How to change the CMAKE of cuda static library from debug to release? - cmake

Question

I'm a new to use CMAKE. I want to use the kernels in cuda as a static library and use " extern "C" void function();" to call it. Finally I will use cmake to compile the whole project. But its running speed in GPU didn't satisfied me. So I used Nsight eclispe to run it severally in debug and release. After analyzing them in NVVP.I found the default mode of cmake in static library is debug mode.
So how can I change the debug mode to release in static library?
Firstly, I create a project in Nsight eclipse.
Below is the example of the file structure that I have.
Test_in_stack
-release
-debug
-src
--GPU.cu
--simpleCUFFT.cu
-lib
--GPU.cuh
--Kernels.h
The content of src/simpleCUFFT.cu is:
// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
// includes, project
// #include <Kernels.h>
#include <GPU.cuh>
#include <cuda_runtime.h>
#include <cufft.h>
#include <cufftXt.h>
#include <helper_functions.h>
#include <helper_cuda.h>
#include <device_functions.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <thrust/sort.h>
#include <thrust/generate.h>
#include <thrust/sequence.h>
#include <thrust/device_ptr.h>
#include <thrust/extrema.h>
#include <thrust/execution_policy.h>
#include <thrust/equal.h>
#include <thrust/for_each.h>
// Complex data type
typedef float2 Complex;
#define FFT_NUM 1024
#define RANGE_NUM 1024
#define SIGNAL_SIZE RANGE_NUM*FFT_NUM
extern "C" void GPU_Pro(Complex *h_signal,int *h_count);
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main()
{
Complex *h_signal = (Complex *)malloc(sizeof(Complex) * SIGNAL_SIZE);
int *h_count = (int *)malloc(sizeof(int) * SIGNAL_SIZE);
// Initialize the memory for the signal
for (unsigned int i = 0; i < SIGNAL_SIZE; ++i)
{
h_signal[i].x = rand() / (float)RAND_MAX;
h_signal[i].y = rand() / (float)RAND_MAX;
h_count[i]=i/FFT_NUM;
}
GPU_Pro(h_signal,h_count);
cudaDeviceReset();
}
The content of src/GPU.cu is:
#include <Kernels.h>
#include <GPU.cuh>
#include <cuda_runtime.h>
#include <cufft.h>
#include <cufftXt.h>
#include <helper_functions.h>
#include <helper_cuda.h>
#include <device_functions.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <thrust/sort.h>
#include <thrust/generate.h>
#include <thrust/sequence.h>
#include <thrust/device_ptr.h>
#include <thrust/extrema.h>
#include <thrust/execution_policy.h>
#include <thrust/equal.h>
#include <thrust/for_each.h>
typedef float2 Complex;
#define FFT_NUM 1024
#define RANGE_NUM 1024
#define SIGNAL_SIZE RANGE_NUM*FFT_NUM
void GPU_Pro(Complex *h_signal,int *h_count)
{
Complex *d_signal;
float *d_signal_float;
int *d_count;
cudaMalloc((void **)&d_signal, SIGNAL_SIZE*sizeof(Complex));
cudaMalloc((void **)&d_count, SIGNAL_SIZE*sizeof(int));
cudaMalloc((void **)&d_signal_float, SIGNAL_SIZE*sizeof(float));
cufftHandle plan;
checkCudaErrors(cufftPlan1d(&plan, FFT_NUM, CUFFT_C2C, 1));
dim3 dimblock(32, 32);
dim3 dimgrid(FFT_NUM / 32, RANGE_NUM / 32);
// Copy host memory to device
checkCudaErrors(cudaMemcpy(d_signal, h_signal, SIGNAL_SIZE*sizeof(Complex),
cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_count, h_count, SIGNAL_SIZE*sizeof(int),
cudaMemcpyHostToDevice));
for(int i=0;i<RANGE_NUM;i++)
{
checkCudaErrors(cufftExecC2C(plan, d_signal+i*RANGE_NUM, d_signal+i*RANGE_NUM, CUFFT_FORWARD));
}
MatAbsNaive_float<<<dimgrid,dimblock>>>(d_signal,d_signal_float,FFT_NUM,RANGE_NUM);
thrust::stable_sort_by_key(thrust::device_pointer_cast(d_signal_float),thrust::device_pointer_cast(d_signal_float)+SIGNAL_SIZE,thrust::device_pointer_cast(d_count));
thrust::stable_sort_by_key(thrust::device_pointer_cast(d_count),thrust::device_pointer_cast(d_count)+SIGNAL_SIZE,thrust::device_pointer_cast(d_signal_float));
cudaDeviceReset();
}
The content of lib/Kernels.h is:
/*
* Kernels.h
*
* Created on: Jan 10, 2019
* Author: root
*/
#include "iostream"
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include "stdio.h"
#include <stdlib.h>
#include <string.h>
#include "math.h"
#include <mat.h>
#include "cuComplex.h"
#include "cublas.h"
#include <cufft.h>
#include <cufftXt.h>
#include <time.h>
#include <cublas_v2.h>
__global__ void MatAbsNaive_float(cuComplex *idata, float *odata, int M, int N)
{
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
if ((x < M) && (y < N))
{
odata[x + M*y] = sqrt(idata[x + M*y].x * idata[x + M*y].x + idata[x + M*y].y * idata[x + M*y].y);
}
}
The content of lib/GPU.cuh is:
#ifndef GPU_CUH
#define GPU_CUH
#include <stdio.h>
#include "cuComplex.h"
typedef float2 Complex;
extern "C"
void GPU_Pro(Complex *h_signal,int *h_count);
#endif
The result of NVVP in debug and release are as follows:
debug
release
Then I put the same files in cmake.
Below is the example of the file structure that I have.
Test_in_stack
-CMakeLists(1).txt
-build
-src
--CMakeLists(2).txt
--simpleCUFFT.cpp
-lib
--CMakeLists(3).txt
--GPU.cu
--GPU.cuh
--Kernels.h
The (1)、(2)、(3) are labels, the real file names are both CMakeLists.txt. And the content of simpleCUFFT.cu and simpleCUFFT.cpp are the same.
The content of CMakeLists(1).txt is:
cmake_minimum_required (VERSION 2.6)
PROJECT(GPU_MODE C CXX)
#PROJECT(GPU_MODE)
ADD_SUBDIRECTORY(src bin)
ADD_SUBDIRECTORY(lib)
The content of CMakeLists(2).txt is:
INCLUDE_DIRECTORIES(
${eclipse_home}VSPS/include
/usr/include
${eclipse_home}PetDCPS/include
/user/include/c++
/usr/local/cuda-8.0/include
)
INCLUDE_DIRECTORIES(/root/Chenjie/cuda-workspace/Test_in_stack/lib
/usr/local/cuda-8.0/samples/common/inc
/usr/local/cuda-8.0/include)
LINK_DIRECTORIES(/usr/local/cuda-8.0/lib64/)
SET(CPU_LIST simpleCUFFT.cpp)
FIND_PACKAGE(CUDA REQUIRED)
SET(EXTRA_LIBS ${EXTRA_LIBS} gpu ${CUDA_LIBRARIES})
ADD_EXECUTABLE(CPUProcessTest ${CPU_LIST})
SET(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_SOURCE_DIR}/src)
TARGET_LINK_LIBRARIES(CPUProcessTest optimized ${EXTRA_LIBS} vsip_c)
The content of CMakeLists(3).txt is:
#for cuda
PROJECT(gpu)
CMAKE_MINIMUM_REQUIRED(VERSION 2.8)
INCLUDE_DIRECTORIES(/root/Chenjie/cuda-workspace/Test_in_stack/lib
/usr/local/cuda-8.0/samples/common/inc
/usr/local/cuda-8.0/include)
FIND_PACKAGE(CUDA QUIET REQUIRED)
#SET(CUDA_NVCC_FLAGS -03;-G;-g)
SET(CUDA_NVCC_FLAGS -gencode arch=compute_52,code=sm_52;-G;-g;-lcufft;-lcudart;-lcublas)
SET(CMAKE_CUDA_FLAGS ${CUDA_NVCC_FLAGS_RELEASE})
FILE(GLOB_RECURSE CURRENT_HEADERS *.h *.hpp *.cuh)
FILE(GLOB CURRENT_SOURCES *.cpp *.cu)
SOURCE_GROUP("Include" FILES ${CURRENT_HEADERS})
SOURCE_GROUP("Source" FILES ${CURRENT_SOURCES})
INCLUDE_DIRECTORIES(/usr/local/cuda-8.0/include)
LINK_DIRECTORIES(/usr/local/cuda-8.0/lib64/)
LINK_LIBRARIES(cufft cublas)
#TARGET_LINK_LIBRARIES(gpu ${CUDA_LIBRARIES})
#CUDA_ADD_LIBRARY(gpu SHARED ${CURRENT_HEADERS} ${CURRENT_SOURCES})
CUDA_ADD_LIBRARY(gpu STATIC ${CURRENT_HEADERS} ${CURRENT_SOURCES} ${CUDA_LIBRARIES} ${CUDA_CUFFT_LIBRARIES})
I used the command line in /built as follows:
cmake -DCMAKE_BUILD_TYPE=Release ..
make
But it didn't work. It seems it still run in debug as the NVVP result shows:
cmake result
So how can I change to compile flag to release in static library of cuda.
I'm using Red Hat Enterprise Linux Server 7.1(Maipo)、cuda 8.0 、cmake version 2.8.12.2、GNU Make 3.82.
Update in 2019.01.12
I added MESSAGE(STATUS "Build type:" ${CMAKE_BUILD_TYPE}") in CMakeLists(2).txt. And the result is:
[root#node2 build]# cmake -DCMAKE_BUILD_TYPE=Release ..
-- Build type: Release
-- Configuring done
-- Generating done
But the result in NVVP didn't change.

Answer 1

Well， I have found out a way to solve it.
I change the line 10 and line 11 in CMakeLists(3).txt to
SET(CUDA_NVCC_FLAGS -gencode arch=compute_52,code=sm_52;-lcufft;-lcudart;-lcublas)
SET(CMAKE_CUDA_FLAGS ${CUDA_NVCC_FLAGS} -O3 -DNDEBUG)
After doing
cmake -DCMAKE_BUILD_TYPE=Release ..
make clean
make
The result in NVVP shows it is compiled with Release mode.