diff --git a/examples_ptx/common_gpu.mk b/examples_ptx/common_gpu.mk index e02e5b95..313768e8 100644 --- a/examples_ptx/common_gpu.mk +++ b/examples_ptx/common_gpu.mk @@ -5,7 +5,7 @@ CXX=g++ -ffast-math CXXFLAGS=-O3 -I$(CUDATK)/include -Iobjs_gpu/ -D_CUDA_ # NVCC=nvcc -NVCC_FLAGS=-O3 -arch=sm_35 -D_CUDA_ -I../ -Xptxas=-v -Iobjs_gpu/ +NVCC_FLAGS+=-O3 -arch=sm_35 -D_CUDA_ -I../ -Xptxas=-v -Iobjs_gpu/ ifdef PTXCC_REGMAX NVCC_FLAGS += --maxrregcount=$(PTXCC_REGMAX) endif diff --git a/examples_ptx/radixSort/Makefile_gpu b/examples_ptx/radixSort/Makefile_gpu index 4e101992..ad48a011 100644 --- a/examples_ptx/radixSort/Makefile_gpu +++ b/examples_ptx/radixSort/Makefile_gpu @@ -1,6 +1,8 @@ PROG=radixSort ISPC_SRC=radixSort.ispc + CU_SRC=radixSort.cu +NVCC_FLAGS=-Xptxas=-O1 CXX_SRC=radixSort.cpp radixSort.cpp PTXCC_REGMAX=64