fixed ISPCSync functionality

builtins/target-nvptx64.ll

@@ -102,7 +102,8 @@ define i32 @__lanemask_lt() nounwind readnone alwaysinline
 
 define i8* @ISPCAlloc(i8**, i64, i32) nounwind alwaysinline
 {
-  ret i8* null
+  %ptr = inttoptr i64 1 to i8*
+  ret i8* %ptr
 }
 
 ;; this call allocate parameter buffer for kernel launch

ctx.cpp

@@ -3673,6 +3673,8 @@ FunctionEmitContext::LaunchInst(llvm::Value *callee,
 
 void
 FunctionEmitContext::SyncInst() {
+    if (!g->target->isPTX())
+    {
       llvm::Value *launchGroupHandle = LoadInst(launchGroupHandlePtr);
       llvm::Value *nullPtrValue =
         llvm::Constant::getNullValue(LLVMTypes::VoidPointerType);
@@ -3697,6 +3699,18 @@ FunctionEmitContext::SyncInst() {
 
       SetCurrentBasicBlock(bPostSync);
     }
+    else
+    {
+      llvm::Value *launchGroupHandle = LoadInst(launchGroupHandlePtr);
+      llvm::Value *nullPtrValue =
+        llvm::Constant::getNullValue(LLVMTypes::VoidPointerType);
+      llvm::Function *fsync = m->module->getFunction("ISPCSync");
+      if (fsync == NULL)
+        FATAL("Couldn't find ISPCSync declaration?!");
+      CallInst(fsync, NULL, launchGroupHandle, "");
+      StoreInst(nullPtrValue, launchGroupHandlePtr);
+    }
+}
 
 
 /** When we gathering from or scattering to a varying atomic type, we need

examples_cuda/aobench/ao1.ispc

@@ -313,5 +313,4 @@ export void ao_ispc_tasks(uniform int w, uniform int h, uniform int nsubsamples,
   const uniform int ntilex = (w+TILEX-1)/TILEX;
   const uniform int ntiley = (h+TILEY-1)/TILEY;
   launch[ntilex,ntiley] ao_task(w, h, nsubsamples, image);
-  sync;
 }