diff --git a/builtins/target-nvptx64.ll b/builtins/target-nvptx64.ll
index 063fdc7b..ae8cb604 100644
--- a/builtins/target-nvptx64.ll
+++ b/builtins/target-nvptx64.ll
@@ -510,18 +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);
-  %r = fdiv float 1.,%0
-  ret float %r
+;  %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 = 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 = 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 6a11cbc5..74618d08 100644
--- a/examples_cuda/aobench/Makefile_gpu
+++ b/examples_cuda/aobench/Makefile_gpu
@@ -8,7 +8,7 @@ LD=g++
 LDFLAGS=-lcuda
 
 ISPC=ispc
-ISPCFLAGS=-O3 --math-lib=default --target=nvptx64
+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
diff --git a/examples_cuda/aobench/ao_cu.cpp b/examples_cuda/aobench/ao_cu.cpp
index 7ec9c1a3..a8ba598b 100755
--- a/examples_cuda/aobench/ao_cu.cpp
+++ b/examples_cuda/aobench/ao_cu.cpp
@@ -191,7 +191,7 @@ CUmodule loadModule(const char * module)
     optionVals[5] = (void*) 1;
     // Max # of registers/pthread
     options[6] = CU_JIT_MAX_REGISTERS;
-    int jitRegCount = 48;
+    int jitRegCount = 64;
     optionVals[6] = (void *)(size_t)jitRegCount;
 
     // Create a pending linker invocation