+added approx rcp/rsqrt/rtz with ftz=true

2013-11-14 22:17:57 +01:00
parent 2c8afde6d9
commit f12826bac5
3 changed files with 9 additions and 6 deletions
--- 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
--- 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