diff --git a/builtins/target-nvptx64.ll b/builtins/target-nvptx64.ll
index ae8cb604..8bceace7 100644
--- a/builtins/target-nvptx64.ll
+++ b/builtins/target-nvptx64.ll
@@ -510,21 +510,21 @@ declare double    @llvm.nvvm.sqrt.d(double %f) nounwind readonly alwaysinline
 define  float @__rcp_uniform_float(float) nounwind readonly alwaysinline {
 ;    uniform float iv = extract(__rcp_u(v), 0);
 ;    return iv * (2. - v * iv);
-;  %ret = fdiv float 1.,%0
-  %ret = tail call float asm sideeffect "rcp.approx.ftz.f32  $0, $1;", "=f,f"(float %0) nounwind readnone alwaysinline
+  %ret = fdiv float 1.,%0
+;  %ret = tail call float asm sideeffect "rcp.approx.ftz.f32  $0, $1;", "=f,f"(float %0) nounwind readnone alwaysinline
   ret float %ret
 }
 ;; declare float @__sqrt_uniform_float(float) nounwind readnone 
 define  float @__sqrt_uniform_float(float) nounwind readonly alwaysinline {
-  ;;%ret = call float @llvm.nvvm.sqrt.f(float %0)
-  %ret = tail call float asm sideeffect "sqrt.approx.ftz.f32  $0, $1;", "=f,f"(float %0) nounwind readnone alwaysinline
+  %ret = call float @llvm.nvvm.sqrt.f(float %0)
+;  %ret = tail call float asm sideeffect "sqrt.approx.ftz.f32  $0, $1;", "=f,f"(float %0) nounwind readnone alwaysinline
   ret float %ret
 }
 ;; declare float @__rsqrt_uniform_float(float) nounwind readnone 
 define  float @__rsqrt_uniform_float(float) nounwind readonly alwaysinline 
 {
-;;  %ret = call float @llvm.nvvm.rsqrt.approx.f(float %0)
-  %ret = tail call float asm sideeffect "rsqrt.approx.ftz.f32  $0, $1;", "=f,f"(float %0) nounwind readnone alwaysinline
+  %ret = call float @llvm.nvvm.rsqrt.approx.f(float %0)
+;  %ret = tail call float asm sideeffect "rsqrt.approx.ftz.f32  $0, $1;", "=f,f"(float %0) nounwind readnone alwaysinline
   ret float %ret
 }
 
diff --git a/examples_cuda/aobench/Makefile_gpu b/examples_cuda/aobench/Makefile_gpu
index 74618d08..54cae93e 100644
--- a/examples_cuda/aobench/Makefile_gpu
+++ b/examples_cuda/aobench/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