int
Lucas::normalize2_runCLKernels ()
{
cl_int status;
/*
* Kernel runs over complete output matrix with blocks of blockSize x blockSize
* running concurrently
*/
size_t globalThreads = Nn / threads;
size_t localThreads = 256;
cl_int eventStatus = CL_QUEUED;
// Set input data to matrix A and matrix B
cl::Event ndrEvt;
// Set appropriate arguments to the kernel
status = normalize2_kernel.setArg (0, g_x);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_x)");
status = normalize2_kernel.setArg (1, threads);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (threads)");
status = normalize2_kernel.setArg (2, bigAB);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (bigAB)");
status = normalize2_kernel.setArg (3, bigAB);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (bigAB)");
status = normalize2_kernel.setArg (4, g_carry);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_carry)");
status = normalize2_kernel.setArg (5, Nn);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (Nn)");
status = normalize2_kernel.setArg (6, g_inv2);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_inv2)");
status = normalize2_kernel.setArg (7, g_ttp2);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_ttp2)");
status = normalize2_kernel.setArg (8, g_ttmp2);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_ttmp2)");
status = normalize2_kernel.setArg (9, g_inv3);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_inv3)");
status = normalize2_kernel.setArg (10, g_ttp3);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_ttp3)");
status = normalize2_kernel.setArg (11, g_ttmp3);
CHECK_OPENCL_ERROR (status, "cl::setArg failed. (g_ttmp3)");
/*
* Enqueue a kernel run call.
*/
// Each thread calculates 2 gaussian numbers
cl::NDRange gThreads (globalThreads);
cl::NDRange lThreads (localThreads);
status = commandQueue.enqueueNDRangeKernel (normalize2_kernel,
cl::NullRange,
gThreads, lThreads, 0, &ndrEvt);
CHECK_OPENCL_ERROR (status, "CommandQueue::enqueueNDRangeKernel() failed.");
status = commandQueue.flush ();
CHECK_OPENCL_ERROR (status, "cl::CommandQueue.flush failed.");
return SDK_SUCCESS;
}
int
MatrixMulDouble::runCLKernels(void)
{
cl_int status;
/*
* Kernel runs over complete output matrix with blocks of blockSize x blockSize
* running concurrently
*/
size_t globalThreads[2]= {widthB / 4, heightA / 4};
size_t localThreads[2] = {blockSize, blockSize};
cl_int eventStatus = CL_QUEUED;
// Set input data to matrix A and matrix B
cl::Event inMapEvtA;
cl::Event inMapEvtB;
cl::Event inUnmapEvtA;
cl::Event inUnmapEvtB;
cl::Event outMapEvt;
cl::Event outUnmapEvt;
cl::Event ndrEvt;
void* inMapPtrA = NULL;
void* inMapPtrB = NULL;
void* outMapPtr = NULL;
inMapPtrA = commandQueue.enqueueMapBuffer(
inputBufA,
CL_FALSE,
CL_MAP_WRITE,
0,
widthA * heightA * sizeof(cl_double),
NULL,
&inMapEvtA,
&status);
CHECK_OPENCL_ERROR(status,
"cl::CommandQueue.enqueueMapBuffer failed. (inputBufA)");
inMapPtrB = commandQueue.enqueueMapBuffer(
inputBufB,
CL_FALSE,
CL_MAP_WRITE,
0,
widthB * heightB * sizeof(cl_double),
NULL,
&inMapEvtB,
&status);
CHECK_OPENCL_ERROR(status,
"cl::CommandQueue.enqueueMapBuffer failed. (inputBufB)");
status = commandQueue.flush();
CHECK_OPENCL_ERROR(status, "cl::CommandQueue.flush failed.");
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = inMapEvtA.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
memcpy(inMapPtrA, inputA, sizeof(cl_double) * widthA * heightA);
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = inMapEvtB.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
memcpy(inMapPtrB, inputB, sizeof(cl_double) * widthB * heightB);
status = commandQueue.enqueueUnmapMemObject(
inputBufA,
inMapPtrA,
NULL,
&inUnmapEvtA);
CHECK_OPENCL_ERROR(status,
"cl::CommandQueue.enqueueUnmapMemObject (inputBufA).");
status = commandQueue.enqueueUnmapMemObject(
inputBufB,
inMapPtrB,
NULL,
&inUnmapEvtB);
CHECK_OPENCL_ERROR(status,
"cl::CommandQueue.enqueueUnmapMemObject (inputBufB)");
status = commandQueue.flush();
CHECK_OPENCL_ERROR(status, "cl::CommandQueue.flush failed.");
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = inUnmapEvtA.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = inUnmapEvtB.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
// Set appropriate arguments to the kernel
// output array as the 1st argument : stores product of inputA and inputB
status = kernel.setArg(0, inputBufA);
CHECK_OPENCL_ERROR(status, "cl::setArg failed. (inputBufA)");
status = kernel.setArg(1, inputBufB);
CHECK_OPENCL_ERROR(status, "cl::setArg failed. (inputBufB)");
status = kernel.setArg(2, outputBuf);
CHECK_OPENCL_ERROR(status, "cl::setArg failed. (outputBuf)");
// widthA of the inputA matrix as 4th argument - widthA
status = kernel.setArg(3, widthA);
CHECK_OPENCL_ERROR(status, "cl::setArg failed. (widthA)");
// Set local memory argument if Scratchpad is available
if(lds)
{
status = kernel.setArg(
4,
(blockSize * 4) * (blockSize * 4) * sizeof(cl_double),
NULL);
CHECK_OPENCL_ERROR(status, "cl::setArg failed. (local memory)");
}
else
{
status = kernel.setArg(4, sizeof(cl_int), &widthB);
CHECK_OPENCL_ERROR(status, "cl::setArg failed. (widthB)");
}
/*
* Enqueue a kernel run call.
*/
// Each thread calculates 2 gaussian numbers
cl::NDRange gThreads(globalThreads[0], globalThreads[1]);
cl::NDRange lThreads(localThreads[0], localThreads[1]);
status = commandQueue.enqueueNDRangeKernel(kernel,
cl::NullRange,
gThreads,
lThreads,
0,
&ndrEvt);
CHECK_OPENCL_ERROR(status, "CommandQueue::enqueueNDRangeKernel() failed.");
status = commandQueue.flush();
CHECK_OPENCL_ERROR(status, "cl::CommandQueue.flush failed.");
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = ndrEvt.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
if(!eAppGFLOPS)
{
// Calculate performance
cl_ulong startTime;
cl_ulong endTime;
status = ndrEvt.getProfilingInfo<cl_ulong>(
CL_PROFILING_COMMAND_START,
&startTime);
CHECK_OPENCL_ERROR(status, "cl::Event.getProfilingInfo failed.(startTime).");
status = ndrEvt.getProfilingInfo<cl_ulong>(
CL_PROFILING_COMMAND_END,
&endTime);
CHECK_OPENCL_ERROR(status, "cl::Event.getProfilingInfo failed.(endTime).");
// Print performance numbers
double sec = 1e-9 * (endTime - startTime);
kernelTime += sec;
}
outMapPtr = commandQueue.enqueueMapBuffer(
outputBuf,
CL_FALSE,
CL_MAP_READ,
0,
widthB * heightA * sizeof(cl_double),
NULL,
&outMapEvt,
&status);
CHECK_OPENCL_ERROR(status,
"cl::CommandQueue.enqueueMapBuffer failed. (outputBuf).");
status = commandQueue.flush();
CHECK_OPENCL_ERROR(status, "cl::CommandQueue.flush failed.");
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = outMapEvt.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
memcpy(output, outMapPtr, sizeof(cl_double) * widthB * heightA);
status = commandQueue.enqueueUnmapMemObject(
outputBuf,
outMapPtr,
NULL,
&outUnmapEvt);
CHECK_OPENCL_ERROR(status,
"cl::CommandQueue.enqueueUnmapMemObject (outputBuf)");
status = commandQueue.flush();
CHECK_OPENCL_ERROR(status, "cl::CommandQueue.flush failed.");
eventStatus = CL_QUEUED;
while(eventStatus != CL_COMPLETE)
{
status = outUnmapEvt.getInfo<cl_int>(
CL_EVENT_COMMAND_EXECUTION_STATUS,
&eventStatus);
CHECK_OPENCL_ERROR(status,
"cl:Event.getInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) failed.");
}
return SDK_SUCCESS;
}