cuFFT¶
简介¶
The cuFFT is a CUDA Fast Fourier Transform library consisting of two components: cuFFT and cuFFTW. The cuFFT library provides high performance on NVIDIA GPUs, and the cuFFTW library is a porting tool to use FFTW on NVIDIA GPUs.
cuFFT使用说明¶
思源一号上的cuFFT¶
先创建一个目录cuFFTtest并进入该目录：
mkdir cuFFTtest
cd cuFFTtest
在该目录下创建如下测试文件cuFFTtest.cu：
// In this example a one-dimensional complex-to-complex transform is applied to the input data.
//Afterwards an inverse transform is performed on the computed frequency domain representation.
#include <complex>
#include <iostream>
#include <random>
#include <vector>
#include <cuda_runtime.h>
#include <cufftXt.h>
// #include "cufft_utils.h"
// CUDA API error checking
#define CUDA_RT_CALL(call) \
{ \
auto status = static_cast<cudaError_t>(call); \
if (status != cudaSuccess) \
fprintf(stderr, \
"ERROR: CUDA RT call \"%s\" in line %d of file %s failed " \
"with " \
"%s (%d).\n", \
#call, \
__LINE__, \
__FILE__, \
cudaGetErrorString(status), \
status); \
}
// cufft API error chekcing
#define CUFFT_CALL(call) \
{ \
auto status = static_cast<cufftResult>(call); \
if (status != CUFFT_SUCCESS) \
fprintf(stderr, \
"ERROR: CUFFT call \"%s\" in line %d of file %s failed " \
"with " \
"code (%d).\n", \
#call, \
__LINE__, \
__FILE__, \
status); \
}
int main(int argc, char *argv[])
{
cufftHandle plan;
cudaStream_t stream = NULL;
int n = 8;
int batch_size = 2;
int fft_size = batch_size * n;
using scalar_type = float;
using data_type = std::complex<scalar_type>;
std::vector<data_type> data(fft_size);
for (int i = 0; i < fft_size; i++)
{
data[i] = data_type(i, -i);
}
std::printf("Input array:\n");
for (auto &i : data)
{
std::printf("%f + %fj\n", i.real(), i.imag());
}
std::printf("=====\n");
cufftComplex *d_data = nullptr;
CUFFT_CALL(cufftCreate(&plan));
CUFFT_CALL(cufftPlan1d(&plan, data.size(), CUFFT_C2C, batch_size));
CUDA_RT_CALL(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
CUFFT_CALL(cufftSetStream(plan, stream));
// Create device data arrays
CUDA_RT_CALL(cudaMalloc(reinterpret_cast<void **>(&d_data), sizeof(data_type) * data.size()));
CUDA_RT_CALL(cudaMemcpyAsync(d_data, data.data(), sizeof(data_type) * data.size(), cudaMemcpyHostToDevice, stream));
/*
* Note:
* Identical pointers to data and output arrays implies in-place transformation
*/
CUFFT_CALL(cufftExecC2C(plan, d_data, d_data, CUFFT_FORWARD));
CUFFT_CALL(cufftExecC2C(plan, d_data, d_data, CUFFT_INVERSE));
CUDA_RT_CALL(cudaMemcpyAsync(data.data(), d_data, sizeof(data_type) * data.size(), cudaMemcpyDeviceToHost, stream));
CUDA_RT_CALL(cudaStreamSynchronize(stream));
std::printf("Output array:\n");
for (auto &i : data)
{
std::printf("%f + %fj\n", i.real(), i.imag());
}
std::printf("=====\n");
/* free resources */
CUDA_RT_CALL(cudaFree(d_data))
CUFFT_CALL(cufftDestroy(plan));
CUDA_RT_CALL(cudaStreamDestroy(stream));
CUDA_RT_CALL(cudaDeviceReset());
return EXIT_SUCCESS;
}
在该目录下创建如下作业提交脚本cuFFTtest.slurm:
#!/bin/bash
#SBATCH --job-name=cuFFTtest # 作业名
#SBATCH --partition=a100 # a100 队列
#SBATCH -N 1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=1 # 1:1 的 GPU:CPU 配比
#SBATCH --gres=gpu:1 # 1 块 GPU
#SBATCH --output=%j.out
#SBATCH --error=%j.err
module load cuda/11.5.0
module load gcc/11.2.0
nvcc cuFFTtest.cu -o cuFFTtest -lcufft
./cuFFTtest
使用如下命令提交作业：
sbatch cuFFTtest.slurm
作业完成后在.out文件中可看到如下结果：
Input array:
0.000000 + 0.000000j
1.000000 + -1.000000j
2.000000 + -2.000000j
3.000000 + -3.000000j
4.000000 + -4.000000j
5.000000 + -5.000000j
6.000000 + -6.000000j
7.000000 + -7.000000j
8.000000 + -8.000000j
9.000000 + -9.000000j
10.000000 + -10.000000j
11.000000 + -11.000000j
12.000000 + -12.000000j
13.000000 + -13.000000j
14.000000 + -14.000000j
15.000000 + -15.000000j
=====
Output array:
0.000004 + 0.000000j
16.000015 + -16.000004j
32.000004 + -32.000004j
48.000004 + -48.000004j
64.000000 + -64.000000j
80.000000 + -80.000008j
96.000000 + -96.000015j
112.000015 + -112.000000j
128.000000 + -128.000000j
144.000000 + -144.000000j
160.000000 + -160.000000j
176.000000 + -176.000000j
192.000000 + -192.000000j
208.000000 + -208.000000j
224.000000 + -223.999985j
239.999985 + -240.000000j
=====
pi2.0上的cuFFT¶
此步骤和上文完全相同；
此步骤和上文完全相同；
在该目录下创建如下作业提交脚本cuFFTtest.slurm:
#!/bin/bash
#SBATCH --job-name=cuFFTtest # 作业名
#SBATCH --partition=dgx2 # dgx2 队列
#SBATCH -N 1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=1 # 1:1 的 GPU:CPU 配比
#SBATCH --gres=gpu:1 # 1 块 GPU
#SBATCH --output=%j.out
#SBATCH --error=%j.err
module load cuda/11.6.2-gcc-8.3.0
module load gcc/8.3.0
nvcc cuFFTtest.cu -o cuFFTtest -lcufft
./cuFFTtest
使用如下命令提交作业：
sbatch cuFFTtest.slurm
作业完成后在.out文件中可看到如下结果：
Input array:
0.000000 + 0.000000j
1.000000 + -1.000000j
2.000000 + -2.000000j
3.000000 + -3.000000j
4.000000 + -4.000000j
5.000000 + -5.000000j
6.000000 + -6.000000j
7.000000 + -7.000000j
8.000000 + -8.000000j
9.000000 + -9.000000j
10.000000 + -10.000000j
11.000000 + -11.000000j
12.000000 + -12.000000j
13.000000 + -13.000000j
14.000000 + -14.000000j
15.000000 + -15.000000j
=====
Output array:
0.000004 + 0.000000j
16.000015 + -16.000004j
32.000004 + -32.000004j
48.000004 + -48.000004j
64.000000 + -64.000000j
80.000000 + -80.000008j
96.000000 + -96.000015j
112.000015 + -112.000000j
128.000000 + -128.000000j
144.000000 + -144.000000j
160.000000 + -160.000000j
176.000000 + -176.000000j
192.000000 + -192.000000j
208.000000 + -208.000000j
224.000000 + -223.999985j
239.999985 + -240.000000j
=====
参考资料¶
最后更新:
2024 年 11 月 14 日