pytorch中使用cuda扩展

以下面这个例子作为教程，实现功能是element-wise add；

（pytorch中想调用cuda模块，还是用另外使用C编写接口脚本）

第一步：cuda编程的源文件和头文件

// mathutil_cuda_kernel.cu
// 头文件，最后一个是cuda特有的
#include <curand.h>
#include <stdio.h>
#include <math.h>
#include <float.h>
#include "mathutil_cuda_kernel.h"

// 获取GPU线程通道信息
dim3 cuda_gridsize(int n)
{
    int k = (n - 1) / BLOCK + 1;
    int x = k;
    int y = 1;
    if(x > 65535) {
        x = ceil(sqrt(k));
        y = (n - 1) / (x * BLOCK) + 1;
    }
    dim3 d(x, y, 1);
    return d;
}
// 这个函数是cuda执行函数，可以看到细化到了每一个元素
__global__ void broadcast_sum_kernel(float *a, float *b, int x, int y, int size)
{
    int i = (blockIdx.x + blockIdx.y * gridDim.x) * blockDim.x + threadIdx.x;
    if(i >= size) return;
    int j = i % x; i = i / x;
    int k = i % y;
    a[IDX2D(j, k, y)] += b[k];
}


// 这个函数是与c语言函数链接的接口函数
void broadcast_sum_cuda(float *a, float *b, int x, int y, cudaStream_t stream)
{
    int size = x * y;
    cudaError_t err;
    
    // 上面定义的函数
    broadcast_sum_kernel<<<cuda_gridsize(size), BLOCK, 0, stream>>>(a, b, x, y, size);

    err = cudaGetLastError();
    if (cudaSuccess != err)
    {
        fprintf(stderr, "CUDA kernel failed : %s
", cudaGetErrorString(err));
        exit(-1);
    }
}

#ifndef _MATHUTIL_CUDA_KERNEL
#define _MATHUTIL_CUDA_KERNEL

#define IDX2D(i, j, dj) (dj * i + j)
#define IDX3D(i, j, k, dj, dk) (IDX2D(IDX2D(i, j, dj), k, dk))

#define BLOCK 512
#define MAX_STREAMS 512

#ifdef __cplusplus
extern "C" {
#endif

void broadcast_sum_cuda(float *a, float *b, int x, int y, cudaStream_t stream);

#ifdef __cplusplus
}
#endif

#endif

第二步：C编程的源文件和头文件（接口函数）

// mathutil_cuda.c
// THC是pytorch底层GPU库
#include <THC/THC.h>
#include "mathutil_cuda_kernel.h"

extern THCState *state;

int broadcast_sum(THCudaTensor *a_tensor, THCudaTensor *b_tensor, int x, int y)
{
    float *a = THCudaTensor_data(state, a_tensor);
    float *b = THCudaTensor_data(state, b_tensor);
    cudaStream_t stream = THCState_getCurrentStream(state);

    // 这里调用之前在cuda中编写的接口函数
    broadcast_sum_cuda(a, b, x, y, stream);

    return 1;
}

int broadcast_sum(THCudaTensor *a_tensor, THCudaTensor *b_tensor, int x, int y);

第三步：编译，先编译cuda模块，再编译接口函数模块（不能放在一起同时编译）

nvcc -c -o mathutil_cuda_kernel.cu.o mathutil_cuda_kernel.cu -x cu -Xcompiler -fPIC -arch=sm_52

import os
import torch
from torch.utils.ffi import create_extension

this_file = os.path.dirname(__file__)

sources = []
headers = []
defines = []
with_cuda = False

if torch.cuda.is_available():
    print('Including CUDA code.')
    sources += ['src/mathutil_cuda.c']
    headers += ['src/mathutil_cuda.h']
    defines += [('WITH_CUDA', None)]
    with_cuda = True

this_file = os.path.dirname(os.path.realpath(__file__))

extra_objects = ['src/mathutil_cuda_kernel.cu.o']   # 这里是编译好后的.o文件位置
extra_objects = [os.path.join(this_file, fname) for fname in extra_objects]


ffi = create_extension(
    '_ext.cuda_util',
    headers=headers,
    sources=sources,
    define_macros=defines,
    relative_to=__file__,
    with_cuda=with_cuda,
    extra_objects=extra_objects
)

if __name__ == '__main__':
    ffi.build()

第四步：调用cuda模块

from _ext import cuda_util  #从对应路径中调用编译好的模块

a = torch.randn(3, 5).cuda()
b = torch.randn(3, 1).cuda()
mathutil.broadcast_sum(a, b, *map(int, a.size()))

# 上面等价于下面的效果：

a = torch.randn(3, 5)
b = torch.randn(3, 1)
a += b