OpenCL: array of arrays of variable lengths - c++

Question

I am trying to process an array of arrays of variables lengths with OpenCL 1.2 in C++. In each instance (workitem?) I want to process one sub array.
Below I've tried to treat the array of arrays as a 1D array, but it does not work - random parts of the data are not processes.
Host:
vector<cl::Platform> platforms; cl::Platform::get(&platforms); _ASSERT(platforms.size() > 0); auto platform = platforms.front(); //get the platform
std::vector<cl::Device> devices; platform.getDevices(CL_DEVICE_TYPE_GPU, &devices); _ASSERT(devices.size() > 0); auto device = devices.front(); // get the device
std::ifstream myFile("DynMultiDimArr.cl"); string src(istreambuf_iterator<char>(myFile), (istreambuf_iterator<char>())); cl::Program::Sources sources(1, std::make_pair(src.c_str(), src.length() + 1)); //create program from cl file
cl::Context context(device);
cl::Program program(context, sources);
auto err = program.build(); if (err!=0) printf("%s\n",program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device).c_str() );
cl::CommandQueue queue(context, device);
int lens[10] = { 5,7,9,6,21,12,4,18,15,10 }, *idx=new int[10], totSize=0, c=0;
for (int i = 0; i < 10; i++) totSize += lens[i];
double *dat = new double[totSize], **myDat = new double *[10]; // array of arrays of different lengths
for (int i = 0; i < 10; i++) {
idx[i] = c;
myDat[i] = dat + c;
for (int j = 0; j < lens[i]; j++) myDat[i][j] = c++;
}
cl::Buffer inBuf(context, CL_MEM_READ_ONLY | CL_MEM_HOST_NO_ACCESS | CL_MEM_COPY_HOST_PTR, sizeof(double)*totSize, dat, &err);
cl::Buffer iBuf(context, CL_MEM_READ_ONLY | CL_MEM_HOST_NO_ACCESS | CL_MEM_COPY_HOST_PTR, sizeof(int)*10, lens, &err);
cl::Buffer lBuf(context, CL_MEM_READ_ONLY | CL_MEM_HOST_NO_ACCESS | CL_MEM_COPY_HOST_PTR, sizeof(int)*10, idx, &err);
cl::Buffer outBuf(context, CL_MEM_WRITE_ONLY | CL_MEM_HOST_READ_ONLY, sizeof(double)*totSize, nullptr, &err);
cl::Kernel kernel(program, "processSubArr");
err = kernel.setArg(0, inBuf);
err = kernel.setArg(1, lBuf);
err = kernel.setArg(2, iBuf);
err = kernel.setArg(3, outBuf);
err=queue.enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(10));
err=queue.enqueueReadBuffer(outBuf, CL_FALSE, 0, sizeof(double)*totSize, dat);
cl::finish();
Kernel:
__kernel void processSubArr(__global double* data, __global int* len, __global int* idx, __global double* outData) {
for (int i=0;i<len[get_global_id(0)];i++)
outData[idx[get_global_id(0)]+i] = data[idx[get_global_id(0)]+i]+1000;
}
This is test code only. In my real problem I have to pass 8 arrays of arrays (all same dimensions). First dimension is 105 to 106 long, second 1 to 100 long. The kernel code is ~100 lines of code, calculating turbulent eddy viscosity and diffusivity on each sub array with the k-epsilon method.
Is this the way or am I on the completely wrong path? I'm new to OpenCL - any help will be much appreciated.

Answer 1

Updated code that works. I have no clue about the performance of this method.
vector<cl::Platform> platforms; cl::Platform::get(&platforms); _ASSERT(platforms.size() > 0); auto platform = platforms.front(); //get the platform
std::vector<cl::Device> devices; platform.getDevices(CL_DEVICE_TYPE_GPU, &devices); _ASSERT(devices.size() > 0); auto device = devices.front(); // get the device
std::ifstream myFile("DynMultiDimArr.cl"); string src(istreambuf_iterator<char>(myFile), (istreambuf_iterator<char>())); cl::Program::Sources sources(1, std::make_pair(src.c_str(), src.length() + 1)); //create program from cl file
cl::Context context(device);
cl::Program program(context, sources);
auto err = program.build(); if (err!=0) printf("%s\n",program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device).c_str() );
cl::CommandQueue queue(context, device);
int lens[10] = { 5,7,9,6,21,12,4,18,15,10 }, *idx=new int[10], totSize=0, c=0;
for (int i = 0; i < 10; i++) totSize += lens[i];
double *dat = new double[totSize], **myDat = new double *[10]; // array of arrays of different lengths
for (int i = 0; i < 10; i++) {
idx[i] = c;
myDat[i] = dat + c;
for (int j = 0; j < lens[i]; j++) myDat[i][j] = c++;
}
cl::Buffer inBuf(context, CL_MEM_READ_ONLY | CL_MEM_HOST_NO_ACCESS | CL_MEM_COPY_HOST_PTR, sizeof(double)*totSize, dat, &err);
cl::Buffer lBuf(context, CL_MEM_READ_ONLY | CL_MEM_HOST_NO_ACCESS | CL_MEM_COPY_HOST_PTR, sizeof(int)*10, lens, &err);
cl::Buffer iBuf(context, CL_MEM_READ_ONLY | CL_MEM_HOST_NO_ACCESS | CL_MEM_COPY_HOST_PTR, sizeof(int)*10, idx, &err);
cl::Buffer outBuf(context, CL_MEM_WRITE_ONLY | CL_MEM_HOST_READ_ONLY, sizeof(double)*totSize, nullptr, &err);
cl::Kernel kernel(program, "processSubArr");
err = kernel.setArg(0, inBuf);
err = kernel.setArg(1, lBuf);
err = kernel.setArg(2, iBuf);
err = kernel.setArg(3, outBuf);
err=queue.enqueueNDRangeKernel(kernel, cl::NullRange, cl::NDRange(10));
err=queue.enqueueReadBuffer(outBuf, CL_FALSE, 0, sizeof(double)*totSize, dat); //queue.enqueueMapBuffer + memcpy faster?
queue.finish();
cl::finish();
for (int i = 0; i < 10; i++) {
int j = 0;
for (j = 0; j < lens[i]-1; j++)
cout << myDat[i][j] << ",";
cout << myDat[i][j] << endl;
}
delete[] dat;