所以，在花了几个小时阅读和理解之后，我终于制作了我的第一个 OpenCL 程序，它实际上做了一些事情，它添加了两个向量并输出到一个文件。
#include <iostream>
#include <vector>
#include <cstdlib>
#include <string>
#include <fstream>
#define __CL_ENABLE_EXCEPTIONS
#include <CL/cl.hpp>
int main(int argc, char *argv[])
{
try
{
// get platforms, devices and display their info.
std::vector<cl::Platform> platforms;
cl::Platform::get(&platforms);
std::vector<cl::Platform>::iterator i=platforms.begin();
std::cout<<"OpenCL \tPlatform : "<<i->getInfo<CL_PLATFORM_NAME>()<<std::endl;
std::cout<<"\tVendor: "<<i->getInfo<CL_PLATFORM_VENDOR>()<<std::endl;
std::cout<<"\tVersion : "<<i->getInfo<CL_PLATFORM_VERSION>()<<std::endl;
std::cout<<"\tExtensions : "<<i->getInfo<CL_PLATFORM_EXTENSIONS>()<<std::endl;
// get devices
std::vector<cl::Device> devices;
i->getDevices(CL_DEVICE_TYPE_ALL,&devices);
int o=99;
std::cout<<"\n\n";
// iterate over available devices
for(std::vector<cl::Device>::iterator j=devices.begin(); j!=devices.end(); j++)
{
std::cout<<"\tOpenCL\tDevice : " << j->getInfo<CL_DEVICE_NAME>()<<std::endl;
std::cout<<"\t\t Type : " << j->getInfo<CL_DEVICE_TYPE>()<<std::endl;
std::cout<<"\t\t Vendor : " << j->getInfo<CL_DEVICE_VENDOR>()<<std::endl;
std::cout<<"\t\t Driver : " << j->getInfo<CL_DRIVER_VERSION>()<<std::endl;
std::cout<<"\t\t Global Mem : " << j->getInfo<CL_DEVICE_GLOBAL_MEM_SIZE>()/(1024*1024)<<" MBytes"<<std::endl;
std::cout<<"\t\t Local Mem : " << j->getInfo<CL_DEVICE_LOCAL_MEM_SIZE>()/1024<<" KBbytes"<<std::endl;
std::cout<<"\t\t Compute Unit : " << j->getInfo<CL_DEVICE_MAX_COMPUTE_UNITS>()<<std::endl;
std::cout<<"\t\t Clock Rate : " << j->getInfo<CL_DEVICE_MAX_CLOCK_FREQUENCY>()<<" MHz"<<std::endl;
}
std::cout<<"\n\n\n";
//MAIN CODE BEGINS HERE
//get Kernel
std::ifstream ifs("vector_add_kernel.cl");
std::string kernelSource((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
std::cout<<kernelSource;
//Create context, select device and command queue.
cl::Context context(devices);
cl::Device &device=devices.front();
cl::CommandQueue cmdqueue(context,device);
// Generate Source vector and push the kernel source in it.
cl::Program::Sources sourceCode;
sourceCode.push_back(std::make_pair(kernelSource.c_str(), kernelSource.size()));
//Generate program using sourceCode
cl::Program program=cl::Program(context, sourceCode);
//Build program..
try
{
program.build(devices);
}
catch(cl::Error &err)
{
std::cerr<<"Building failed, "<<err.what()<<"("<<err.err()<<")"
<<"\nRetrieving build log"
<<"\n Build Log Follows \n"
<<program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(devices.front());
}
//Declare and initialize vectors
std::vector<cl_float>B(993448,1.3);
std::vector<cl_float>C(993448,1.3);
std::vector<cl_float>A(993448,1.3);
cl_int N=A.size();
//Declare and intialize proper work group size and global size. Global size raised to the nearest multiple of workGroupSize.
int workGroupSize=128;
int GlobalSize;
if(N%workGroupSize) GlobalSize=N - N%workGroupSize + workGroupSize;
else GlobalSize=N;
//Declare buffers.
cl::Buffer vecA(context, CL_MEM_READ_WRITE, sizeof(cl_float)*N);
cl::Buffer vecB(context, CL_MEM_READ_ONLY , (B.size())*sizeof(cl_float));
cl::Buffer vecC(context, CL_MEM_READ_ONLY , (C.size())*sizeof(cl_float));
//Write vectors into buffers
cmdqueue.enqueueWriteBuffer(vecB, 0, 0, (B.size())*sizeof(cl_float), &B[0] );
cmdqueue.enqueueWriteBuffer(vecB, 0, 0, (C.size())*sizeof(cl_float), &C[0] );
//Executing kernel
cl::Kernel kernel(program, "vector_add");
cl::KernelFunctor kernel_func=kernel.bind(cmdqueue, cl::NDRange(GlobalSize), cl::NDRange(workGroupSize));
kernel_func(vecA, vecB, vecC, N);
//Reading back values into vector A
cmdqueue.enqueueReadBuffer(vecA,true,0,N*sizeof(cl_float), &A[0]);
cmdqueue.finish();
//Saving into file.
std::ofstream output("vectorAdd.txt");
for(int i=0;i<N;i++) output<<A[i]<<"\n";
}
catch(cl::Error& err)
{
std::cerr << "OpenCL error: " << err.what() << "(" << err.err() <<
")" << std::endl;
return EXIT_FAILURE;
}
return EXIT_SUCCESS;
}
问题是，对于较小的 N 值，我得到的正确结果是 2.6 但对于较大的值，例如上面代码中的值 (993448)，我得到的垃圾输出在 1 和 2.4 之间变化。
这是内核代码：
__kernel void vector_add(__global float *A, __global float *B, __global float *C, int N) {
// Get the index of the current element
int i = get_global_id(0);
//Do the operation
if(i<N) A[i] = C[i] + B[i];
}
更新：好的，看来代码现在可以工作了。我在上面的代码中修复了一些小错误 1）初始化 GlobalSize 的部分已经修复。2）enqueueWriteBuffer 中的愚蠢错误（给定的参数错误）它现在正在输出大 N 值的正确结果。