Code and Test on CUDA

Coding on CUDA

Most memory-related functions work with memory length in bytes, e.g. cudaMallocXxx, cudaMemcpyXxx.
As CUDA Streams are widely used, whenever unsure and there’s need, do synchronisation as cudaDeviceSynchronize().
The CUDA Unified Memory is greatly handy. It can be allocated by cudaMallocManaged() and released by cudaFree().
Below is a sample snippet for nvcc gencode instruction.

# CUDA architecture setting: going with all of them.
# For CUDA < 6.0, comment the *_50 lines for compatibility.
CUDA_ARCH := -gencode arch=compute_30,code=sm_30 \
	-gencode arch=compute_35,code=sm_35 \
	-gencode arch=compute_50,code=sm_50 \
	-gencode arch=compute_50,code=compute_50

Testing on CUDA

Below is a sample snippet of Makefile for making C++ test files for CUDA.

ifeq ($(USE_CUDA), 1)
# All tests for GPU in *_test_gpu.cu
TEST_SRC_GPU = $(wildcard tests/cpp/*_test_gpu.cu)
TEST_GPU = $(patsubst tests/cpp/%_test_gpu.cu, tests/cpp/%_test_gpu, $(TEST_SRC_GPU))
override TEST += $(TEST_GPU)

tests/cpp/%_test_gpu : tests/cpp/%_test_gpu.cu lib/libmxnet.a
  $(NVCC) $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS) -I$(GTEST_INC)" -M -MT tests/cpp/$*_test_gpu $< >tests/cpp/$*_test_gpu.d
  $(NVCC) -c $(NVCCFLAGS) $(CUDA_ARCH) -Xcompiler "$(CFLAGS) -I$(GTEST_INC)" -o $@.o $(filter %.cu, $^)
  $(CXX) -o $@ $@.o $(filter %.a, $^) -L$(GTEST_LIB) -lgtest $(LDFLAGS) -lcudart

endif