added some ptx options

2013-11-15 17:23:22 +01:00
parent 6b65f6d9f4
commit 3454f51d2c
9 changed files with 17 additions and 3 deletions
--- a/examples_cuda/mandelbrot_tasks3d/Makefile_gpu
+++ b/examples_cuda/mandelbrot_tasks3d/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

--- a/examples_cuda/mandelbrot_tasks3d/mandelbrot_tasks3d.ispc
+++ b/examples_cuda/mandelbrot_tasks3d/mandelbrot_tasks3d.ispc
@@ -119,5 +119,6 @@ mandelbrot_ispc(uniform float x0, uniform float y0,
 #endif
      mandelbrot_scanline(x0, dx, y0, dy, width, height, xspan, yspan,
                          maxIterations, output);
+    sync;
 }
 #endif