void computeMIonGPU(SequenceSet& sequence, Matrix<float>& MI, bool GPU)
{
// initializes context and kernel and stores them
OCL ocl(GPU);
cl_int oclError1, oclError2;
timeval start, end;
// memory sizes
size_t sequenceLength = sequence.getSequenceLength();
size_t numSequences = sequence.getNumberOfSequences();
// matrix MI is of size numElements
size_t numElements = sequenceLength * sequenceLength;
size_t sequenceSize = sequence.getNumberOfSequences() * sequenceLength;
size_t onePointProbsSize = sequenceLength * NUMPROTEINCHARS;
// host memory
float * dst = new float[MI.size()];
memset(dst, 0, MI.size());
// device memory for sequences, one point probablities and resulting matrix
cl_mem oclDevSrcSequence, oclDevSrcOnePointProbs, oclDevDstMI;
// size for a work group: each workgroup computes one matrix entry, thus computes the correlation
// one time for each character => 25 work items are sufficient
size_t localWorkSize[2] = { 5, 5 };
if (sequenceLength % localWorkSize[0] != 0) throw std::runtime_error("sequence length ^ 2 not divisable by local work size");
// global work size defines the total amount of threads over all work group, thus needs to be a multiple of the local
// work size in each dimension.
size_t globalWorkSize[2] = { sequenceLength, sequenceLength };
// create buffer on device, one for each input array
oclDevSrcSequence = clCreateBuffer( ocl.oclContext,
CL_MEM_READ_ONLY,
sizeof(cl_uchar) * sequenceSize,
0, &oclError1);
oclDevSrcOnePointProbs = clCreateBuffer(ocl.oclContext,
CL_MEM_READ_ONLY,
sizeof(cl_float) * onePointProbsSize,
0, &oclError2);
oclError1 |= oclError2;
oclDevDstMI = clCreateBuffer( ocl.oclContext,
CL_MEM_WRITE_ONLY,
sizeof(cl_float) * numElements,
0, &oclError2);
oclError1 |= oclError2;
if (oclError1 != CL_SUCCESS) {
std::cout << "error while allocating buffers" << std::endl;
exit(1);
}
// set buffer to appropriate kernel arguments
oclError1 = clSetKernelArg(ocl.oclKernel, 0, sizeof(cl_mem), (void*)&oclDevSrcSequence);
oclError1 |= clSetKernelArg(ocl.oclKernel, 1, sizeof(cl_mem), (void*)&oclDevSrcOnePointProbs);
oclError1 |= clSetKernelArg(ocl.oclKernel, 2, sizeof(cl_mem), (void*)&oclDevDstMI);
oclError1 |= clSetKernelArg(ocl.oclKernel, 3, sizeof(cl_uint), &sequenceLength);
oclError1 |= clSetKernelArg(ocl.oclKernel, 4, sizeof(cl_uint), &numSequences);
if (oclError1 != CL_SUCCESS) {
std::cout << "error while setting arguments: " << ocl.oclErrorString(oclError1) << std::endl;
exit(1);
}
// copy host memory to device, non-blocking copy
oclError1 = clEnqueueWriteBuffer( ocl.oclCmdQueue,
oclDevSrcSequence,
CL_FALSE,
0,
sizeof(cl_uchar) * sequenceSize,
(const void *) sequence.getData(),
0, 0, 0);
oclError1 |= clEnqueueWriteBuffer( ocl.oclCmdQueue,
oclDevSrcOnePointProbs,
CL_FALSE,
0,
sizeof(cl_float) * onePointProbsSize,
(const void *) sequence.getOnePointProbs(),
0, 0, 0);
if (oclError1 != CL_SUCCESS) {
std::cout << "error while writing to device " << ocl.oclErrorString(oclError1) << std::endl;
exit(1);
}
// execute kernel LOOPCOUNT times and measure execution time
// TODO LOOPCOUNT aendern, um Kernel mehrfach auszufuehren
gettimeofday(&start, 0);
for (int i = 0; i < LOOPCOUNT; ++i) {
oclError1 = clEnqueueNDRangeKernel( ocl.oclCmdQueue,
ocl.oclKernel,
2, // dimension
0,
globalWorkSize,
localWorkSize,
0, 0, 0);
if (oclError1 != CL_SUCCESS) {
std::cout << "error while executing kernel: " << ocl.oclErrorString(oclError1) << std::endl;
exit(1);
}
}
// clFinish blocks until all issued commands so far are completed, necessary for computing execution time
oclError1 = clFinish(ocl.oclCmdQueue);
gettimeofday(&end, 0);
// read memory from device, store in temporary array and if no error happend copy to result matrix
oclError1 = clEnqueueReadBuffer( ocl.oclCmdQueue,
oclDevDstMI,
CL_TRUE,
0,
sizeof(cl_float) * numElements,
dst,
0, 0, 0);
if (oclError1 != CL_SUCCESS) {
std::cout << "error while reading from device: " << ocl.oclErrorString(oclError1) << std::endl;
exit(1);
}
std::cout << "execution time: "
<< (end.tv_sec - start.tv_sec ) * 1000 + ( end.tv_usec - start.tv_usec) / 1000
<< " milliseconds" << std::endl;
// fill the matrix with the computed results
MI.copyElements(dst);
// release used memory, can cause really bad crashes otherwise
clReleaseMemObject(oclDevSrcSequence);
clReleaseMemObject(oclDevSrcOnePointProbs);
clReleaseMemObject(oclDevDstMI);
}
