aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

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

Browse Source
This commit is contained in:
Evghenii
2013-11-14 22:17:57 +01:00
parent 2c8afde6d9
commit f12826bac5
3 changed files with 9 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
builtins/target-nvptx64.ll
View File

@@ -510,18 +510,21 @@ declare double @llvm.nvvm.sqrt.d(double %f) nounwind readonly alwaysinline
define float @__rcp_uniform_float(float) nounwind readonly alwaysinline { define float @__rcp_uniform_float(float) nounwind readonly alwaysinline {
; uniform float iv = extract(__rcp_u(v), 0); ; uniform float iv = extract(__rcp_u(v), 0);
; return iv * (2. - v * iv); ; return iv * (2. - v * iv);
%r = fdiv float 1.,%0 ; %ret = fdiv float 1.,%0
ret float %r %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 ;; declare float @__sqrt_uniform_float(float) nounwind readnone
define float @__sqrt_uniform_float(float) nounwind readonly alwaysinline { 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 ret float %ret
} }
;; declare float @__rsqrt_uniform_float(float) nounwind readnone ;; declare float @__rsqrt_uniform_float(float) nounwind readnone
define float @__rsqrt_uniform_float(float) nounwind readonly alwaysinline 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 ret float %ret
} }

2
examples_cuda/aobench/Makefile_gpu
View File

@@ -8,7 +8,7 @@ LD=g++
LDFLAGS=-lcuda LDFLAGS=-lcuda
ISPC=ispc 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 LLVM32 = $(HOME)/usr/local/llvm/bin-3.2
LLVM = $(HOME)/usr/local/llvm/bin-3.3 LLVM = $(HOME)/usr/local/llvm/bin-3.3

2
examples_cuda/aobench/ao_cu.cpp
View File

@@ -191,7 +191,7 @@ CUmodule loadModule(const char * module)
optionVals[5] = (void*) 1; optionVals[5] = (void*) 1;
// Max # of registers/pthread // Max # of registers/pthread
options[6] = CU_JIT_MAX_REGISTERS; options[6] = CU_JIT_MAX_REGISTERS;
int jitRegCount = 48; int jitRegCount = 64;
optionVals[6] = (void *)(size_t)jitRegCount; optionVals[6] = (void *)(size_t)jitRegCount;
// Create a pending linker invocation // Create a pending linker invocation