C++ OpenCL-结果与CPU版本不同
我想使用openCl在两个图像之间进行像素值比较。一幅图像应通过变换矩阵进行“变换” 1）我面临一个问题，openCL版本的结果与CPU版本不一样。 像素值的差异在我的示例图像（imageA：所有像素都是5，imagesB：所有像素都是6）中始终是1，因此1000*1000像素的总数应该是1000000。 CPU版本总是正确的，但openCL版本总是有点不精确，并且有时也会有所不同（例如998895或998829） 2）我遇到的另一个问题是运行时，因为将两个比较像素的差值添加到结果变量需要很长时间。但我的感觉是，它可以通过另一种记忆布局来解决 对我遇到的问题有什么想法吗？ 也许使用二维工作集的方式也会导致错误 谢谢你和亲切的问候 亨德里克 以下是内核： 基本上，它得到了两幅图像和700个变换矩阵（目前所有的矩阵都表示身份）C++ OpenCL-结果与CPU版本不同,c++,image-processing,opencl,C++,Image Processing,Opencl,我想使用openCl在两个图像之间进行像素值比较。一幅图像应通过变换矩阵进行“变换” 1）我面临一个问题，openCL版本的结果与CPU版本不一样。 像素值的差异在我的示例图像（imageA：所有像素都是5，imagesB：所有像素都是6）中始终是1，因此1000*1000像素的总数应该是1000000。 CPU版本总是正确的，但openCL版本总是有点不精确，并且有时也会有所不同（例如998895或998829） 2）我遇到的另一个问题是运行时，因为将两个比较像素的差值添加到结果变量需要很长时
\u内核无效合规性(
__只读图像2D\t图像A，
__只读图像2D\t图像B，
__常数浮点*矩阵，
__全局整数*结果
)
{
对于（int i=0；i<700；i++）
{
大小\u t x=获取全局\u id（0）；
大小y=获取全局id（1）；
float t1=矩阵[0+i*6]；
浮点t2=矩阵[1+i*6]；
浮点数t3=矩阵[2+i*6]；
浮点数t4=矩阵[3+i*6]；
浮点数t5=矩阵[4+i*6]；
浮点数t6=矩阵[5+i*6]；
//计算比较像素的其他坐标
int x_new=x*t1+y*t2+1*t3；
int y_new=x*t4+y*t5+1*t6；
inta=（读取图像i（图像a，（int2）（x，y））.x）；
intb=（读取图像i（图像b，（int2）（x_新，y_新））.x）；
int diff=b-a；
//将两个比较像素中的每一个添加到结果中
结果[i]+=diff；
}
}
#define __CL_ENABLE_EXCEPTIONS
#include <CL/cl.hpp>
#include <utility>
#include <iostream>
#include <fstream>
#include <string>
#include <chrono>
#include <opencv2\core.hpp>
#include <opencv2\imgproc.hpp>
#include <opencv2\highgui.hpp>
using namespace std;
int main(int argc, char** argv) {
//700 transformation matrices
int numberMatrices = 700;
bool opencl = true;
//iamge width
int width = 1000;
//image height
int height = 1000;
//total number of pixels of one image
int size = width*height;
// Create two example images
const int LIST_SIZE = size;
int *imageA = new int[LIST_SIZE];
int *imageB = new int[LIST_SIZE];
for (int i = 0; i < LIST_SIZE; i++) {
//every pixel value of imageA is 5
imageA[i] = 5;
//every pixel value of imageA is 6
imageB[i] = 6;
}
//creation of n transformation matrices
const int MATRIX_SIZE = 6* numberMatrices;
float *indi = new float[MATRIX_SIZE];
//all the matrices are the same
for (int i = 0; i < numberMatrices; i++)
{
//identity matrix
indi[0 + i * 6] = 1;
indi[1 + i * 6] = 0;
indi[2 + i * 6] = 0;
indi[3 + i * 6] = 0;
indi[4 + i * 6] = 1;
indi[5 + i * 6] = 0;
}
//array to save the results of the comparison
const int RESULT_SIZE = numberMatrices;
int *result = new int[RESULT_SIZE];
if (opencl)
{
try {
// Get available platforms
vector<cl::Platform> platforms;
cl::Platform::get(&platforms);
std::cerr << "Platform number is: " << platforms.size() << std::endl;
std::string platformVendor;
platforms[0].getInfo((cl_platform_info)CL_PLATFORM_VENDOR, &platformVendor);
std::cerr << "Platform is by: " << platformVendor << "\n";
// Select the default platform and create a context using this platform and the GPU
cl_context_properties cps[3] = {
CL_CONTEXT_PLATFORM,
(cl_context_properties)(platforms[0])(),
0
};
cl::Context context(CL_DEVICE_TYPE_CPU, cps);
vector<cl::ImageFormat> format;
context.getSupportedImageFormats(CL_MEM_READ_ONLY, CL_MEM_OBJECT_IMAGE2D, &format);
/* for (int i = 0; i < format.size(); i++)
{
cout << "Channel Data Type: " << format.at(i).image_channel_data_type
<< " Channel order: " << format.at(i).image_channel_order << endl;
}*/
// Get a list of devices on this platform
vector<cl::Device> devices = context.getInfo<CL_CONTEXT_DEVICES>();
for (int i = 0; i < devices.size(); i++)
{
cout << "Device: " << devices.at(i).getInfo<CL_DEVICE_NAME>() << endl;
cout << "DOUBLE FP: " << devices.at(i).getInfo<CL_DEVICE_DOUBLE_FP_CONFIG>() << endl;
cout << "Image Max Height: " << devices.at(i).getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>() << endl;
cout << "Image Support: " << devices.at(i).getInfo<CL_DEVICE_IMAGE_SUPPORT>() << endl;
cout << "Local Memory Size: " << devices.at(i).getInfo<CL_DEVICE_LOCAL_MEM_SIZE>() << endl;
cout << "Clock Frequency: " << devices.at(i).getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>() << endl;
cout << "CUs: " << devices.at(i).getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>() << endl;
cout << "Driver: " << devices.at(i).getInfo<CL_DRIVER_VERSION>() << endl;
cout << "Version: " << devices.at(i).getInfo<CL_DEVICE_VERSION>() << endl;
cout << "Work Group: " << devices.at(i).getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>() << endl;
cout << "Items: " << devices.at(i).getInfo<CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS>();
cout << endl;
}
//Create opencl image
cl::Image2D clImage_A = cl::Image2D(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, cl::ImageFormat(CL_RGBA, CL_UNSIGNED_INT8), (size_t)width, (size_t)height, 0, imageA);
cl::Image2D clImage_B = cl::Image2D(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, cl::ImageFormat(CL_RGBA, CL_UNSIGNED_INT8), (size_t)width, (size_t)height, 0, imageB);
// Create a command queue and use the first device
cl::CommandQueue queue = cl::CommandQueue(context, devices[0]);
// Read kernel source file
std::ifstream sourceFile("difference.cl");
std::string sourceCode(
std::istreambuf_iterator<char>(sourceFile),
(std::istreambuf_iterator<char>()));
cl::Program::Sources source(1, std::make_pair(sourceCode.c_str(), sourceCode.length() + 1));
// Make program of the source code in the context
cl::Program program = cl::Program(context, source);
// Build program for these specific devices
program.build(devices);
// Make kernel
cl::Kernel kernel(program, "compliance");
// Create memory buffers
cl::Buffer buffer_matrix = cl::Buffer(context, CL_MEM_READ_ONLY, LIST_SIZE * sizeof(float));
cl::Buffer buffer_result = cl::Buffer(context, CL_MEM_READ_WRITE, RESULT_SIZE * sizeof(int));
// Copy list of results to the memory buffers
queue.enqueueWriteBuffer(buffer_matrix, CL_TRUE, 0, MATRIX_SIZE * sizeof(float), indi);
// Set arguments to kernel
kernel.setArg(0, clImage_A);
kernel.setArg(1, clImage_B);
kernel.setArg(2, buffer_matrix);
kernel.setArg(3, buffer_result);
cl::Event event;
std::cout << "Start OpenCL processing.." << endl;
chrono::high_resolution_clock::time_point t1 = chrono::high_resolution_clock::now();
// Run the kernel n-times on specific ND range
for (int i = 0; i < 1; i++)
{
queue.enqueueNDRangeKernel(
kernel,
cl::NullRange,
cl::NDRange((size_t)width, (size_t)height),
cl::NDRange(1, 1),
NULL,
&event);
cout << i << " ";
event.wait();
}
chrono::high_resolution_clock::time_point t2 = chrono::high_resolution_clock::now();
auto duration_opencl = std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1).count();
std::cout << "OpenCL processing done.." << endl;
std::cout << "Start CPU Processing.." << endl;
// Read buffer_result into result
queue.enqueueReadBuffer(buffer_result, CL_TRUE, 0, RESULT_SIZE * sizeof(int), result);
//cpu version to calculate the difference between the two arryays
t1 = chrono::high_resolution_clock::now();
int different = 0;
int x_new;
int x;
for (int i = 0; i < numberMatrices; i++)
{
different = 0;
for (int n = 0; n < LIST_SIZE; n++)
{
x = imageA[n];
x_new = x;;
int a = imageA[x];
int b = imageB[x_new];
int diff = imageB[x_new] - imageA[x];
different += diff;
}
}
t2 = chrono::high_resolution_clock::now();
auto duration_cpu = std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1).count();
std::cout << "CPU processing done.." << endl;
//output of the results
std::cout << "opencl: diff " << result[0] << endl;
std::cout << "Runtime opencl: " << duration_opencl << endl;
std::cout << "CPU: diff " << different << endl;
std::cout << "Runtime CPU: " << duration_cpu << endl;
double times = (double)duration_cpu / (double)duration_opencl;
std::cout << "OpenCL is " << times << " times faster!!!" << endl;
char c;
std::cin >> c;
}
catch (cl::Error error) {
std::cout << error.what() << "(" << error.err() << ")" << std::endl;
char c;
std::cin >> c;
}
}
return 0;
}
\define\u CL\u ENABLE\u异常
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
使用名称空间std；
int main（int argc，字符**argv）{
//700变换矩阵
整数米=700；
bool opencl=true；
//直径宽度
整数宽度=1000；
//像高
整数高度=1000；
//一幅图像的总像素数
整数大小=宽度*高度；
//创建两个示例图像
const int LIST_SIZE=大小；
int*imageA=新的int[列表大小]；
int*imageB=新int[列表大小]；
对于（int i=0；i结果[i]中的情况下，你有竞争条件吗+=diff；
在OpenCL代码中，您的程序在每个工作项中同时执行此操作？因此，这可能是一个问题。浮点算法在不同的平台上有所不同。您很可能看到了由OpenCL
编译器执行的特定于硬件的优化的效果。据我所知，禁用优化使用-cl opt disable
在这种情况下没有帮助。是的，这可能是一个问题。这也是运行速度慢的原因。您知道保存每个像素对差异总和的另一种方法吗？您可以使用我所知的OpenCL的不同方法：原子添加（结果[I]，差异）
您可以防止在同一个内存位置同时写入。谢谢……虽然我还没有成功实现原子添加。这个问题被称为并行还原。我发现的一个术语是还原。这里有一个指向ppt的链接。我没有做过，但可能有人也有同样的问题：
#define __CL_ENABLE_EXCEPTIONS
#include <CL/cl.hpp>
#include <utility>
#include <iostream>
#include <fstream>
#include <string>
#include <chrono>
#include <opencv2\core.hpp>
#include <opencv2\imgproc.hpp>
#include <opencv2\highgui.hpp>
using namespace std;
int main(int argc, char** argv) {
//700 transformation matrices
int numberMatrices = 700;
bool opencl = true;
//iamge width
int width = 1000;
//image height
int height = 1000;
//total number of pixels of one image
int size = width*height;
// Create two example images
const int LIST_SIZE = size;
int *imageA = new int[LIST_SIZE];
int *imageB = new int[LIST_SIZE];
for (int i = 0; i < LIST_SIZE; i++) {
//every pixel value of imageA is 5
imageA[i] = 5;
//every pixel value of imageA is 6
imageB[i] = 6;
}
//creation of n transformation matrices
const int MATRIX_SIZE = 6* numberMatrices;
float *indi = new float[MATRIX_SIZE];
//all the matrices are the same
for (int i = 0; i < numberMatrices; i++)
{
//identity matrix
indi[0 + i * 6] = 1;
indi[1 + i * 6] = 0;
indi[2 + i * 6] = 0;
indi[3 + i * 6] = 0;
indi[4 + i * 6] = 1;
indi[5 + i * 6] = 0;
}
//array to save the results of the comparison
const int RESULT_SIZE = numberMatrices;
int *result = new int[RESULT_SIZE];
if (opencl)
{
try {
// Get available platforms
vector<cl::Platform> platforms;
cl::Platform::get(&platforms);
std::cerr << "Platform number is: " << platforms.size() << std::endl;
std::string platformVendor;
platforms[0].getInfo((cl_platform_info)CL_PLATFORM_VENDOR, &platformVendor);
std::cerr << "Platform is by: " << platformVendor << "\n";
// Select the default platform and create a context using this platform and the GPU
cl_context_properties cps[3] = {
CL_CONTEXT_PLATFORM,
(cl_context_properties)(platforms[0])(),
0
};
cl::Context context(CL_DEVICE_TYPE_CPU, cps);
vector<cl::ImageFormat> format;
context.getSupportedImageFormats(CL_MEM_READ_ONLY, CL_MEM_OBJECT_IMAGE2D, &format);
/* for (int i = 0; i < format.size(); i++)
{
cout << "Channel Data Type: " << format.at(i).image_channel_data_type
<< " Channel order: " << format.at(i).image_channel_order << endl;
}*/
// Get a list of devices on this platform
vector<cl::Device> devices = context.getInfo<CL_CONTEXT_DEVICES>();
for (int i = 0; i < devices.size(); i++)
{
cout << "Device: " << devices.at(i).getInfo<CL_DEVICE_NAME>() << endl;
cout << "DOUBLE FP: " << devices.at(i).getInfo<CL_DEVICE_DOUBLE_FP_CONFIG>() << endl;
cout << "Image Max Height: " << devices.at(i).getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>() << endl;
cout << "Image Support: " << devices.at(i).getInfo<CL_DEVICE_IMAGE_SUPPORT>() << endl;
cout << "Local Memory Size: " << devices.at(i).getInfo<CL_DEVICE_LOCAL_MEM_SIZE>() << endl;
cout << "Clock Frequency: " << devices.at(i).getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>() << endl;
cout << "CUs: " << devices.at(i).getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>() << endl;
cout << "Driver: " << devices.at(i).getInfo<CL_DRIVER_VERSION>() << endl;
cout << "Version: " << devices.at(i).getInfo<CL_DEVICE_VERSION>() << endl;
cout << "Work Group: " << devices.at(i).getInfo<CL_DEVICE_MAX_WORK_GROUP_SIZE>() << endl;
cout << "Items: " << devices.at(i).getInfo<CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS>();
cout << endl;
}
//Create opencl image
cl::Image2D clImage_A = cl::Image2D(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, cl::ImageFormat(CL_RGBA, CL_UNSIGNED_INT8), (size_t)width, (size_t)height, 0, imageA);
cl::Image2D clImage_B = cl::Image2D(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, cl::ImageFormat(CL_RGBA, CL_UNSIGNED_INT8), (size_t)width, (size_t)height, 0, imageB);
// Create a command queue and use the first device
cl::CommandQueue queue = cl::CommandQueue(context, devices[0]);
// Read kernel source file
std::ifstream sourceFile("difference.cl");
std::string sourceCode(
std::istreambuf_iterator<char>(sourceFile),
(std::istreambuf_iterator<char>()));
cl::Program::Sources source(1, std::make_pair(sourceCode.c_str(), sourceCode.length() + 1));
// Make program of the source code in the context
cl::Program program = cl::Program(context, source);
// Build program for these specific devices
program.build(devices);
// Make kernel
cl::Kernel kernel(program, "compliance");
// Create memory buffers
cl::Buffer buffer_matrix = cl::Buffer(context, CL_MEM_READ_ONLY, LIST_SIZE * sizeof(float));
cl::Buffer buffer_result = cl::Buffer(context, CL_MEM_READ_WRITE, RESULT_SIZE * sizeof(int));
// Copy list of results to the memory buffers
queue.enqueueWriteBuffer(buffer_matrix, CL_TRUE, 0, MATRIX_SIZE * sizeof(float), indi);
// Set arguments to kernel
kernel.setArg(0, clImage_A);
kernel.setArg(1, clImage_B);
kernel.setArg(2, buffer_matrix);
kernel.setArg(3, buffer_result);
cl::Event event;
std::cout << "Start OpenCL processing.." << endl;
chrono::high_resolution_clock::time_point t1 = chrono::high_resolution_clock::now();
// Run the kernel n-times on specific ND range
for (int i = 0; i < 1; i++)
{
queue.enqueueNDRangeKernel(
kernel,
cl::NullRange,
cl::NDRange((size_t)width, (size_t)height),
cl::NDRange(1, 1),
NULL,
&event);
cout << i << " ";
event.wait();
}
chrono::high_resolution_clock::time_point t2 = chrono::high_resolution_clock::now();
auto duration_opencl = std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1).count();
std::cout << "OpenCL processing done.." << endl;
std::cout << "Start CPU Processing.." << endl;
// Read buffer_result into result
queue.enqueueReadBuffer(buffer_result, CL_TRUE, 0, RESULT_SIZE * sizeof(int), result);
//cpu version to calculate the difference between the two arryays
t1 = chrono::high_resolution_clock::now();
int different = 0;
int x_new;
int x;
for (int i = 0; i < numberMatrices; i++)
{
different = 0;
for (int n = 0; n < LIST_SIZE; n++)
{
x = imageA[n];
x_new = x;;
int a = imageA[x];
int b = imageB[x_new];
int diff = imageB[x_new] - imageA[x];
different += diff;
}
}
t2 = chrono::high_resolution_clock::now();
auto duration_cpu = std::chrono::duration_cast<std::chrono::milliseconds>(t2 - t1).count();
std::cout << "CPU processing done.." << endl;
//output of the results
std::cout << "opencl: diff " << result[0] << endl;
std::cout << "Runtime opencl: " << duration_opencl << endl;
std::cout << "CPU: diff " << different << endl;
std::cout << "Runtime CPU: " << duration_cpu << endl;
double times = (double)duration_cpu / (double)duration_opencl;
std::cout << "OpenCL is " << times << " times faster!!!" << endl;
char c;
std::cin >> c;
}
catch (cl::Error error) {
std::cout << error.what() << "(" << error.err() << ")" << std::endl;
char c;
std::cin >> c;
}
}
return 0;
}