C++实战之OpenCL矩阵相乘优化（二）

前言

上一篇文章，分析了简单的矩阵相乘在opencl里面的优化kernel代码，每个work-item只负责计算结果矩阵的一个元素。下一步准备每次计算出结果矩阵的块元素，看看计算时间是如何。

这个矩阵系列参考国外一个大神的教程：

https://cnugteren.github.io/tutorial/pages/page4.html

有每个kernel的详解工程，还有github代码工程。

具体分析

这里引入opencl内存的概念：

比较常见的有:

全局内存 __global 修饰符，通常修饰指向一个数据类型的地址，

本地内存 __local 修饰符。local 定义的变量在一个work-group中是共享的，也就是说一个work-group中的所有work-item都可以通过本地内存来进行通信，

私有内存，private 每个work-item里的内部变量

常量内存， constant

下面是opencl的内存模型：



我们分析一下之前的矩阵相乘的一些性能：

__kernel void hello_kernel(__global const int *a,
__global const int *b,
__global int *result_matrix,int  result_matrix_row,
int  result_matrix_col,int  compute_size)
{
int row = get_global_id(0);
int col = get_global_id(1);

int sum = 0;
for(int i=0;i<compute_size;i++)
{
sum += a[row*compute_size+i] * b[i*result_matrix_col+col];
}

result_matrix[row*result_matrix_col+col] = sum;
}

首先在运行时总共有M*N个work-item同时执行，每个work-item中执行一个size为k（computesize）的for循环，循环里面每次分别load 数组a和b中的一个元素，所以综合起来一个kernel会有 M*N*K*2 个加载global内存的操作，乘以2是因为a，b两个数组。

其次每个work-item计算出结果矩阵的一个元素并保存，所以有M*N个对global内存的 store 的操作。

由上图的内存模型可知这种访问并不是最优的，再同一个work-group中我们可以定义local内存，来减少这种操作。

下面摘自国外博客的配图说明一下这次优化的原理：



其实就是把之前row*col的方式变成了 多个row和col相乘，究其本质还是对应元素相乘再相加。

这边的中心思想是引入work-group分块计算再相加，work-item的大小还是没变为M*N，不同的是在同一个work-group中把global数组A和B的对应值保存在local内存中，之后每个work-item在这个group中访问这个local变量速度会相对访问global较快，后面的大小为k的循环访问的也是local内存，所以在这个点上是被优化了。先看一下代码实现：

__kernel void hello_kernel(const __global int* A,
const __global int* B,
__global int* C, int M,  int N,  int K) {

// Thread identifiers
const int row = get_local_id(0); // Local row ID (max: TS)
const int col = get_local_id(1); // Local col ID (max: TS)
const int globalRow = TS*get_group_id(0) + row; // Row ID of C (0..M)
const int globalCol = TS*get_group_id(1) + col; // Col ID of C (0..N)

// Local memory to fit a tile of TS*TS elements of A and B
__local int Asub[TS][TS];
__local int Bsub[TS][TS];

// Initialise the accumulation register
int acc = 0;

// Loop over all tiles
const int numTiles = K/TS;
for (int t=0; t<numTiles; t++) {

// Load one tile of A and B into local memory
const int tiledRow = TS*t + row;
const int tiledCol = TS*t + col;
Asub[col][row] = A[tiledCol*M + globalRow];
Bsub[col][row] = B[globalCol*K + tiledRow];
printf("Asub[%d][%d]=A[%d]=%d\t",col,row,tiledCol*M + globalRow,A[tiledCol*M + globalRow]);
// Synchronise to make sure the tile is loaded
barrier(CLK_LOCAL_MEM_FENCE);

// Perform the computation for a single tile
for (int k=0; k<TS; k++) {
acc += Asub[k][row] * Bsub[col][k];
//printf("acc[%d][%d]=%d\n",k,row,Asub[k][row]);
}
printf("acc = %d\n",acc);
// Synchronise before loading the next tile
barrier(CLK_LOCAL_MEM_FENCE);
}

// Store the final result in C
C[globalCol*M + globalRow] = acc;
}

下面具体分析一下这个kernel在运行时的的运行情况：

线看一下cpu端的配置：

#define TS 16
size_t globalWorkSize[2];
globalWorkSize[0]= heightA;
globalWorkSize[1]=widthB;
size_t localWorkSize[2] ;
localWorkSize[0]= TS;
localWorkSize[1]= TS;
errNum = clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL,
globalWorkSize, localWorkSize,
0, NULL, NULL);

这边新加了localworksize的参数，并且设置大小为16，这里设置大小是有讲究的：

首先TS 必须为2的幂次方

也就是

localWorkSize[0]*localWorkSize[1] <= CL_DEVICE_MAX_WORK_GROUP_SIZE 要怎么知道自己机器的这个size呢？可以通过

size_t      maxWorkItemPerGroup;
clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(maxWorkItemPerGroup), &maxWorkItemPerGroup, NULL);
printf("maxWorkItemPerGroup: %zd\n", maxWorkItemPerGroup);

我这边打印的结果是256，也就是说我这边group的size最大只能设置到16.（16*16=256）

接下来看kernel的实现细节：

const int row = get_local_id(0); // Local row ID (max: TS)
const int col = get_local_id(1); // Local col ID (max: TS)

get_local_id 这一组操作主要是获取work-group中当前work-item所在的2d索引。

const int globalRow = TS*get_group_id(0) + row; // Row ID of C (0..M)
const int globalCol = TS*get_group_id(1) + col; // Col ID of C (0..N)

这个是通过当前work-item所在的group-id和自己在此group中的索引计算出，当前work-item在全局的索引。get_group_id是获取当前work-item所在work-group的id。

__local int Asub[TS][TS];
__local int Bsub[TS][TS];

定义local 内存，在同一个work-group对所有work-item可见。

const int numTiles = K/TS;

这个是一个work-item 需要循环计算的group的数量，这边可以知道，K也要为TS的倍数才行。

for (int t=0; t<numTiles; t++) {

// Load one tile of A and B into local memory
const int tiledRow = TS*t + row;
const int tiledCol = TS*t + col;
Asub[col][row] = A[tiledCol*M + globalRow];
Bsub[col][row] = B[globalCol*K + tiledRow];
printf("Asub[%d][%d]=A[%d]=%d\t",col,row,tiledCol*M + globalRow,A[tiledCol*M + globalRow]);
// Synchronise to make sure the tile is loaded
barrier(CLK_LOCAL_MEM_FENCE);

// Perform the computation for a single tile
for (int k=0; k<TS; k++) {
acc += Asub[k][row] * Bsub[col][k];
//printf("acc[%d][%d]=%d\n",k,row,Asub[k][row]);
}
printf("acc = %d\n",acc);
// Synchronise before loading the next tile
barrier(CLK_LOCAL_MEM_FENCE);
}

主要核心就是这个for循环，循环一进来首先计算此时work-item在当前块的索引位置

然后开始从global内存中把数组A和B 中每块大小为16*16的值储存到本地内存上。用串行的思想去看这段代码，会比较困难。这边有个barrier(CLK_LOCAL_MEM_FENCE); 关键语句，作用就是用来再work-group中同步所有work-item。也就是说只有当前work-group中所有的work-item到达这个点，换个意思就是要保证Asub和Bsub两个大小为16*16大小本地内存被赋值完毕，16*16个work-item必须全部达到这个点，才会继续下去执行。

接下去是一个k循环，前面已经得到了A和B的两个子矩阵并被保存在本地内存中，通过行列相乘相加得到一个子矩阵上的结果，一个work-item一样也只计算出一个元素，一个work-group计算出结果矩阵对应的子矩阵全部元素。

接下去又是一个同步：这个同步是保证这一个分块或者说group全部计算完毕，再去load下一个分块。一个大循环结束后，就计算出结果矩阵对应的一个元素了，把它保存在global内存中：

// Store the final result in C
C[globalCol*M + globalRow] = acc;

下面是主代码：

//
//  main.cpp
//  OpenCL
//
//  Created by wmy on 2017/9/19.
//  Copyright © 2017年 wmy. All rights reserved.
//

#include <OpenCL/OpenCL.h>
#include <iostream>
#include <fstream>
#include <sstream>
#include <unistd.h>
#include <sys/time.h>
#include<time.h>
#include<stdio.h>
#include<stdlib.h>
#include <mach/mach_time.h>
#include <boost/algorithm/string.hpp>

using namespace std;

//const int ARRAY_SIZE = 100000;

//4*3---3*5

const int midle = 32;
const int heightA = 32;

const int widthB = 32;
//const int heightB = 3;

//一、 选择OpenCL平台并创建一个上下文
cl_context CreateContext()
{
cl_int errNum;
cl_uint numPlatforms;
cl_platform_id firstPlatformId;
cl_context context = NULL;

//选择可用的平台中的第一个
errNum = clGetPlatformIDs(1, &firstPlatformId, &numPlatforms);
if (errNum != CL_SUCCESS || numPlatforms <= 0)
{
std::cerr << "Failed to find any OpenCL platforms." << std::endl;
return NULL;
}

//创建一个OpenCL上下文环境
cl_context_properties contextProperties[] =
{
CL_CONTEXT_PLATFORM,
(cl_context_properties)firstPlatformId,
0
};
context = clCreateContextFromType(contextProperties, CL_DEVICE_TYPE_GPU,
NULL, NULL, &errNum);

return context;
}

//二、 创建设备并创建命令队列
cl_command_queue CreateCommandQueue(cl_context context, cl_device_id *device)
{
cl_int errNum;
cl_device_id *devices;
cl_command_queue commandQueue = NULL;
size_t deviceBufferSize = -1;

// 获取设备缓冲区大小
errNum = clGetContextInfo(context, CL_CONTEXT_DEVICES, 0, NULL, &deviceBufferSize);

if (deviceBufferSize <= 0)
{
std::cerr << "No devices available.";
return NULL;
}

// 为设备分配缓存空间
devices = new cl_device_id[deviceBufferSize / sizeof(cl_device_id)];
printf("deviceBufferSize / sizeof(cl_device_id)=%ld\n",deviceBufferSize / sizeof(cl_device_id));
errNum = clGetContextInfo(context, CL_CONTEXT_DEVICES, deviceBufferSize, devices, NULL);

//    size_t      valueSize;
//    clGetDeviceInfo(devices[0], CL_DEVICE_NAME, 0, NULL, &valueSize);
//    char* value = (char*) malloc(valueSize);
//    clGetDeviceInfo(devices[0], CL_DEVICE_NAME, valueSize, value, NULL);
//    printf("Device1 Name: %s\n", value);
//    free(value);
//
//    clGetDeviceInfo(devices[1], CL_DEVICE_NAME, 0, NULL, &valueSize);
//    value = (char*) malloc(valueSize);
//    clGetDeviceInfo(devices[1], CL_DEVICE_NAME, valueSize, value, NULL);
//    printf("Device2 Name: %s\n", value);
//    free(value);

//选取可用设备中的第一个
commandQueue = clCreateCommandQueue(context, devices[1], 0, NULL);

*device = devices[0];
delete[] devices;
return commandQueue;
}

// 三、创建和构建程序对象
cl_program CreateProgram(cl_context context, cl_device_id device, const char* fileName)
{
cl_int errNum;
cl_program program;

std::ifstream kernelFile(fileName, std::ios::in);
if (!kernelFile.is_open())
{
std::cerr << "Failed to open file for reading: " << fileName << std::endl;
return NULL;
}

std::ostringstream oss;
oss << kernelFile.rdbuf();

std::string srcStdStr = oss.str();
const char *srcStr = srcStdStr.c_str();
program = clCreateProgramWithSource(context, 1,
(const char**)&srcStr,
NULL, NULL);

errNum = clBuildProgram(program, 0, NULL, NULL, NULL, NULL);

return program;
}

//创建和构建程序对象
bool CreateMemObjects(cl_context context, cl_mem memObjects[3],
int *a, int *b)
{
memObjects[0] = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(int) * midle*heightA, a, NULL);
memObjects[1] = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(int) * widthB*midle, b, NULL);
memObjects[2] = clCreateBuffer(context, CL_MEM_READ_WRITE,
sizeof(int) * widthB*heightA, NULL, NULL);
return true;
}

// 释放OpenCL资源
void Cleanup(cl_context context, cl_command_queue commandQueue,
cl_program program, cl_kernel kernel, cl_mem memObjects[3])
{
for (int i = 0; i < 3; i++)
{
if (memObjects[i] != 0)
clReleaseMemObject(memObjects[i]);
}
if (commandQueue != 0)
clReleaseCommandQueue(commandQueue);

if (kernel != 0)
clReleaseKernel(kernel);

if (program != 0)
clReleaseProgram(program);

if (context != 0)
clReleaseContext(context);
}

void checkError(cl_int error, int line) {
if (error != CL_SUCCESS) {
switch (error) {
case CL_DEVICE_NOT_FOUND:                 printf("-- Error at %d:  Device not found.\n", line); break;
case CL_DEVICE_NOT_AVAILABLE:             printf("-- Error at %d:  Device not available\n", line); break;
case CL_COMPILER_NOT_AVAILABLE:           printf("-- Error at %d:  Compiler not available\n", line); break;
case CL_MEM_OBJECT_ALLOCATION_FAILURE:    printf("-- Error at %d:  Memory object allocation failure\n", line); break;
case CL_OUT_OF_RESOURCES:                 printf("-- Error at %d:  Out of resources\n", line); break;
case CL_OUT_OF_HOST_MEMORY:               printf("-- Error at %d:  Out of host memory\n", line); break;
case CL_PROFILING_INFO_NOT_AVAILABLE:     printf("-- Error at %d:  Profiling information not available\n", line); break;
case CL_MEM_COPY_OVERLAP:                 printf("-- Error at %d:  Memory copy overlap\n", line); break;
case CL_IMAGE_FORMAT_MISMATCH:            printf("-- Error at %d:  Image format mismatch\n", line); break;
case CL_IMAGE_FORMAT_NOT_SUPPORTED:       printf("-- Error at %d:  Image format not supported\n", line); break;
case CL_BUILD_PROGRAM_FAILURE:            printf("-- Error at %d:  Program build failure\n", line); break;
case CL_MAP_FAILURE:                      printf("-- Error at %d:  Map failure\n", line); break;
case CL_INVALID_VALUE:                    printf("-- Error at %d:  Invalid value\n", line); break;
case CL_INVALID_DEVICE_TYPE:              printf("-- Error at %d:  Invalid device type\n", line); break;
case CL_INVALID_PLATFORM:                 printf("-- Error at %d:  Invalid platform\n", line); break;
case CL_INVALID_DEVICE:                   printf("-- Error at %d:  Invalid device\n", line); break;
case CL_INVALID_CONTEXT:                  printf("-- Error at %d:  Invalid context\n", line); break;
case CL_INVALID_QUEUE_PROPERTIES:         printf("-- Error at %d:  Invalid queue properties\n", line); break;
case CL_INVALID_COMMAND_QUEUE:            printf("-- Error at %d:  Invalid command queue\n", line); break;
case CL_INVALID_HOST_PTR:                 printf("-- Error at %d:  Invalid host pointer\n", line); break;
case CL_INVALID_MEM_OBJECT:               printf("-- Error at %d:  Invalid memory object\n", line); break;
case CL_INVALID_IMAGE_FORMAT_DESCRIPTOR:  printf("-- Error at %d:  Invalid image format descriptor\n", line); break;
case CL_INVALID_IMAGE_SIZE:               printf("-- Error at %d:  Invalid image size\n", line); break;
case CL_INVALID_SAMPLER:                  printf("-- Error at %d:  Invalid sampler\n", line); break;
case CL_INVALID_BINARY:                   printf("-- Error at %d:  Invalid binary\n", line); break;
case CL_INVALID_BUILD_OPTIONS:            printf("-- Error at %d:  Invalid build options\n", line); break;
case CL_INVALID_PROGRAM:                  printf("-- Error at %d:  Invalid program\n", line); break;
case CL_INVALID_PROGRAM_EXECUTABLE:       printf("-- Error at %d:  Invalid program executable\n", line); break;
case CL_INVALID_KERNEL_NAME:              printf("-- Error at %d:  Invalid kernel name\n", line); break;
case CL_INVALID_KERNEL_DEFINITION:        printf("-- Error at %d:  Invalid kernel definition\n", line); break;
case CL_INVALID_KERNEL:                   printf("-- Error at %d:  Invalid kernel\n", line); break;
case CL_INVALID_ARG_INDEX:                printf("-- Error at %d:  Invalid argument index\n", line); break;
case CL_INVALID_ARG_VALUE:                printf("-- Error at %d:  Invalid argument value\n", line); break;
case CL_INVALID_ARG_SIZE:                 printf("-- Error at %d:  Invalid argument size\n", line); break;
case CL_INVALID_KERNEL_ARGS:              printf("-- Error at %d:  Invalid kernel arguments\n", line); break;
case CL_INVALID_WORK_DIMENSION:           printf("-- Error at %d:  Invalid work dimensionsension\n", line); break;
case CL_INVALID_WORK_GROUP_SIZE:          printf("-- Error at %d:  Invalid work group size\n", line); break;
case CL_INVALID_WORK_ITEM_SIZE:           printf("-- Error at %d:  Invalid work item size\n", line); break;
case CL_INVALID_GLOBAL_OFFSET:            printf("-- Error at %d:  Invalid global offset\n", line); break;
case CL_INVALID_EVENT_WAIT_LIST:          printf("-- Error at %d:  Invalid event wait list\n", line); break;
case CL_INVALID_EVENT:                    printf("-- Error at %d:  Invalid event\n", line); break;
case CL_INVALID_OPERATION:                printf("-- Error at %d:  Invalid operation\n", line); break;
case CL_INVALID_GL_OBJECT:                printf("-- Error at %d:  Invalid OpenGL object\n", line); break;
case CL_INVALID_BUFFER_SIZE:              printf("-- Error at %d:  Invalid buffer size\n", line); break;
case CL_INVALID_MIP_LEVEL:                printf("-- Error at %d:  Invalid mip-map level\n", line); break;
case -1024:                               printf("-- Error at %d:  *clBLAS* Functionality is not implemented\n", line); break;
case -1023:                               printf("-- Error at %d:  *clBLAS* Library is not initialized yet\n", line); break;
case -1022:                               printf("-- Error at %d:  *clBLAS* Matrix A is not a valid memory object\n", line); break;
case -1021:                               printf("-- Error at %d:  *clBLAS* Matrix B is not a valid memory object\n", line); break;
case -1020:                               printf("-- Error at %d:  *clBLAS* Matrix C is not a valid memory object\n", line); break;
case -1019:                               printf("-- Error at %d:  *clBLAS* Vector X is not a valid memory object\n", line); break;
case -1018:                               printf("-- Error at %d:  *clBLAS* Vector Y is not a valid memory object\n", line); break;
case -1017:                               printf("-- Error at %d:  *clBLAS* An input dimension (M,N,K) is invalid\n", line); break;
case -1016:                               printf("-- Error at %d:  *clBLAS* Leading dimension A must not be less than the size of the first dimension\n", line); break;
case -1015:                               printf("-- Error at %d:  *clBLAS* Leading dimension B must not be less than the size of the second dimension\n", line); break;
case -1014:                               printf("-- Error at %d:  *clBLAS* Leading dimension C must not be less than the size of the third dimension\n", line); break;
case -1013:                               printf("-- Error at %d:  *clBLAS* The increment for a vector X must not be 0\n", line); break;
case -1012:                               printf("-- Error at %d:  *clBLAS* The increment for a vector Y must not be 0\n", line); break;
case -1011:                               printf("-- Error at %d:  *clBLAS* The memory object for Matrix A is too small\n", line); break;
case -1010:                               printf("-- Error at %d:  *clBLAS* The memory object for Matrix B is too small\n", line); break;
case -1009:                               printf("-- Error at %d:  *clBLAS* The memory object for Matrix C is too small\n", line); break;
case -1008:                               printf("-- Error at %d:  *clBLAS* The memory object for Vector X is too small\n", line); break;
case -1007:                               printf("-- Error at %d:  *clBLAS* The memory object for Vector Y is too small\n", line); break;
case -1001:                               printf("-- Error at %d:  Code -1001: no GPU available?\n", line); break;
default:                                  printf("-- Error at %d:  Unknown with code %d\n", line, error);
}
exit(1);
}
}
#define TIMES 10
#define TS 16

int main(int argc, char** argv)
{

cl_context context = 0;
cl_command_queue commandQueue = 0;
cl_program program = 0;
cl_device_id device = 0;
cl_kernel kernel = 0;
cl_mem memObjects[3] = { 0, 0, 0 };
cl_int errNum;
// uint64_t t1,t2,t3;
clock_t t1,t2,t3,t4;

const char* filename = "/Users/wangmingyong/Projects/OpenCL/OpenCL/HelloWorld.cl";
// 一、选择OpenCL平台并创建一个上下文
context = CreateContext();

// 二、 创建设备并创建命令队列
commandQueue = CreateCommandQueue(context, &device);

size_t      maxWorkItemPerGroup;
clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(maxWorkItemPerGroup), &maxWorkItemPerGroup, NULL);
printf("maxWorkItemPerGroup: %zd\n", maxWorkItemPerGroup);
size_t      valueSize;
clGetDeviceInfo(device, CL_DEVICE_NAME, 0, NULL, &valueSize);
char* value = (char*) malloc(valueSize);
clGetDeviceInfo(device, CL_DEVICE_NAME, valueSize, value, NULL);
printf("Device Name: %s\n", value);
free(value);

//创建和构建程序对象
program = CreateProgram(context, device, filename);//"HelloWorld.cl");

// 四、 创建OpenCL内核并分配内存空间
kernel = clCreateKernel(program, "hello_kernel", NULL);

//创建要处理的数据
int result[widthB*heightA]{0};
int a[midle*heightA];
int b[widthB*midle];
for (int i = 0; i < heightA; i++)
{
for(int j = 0;j < midle;j++)
{
a[i*midle+j]=2;//10.0f * ((int) rand() / (int) RAND_MAX);
}

}

for (int k = 0; k < midle; k++)
{
for(int m = 0;m < widthB;m++)
{
b[k*widthB+m]=3;//10.0f * ((int) rand() / (int) RAND_MAX);
}

}

t1 = clock();  //mach_absolute_time();
//printf("t1 = %.8f\n",(double)t1);

for(int tt=0;tt<TIMES;tt++){
for(int l=0;l<heightA;l++){
for(int n = 0;n<widthB;n++){
for(int q=0;q<midle;q++){
result[l*widthB+n] +=a [l*midle+q]*b[q*widthB+n];

}
//std::cout<<"r = "<<result[l*widthB+n]<<std::endl;
}
}
}
t2 = clock(); //mach_absolute_time();
//printf("t2 = %.8f\n",(double)t2);

//创建内存对象
if (!CreateMemObjects(context, memObjects, a, b))
{
Cleanup(context, commandQueue, program, kernel, memObjects);
return 1;
}

// 五、 设置内核数据并执行内核
errNum = clSetKernelArg(kernel, 0, sizeof(cl_mem), &memObjects[0]);
errNum |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &memObjects[1]);
errNum |= clSetKernelArg(kernel, 2, sizeof(cl_mem), &memObjects[2]);
errNum |= clSetKernelArg(kernel, 3, sizeof(int), &heightA);
errNum |= clSetKernelArg(kernel, 4, sizeof(int), &widthB);
errNum |= clSetKernelArg(kernel, 5, sizeof(int), &midle);

size_t globalWorkSize[2];
globalWorkSize[0]= heightA;
globalWorkSize[1]=widthB;
size_t localWorkSize[2] ;
localWorkSize[0]= TS;
localWorkSize[1]= TS;
t3 = clock();
for(int run=0;run<TIMES;run++){
errNum = clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL,
globalWorkSize, localWorkSize,
0, NULL, NULL);
checkError(errNum,__LINE__);
if(errNum == CL_SUCCESS)
cout<<"enqueue success!"<<endl;
else
printf("errNum= %d\n",errNum);

//mach_absolute_time();
// 六、 读取执行结果并释放OpenCL资源
errNum = clEnqueueReadBuffer(commandQueue, memObjects[2], CL_TRUE,
0, widthB*heightA * sizeof(int), result,
0, NULL, NULL);
//    for(int p=0;p<20;p++){
//        cout<<"new ="<<result[p];
//    }

}

t4 = clock();

printf("cpu t = %.8f\n",(float)(t2-t1)/CLOCKS_PER_SEC/TIMES);
printf("gpu t = %.8f \n",(double)(t4-t3)/CLOCKS_PER_SEC/TIMES);

std::cout << std::endl;
std::cout << "Executed program succesfully." << std::endl;
getchar();
Cleanup(context, commandQueue, program, kernel, memObjects);

return 0;
}

下面是时间性能分析：

维度	cpu	gpu
32*32	0.00010410	0.00040870
128*128	0.00676160	0.00040980
512*512	0.52244419	0.00058840

这是跟之前的kernel性能相比：

维度	gpu1	gpu2
32*32	0.00029130	0.00040870
128*128	0.00036250	0.00040980
512*512	0.00056370	0.00058840

貌似时间都差不多，我这边把readbuffer的操作去掉发现时间少了很多，但是跟前一个kernel的都在同一个数量级差不多的时间，我这边维数改到1024程序就会报错，所以验证不了高维度的性能。后续跟踪下程序为什么限制到了1024，不知是否是机器的原因。

kernel1

512 gpu t = 0.00000460

128 gpu t = 0.00000400

32 0.00000360

kernel2

512 0.00000320

128 gpu t = 0.00000310

32 0.00000370