}

examples/tasks_gcd.cpp

@@ -1,126 +0,0 @@
-/*
-  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 "taskinfo.h"
-
-#if defined(_WIN32) || defined(_WIN64)
-#define ISPC_IS_WINDOWS
-#elif defined(__linux__)
-#define ISPC_IS_LINUX
-#elif defined(__APPLE__)
-#define ISPC_IS_APPLE
-#endif
-
-/* A simple task system for ispc programs based on Apple's Grand Central
-   Dispatch. */
-#include <dispatch/dispatch.h>
-#include <stdio.h>
-#include <stdint.h>
-#include <stdlib.h>
-
-static int initialized = 0;
-static volatile int32_t lock = 0;
-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();
-    void *ISPCMalloc(int64_t size, int32_t alignment);
-    void ISPCFree(void *ptr);
-}
-
-
-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;
-    TaskFuncType func = (TaskFuncType)(taskInfo->func);
-
-    // Actually run the task
-    func(taskInfo->data, threadIndex, threadCount);
-}
-
-
-void ISPCLaunch(void *func, void *data) {
-    if (!initialized) {
-        while (1) {
-            if (lAtomicCompareAndSwap32(&lock, 1, 0) == 0) {
-                if (!initialized) {
-                    gcdQueue = dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_DEFAULT, 0);
-                    gcdGroup = dispatch_group_create();
-                    lInitTaskInfo();
-                    __asm__ __volatile__("mfence":::"memory");
-                    initialized = 1;
-                }
-                lock = 0;
-                break;
-            }
-        }
-    }
-
-    TaskInfo *ti = lGetTaskInfo();
-    ti->func = func;
-    ti->data = data;
-    dispatch_group_async_f(gcdGroup, gcdQueue, ti, lRunTask);
-}
-
-
-void ISPCSync() {
-    if (!initialized)
-        return;
-
-    // Wait for all of the tasks in the group to complete before returning
-    dispatch_group_wait(gcdGroup, DISPATCH_TIME_FOREVER);
-
-    lResetTaskInfo();
-}
-
-void *ISPCMalloc(int64_t size, int32_t alignment) {
-    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;
-}
-
-
-void ISPCFree(void *ptr) {
-    free(((void**)ptr)[-1]);
-}
-

examples/tasks_pthreads.cpp

@@ -1,339 +0,0 @@
-/*
-  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.  
-*/
-
-#if defined(_WIN32) || defined(_WIN64)
-#define ISPC_IS_WINDOWS
-#elif defined(__linux__)
-#define ISPC_IS_LINUX
-#elif defined(__APPLE__)
-#define ISPC_IS_APPLE
-#endif
-
-#include "taskinfo.h"
-#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>
-#ifdef ISPC_IS_LINUX
-#include <malloc.h>
-#endif
-
-static int initialized = 0;
-static volatile int32_t lock = 0;
-
-static int nThreads;
-static pthread_t *threads;
-static pthread_mutex_t taskQueueMutex;
-static int nextTaskToRun;
-static sem_t *workerSemaphore;
-static uint32_t numUnfinishedTasks;
-static pthread_mutex_t tasksRunningConditionMutex;
-static pthread_cond_t tasksRunningCondition;
-
-// 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);
-}
-
-static void *lTaskEntry(void *arg);
-
-/** Figure out how many CPU cores there are in the system
- */
-static int
-lNumCPUCores() {
-    return sysconf(_SC_NPROCESSORS_ONLN);
-}
-
-
-static void
-lTasksInit() {
-    nThreads = lNumCPUCores();
-
-    threads = (pthread_t *)malloc(nThreads * sizeof(pthread_t));
-
-    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, "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);
-    }
-
-    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, (void *)(i));
-        if (err != 0) {
-            fprintf(stderr, "Error creating pthread %d: %s\n", i, strerror(err));
-            exit(1);
-        }
-    }
-}
-
-
-void
-ISPCLaunch(void *f, void *d) {
-    int err;
-
-    if (!initialized) {
-        while (1) {
-            if (lAtomicCompareAndSwap32(&lock, 1, 0) == 0) {
-                if (!initialized) {
-                    lTasksInit();
-                    __asm__ __volatile__("mfence":::"memory");
-                    initialized = 1;
-                }
-                lock = 0;
-                break;
-            }
-        }
-    }
-
-    //
-    // Acquire mutex, add task
-    //
-    if ((err = pthread_mutex_lock(&taskQueueMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    // Need a mutex here to ensure we get this filled in before a worker
-    // grabs it and starts running...
-    TaskInfo *ti = lGetTaskInfo();
-    ti->func = f;
-    ti->data = 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);
-    }
-
-    // FIXME: is this redundant with nextTaskInfoCoordinate?
-    ++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)((int64_t)arg);
-    int threadCount = nThreads;
-    TaskFuncType func;
-
-    while (1) {
-        int err;
-        if ((err = sem_wait(workerSemaphore)) != 0) {
-            fprintf(stderr, "Error from sem_wait: %s\n", strerror(err));
-            exit(1);
-        }
-
-        //
-        // 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 (nextTaskToRun == nextTaskInfoCoordinate) {
-            //
-            // 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;
-        }
-
-        int runCoord = nextTaskToRun++;
-        int index = (runCoord >> LOG_TASK_QUEUE_CHUNK_SIZE);
-        int offset = runCoord & (TASK_QUEUE_CHUNK_SIZE-1);
-        TaskInfo *myTask = &taskInfo[index][offset];
-
-        if ((err = pthread_mutex_unlock(&taskQueueMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_mutex_unlock: %s\n", strerror(err));
-            exit(1);
-        }
-
-        //
-        // Do work for _myTask_
-        //
-        func = (TaskFuncType)myTask->func;
-        func(myTask->data, 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);
-        }
-
-        // FIXME: can this be a comparison of (nextTaskToRun == nextTaskInfoCoordinate)?
-        // (I don't think so--think there is a race...)
-        int unfinished = --numUnfinishedTasks;
-        if (unfinished == 0) {
-            //
-            // Signal the "no more tasks are running" condition if all of
-            // them are done.
-            //
-            int err;
-            if ((err = pthread_cond_signal(&tasksRunningCondition)) != 0) {
-                fprintf(stderr, "Error from pthread_cond_signal: %s\n", strerror(err));
-                exit(1);
-            }
-        }
-
-        if ((err = pthread_mutex_unlock(&tasksRunningConditionMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-            exit(1);
-        }
-    }
-
-    pthread_exit(NULL);
-    return 0;
-}
-
-
-void ISPCSync() {
-    int err;
-    if ((err = pthread_mutex_lock(&tasksRunningConditionMutex)) != 0) {
-        fprintf(stderr, "Error from pthread_mutex_lock: %s\n", strerror(err));
-        exit(1);
-    }
-
-    // As long as there are tasks running, wait on the condition variable;
-    // doing so causes this thread to go to sleep until someone signals on
-    // the tasksRunningCondition condition variable.
-    while (numUnfinishedTasks > 0) {
-        if ((err = pthread_cond_wait(&tasksRunningCondition, 
-                                     &tasksRunningConditionMutex)) != 0) {
-            fprintf(stderr, "Error from pthread_cond_wait: %s\n", strerror(err));
-            exit(1);
-        }
-    }
-    
-    lResetTaskInfo();
-    nextTaskToRun = 0;
-
-    // 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);
-    }
-}
-
-
-void *ISPCMalloc(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
-}
-
-
-void ISPCFree(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
-}
-

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/volume_rendering/Makefile

@@ -1,14 +1,8 @@
 
 ARCH = $(shell uname)
 
-TASK_CXX=../tasks_pthreads.cpp
+TASK_CXX=../tasksys.cpp
 TASK_LIB=-lpthread
-
-ifeq ($(ARCH), Darwin)
-  TASK_CXX=../tasks_gcd.cpp
-  TASK_LIB=
-endif
-
 TASK_OBJ=$(addprefix objs/, $(subst ../,, $(TASK_CXX:.cpp=.o)))
 
 CXX=g++

examples/volume_rendering/volume.ispc

@@ -343,11 +343,20 @@ volume_tile(uniform int x0, uniform int y0, uniform int x1,
 
 
 task void
-volume_task(uniform int x0, uniform int y0, uniform int x1,
-            uniform int y1, uniform float density[], uniform int nVoxels[3], 
+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);
 }
@@ -370,9 +379,7 @@ volume_ispc_tasks(uniform float density[], uniform int nVoxels[3],
                   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;
-    for (uniform int y = 0; y < height; y += dy)
-        for (uniform int x = 0; x < width; x += dx)
-            launch < volume_task(x, y, x+dx, y+dy, density, nVoxels, 
-                                 raster2camera, camera2world, width, height, 
-                                 image) >;
+    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

@@ -143,7 +143,7 @@
   <ItemGroup>
     <ClCompile Include="volume.cpp" />
     <ClCompile Include="volume_serial.cpp" />
-    <ClCompile Include="../tasks_concrt.cpp" />
+    <ClCompile Include="../tasksys.cpp" />
   </ItemGroup>
   <ItemGroup>
     <CustomBuild Include="volume.ispc">

expr.cpp

@@ -1189,10 +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;
@@ -1305,6 +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:
@@ -2202,7 +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
@@ -2216,44 +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();
+    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);
 }
 
 
@@ -2379,8 +2401,12 @@ FunctionCallExpr::GetValue(FunctionEmitContext *ctx) const {
 
     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 intrinsics, builtins, and extern
@@ -2456,10 +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.");
@@ -4103,7 +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
@@ -4163,9 +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;
@@ -5193,9 +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;
 }
@@ -5261,14 +5297,8 @@ 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;
-    }
-
-    return ctx->CallInst(fsync, noArg, "");
+    ctx->SyncInst();
+    return NULL;
 }
 
 

expr.h

@@ -250,7 +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;
@@ -263,9 +264,10 @@ public:
     Expr *func;
     ExprList *args;
     bool isLaunch;
+    Expr *launchCountExpr;
 
 private:
-    void resolveFunctionOverloads();
+    void resolveFunctionOverloads(bool exactMatchOnly);
     bool tryResolve(bool (*matchFunc)(Expr *, const Type *));
 };
 
@@ -567,7 +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;
@@ -581,6 +583,9 @@ public:
 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. */

ispc.cpp

@@ -241,7 +241,9 @@ Target::GetTargetMachine() const {
     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);
 

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">
@@ -31,12 +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 -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</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 -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</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" />
@@ -61,9 +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>
@@ -194,7 +196,7 @@
       <PrecompiledHeader>NotUsing</PrecompiledHeader>
       <WarningLevel>Level3</WarningLevel>
       <Optimization>Disabled</Optimization>
-      <PreprocessorDefinitions>NOMINMAX;LLVM_2_9</PreprocessorDefinitions>
+      <PreprocessorDefinitions>NOMINMAX;LLVM_3_0</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>$(LLVM_INSTALL_DIR)\include;.;.\winstuff;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4146;4800;4996;4355;4624</DisableSpecificWarnings>
     </ClCompile>
@@ -202,7 +204,7 @@
       <SubSystem>Console</SubSystem>
       <GenerateDebugInformation>true</GenerateDebugInformation>
       <AdditionalLibraryDirectories>$(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'">
@@ -212,7 +214,7 @@
       <Optimization>MaxSpeed</Optimization>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <IntrinsicFunctions>true</IntrinsicFunctions>
-      <PreprocessorDefinitions>NOMINMAX;LLVM_2_9</PreprocessorDefinitions>
+      <PreprocessorDefinitions>NOMINMAX;LLVM_3_0</PreprocessorDefinitions>
       <AdditionalIncludeDirectories>$(LLVM_INSTALL_DIR)\include;.;.\winstuff;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <DisableSpecificWarnings>4146;4800;4996;4355;4624</DisableSpecificWarnings>
     </ClCompile>
@@ -222,10 +224,10 @@
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
       <OptimizeReferences>true</OptimizeReferences>
       <AdditionalLibraryDirectories>$(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

@@ -75,7 +75,6 @@ extern "C" {
 #include <llvm/Instructions.h>
 #include <llvm/ExecutionEngine/ExecutionEngine.h>
 #if defined(LLVM_3_0) || defined(LLVM_3_0svn)
-  #include <llvm/ExecutionEngine/MCJIT.h>
   #include <llvm/Support/TargetRegistry.h>
   #include <llvm/Support/TargetSelect.h>
 #else
@@ -99,24 +98,27 @@ extern "C" {
 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 *ISPCMalloc(int64_t size, int32_t alignment) {
+void *ISPCAlloc(void **handle, int64_t size, int32_t alignment) {
+    *handle = (void *)0xdeadbeef;
+    // leak time!
 #ifdef ISPC_IS_WINDOWS
     return _aligned_malloc(size, alignment);
 #endif
@@ -134,18 +136,6 @@ void *ISPCMalloc(int64_t size, int32_t alignment) {
 }
 
 
-void ISPCFree(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 usage(int ret) {
     fprintf(stderr, "usage: ispc_test\n");
     fprintf(stderr, "\t[-h/--help]\tprint help\n");
@@ -218,8 +208,7 @@ static bool lRunTest(const char *fn) {
         ee->addGlobalMapping(func, (void *)FUNC)
     DO_FUNC(ISPCLaunch, "ISPCLaunch");
     DO_FUNC(ISPCSync, "ISPCSync");
-    DO_FUNC(ISPCMalloc, "ISPCMalloc");
-    DO_FUNC(ISPCFree, "ISPCFree");
+    DO_FUNC(ISPCAlloc, "ISPCAlloc");
     DO_FUNC(putchar, "putchar");
     DO_FUNC(printf, "printf");
     DO_FUNC(fflush, "fflush");

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

@@ -96,7 +96,9 @@ static void usage(int ret) {
     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");
 #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");
@@ -302,8 +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);

module.cpp

@@ -78,6 +78,7 @@
 #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>
 #include <llvm/Support/ToolOutputFile.h>
@@ -626,6 +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
@@ -653,18 +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);
-#else
-        // We also do this for AVX... (See discussion in
-        // FunctionEmitContext::LaunchInst().)
-        if (g->target.isa == Target::AVX)
-            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
@@ -1388,23 +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");
@@ -1417,7 +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();
 }
-

opt.cpp

@@ -2568,7 +2568,7 @@ llvm::RegisterPass<MakeInternalFuncsStaticPass>
 bool
 MakeInternalFuncsStaticPass::runOnModule(llvm::Module &module) {
     const char *names[] = {
-        "__do_print", "__fast_masked_vload",
+        "__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

@@ -17,6 +17,7 @@ import random
 import string
 import mutex
 import subprocess
+import platform
 
 parser = OptionParser()
 parser.add_option("-r", "--random-shuffle", dest="random", help="Randomly order tests",
@@ -137,8 +138,10 @@ def run_tasks_from_queue(queue):
                     gcc_arch = '-m32'
                 else:
                     gcc_arch = '-m64'
-                gcc_cmd = "g++ -Wl,-no_pie %s test_static.cpp -DTEST_SIG=%d %s.o -o %s" % \
+                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"
                     

stdlib.ispc

@@ -369,7 +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));
@@ -427,7 +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));
@@ -471,21 +471,21 @@ static inline uniform unsigned int64 reduce_max(unsigned int64 v) {
     return __reduce_max_uint64(__mask ? v : 0);
 }
 
-#define REDUCE_EQUAL(TYPE, FUNCTYPE)                               \
+#define REDUCE_EQUAL(TYPE, FUNCTYPE, MASKTYPE)                     \
 static inline uniform bool reduce_equal(TYPE v) {                  \
     uniform TYPE unusedValue;                                      \
-    return __reduce_equal_##FUNCTYPE(v, unusedValue, (int32)__mask); \
+    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, (int32)__mask);       \
+    return __reduce_equal_##FUNCTYPE(v, value, (MASKTYPE)__mask);       \
 }
 
-REDUCE_EQUAL(int32, int32)
-REDUCE_EQUAL(unsigned int32, int32)
-REDUCE_EQUAL(float, float)
-REDUCE_EQUAL(int64, int64)
-REDUCE_EQUAL(unsigned int64, int64)
-REDUCE_EQUAL(double, double)
+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);
@@ -549,23 +549,32 @@ static unsigned int64 exclusive_scan_or(unsigned int64 v) {
 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();
 }
 
 ///////////////////////////////////////////////////////////////////////////
@@ -581,24 +590,38 @@ static inline TA atomic_##OPA##_global(uniform reference TA ref, TA value) { \
     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 DEFINE_ATOMIC_MINMAX_OP(TA,TB,OPA,OPB)                          \
+#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##_##TB##_global(ref, oneval, __mask);      \
+        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)
-DEFINE_ATOMIC_MINMAX_OP(int32,int32,max,max)
+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)
@@ -606,56 +629,63 @@ 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,int32)
-DEFINE_ATOMIC_OP(unsigned int32,int32,subtract,sub,int32)
-DEFINE_ATOMIC_MINMAX_OP(unsigned int32,uint32,min,umin)
-DEFINE_ATOMIC_MINMAX_OP(unsigned int32,uint32,max,umax)
-DEFINE_ATOMIC_OP(unsigned int32,int32,and,and,int32)
-DEFINE_ATOMIC_OP(unsigned int32,int32,or,or,int32)
-DEFINE_ATOMIC_OP(unsigned int32,int32,xor,xor,int32)
-DEFINE_ATOMIC_OP(unsigned int32,int32,swap,swap,int32)
+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,int32)
 
-DEFINE_ATOMIC_OP(int64,int64,add,add,int64)
-DEFINE_ATOMIC_OP(int64,int64,subtract,sub,int64)
-DEFINE_ATOMIC_MINMAX_OP(int64,int64,min,min)
-DEFINE_ATOMIC_MINMAX_OP(int64,int64,max,max)
-DEFINE_ATOMIC_OP(int64,int64,and,and,int64)
-DEFINE_ATOMIC_OP(int64,int64,or,or,int64)
-DEFINE_ATOMIC_OP(int64,int64,xor,xor,int64)
+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,int64)
-DEFINE_ATOMIC_OP(unsigned int64,int64,subtract,sub,int64)
-DEFINE_ATOMIC_MINMAX_OP(unsigned int64,uint64,min,umin)
-DEFINE_ATOMIC_MINMAX_OP(unsigned int64,uint64,max,umax)
-DEFINE_ATOMIC_OP(unsigned int64,int64,and,and,int64)
-DEFINE_ATOMIC_OP(unsigned int64,int64,or,or,int64)
-DEFINE_ATOMIC_OP(unsigned int64,int64,xor,xor,int64)
-DEFINE_ATOMIC_OP(unsigned int64,int64,swap,swap,int32)
+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,int32)
 
 #undef DEFINE_ATOMIC_OP
 
-#define ATOMIC_DECL_CMPXCHG(TA, TB)                                        \
+#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
 

test_static.cpp

@@ -58,23 +58,26 @@ extern "C" {
     extern void f_di(float *result, double *a, int *b);
     extern void result(float *val);
     
-    void ISPCLaunch(void *f, void *d);
-    void ISPCSync();
-    void *ISPCMalloc(int64_t size, int32_t alignment);
-    void ISPCFree(void *ptr);
+    void ISPCLaunch(void **handlePtr, void *f, void *d, int);
+    void ISPCSync(void *handle);
+    void *ISPCAlloc(void **handlePtr, int64_t size, int32_t alignment);
 }
 
-void ISPCLaunch(void *f, void *d) {
-    typedef void (*TaskFuncType)(void *, int, int);
+void ISPCLaunch(void **handle, void *f, void *d, int count) {
+    *handle = (void *)0xdeadbeef;
+    typedef void (*TaskFuncType)(void *, int, int, int, int);
     TaskFuncType func = (TaskFuncType)f;
-    func(d, 0, 1);
+    for (int i = 0; i < count; ++i)
+        func(d, 0, 1, i, count);
 }
 
-void ISPCSync() {
+void ISPCSync(void *) {
 }
 
 
-void *ISPCMalloc(int64_t size, int32_t alignment) {
+void *ISPCAlloc(void **handle, int64_t size, int32_t alignment) {
+    *handle = (void *)0xdeadbeef;
+    // and now, we leak...
 #ifdef ISPC_IS_WINDOWS
     return _aligned_malloc(size, alignment);
 #endif
@@ -92,18 +95,6 @@ void *ISPCMalloc(int64_t size, int32_t alignment) {
 }
 
 
-void ISPCFree(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
-}
-
 
 int main(int argc, char *argv[]) {
     int w = width();

tests/atomics-1.ispc

@@ -5,7 +5,8 @@ uniform unsigned int32 s = 0;
 
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     float a = aFOO[programIndex]; 
-    float b = atomic_add_global(s, 1);
+    float delta = 1;
+    float b = atomic_add_global(s, delta);
     RET[programIndex] = reduce_add(b);
 }
 

tests/atomics-10.ispc

@@ -6,8 +6,9 @@ uniform unsigned int32 s = 0;
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     float a = aFOO[programIndex]; 
     float b = 0;
+    float delta = 1;
     if (programIndex < 2)
-        b = atomic_add_global(s, 1);
+        b = atomic_add_global(s, delta);
     RET[programIndex] = s;
 }
 

tests/atomics-2.ispc

@@ -5,7 +5,8 @@ uniform int64 s = 0;
 
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     float a = aFOO[programIndex]; 
-    float b = atomic_add_global(s, 1);
+    float delta = 1;
+    float b = atomic_add_global(s, delta);
     RET[programIndex] = reduce_add(b);
 }
 

tests/atomics-3.ispc

@@ -5,7 +5,8 @@ uniform int32 s = 0xff;
 
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     float a = aFOO[programIndex]; 
-    float b = atomic_xor_global(s, 0xfffffff0);
+    int32 bits = 0xfffffff0;
+    float b = atomic_xor_global(s, bits);
     RET[programIndex] = s;
 }
 

tests/atomics-9.ispc

@@ -6,8 +6,9 @@ uniform unsigned int32 s = 0;
 export void f_f(uniform float RET[], uniform float aFOO[]) {
     float a = aFOO[programIndex]; 
     float b = 0;
+    int32 delta = 1;
     if (programIndex < 2)
-        b = atomic_add_global(s, 1);
+        b = atomic_add_global(s, delta);
     RET[programIndex] = reduce_add(b);
 }
 

tests/atomics-uniform-1.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform unsigned int32 s = 10;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform unsigned int32 b = atomic_add_global(s, 1);
+    RET[programIndex] = s;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 11;
+}

tests/atomics-uniform-2.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform unsigned int32 s = 0b1010;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform unsigned int32 b = atomic_or_global(s, 1);
+    RET[programIndex] = s;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 0b1011;
+}

tests/atomics-uniform-3.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform unsigned int32 s = 0b1010;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform unsigned int32 b = atomic_or_global(s, 1);
+    RET[programIndex] = b;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 0b1010;
+}

tests/atomics-uniform-4.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform unsigned int32 s = 0xffff;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform unsigned int32 b = atomic_min_global(s, 1);
+    RET[programIndex] = b;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 0xffff;
+}

tests/atomics-uniform-5.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform unsigned int32 s = 0xffff;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform unsigned int32 b = atomic_min_global(s, 1);
+    RET[programIndex] = s;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 1;
+}

tests/atomics-uniform-6.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform float s = 100.;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform float b = atomic_swap_global(s, 1.);
+    RET[programIndex] = s;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 1.;
+}

tests/atomics-uniform-7.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform float s = 100.;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform float b = atomic_swap_global(s, 1.);
+    RET[programIndex] = b;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 100.;
+}

tests/atomics-uniform-8.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform float s = 100.;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform float b = atomic_compare_exchange_global(s, 1., -100.);
+    RET[programIndex] = b;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 100.;
+}

tests/atomics-uniform-9.ispc

@@ -0,0 +1,14 @@
+
+export uniform int width() { return programCount; }
+
+uniform int64 s = 100.;
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    float a = aFOO[programIndex]; 
+    uniform int64 b = atomic_compare_exchange_global(s, 100, -100);
+    RET[programIndex] = s;
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = -100.;
+}

tests/launch-1.ispc

@@ -0,0 +1,26 @@
+
+export uniform int width() { return programCount; }
+
+
+static uniform float array[10000];
+
+task void x(float f) {
+    uniform int j;
+    uniform int i = taskIndex;
+    array[i] = i / 10000.;
+    cfor (j = 0; j < 10000; ++j)
+        array[i] = sin(array[i]);
+    if (array[i] < .02)
+        array[i] = i;
+}
+export void f_f(uniform float RET[], uniform float fFOO[]) { 
+    float f = fFOO[programIndex];
+    launch[10000] < x(f) >;
+    sync;
+    RET[programIndex] = array[9999];
+}
+
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 9999.000000;
+}

tests/launch-4.ispc

@@ -0,0 +1,18 @@
+
+export uniform int width() { return programCount; }
+
+
+static float array[6];
+task void x(uniform int i, float f) {
+    array[i] = f;
+}
+export void f_fu(uniform float RET[], uniform float fFOO[], uniform float fu) { 
+    float f = fFOO[programIndex];
+    launch[1] < x(fu, f) >;
+    sync;
+    RET[programIndex] = array[5];
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 1+programIndex;
+}

tests/launch-7.ispc

@@ -0,0 +1,19 @@
+
+export uniform int width() { return programCount; }
+
+
+static uniform float array[10];
+task void foo(uniform float f) { array[0] = f; }
+task void foo(uniform float f, uniform int i) { array[i] = f; }
+
+export void f_v(uniform float RET[]) { 
+    launch[1] < foo(12.) >;
+    launch[1] < foo(-1., 1) >;
+    sync;
+    RET[programIndex] = array[0] + array[1];
+}
+
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 11.000000;
+}

type.cpp

@@ -1833,6 +1833,7 @@ FunctionType::LLVMFunctionType(llvm::LLVMContext *ctx, bool includeMask) const {
     for (unsigned int i = 0; i < argTypes.size(); ++i) {
         if (!argTypes[i])
             return NULL;
+        assert(argTypes[i] != AtomicType::Void);
 
         LLVM_TYPE_CONST llvm::Type *t = argTypes[i]->LLVMType(ctx);
         if (!t)
@@ -1855,6 +1856,8 @@ FunctionType::LLVMFunctionType(llvm::LLVMContext *ctx, bool includeMask) const {
         callTypes.push_back(llvm::PointerType::getUnqual(st));
         callTypes.push_back(LLVMTypes::Int32Type); // threadIndex
         callTypes.push_back(LLVMTypes::Int32Type); // threadCount
+        callTypes.push_back(LLVMTypes::Int32Type); // taskIndex
+        callTypes.push_back(LLVMTypes::Int32Type); // taskCount
     }
     else
         // Otherwise we already have the types of the arguments 

util.cpp

@@ -344,6 +344,10 @@ StringEditDistance(const std::string &str1, const std::string &str2, int maxDist
 
 std::vector<std::string> 
 MatchStrings(const std::string &str, const std::vector<std::string> &options) {
+    if (str.size() == 0 || (str.size() == 1 && !isalpha(str[0])))
+        // don't even try...
+        return std::vector<std::string>();
+
     const int maxDelta = 2;
     std::vector<std::string> matches[maxDelta+1];