|
技术背景
在前面一篇文章中,我们介绍过Cython+CUDA框架下实现一个简单的Gather算子的方法。这里演示Gather算子的升级版本实现——BatchGather算子。不过这里只是加了一个Batch维度,并没有添加其他的维度,例如Dimension维度,在这里暂不考虑。
CUDA头文件
这里我们保留了原本的Gather部分,只添加一个BatchGather的运算,以下为cuda_index.cuh的内容:
#include <stdio.h>extern "C" float Gather(float *source, int *index, float *res, int N, int M);extern "C" float BatchGather(float *source, int *index, float *res, int N, int M, int B);BatchGather只是在Gather的基础上加了一个B的维度。除了CUDA算子本身的头文件之外,这里我们还使用到了异常捕获头文件error.cuh:
#pragma once#include <stdio.h>#define CHECK(call) do{const cudaError_t error_code = call; if (error_code != cudaSuccess){printf("CUDA Error:\n"); printf(" File: %s\n", __FILE__); printf(" Line: %d\n", __LINE__); printf(" Error code: %d\n", error_code); printf(" Error text: %s\n", cudaGetErrorString(error_code)); exit(1);}} while (0)其中的宏可用于检测CUDA函数所抛出的异常。另外还有一个用于统计CUDA函数运行时长的头文件:
#pragma once#include <stdio.h>#include <cuda_runtime.h>// 宏定义,用于测量CUDA函数的执行时间#define TIME_CUDA_FUNCTION(func) \ do { \ cudaEvent_t start, stop; \ float elapsedTime; \ cudaEventCreate(&start); \ cudaEventCreate(&stop); \ cudaEventRecord(start, NULL); \ \ func; \ \ cudaEventRecord(stop, NULL); \ cudaEventSynchronize(stop); \ cudaEventElapsedTime(&elapsedTime, start, stop); \ printf("Time taken by function %s is: %f ms\n", #func, elapsedTime); \ \ cudaEventDestroy(start); \ cudaEventDestroy(stop); \ } while (0)// 宏定义,用于测量CUDA函数的执行时间并返回该时间#define GET_CUDA_TIME(func) \ ({ \ cudaEvent_t start, stop; \ float elapsedTime = 0.0f; \ cudaEventCreate(&start); \ cudaEventCreate(&stop); \ cudaEventRecord(start, NULL); \ \ func; \ \ cudaEventRecord(stop, NULL); \ cudaEventSynchronize(stop); \ cudaEventElapsedTime(&elapsedTime, start, stop); \ \ cudaEventDestroy(start); \ cudaEventDestroy(stop); \ \ elapsedTime; \ })可选择直接打印时长,也可以选择返回时长的float值。
CUDA文件
接下来就是正式的CUDA函数内容cuda_index.cu:
// nvcc -shared ./cuda_index.cu -Xcompiler -fPIC -o ./libcuindex.so#include <stdio.h>#include "cuda_index.cuh"#include "error.cuh"#include "record.cuh"__global__ void GatherKernel(float *source, int *index, float *res, int N){ int idx = blockIdx.x * blockDim.x + threadIdx.x; if (idx < N){ res[idx] = source[index[idx]]; }}extern "C" float Gather(float *source, int *index, float *res, int N, int M){ float *souce_device, *res_device; int *index_device; CHECK(cudaMalloc((void **)&souce_device, M * sizeof(float))); CHECK(cudaMalloc((void **)&res_device, N * sizeof(float))); CHECK(cudaMalloc((void **)&index_device, N * sizeof(int))); CHECK(cudaMemcpy(souce_device, source, M * sizeof(float), cudaMemcpyHostToDevice)); CHECK(cudaMemcpy(res_device, res, N * sizeof(float), cudaMemcpyHostToDevice)); CHECK(cudaMemcpy(index_device, index, N * sizeof(int), cudaMemcpyHostToDevice)); int block_size = 1024; int grid_size = (N + block_size - 1) / block_size; float timeTaken = GET_CUDA_TIME((GatherKernel<<<grid_size, block_size>>>(souce_device, index_device, res_device, N))); CHECK(cudaGetLastError()); CHECK(cudaDeviceSynchronize()); CHECK(cudaMemcpy(res, res_device, N * sizeof(float), cudaMemcpyDeviceToHost)); CHECK(cudaFree(souce_device)); CHECK(cudaFree(index_device)); CHECK(cudaDeviceSynchronize()); CHECK(cudaFree(res_device)); CHECK(cudaDeviceReset()); return timeTaken;}__global__ void BatchGatherKernel(float *source, int *index, float *res, int N, int M, int B){ int idx = blockIdx.x * blockDim.x + threadIdx.x; if (idx < N*B){ int batch_idx = idx / N; int source_idx = batch_idx * M + index[idx]; res[idx] = source[source_idx]; }}extern "C" float BatchGather(float *source, int *index, float *res, int N, int M, int B){ float *souce_device, *res_device; int *index_device; CHECK(cudaMalloc((void **)&souce_device, B * M * sizeof(float))); CHECK(cudaMalloc((void **)&res_device, B * N * sizeof(float))); CHECK(cudaMalloc((void **)&index_device, B * N * sizeof(int))); CHECK(cudaMemcpy(souce_device, source, B * M * sizeof(float), cudaMemcpyHostToDevice)); CHECK(cudaMemcpy(res_device, res, B * N * sizeof(float), cudaMemcpyHostToDevice)); CHECK(cudaMemcpy(index_device, index, B * N * sizeof(int), cudaMemcpyHostToDevice)); int block_size = 1024; int grid_size = (B * N + block_size - 1) / block_size; float timeTaken = GET_CUDA_TIME((BatchGatherKernel<<<grid_size, block_size>>>(souce_device, index_device, res_device, N, M, B))); CHECK(cudaGetLastError()); CHECK(cudaDeviceSynchronize()); CHECK(cudaMemcpy(res, res_device, B * N * sizeof(float), cudaMemcpyDeviceToHost)); CHECK(cudaFree(souce_device)); CHECK(cudaFree(index_device)); CHECK(cudaDeviceSynchronize()); CHECK(cudaFree(res_device)); CHECK(cudaDeviceReset()); return timeTaken;}这里传入到CUDA之前,我们需要在Cython或者Python中把相关的数据压缩为一维,所以传入CUDA函数的是一个一维的指针。相比于单一的Gather操作,BatchGather中的几个输入含义有所变化,例如N表示的是单Batch的Index长度,M表示的是单Batch的源数组长度。
Cython文件
对于一个新的Batch函数来说,我们需要构建一个新的Cython调用函数wrapper.pyx:
# cythonize -i -f wrapper.pyximport numpy as npcimport numpy as npcimport cythoncdef extern from "<dlfcn.h>" nogil: void *dlopen(const char *, int) char *dlerror() void *dlsym(void *, const char *) int dlclose(void *) enum: RTLD_LAZYctypedef float (*GatherFunc)(float *source, int *index, float *res, int N, int M) noexcept nogilctypedef float (*BatchGatherFunc)(float *source, int *index, float *res, int N, int M, int B) noexcept nogilcdef void* handle = dlopen('/path/to/libcuindex.so', RTLD_LAZY)@cython.boundscheck(False)@cython.wraparound(False)cpdef float[:] cuda_gather(float[:] x, int[:] idx): cdef: GatherFunc Gather float timeTaken int N = idx.shape[0] int M = x.shape[0] float[:] res = np.zeros((N, ), dtype=np.float32) Gather = <GatherFunc>dlsym(handle, "Gather") timeTaken = Gather(&x[0], &idx[0], &res[0], N, M) print (timeTaken) return res@cython.boundscheck(False)@cython.wraparound(False)cpdef float[:] batch_cuda_gather(float[:] x, int[:] idx, int B): cdef: BatchGatherFunc BatchGather float timeTaken int N = idx.shape[0] // B int M = x.shape[0] // B float[:] res = np.zeros((B*N, ), dtype=np.float32) BatchGather = <BatchGatherFunc>dlsym(handle, "BatchGather") timeTaken = BatchGather(&x[0], &idx[0], &res[0], N, M, B) print (timeTaken) return reswhile not True: dlclose(handle)这里我们还是接受一维的数组,多引入一个Batch维度的参数B,其他的都是一样的。
Python调用文件
最后是用来调用的最上层Python端的代码test_gather.py:
import numpy as npnp.random.seed(0)from wrapper import batch_cuda_gatherB = 2M = 1024 * 1024 * 128N = 1024 * 1024x = np.random.random((M*B,)).astype(np.float32)idx = np.random.randint(0, M, (N*B,)).astype(np.int32)np_res = np.zeros((B, N), dtype=np.float32)for i in range(B): np_res = x.reshape((B,-1))[idx.reshape((B, -1))]np_res = np_res.reshape(-1)res = np.asarray(batch_cuda_gather(x, idx, B))print (res.shape)print ((res==np_res).sum())为了方便处理,在构建数据的时候,我们直接在生成数据阶段就生成一维的数据,然后直接调用Cython函数进行CUDA相关运算。
运行方法
首先将CUDA文件编译成动态链接库,使其可以在Cython中被调用。然后将Cython文件编译成动态链接库,使其可以在Python中被调用。最后运行Python代码即可:
$ nvcc -shared ./cuda_index.cu -Xcompiler -fPIC -o ./libcuindex.so$ cythonize -i -f wrapper.pyx$ python3 test_gather.py运行结果如下:
0.9606080055236816(2097152,)2097152这表示CUDA核函数部分的运行时长为0.96ms,输入的数组总长度为2097152,跟numpy版本的数组索引实现对比之后,得到2097152个相同的元素。也就是说,计算结果跟numpy的计算结果是一致的,以此来校验CUDA部分的运算结果。
总结概要
以学习CUDA为目的,接上一篇关于Cython与CUDA架构下的Gather算子实现,这里我们加一个Batch的维度,做一个BatchGather的简单实现。
版权声明
本文首发链接为:https://www.cnblogs.com/dechinphy/p/cython-cuda-batchgather.html
作者ID:DechinPhy
更多原著文章:https://www.cnblogs.com/dechinphy/
请博主喝咖啡:https://www.cnblogs.com/dechinphy/gallery/image/379634.html |
|