added some ptx options

examples_cuda/volume_rendering/Makefile_gpu

@@ -13,6 +13,8 @@ ISPCFLAGS=-O3 --math-lib=default --target=nvptx64 --opt=fast-math
 LLVM32 = $(HOME)/usr/local/llvm/bin-3.2
 LLVM   = $(HOME)/usr/local/llvm/bin-3.3
 PTXGEN = $(HOME)/ptxgen
+PTXGEN += -opt=3
+PTXGEN += -ftz=1 -prec-div=0 -prec-sqrt=0 -fma=1
 
 LLVM32DIS=$(LLVM32)/bin/llvm-dis
 

examples_cuda/volume_rendering/volume1.ispc

@@ -419,4 +419,5 @@ volume_ispc_tasks(uniform float density[], uniform int nVoxels[3],
     uniform int nTasks = ((width+(dx-1))/dx) * ((height+(dy-1))/dy);
     launch[nTasks] volume_task(density, nVoxels, raster2camera, camera2world, 
                                width, height, image);
+    sync;
 }