int main(int argc, char **argv)
{
bool bTestResult = true;
shrQAStart(argc, argv);
// Start the log
shrSetLogFileName(shrLogFile);
shrLog("%s Starting...\n\n", argv[0]);
// Check help flag
if (shrCheckCmdLineFlag(argc, (const char **)argv, "help")) {
shrLog("Displaying help on console\n");
showHelp(argc, (const char **)argv);
} else {
// Execute
bTestResult = runTest(argc, (const char **)argv);
oclCheckErrorEX(bTestResult, true, NULL);
}
// Finish
shrQAFinishExit( argc, (const char **)argv, (bTestResult ? QA_PASSED : QA_FAILED) );
}
// Main program
//*****************************************************************************
int main(int argc, char** argv)
{
// Locals used with command line args
int p = 256; // workgroup X dimension
int q = 1; // workgroup Y dimension
pArgc = &argc;
pArgv = argv;
shrQAStart(argc, argv);
// latch the executable path for other funcs to use
cExecutablePath = argv[0];
// start logs and show command line help
shrSetLogFileName ("oclNbody.txt");
shrLog("%s Starting...\n\n", cExecutablePath);
shrLog("Command line switches:\n");
shrLog(" --qatest\t\tCheck correctness of GPU execution and measure performance)\n");
shrLog(" --noprompt\t\tQuit simulation automatically after a brief period\n");
shrLog(" --n=<numbodies>\tSpecify # of bodies to simulate (default = %d)\n", numBodies);
shrLog(" --double\t\tUse double precision floating point values for simulation\n");
shrLog(" --p=<workgroup X dim>\tSpecify X dimension of workgroup (default = %d)\n", p);
shrLog(" --q=<workgroup Y dim>\tSpecify Y dimension of workgroup (default = %d)\n\n", q);
// Get command line arguments if there are any and set vars accordingly
if (argc > 0)
{
shrGetCmdLineArgumenti(argc, (const char**)argv, "p", &p);
shrGetCmdLineArgumenti(argc, (const char**)argv, "q", &q);
shrGetCmdLineArgumenti(argc, (const char**)argv, "n", &numBodies);
bDouble = (shrTRUE == shrCheckCmdLineFlag(argc, (const char**)argv, "double"));
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
bQATest = shrCheckCmdLineFlag(argc, (const char**)argv, "qatest");
}
//Get the NVIDIA platform
cl_int ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clGetPlatformID...\n\n");
if (bDouble)
{
shrLog("Double precision execution...\n\n");
}
else
{
shrLog("Single precision execution...\n\n");
}
flopsPerInteraction = bDouble ? 30 : 20;
//Get all the devices
shrLog("Get the Device info and select Device...\n");
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Set target device and Query number of compute units on uiTargetDevice
shrLog(" # of Devices Available = %u\n", uiNumDevices);
if(shrGetCmdLineArgumentu(argc, (const char**)argv, "device", &uiTargetDevice)== shrTRUE)
{
uiTargetDevice = CLAMP(uiTargetDevice, 0, (uiNumDevices - 1));
}
shrLog(" Using Device %u, ", uiTargetDevice);
oclPrintDevName(LOGBOTH, cdDevices[uiTargetDevice]);
cl_uint uiNumComputeUnits;
clGetDeviceInfo(cdDevices[uiTargetDevice], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(uiNumComputeUnits), &uiNumComputeUnits, NULL);
shrLog(" # of Compute Units = %u\n", uiNumComputeUnits);
//Create the context
shrLog("clCreateContext...\n");
cxContext = clCreateContext(0, uiNumDevsUsed, &cdDevices[uiTargetDevice], NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create a command-queue
shrLog("clCreateCommandQueue...\n\n");
cqCommandQueue = clCreateCommandQueue(cxContext, cdDevices[uiTargetDevice], CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Log and config for number of bodies
shrLog("Number of Bodies = %d\n", numBodies);
switch (numBodies)
{
case 1024:
activeParams.m_clusterScale = 1.52f;
activeParams.m_velocityScale = 2.f;
break;
case 2048:
activeParams.m_clusterScale = 1.56f;
activeParams.m_velocityScale = 2.64f;
break;
case 4096:
activeParams.m_clusterScale = 1.68f;
activeParams.m_velocityScale = 2.98f;
break;
case 7680:
case 8192:
activeParams.m_clusterScale = 1.98f;
activeParams.m_velocityScale = 2.9f;
break;
default:
case 15360:
case 16384:
activeParams.m_clusterScale = 1.54f;
activeParams.m_velocityScale = 8.f;
break;
case 30720:
case 32768:
activeParams.m_clusterScale = 1.44f;
activeParams.m_velocityScale = 11.f;
break;
}
if ((q * p) > 256)
{
p = 256 / q;
shrLog("Setting p=%d to maintain %d threads per block\n", p, 256);
}
if ((q == 1) && (numBodies < p))
{
p = numBodies;
shrLog("Setting p=%d because # of bodies < p\n", p);
}
shrLog("Workgroup Dims = (%d x %d)\n\n", p, q);
// Initialize OpenGL items if using GL
if (bQATest == shrFALSE)
{
assert(0);
/*
shrLog("Calling InitGL...\n");
InitGL(&argc, argv);
*/
}
else
{
shrLog("Skipping InitGL...\n");
}
// CL/GL interop disabled
bUsePBO = (false && (bQATest == shrFALSE));
InitNbody(cdDevices[uiTargetDevice], cxContext, cqCommandQueue, numBodies, p, q, bUsePBO, bDouble);
ResetSim(nbody, numBodies, NBODY_CONFIG_SHELL, bUsePBO);
// init timers
shrDeltaT(DEMOTIME); // timer 0 is for timing demo periods
shrDeltaT(FUNCTIME); // timer 1 is for logging function delta t's
shrDeltaT(FPSTIME); // timer 2 is for fps measurement
// Standard simulation
if (bQATest == shrFALSE)
{
assert(0);
/*
shrLog("Running standard oclNbody simulation...\n\n");
glutDisplayFunc(DisplayGL);
glutReshapeFunc(ReshapeGL);
glutMouseFunc(MouseGL);
glutMotionFunc(MotionGL);
glutKeyboardFunc(KeyboardGL);
glutSpecialFunc(SpecialGL);
glutIdleFunc(IdleGL);
glutMainLoop();
*/
}
// Compare to host, profile and write out file for regression analysis
if (bQATest == shrTRUE) {
bool bTestResults = false;
shrLog("Running oclNbody Results Comparison...\n\n");
bTestResults = CompareResults(numBodies);
//shrLog("Profiling oclNbody...\n\n");
//RunProfiling(100, (unsigned int)(p * q)); // 100 iterations
shrQAFinish(argc, (const char **)argv, bTestResults ? QA_PASSED : QA_FAILED);
} else {
// Cleanup/exit
bNoPrompt = shrTRUE;
shrQAFinish2(false, *pArgc, (const char **)pArgv, QA_PASSED);
}
Cleanup(EXIT_SUCCESS);
}
// Main function
// *********************************************************************
int main(int argc, char **argv)
{
gp_argc = &argc;
gp_argv = &argv;
shrQAStart(argc, argv);
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clGetPlatformID...\n");
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clGetPlatformID...\n");
//Get all the devices
cl_uint uiNumDevices = 0; // Number of devices available
cl_uint uiTargetDevice = 0; // Default Device to compute on
cl_uint uiNumComputeUnits; // Number of compute units (SM's on NV GPU)
shrLog("Get the Device info and select Device...\n");
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
// Get command line device options and config accordingly
shrLog(" # of Devices Available = %u\n", uiNumDevices);
if(shrGetCmdLineArgumentu(argc, (const char**)argv, "device", &uiTargetDevice)== shrTRUE)
{
uiTargetDevice = CLAMP(uiTargetDevice, 0, (uiNumDevices - 1));
}
shrLog(" Using Device %u: ", uiTargetDevice);
oclPrintDevName(LOGBOTH, cdDevices[uiTargetDevice]);
ciErrNum = clGetDeviceInfo(cdDevices[uiTargetDevice], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(uiNumComputeUnits), &uiNumComputeUnits, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("\n # of Compute Units = %u\n", uiNumComputeUnits);
// get command line arg for quick test, if provided
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
// start logs
cExecutableName = argv[0];
shrSetLogFileName ("oclDotProduct.txt");
shrLog("%s Starting...\n\n# of float elements per Array \t= %u\n", argv[0], iNumElements);
// set and log Global and Local work size dimensions
szLocalWorkSize = 256;
szGlobalWorkSize = shrRoundUp((int)szLocalWorkSize, iNumElements); // rounded up to the nearest multiple of the LocalWorkSize
shrLog("Global Work Size \t\t= %u\nLocal Work Size \t\t= %u\n# of Work Groups \t\t= %u\n\n",
szGlobalWorkSize, szLocalWorkSize, (szGlobalWorkSize % szLocalWorkSize + szGlobalWorkSize/szLocalWorkSize));
// Allocate and initialize host arrays
shrLog( "Allocate and Init Host Mem...\n");
srcA = (void *)malloc(sizeof(cl_float4) * szGlobalWorkSize);
srcB = (void *)malloc(sizeof(cl_float4) * szGlobalWorkSize);
dst = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
Golden = (void *)malloc(sizeof(cl_float) * iNumElements);
shrFillArray((float*)srcA, 4 * iNumElements);
shrFillArray((float*)srcB, 4 * iNumElements);
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Get a GPU device
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &cdDevices[uiTargetDevice], NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create the context
cxGPUContext = clCreateContext(0, 1, &cdDevices[uiTargetDevice], NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create a command-queue
shrLog("clCreateCommandQueue...\n");
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevices[uiTargetDevice], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
shrLog("clCreateBuffer (SrcA, SrcB and Dst in Device GMEM)...\n");
cmDevSrcA = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize * 4, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmDevSrcB = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize * 4, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmDevDst = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Read the OpenCL kernel in from source file
shrLog("oclLoadProgSource (%s)...\n", cSourceFile);
cPathAndName = shrFindFilePath(cSourceFile, argv[0]);
oclCheckErrorEX(cPathAndName != NULL, shrTRUE, pCleanup);
cSourceCL = oclLoadProgSource(cPathAndName, "", &szKernelLength);
oclCheckErrorEX(cSourceCL != NULL, shrTRUE, pCleanup);
// Create the program
shrLog("clCreateProgramWithSource...\n");
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErrNum);
// Build the program with 'mad' Optimization option
#ifdef MAC
char* flags = "-cl-fast-relaxed-math -DMAC";
#else
char* flags = "-cl-fast-relaxed-math";
#endif
shrLog("clBuildProgram...\n");
ciErrNum = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclDotProduct.ptx");
Cleanup(EXIT_FAILURE);
}
// Create the kernel
shrLog("clCreateKernel (DotProduct)...\n");
ckKernel = clCreateKernel(cpProgram, "DotProduct", &ciErrNum);
// Set the Argument values
shrLog("clSetKernelArg 0 - 3...\n\n");
ciErrNum = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void*)&cmDevSrcA);
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void*)&cmDevSrcB);
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void*)&cmDevDst);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(cl_int), (void*)&iNumElements);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// --------------------------------------------------------
// Core sequence... copy input data to GPU, compute, copy results back
// Asynchronous write of data to GPU device
shrLog("clEnqueueWriteBuffer (SrcA and SrcB)...\n");
ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcA, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize * 4, srcA, 0, NULL, NULL);
ciErrNum |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcB, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize * 4, srcB, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Launch kernel
shrLog("clEnqueueNDRangeKernel (DotProduct)...\n");
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Read back results and check accumulated errors
shrLog("clEnqueueReadBuffer (Dst)...\n\n");
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmDevDst, CL_TRUE, 0, sizeof(cl_float) * szGlobalWorkSize, dst, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Compute and compare results for golden-host and report errors and pass/fail
shrLog("Comparing against Host/C++ computation...\n\n");
DotProductHost ((const float*)srcA, (const float*)srcB, (float*)Golden, iNumElements);
shrBOOL bMatch = shrComparefet((const float*)Golden, (const float*)dst, (unsigned int)iNumElements, 0.0f, 0);
// Cleanup and leave
Cleanup (EXIT_SUCCESS);
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
shrQAStart(argc, argv);
// start logs
shrSetLogFileName ("oclSimpleMultiGPU.txt");
shrLog("%s Starting, Array = %u float values...\n\n", argv[0], DATA_N);
// OpenCL
cl_platform_id cpPlatform;
cl_uint ciDeviceCount;
cl_device_id* cdDevices;
cl_context cxGPUContext;
cl_device_id cdDevice; // GPU device
int deviceNr[MAX_GPU_COUNT];
cl_command_queue commandQueue[MAX_GPU_COUNT];
cl_mem d_Data[MAX_GPU_COUNT];
cl_mem d_Result[MAX_GPU_COUNT];
cl_program cpProgram;
cl_kernel reduceKernel[MAX_GPU_COUNT];
cl_event GPUDone[MAX_GPU_COUNT];
cl_event GPUExecution[MAX_GPU_COUNT];
size_t programLength;
cl_int ciErrNum;
char cDeviceName [256];
cl_mem h_DataBuffer;
// Vars for reduction results
float h_SumGPU[MAX_GPU_COUNT * ACCUM_N];
float *h_Data;
double sumGPU;
double sumCPU, dRelError;
// allocate and init host buffer with with some random generated input data
h_Data = (float *)malloc(DATA_N * sizeof(float));
shrFillArray(h_Data, DATA_N);
// start timer & logs
shrLog("Setting up OpenCL on the Host...\n\n");
shrDeltaT(1);
// Annotate profiling state
#ifdef GPU_PROFILING
shrLog("OpenCL Profiling is enabled...\n\n");
#endif
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clGetPlatformID...\n");
//Get the devices
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &ciDeviceCount);
oclCheckError(ciErrNum, CL_SUCCESS);
cdDevices = (cl_device_id *)malloc(ciDeviceCount * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, ciDeviceCount, cdDevices, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clGetDeviceIDs...\n");
//Create the context
cxGPUContext = clCreateContext(0, ciDeviceCount, cdDevices, NULL, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateContext...\n");
// Set up command queue(s) for GPU's specified on the command line or all GPU's
if(shrCheckCmdLineFlag(argc, (const char **)argv, "device"))
{
// User specified GPUs
int ciMaxDeviceID = ciDeviceCount-1;
ciDeviceCount = 0;
char* deviceList;
char* deviceStr;
char* next_token;
shrGetCmdLineArgumentstr(argc, (const char **)argv, "device", &deviceList);
#ifdef WIN32
deviceStr = strtok_s (deviceList," ,.-", &next_token);
#else
deviceStr = strtok (deviceList," ,.-");
#endif
// Create command queues for all Requested GPU's
while(deviceStr != NULL)
{
// get & log device index # and name
deviceNr[ciDeviceCount] = atoi(deviceStr);
if( deviceNr[ciDeviceCount] > ciMaxDeviceID ) {
shrLog(" Invalid user specified device ID: %d\n", deviceNr[ciDeviceCount]);
return 1;
}
cdDevice = oclGetDev(cxGPUContext, deviceNr[ciDeviceCount]);
ciErrNum = clGetDeviceInfo(cdDevice, CL_DEVICE_NAME, sizeof(cDeviceName), cDeviceName, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog(" Device %i: %s\n\n", deviceNr[ciDeviceCount], cDeviceName);
// create a command que
commandQueue[ciDeviceCount] = clCreateCommandQueue(cxGPUContext, cdDevice, CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateCommandQueue\n");
++ciDeviceCount;
#ifdef WIN32
deviceStr = strtok_s (NULL," ,.-", &next_token);
#else
deviceStr = strtok (NULL," ,.-");
#endif
}
free(deviceList);
}
else
{
// Find out how many GPU's to compute on all available GPUs
size_t nDeviceBytes;
ciErrNum = clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &nDeviceBytes);
oclCheckError(ciErrNum, CL_SUCCESS);
ciDeviceCount = (cl_uint)nDeviceBytes/sizeof(cl_device_id);
for(unsigned int i = 0; i < ciDeviceCount; ++i )
{
// get & log device index # and name
deviceNr[i] = i;
cdDevice = oclGetDev(cxGPUContext, i);
ciErrNum = clGetDeviceInfo(cdDevice, CL_DEVICE_NAME, sizeof(cDeviceName), cDeviceName, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog(" Device %i: %s\n", i, cDeviceName);
// create a command que
commandQueue[i] = clCreateCommandQueue(cxGPUContext, cdDevice, CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateCommandQueue\n\n");
}
}
// Load the OpenCL source code from the .cl file
const char* source_path = shrFindFilePath("simpleMultiGPU.cl", argv[0]);
char *source = oclLoadProgSource(source_path, "", &programLength);
oclCheckError(source != NULL, shrTRUE);
shrLog("oclLoadProgSource\n");
// Create the program for all GPUs in the context
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&source, &programLength, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateProgramWithSource\n");
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-fast-relaxed-math", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclSimpleMultiGPU.ptx");
oclCheckError(ciErrNum, CL_SUCCESS);
}
shrLog("clBuildProgram\n");
// Create host buffer with page-locked memory
h_DataBuffer = clCreateBuffer(cxGPUContext, CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR,
DATA_N * sizeof(float), h_Data, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateBuffer (Page-locked Host)\n\n");
// Create buffers for each GPU, with data divided evenly among GPU's
int sizePerGPU = DATA_N / ciDeviceCount;
int workOffset[MAX_GPU_COUNT];
int workSize[MAX_GPU_COUNT];
workOffset[0] = 0;
for(unsigned int i = 0; i < ciDeviceCount; ++i )
{
workSize[i] = (i != (ciDeviceCount - 1)) ? sizePerGPU : (DATA_N - workOffset[i]);
// Input buffer
d_Data[i] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, workSize[i] * sizeof(float), NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateBuffer (Input)\t\tDev %i\n", i);
// Copy data from host to device
ciErrNum = clEnqueueCopyBuffer(commandQueue[i], h_DataBuffer, d_Data[i], workOffset[i] * sizeof(float),
0, workSize[i] * sizeof(float), 0, NULL, NULL);
shrLog("clEnqueueCopyBuffer (Input)\tDev %i\n", i);
// Output buffer
d_Result[i] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, ACCUM_N * sizeof(float), NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateBuffer (Output)\t\tDev %i\n", i);
// Create kernel
reduceKernel[i] = clCreateKernel(cpProgram, "reduce", &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateKernel\t\t\tDev %i\n", i);
// Set the args values and check for errors
ciErrNum |= clSetKernelArg(reduceKernel[i], 0, sizeof(cl_mem), &d_Result[i]);
ciErrNum |= clSetKernelArg(reduceKernel[i], 1, sizeof(cl_mem), &d_Data[i]);
ciErrNum |= clSetKernelArg(reduceKernel[i], 2, sizeof(int), &workSize[i]);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clSetKernelArg\t\t\tDev %i\n\n", i);
workOffset[i + 1] = workOffset[i] + workSize[i];
}
// Set # of work items in work group and total in 1 dimensional range
size_t localWorkSize[] = {THREAD_N};
size_t globalWorkSize[] = {ACCUM_N};
// Start timer and launch reduction kernel on each GPU, with data split between them
shrLog("Launching Kernels on GPU(s)...\n\n");
for(unsigned int i = 0; i < ciDeviceCount; i++)
{
ciErrNum = clEnqueueNDRangeKernel(commandQueue[i], reduceKernel[i], 1, 0, globalWorkSize, localWorkSize,
0, NULL, &GPUExecution[i]);
oclCheckError(ciErrNum, CL_SUCCESS);
}
// Copy result from device to host for each device
for(unsigned int i = 0; i < ciDeviceCount; i++)
{
ciErrNum = clEnqueueReadBuffer(commandQueue[i], d_Result[i], CL_FALSE, 0, ACCUM_N * sizeof(float),
h_SumGPU + i * ACCUM_N, 0, NULL, &GPUDone[i]);
oclCheckError(ciErrNum, CL_SUCCESS);
}
// Synchronize with the GPUs and do accumulated error check
clWaitForEvents(ciDeviceCount, GPUDone);
shrLog("clWaitForEvents complete...\n\n");
// Aggregate results for multiple GPU's and stop/log processing time
sumGPU = 0;
for(unsigned int i = 0; i < ciDeviceCount * ACCUM_N; i++)
{
sumGPU += h_SumGPU[i];
}
// Print Execution Times for each GPU
#ifdef GPU_PROFILING
shrLog("Profiling Information for GPU Processing:\n\n");
for(unsigned int i = 0; i < ciDeviceCount; i++)
{
cdDevice = oclGetDev(cxGPUContext, deviceNr[i]);
clGetDeviceInfo(cdDevice, CL_DEVICE_NAME, sizeof(cDeviceName), cDeviceName, NULL);
shrLog("Device %i : %s\n", deviceNr[i], cDeviceName);
shrLog(" Reduce Kernel : %.5f s\n", executionTime(GPUExecution[i]));
shrLog(" Copy Device->Host : %.5f s\n\n\n", executionTime(GPUDone[i]));
}
#endif
// Run the computation on the Host CPU and log processing time
shrLog("Launching Host/CPU C++ Computation...\n\n");
sumCPU = 0;
for(unsigned int i = 0; i < DATA_N; i++)
{
sumCPU += h_Data[i];
}
// Check GPU result against CPU result
dRelError = 100.0 * fabs(sumCPU - sumGPU) / fabs(sumCPU);
shrLog("Comparing against Host/C++ computation...\n");
shrLog(" GPU sum: %f\n CPU sum: %f\n", sumGPU, sumCPU);
shrLog(" Relative Error (100.0 * Error / Golden) = %f \n\n", dRelError);
// cleanup
free(source);
free(h_Data);
for(unsigned int i = 0; i < ciDeviceCount; ++i )
{
clReleaseKernel(reduceKernel[i]);
clReleaseCommandQueue(commandQueue[i]);
}
clReleaseProgram(cpProgram);
clReleaseContext(cxGPUContext);
// finish
shrQAFinishExit(argc, (const char **)argv, (dRelError < 1e-4) ? QA_PASSED : QA_FAILED);
}
// Main function
// *********************************************************************
int ymain(int argc, char **argv)
{
shrQAStart(argc, argv);
// get command line arg for quick test, if provided
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
// start logs
cExecutableName = argv[0];
shrSetLogFileName ("oclVectorAdd2.txt");
shrLog("%s Starting...\n\n# of float elements per Array \t= %i\n", argv[0], iNumElements);
// set and log Global and Local work size dimensions
szLocalWorkSize = 256;
szGlobalWorkSize = shrRoundUp((int)szLocalWorkSize, iNumElements); // rounded up to the nearest multiple of the LocalWorkSize
shrLog("Global Work Size \t\t= %u\nLocal Work Size \t\t= %u\n# of Work Groups \t\t= %u\n\n",
szGlobalWorkSize, szLocalWorkSize, (szGlobalWorkSize % szLocalWorkSize + szGlobalWorkSize/szLocalWorkSize));
// Allocate and initialize host arrays
shrLog( "Allocate and Init Host Mem...\n");
srcA = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
srcB = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
dst = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
Golden = (void *)malloc(sizeof(cl_float) * iNumElements);
shrFillArray((float*)srcA, iNumElements);
shrFillArray((float*)srcB, iNumElements);
//Get an OpenCL platform
ciErr1 = clGetPlatformIDs(1, &cpPlatform, NULL);
shrLog("clGetPlatformID...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clGetPlatformID, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
//Get the devices
ciErr1 = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &cdDevice, NULL);
shrLog("clGetDeviceIDs...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clGetDeviceIDs, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
//Create the context
cxGPUContext = clCreateContext(0, 1, &cdDevice, NULL, NULL, &ciErr1);
shrLog("clCreateContext...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateContext, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevice, 0, &ciErr1);
shrLog("clCreateCommandQueue...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateCommandQueue, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
cmDevSrcA = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr1);
cmDevSrcB = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
cmDevDst = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
shrLog("clCreateBuffer...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Read the OpenCL kernel in from source file
shrLog("oclLoadProgSource (%s)...\n", cSourceFile);
cPathAndName = shrFindFilePath(cSourceFile, argv[0]);
shrLog("Looking for: %s in Path: %s\n", cSourceFile, argv[0]);
cSourceCL = oclLoadProgSource(cPathAndName, "", &szKernelLength);
// Create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErr1);
shrLog("clCreateProgramWithSource...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateProgramWithSource, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Build the program with 'mad' Optimization option
#ifdef MAC
char* flags = "-cl-fast-relaxed-math -DMAC";
#else
char* flags = "-cl-fast-relaxed-math";
#endif
ciErr1 = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL);
shrLog("clBuildProgram...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clBuildProgram, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Create the kernel
ckKernel = clCreateKernel(cpProgram, "VectorAdd", &ciErr1);
shrLog("clCreateKernel (VectorAdd)...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Set the Argument values
ciErr1 = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void*)&cmDevSrcA);
ciErr1 |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void*)&cmDevSrcB);
ciErr1 |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void*)&cmDevDst);
ciErr1 |= clSetKernelArg(ckKernel, 3, sizeof(cl_int), (void*)&iNumElements);
shrLog("clSetKernelArg 0 - 3...\n\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clSetKernelArg, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// --------------------------------------------------------
// Start Core sequence... copy input data to GPU, compute, copy results back
// Asynchronous write of data to GPU device
ciErr1 = clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcA, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, srcA, 0, NULL, NULL);
ciErr1 |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcB, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, srcB, 0, NULL, NULL);
shrLog("clEnqueueWriteBuffer (SrcA and SrcB)...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clEnqueueWriteBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Launch kernel
ciErr1 = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, NULL);
shrLog("clEnqueueNDRangeKernel (VectorAdd)...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clEnqueueNDRangeKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Synchronous/blocking read of results, and check accumulated errors
ciErr1 = clEnqueueReadBuffer(cqCommandQueue, cmDevDst, CL_TRUE, 0, sizeof(cl_float) * szGlobalWorkSize, dst, 0, NULL, NULL);
shrLog("clEnqueueReadBuffer (Dst)...\n\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clEnqueueReadBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
//--------------------------------------------------------
// Compute and compare results for golden-host and report errors and pass/fail
shrLog("Comparing against Host/C++ computation...\n\n");
VectorAddHost ((const float*)srcA, (const float*)srcB, (float*)Golden, iNumElements);
shrBOOL bMatch = shrComparefet((const float*)Golden, (const float*)dst, (unsigned int)iNumElements, 0.0f, 0);
// Cleanup and leave
Cleanup (argc, argv, (bMatch == shrTRUE) ? EXIT_SUCCESS : EXIT_FAILURE);
}
// Main function
// *********************************************************************
int main( int argc, const char** argv)
{
shrQAStart(argc, (char **)argv);
// start logs
shrSetLogFileName ("oclReduction.txt");
shrLog("%s Starting...\n\n", argv[0]);
char *typeChoice;
shrGetCmdLineArgumentstr(argc, argv, "type", &typeChoice);
// determine type of array from command line args
if (0 == typeChoice)
{
typeChoice = (char*)malloc(7 * sizeof(char));
#ifdef WIN32
strcpy_s(typeChoice, 7 * sizeof(char) + 1, "int");
#else
strcpy(typeChoice, "int");
#endif
}
ReduceType datatype = REDUCE_INT;
#ifdef WIN32
if (!_strcmpi(typeChoice, "float"))
datatype = REDUCE_FLOAT;
else if (!_strcmpi(typeChoice, "double"))
datatype = REDUCE_DOUBLE;
else
datatype = REDUCE_INT;
#else
if (!strcmp(typeChoice, "float"))
datatype = REDUCE_FLOAT;
else if (!strcmp(typeChoice, "double"))
datatype = REDUCE_DOUBLE;
else
datatype = REDUCE_INT;
#endif
shrLog("Reducing array of type %s.\n", typeChoice);
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckError(ciErrNum, CL_SUCCESS);
//Get the devices
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckError(ciErrNum, CL_SUCCESS);
cl_device_id *cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
//Create the context
cxGPUContext = clCreateContext(0, uiNumDevices, cdDevices, NULL, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// get and log the device info
if( shrCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
int device_nr = 0;
shrGetCmdLineArgumenti(argc, (const char**)argv, "device", &device_nr);
if( device_nr < uiNumDevices ) {
device = oclGetDev(cxGPUContext, device_nr);
} else {
shrLog("Invalid Device %d Requested.\n", device_nr);
shrExitEX(argc, argv, EXIT_FAILURE);
}
} else {
device = oclGetMaxFlopsDev(cxGPUContext);
}
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, device, 0, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
source_path = shrFindFilePath("oclReduction_kernel.cl", argv[0]);
bool bSuccess = false;
switch (datatype)
{
default:
case REDUCE_INT:
bSuccess = runTest<int>( argc, argv, datatype);
break;
case REDUCE_FLOAT:
bSuccess = runTest<float>( argc, argv, datatype);
break;
}
// finish
shrQAFinishExit(argc, (const char **)argv, bSuccess ? QA_PASSED : QA_FAILED);
}
int main( int argc, char** argv )
{
bool setQuit = false;
int capWidth = gw;
int capHeight = gh;
state.use_IPP = false;
state.bqatest = false;
unsigned int devN = 0;
if( shrGetCmdLineArgumentu(argc, (const char **)argv, "device", &devN ) ) {
printf("Using device %d\n", devN);
}
if( shrCheckCmdLineFlag( argc, (const char **)argv, "qatest") ) {
printf("QA test mode.\n");
capture = NULL;
state.bqatest = true;
} else {
printf("Attempting to initialize camera\n");
if( argc == 1 || (argc == 2 && strlen(argv[1]) == 1 && isdigit(argv[1][0])))
capture = cvCaptureFromCAM( argc == 2 ? argv[1][0] - '0' : 0 );
else if( argc == 2 )
capture = cvCaptureFromAVI( argv[1] );
}
if( !capture )
{
fprintf(stderr,"Could not initialize capturing...\n");
fprintf(stderr,"Attempting to use PGM files..\n");
} else {
printf("Camera Initialized\n");
printf("Setting Size\n");
#ifdef _MSC_VER
Sleep(5000);
#else
sleep(5); // pause 5 seconds before setting size on Mac , otherwise unstable
#endif
cvSetCaptureProperty(capture, CV_CAP_PROP_FRAME_WIDTH, gw);
cvSetCaptureProperty(capture, CV_CAP_PROP_FRAME_HEIGHT, gh);
}
initGlut(argc,argv, gw, gh);
initGLData(gw, gh, NULL);
// Initialize openCL
// loads the CL functions from the .cl files
// also loads reference images../data/minicooper/frame10.pgm
clCtx = initOCLFlow(vbo,devN);
if( !capture ) {
if( images[0].w != gw || images[0].h != gh || images[1].w != gw || images[1].h != gh ) {
fprintf(stderr, "Bad image sizes supplied. Please use %d x %d images\n", gw, gh );
}
// load the file into the texture (actually initOCLFLow loaded them into GPU but discarded them
// so this is justa quick way to load them into the texture as well)
unsigned int w, h;
unsigned char *image_ub = NULL;
shrLoadPGMub( "data/minicooper/frame10.pgm", (unsigned char **)&image_ub, &w, &h );
glBindTexture(GL_TEXTURE_RECTANGLE_NV, tex);
glTexSubImage2D(GL_TEXTURE_RECTANGLE_NV, 0, 0, 0, gw, gh, GL_LUMINANCE, GL_UNSIGNED_BYTE, image_ub );
}
glutMainLoop();
return 0;
}
bool runTest(int argc, const char **argv)
{
bool ok = true;
float *host_output;
float *device_output;
float *input;
float *coeff;
int defaultDim;
int dimx;
int dimy;
int dimz;
int outerDimx;
int outerDimy;
int outerDimz;
int radius;
int timesteps;
size_t volumeSize;
memsize_t memsize;
const float lowerBound = 0.0f;
const float upperBound = 1.0f;
// Determine default dimensions
shrLog("Set-up, based upon target device GMEM size...\n");
if (ok)
{
// Get the memory size of the target device
shrLog(" getTargetDeviceGlobalMemSize\n");
ok = getTargetDeviceGlobalMemSize(&memsize, argc, argv);
}
if (ok)
{
// We can never use all the memory so to keep things simple we aim to
// use around half the total memory
memsize /= 2;
// Most of our memory use is taken up by the input and output buffers -
// two buffers of equal size - and for simplicity the volume is a cube:
// dim = floor( (N/2)^(1/3) )
defaultDim = (int)floor(pow((memsize / (2.0 * sizeof(float))), 1.0/3.0));
// By default, make the volume edge size an integer multiple of 128B to
// improve performance by coalescing memory accesses, in a real
// application it would make sense to pad the lines accordingly
int roundTarget = 128 / sizeof(float);
defaultDim = defaultDim / roundTarget * roundTarget;
defaultDim -= k_radius_default * 2;
// Check dimension is valid
if (defaultDim < k_dim_min)
{
shrLogEx(LOGBOTH | ERRORMSG, -1000, STDERROR);
shrLog("\tinsufficient device memory (maximum volume on device is %d, must be between %d and %d).\n", defaultDim, k_dim_min, k_dim_max);
ok = false;
}
else if (defaultDim > k_dim_max)
{
defaultDim = k_dim_max;
}
}
// For QA testing, override default volume size
if (ok)
{
if (shrCheckCmdLineFlag(argc, argv, "qatest"))
{
defaultDim = MIN(defaultDim, k_dim_qa);
}
}
// Parse command line arguments
if (ok)
{
char *dim = 0;
if (shrGetCmdLineArgumentstr(argc, argv, "dimx", &dim))
{
dimx = (int)atoi(dim);
if (dimx < k_dim_min || dimx > k_dim_max)
{
shrLogEx(LOGBOTH | ERRORMSG, -1001, STDERROR);
shrLog("\tdimx out of range (%d requested, must be between %d and %d), see header files for details.\n", dimx, k_dim_min, k_dim_max);
ok = false;
}
}
else
{
dimx = defaultDim;
}
if (shrGetCmdLineArgumentstr(argc, argv, "dimy", &dim))
{
dimy = (int)atoi(dim);
if (dimy < k_dim_min || dimy > k_dim_max)
{
shrLogEx(LOGBOTH | ERRORMSG, -1002, STDERROR);
shrLog("\tdimy out of range (%d requested, must be between %d and %d), see header files for details.\n", dimy, k_dim_min, k_dim_max);
ok = false;
}
}
else
{
dimy = defaultDim;
}
if (shrGetCmdLineArgumentstr(argc, argv, "dimz", &dim))
{
dimz = (int)atoi(dim);
if (dimz < k_dim_min || dimz > k_dim_max)
{
shrLogEx(LOGBOTH | ERRORMSG, -1003, STDERROR);
shrLog("\tdimz out of range (%d requested, must be between %d and %d), see header files for details.\n", dimz, k_dim_min, k_dim_max);
ok = false;
}
}
else
{
dimz = defaultDim;
}
if (shrGetCmdLineArgumentstr(argc, argv, "radius", &dim))
{
radius = (int)atoi(dim);
if (radius < k_radius_min || radius >= k_radius_max)
{
shrLogEx(LOGBOTH | ERRORMSG, -1004, STDERROR);
shrLog("\tradius out of range (%d requested, must be between %d and %d), see header files for details.\n", radius, k_radius_min, k_radius_max);
ok = false;
}
}
else
{
radius = k_radius_default;
}
if (shrGetCmdLineArgumentstr(argc, argv, "timesteps", &dim))
{
timesteps = (int)atoi(dim);
if (timesteps < k_timesteps_min || radius >= k_timesteps_max)
{
shrLogEx(LOGBOTH | ERRORMSG, -1005, STDERROR);
shrLog("\ttimesteps out of range (%d requested, must be between %d and %d), see header files for details.\n", timesteps, k_timesteps_min, k_timesteps_max);
ok = false;
}
}
else
{
timesteps = k_timesteps_default;
}
if (dim)
free(dim);
}
// Determine volume size
if (ok)
{
outerDimx = dimx + 2 * radius;
outerDimy = dimy + 2 * radius;
outerDimz = dimz + 2 * radius;
volumeSize = outerDimx * outerDimy * outerDimz;
}
// Allocate memory
if (ok)
{
shrLog(" calloc host_output\n");
if ((host_output = (float *)calloc(volumeSize, sizeof(float))) == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -1006, STDERROR);
shrLog("\tInsufficient memory for host_output calloc, please try a smaller volume (use --help for syntax).\n");
ok = false;
}
}
if (ok)
{
shrLog(" malloc input\n");
if ((input = (float *)malloc(volumeSize * sizeof(float))) == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -1007, STDERROR);
shrLog("\tInsufficient memory for input malloc, please try a smaller volume (use --help for syntax).\n");
ok = false;
}
}
if (ok)
{
shrLog(" malloc coeff\n");
if ((coeff = (float *)malloc((radius + 1) * sizeof(float))) == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -1008, STDERROR);
shrLog("\tInsufficient memory for coeff malloc, please try a smaller volume (use --help for syntax).\n");
ok = false;
}
}
// Create coefficients
if (ok)
{
for (int i = 0 ; i <= radius ; i++)
{
coeff[i] = 0.1f;
}
}
// Generate data
if (ok)
{
shrLog(" generateRandomData\n\n");
generateRandomData(input, outerDimx, outerDimy, outerDimz, lowerBound, upperBound);
}
if (ok)
{
shrLog("FDTD on %d x %d x %d volume with symmetric filter radius %d for %d timesteps...\n\n", dimx, dimy, dimz, radius, timesteps);
}
// Execute on the host
if (ok)
{
shrLog("fdtdReference...\n");
ok = fdtdReference(host_output, input, coeff, dimx, dimy, dimz, radius, timesteps);
shrLog("fdtdReference complete\n");
}
// Allocate memory
if (ok)
{
shrLog(" calloc device_output\n");
if ((device_output = (float *)calloc(volumeSize, sizeof(float))) == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -1009, STDERROR);
shrLog("\tInsufficient memory for device output calloc, please try a smaller volume (use --help for syntax).\n");
ok = false;
}
}
// Execute on the device
if (ok)
{
shrLog("fdtdGPU...\n");
ok = fdtdGPU(device_output, input, coeff, dimx, dimy, dimz, radius, timesteps, argc, argv);
shrLog("fdtdGPU complete\n");
}
// Compare the results
if (ok)
{
float tolerance = 0.0001f;
shrLog("\nCompareData (tolerance %f)...\n", tolerance);
ok = compareData(device_output, host_output, dimx, dimy, dimz, radius, tolerance);
}
return ok;
}
// Main function
// *********************************************************************
int main(int argc, char** argv)
{
shrQAStart(argc, argv);
int use_gpu = 0;
for(int i = 0; i < argc && argv; i++)
{
if(!argv[i])
continue;
if(strstr(argv[i], "cpu"))
use_gpu = 0;
else if(strstr(argv[i], "gpu"))
use_gpu = 1;
}
// start logs
shrSetLogFileName ("oclDXTCompression.txt");
shrLog("%s Starting...\n\n", argv[0]);
cl_platform_id cpPlatform = NULL;
cl_uint uiNumDevices = 0;
cl_device_id *cdDevices = NULL;
cl_context cxGPUContext;
cl_command_queue cqCommandQueue;
cl_program cpProgram;
cl_kernel ckKernel;
cl_mem cmMemObjs[3];
cl_mem cmAlphaTable4, cmProds4;
cl_mem cmAlphaTable3, cmProds3;
size_t szGlobalWorkSize[1];
size_t szLocalWorkSize[1];
cl_int ciErrNum;
// Get the path of the filename
char *filename;
if (shrGetCmdLineArgumentstr(argc, (const char **)argv, "image", &filename)) {
image_filename = filename;
}
// load image
const char* image_path = shrFindFilePath(image_filename, argv[0]);
oclCheckError(image_path != NULL, shrTRUE);
shrLoadPPM4ub(image_path, (unsigned char **)&h_img, &width, &height);
oclCheckError(h_img != NULL, shrTRUE);
shrLog("Loaded '%s', %d x %d pixels\n\n", image_path, width, height);
// Convert linear image to block linear.
const uint memSize = width * height * sizeof(cl_uint);
uint* block_image = (uint*)malloc(memSize);
// Convert linear image to block linear.
for(uint by = 0; by < height/4; by++) {
for(uint bx = 0; bx < width/4; bx++) {
for (int i = 0; i < 16; i++) {
const int x = i & 3;
const int y = i / 4;
block_image[(by * width/4 + bx) * 16 + i] =
((uint *)h_img)[(by * 4 + y) * 4 * (width/4) + bx * 4 + x];
}
}
}
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckError(ciErrNum, CL_SUCCESS);
// Get the platform's GPU devices
ciErrNum = clGetDeviceIDs(cpPlatform, use_gpu?CL_DEVICE_TYPE_GPU:CL_DEVICE_TYPE_CPU, 0, NULL, &uiNumDevices);
oclCheckError(ciErrNum, CL_SUCCESS);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, use_gpu?CL_DEVICE_TYPE_GPU:CL_DEVICE_TYPE_CPU, uiNumDevices, cdDevices, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
// Create the context
cxGPUContext = clCreateContext(0, uiNumDevices, cdDevices, NULL, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// get and log device
cl_device_id device;
if( shrCheckCmdLineFlag(argc, (const char **)argv, "device") ) {
int device_nr = 0;
shrGetCmdLineArgumenti(argc, (const char **)argv, "device", &device_nr);
device = oclGetDev(cxGPUContext, device_nr);
if( device == (cl_device_id)-1 ) {
shrLog(" Invalid GPU Device: devID=%d. %d valid GPU devices detected\n\n", device_nr, uiNumDevices);
shrLog(" exiting...\n");
return -1;
}
} else {
device = oclGetMaxFlopsDev(cxGPUContext);
}
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, device, 0, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Memory Setup
// Constants
cmAlphaTable4 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_float), (void*)&alphaTable4[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
cmProds4 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_int), (void*)&prods4[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
cmAlphaTable3 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_float), (void*)&alphaTable3[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
cmProds3 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_int), (void*)&prods3[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Compute permutations.
cl_uint permutations[1024];
computePermutations(permutations);
// Upload permutations.
cmMemObjs[0] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(cl_uint) * 1024, permutations, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Image
cmMemObjs[1] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, memSize, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Result
const uint compressedSize = (width / 4) * (height / 4) * 8;
cmMemObjs[2] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, compressedSize, NULL , &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
unsigned int * h_result = (uint*)malloc(compressedSize);
// Program Setup
size_t program_length;
const char* source_path = shrFindFilePath("DXTCompression.cl", argv[0]);
oclCheckError(source_path != NULL, shrTRUE);
char *source = oclLoadProgSource(source_path, "", &program_length);
oclCheckError(source != NULL, shrTRUE);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,
(const char **) &source, &program_length, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-fast-relaxed-math", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclDXTCompression.ptx");
oclCheckError(ciErrNum, CL_SUCCESS);
}
// create the kernel
ckKernel = clCreateKernel(cpProgram, "compress", &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// set the args values
ciErrNum = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void *) &cmMemObjs[0]);
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void *) &cmMemObjs[1]);
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void *) &cmMemObjs[2]);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(cl_mem), (void*)&cmAlphaTable4);
ciErrNum |= clSetKernelArg(ckKernel, 4, sizeof(cl_mem), (void*)&cmProds4);
ciErrNum |= clSetKernelArg(ckKernel, 5, sizeof(cl_mem), (void*)&cmAlphaTable3);
ciErrNum |= clSetKernelArg(ckKernel, 6, sizeof(cl_mem), (void*)&cmProds3);
oclCheckError(ciErrNum, CL_SUCCESS);
// Copy input data host to device
clEnqueueWriteBuffer(cqCommandQueue, cmMemObjs[1], CL_FALSE, 0, sizeof(cl_uint) * width * height, block_image, 0,0,0);
// Determine launch configuration and run timed computation numIterations times
int blocks = ((width + 3) / 4) * ((height + 3) / 4); // rounds up by 1 block in each dim if %4 != 0
// Restrict the numbers of blocks to launch on low end GPUs to avoid kernel timeout
cl_uint compute_units;
clGetDeviceInfo(device, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(compute_units), &compute_units, NULL);
int blocksPerLaunch = MIN(blocks, 768 * (int)compute_units);
// set work-item dimensions
szGlobalWorkSize[0] = blocksPerLaunch * NUM_THREADS;
szLocalWorkSize[0]= NUM_THREADS;
#ifdef GPU_PROFILING
shrLog("\nRunning DXT Compression on %u x %u image...\n", width, height);
shrLog("\n%u Workgroups, %u Work Items per Workgroup, %u Work Items in NDRange...\n\n",
blocks, NUM_THREADS, blocks * NUM_THREADS);
int numIterations = 50;
for (int i = -1; i < numIterations; ++i) {
if (i == 0) { // start timing only after the first warmup iteration
clFinish(cqCommandQueue); // flush command queue
shrDeltaT(0); // start timer
}
#endif
// execute kernel
for( int j=0; j<blocks; j+= blocksPerLaunch ) {
clSetKernelArg(ckKernel, 7, sizeof(int), &j);
szGlobalWorkSize[0] = MIN( blocksPerLaunch, blocks-j ) * NUM_THREADS;
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL,
szGlobalWorkSize, szLocalWorkSize,
0, NULL, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
}
#ifdef GPU_PROFILING
}
clFinish(cqCommandQueue);
double dAvgTime = shrDeltaT(0) / (double)numIterations;
shrLogEx(LOGBOTH | MASTER, 0, "oclDXTCompression, Throughput = %.4f MPixels/s, Time = %.5f s, Size = %u Pixels, NumDevsUsed = %i, Workgroup = %d\n",
(1.0e-6 * (double)(width * height)/ dAvgTime), dAvgTime, (width * height), 1, szLocalWorkSize[0]);
#endif
// blocking read output
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmMemObjs[2], CL_TRUE, 0,
compressedSize, h_result, 0, NULL, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
// Write DDS file.
FILE* fp = NULL;
char output_filename[1024];
#ifdef WIN32
strcpy_s(output_filename, 1024, image_path);
strcpy_s(output_filename + strlen(image_path) - 3, 1024 - strlen(image_path) + 3, "dds");
fopen_s(&fp, output_filename, "wb");
#else
strcpy(output_filename, image_path);
strcpy(output_filename + strlen(image_path) - 3, "dds");
fp = fopen(output_filename, "wb");
#endif
oclCheckError(fp != NULL, shrTRUE);
DDSHeader header;
header.fourcc = FOURCC_DDS;
header.size = 124;
header.flags = (DDSD_WIDTH|DDSD_HEIGHT|DDSD_CAPS|DDSD_PIXELFORMAT|DDSD_LINEARSIZE);
header.height = height;
header.width = width;
header.pitch = compressedSize;
header.depth = 0;
header.mipmapcount = 0;
memset(header.reserved, 0, sizeof(header.reserved));
header.pf.size = 32;
header.pf.flags = DDPF_FOURCC;
header.pf.fourcc = FOURCC_DXT1;
header.pf.bitcount = 0;
header.pf.rmask = 0;
header.pf.gmask = 0;
header.pf.bmask = 0;
header.pf.amask = 0;
header.caps.caps1 = DDSCAPS_TEXTURE;
header.caps.caps2 = 0;
header.caps.caps3 = 0;
header.caps.caps4 = 0;
header.notused = 0;
fwrite(&header, sizeof(DDSHeader), 1, fp);
fwrite(h_result, compressedSize, 1, fp);
fclose(fp);
// Make sure the generated image matches the reference image (regression check)
shrLog("\nComparing against Host/C++ computation...\n");
const char* reference_image_path = shrFindFilePath(refimage_filename, argv[0]);
oclCheckError(reference_image_path != NULL, shrTRUE);
// read in the reference image from file
#ifdef WIN32
fopen_s(&fp, reference_image_path, "rb");
#else
fp = fopen(reference_image_path, "rb");
#endif
oclCheckError(fp != NULL, shrTRUE);
fseek(fp, sizeof(DDSHeader), SEEK_SET);
uint referenceSize = (width / 4) * (height / 4) * 8;
uint * reference = (uint *)malloc(referenceSize);
fread(reference, referenceSize, 1, fp);
fclose(fp);
// compare the reference image data to the sample/generated image
float rms = 0;
for (uint y = 0; y < height; y += 4)
{
for (uint x = 0; x < width; x += 4)
{
// binary comparison of data
uint referenceBlockIdx = ((y/4) * (width/4) + (x/4));
uint resultBlockIdx = ((y/4) * (width/4) + (x/4));
int cmp = compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
// log deviations, if any
if (cmp != 0.0f)
{
compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
shrLog("Deviation at (%d, %d):\t%f rms\n", x/4, y/4, float(cmp)/16/3);
}
rms += cmp;
}
}
rms /= width * height * 3;
shrLog("RMS(reference, result) = %f\n\n", rms);
// Free OpenCL resources
oclDeleteMemObjs(cmMemObjs, 3);
clReleaseMemObject(cmAlphaTable4);
clReleaseMemObject(cmProds4);
clReleaseMemObject(cmAlphaTable3);
clReleaseMemObject(cmProds3);
clReleaseKernel(ckKernel);
clReleaseProgram(cpProgram);
clReleaseCommandQueue(cqCommandQueue);
clReleaseContext(cxGPUContext);
// Free host memory
free(source);
free(h_img);
// finish
shrQAFinishExit(argc, (const char **)argv, (rms <= ERROR_THRESHOLD) ? QA_PASSED : QA_FAILED);
}
// Main function
// *********************************************************************
int main(int argc, char **argv)
{
//////////////////////////////////////////////////////////////////////////
unsigned int count = iNumElements;
int k = 8;
unsigned int random_seed, random_seed2;
srand( (unsigned)time( NULL ) );
random_seed = rand();
random_seed2 = rand();
//////////////////////////////////////////////////////////////////////////
// get command line arg for quick test, if provided
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
// start logs
shrSetLogFileName ("oclVectorAdd.txt");
shrLog("%s Starting...\n\n# of float elements per Array \t= %i\n", argv[0], iNumElements);
// set and log Global and Local work size dimensions
szLocalWorkSize = 256;
szGlobalWorkSize = shrRoundUp((int)szLocalWorkSize, iNumElements); // rounded up to the nearest multiple of the LocalWorkSize
shrLog("Global Work Size \t\t= %u\nLocal Work Size \t\t= %u\n# of Work Groups \t\t= %u\n\n",
szGlobalWorkSize, szLocalWorkSize, (szGlobalWorkSize % szLocalWorkSize + szGlobalWorkSize/szLocalWorkSize));
// Allocate and initialize host arrays
shrLog( "Allocate and Init Host Mem...\n");
srcA = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
srcB = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
dst = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
Golden = (void *)malloc(sizeof(cl_float) * iNumElements);
shrFillArray((float*)srcA, iNumElements);
shrFillArray((float*)srcB, iNumElements);
//////////////////////////////////////////////////////////////////////////
float *scalar_value = new float[count];
float *gradient_magnitude = new float[count];
float *second_derivative_magnitude = new float[count];
unsigned char *label_ptr = new unsigned char[count];
shrFillArray(scalar_value, count);
shrFillArray(gradient_magnitude, count);
shrFillArray(second_derivative_magnitude, count);
//////////////////////////////////////////////////////////////////////////
//Get an OpenCL platform
ciErr1 = clGetPlatformIDs(1, &cpPlatform, NULL);
shrLog("clGetPlatformID...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clGetPlatformID, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
//Get the devices
ciErr1 = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &cdDevice, NULL);
shrLog("clGetDeviceIDs...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clGetDeviceIDs, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
//Create the context
cxGPUContext = clCreateContext(0, 1, &cdDevice, NULL, NULL, &ciErr1);
shrLog("clCreateContext...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateContext, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevice, 0, &ciErr1);
shrLog("clCreateCommandQueue...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateCommandQueue, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
cmDevSrcA = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr1);
cmDevSrcB = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
cmDevDst = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
//////////////////////////////////////////////////////////////////////////
cmDevSrc_scalar_value = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr1);
cmDevSrc_gradient_magnitude = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
cmDevSrc_second_derivative_magnitude = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
cmDevDst_label_ptr = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErr2);
ciErr1 |= ciErr2;
//////////////////////////////////////////////////////////////////////////
shrLog("clCreateBuffer...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Read the OpenCL kernel in from source file
shrLog("oclLoadProgSource (%s)...\n", cSourceFile);
cPathAndName = shrFindFilePath(cSourceFile, argv[0]);
cSourceCL = oclLoadProgSource(cPathAndName, "", &szKernelLength);
printf("%s\n%s\n", cSourceFile, cPathAndName);
// Create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErr1);
shrLog("clCreateProgramWithSource...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateProgramWithSource, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Build the program with 'mad' Optimization option
#ifdef MAC
char* flags = "-cl-fast-relaxed-math -DMAC";
#else
char* flags = "-cl-fast-relaxed-math";
#endif
ciErr1 = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL);
shrLog("clBuildProgram...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clBuildProgram, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Create the kernel
ckKernel = clCreateKernel(cpProgram, "k_means", &ciErr1);
shrLog("clCreateKernel (VectorAdd)...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clCreateKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Set the Argument values
//ciErr1 = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void*)&cmDevSrcA);
//ciErr1 |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void*)&cmDevSrcB);
//ciErr1 |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void*)&cmDevDst);
//ciErr1 |= clSetKernelArg(ckKernel, 3, sizeof(cl_int), (void*)&iNumElements);
//////////////////////////////////////////////////////////////////////////
// __global const float *scalar_value, __global const float *gradient_magnitude, __global const float *second_derivative_magnitude, __global unsigned char *label_ptr, __global const unsigned int count, __global const int k, __global const unsigned int random_seed, __global const unsigned int random_seed2
ciErr1 = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void*)&cmDevSrc_scalar_value);
ciErr1 |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void*)&cmDevSrc_gradient_magnitude);
ciErr1 |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void*)&cmDevSrc_second_derivative_magnitude);
ciErr1 |= clSetKernelArg(ckKernel, 3, sizeof(cl_mem), (void*)&cmDevDst_label_ptr);
ciErr1 |= clSetKernelArg(ckKernel, 4, sizeof(cl_uint), (void*)&count);
ciErr1 |= clSetKernelArg(ckKernel, 5, sizeof(cl_uint), (void*)&k);
ciErr1 |= clSetKernelArg(ckKernel, 6, sizeof(cl_uint), (void*)&random_seed);
ciErr1 |= clSetKernelArg(ckKernel, 7, sizeof(cl_uint), (void*)&random_seed2);
//////////////////////////////////////////////////////////////////////////
shrLog("clSetKernelArg 0 - 3...\n\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clSetKernelArg, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// --------------------------------------------------------
// Start Core sequence... copy input data to GPU, compute, copy results back
// Asynchronous write of data to GPU device
//ciErr1 = clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcA, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, srcA, 0, NULL, NULL);
//ciErr1 |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcB, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, srcB, 0, NULL, NULL);
//////////////////////////////////////////////////////////////////////////
ciErr1 = clEnqueueWriteBuffer(cqCommandQueue, cmDevSrc_scalar_value, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, scalar_value, 0, NULL, NULL);
ciErr1 |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrc_gradient_magnitude, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, gradient_magnitude, 0, NULL, NULL);
ciErr1 |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrc_second_derivative_magnitude, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize, second_derivative_magnitude, 0, NULL, NULL);
//////////////////////////////////////////////////////////////////////////
shrLog("clEnqueueWriteBuffer (SrcA and SrcB)...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clEnqueueWriteBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Launch kernel
ciErr1 = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, NULL);
shrLog("clEnqueueNDRangeKernel (VectorAdd)...\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clEnqueueNDRangeKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
// Synchronous/blocking read of results, and check accumulated errors
//ciErr1 = clEnqueueReadBuffer(cqCommandQueue, cmDevDst, CL_TRUE, 0, sizeof(cl_float) * szGlobalWorkSize, dst, 0, NULL, NULL);
//////////////////////////////////////////////////////////////////////////
ciErr1 = clEnqueueReadBuffer(cqCommandQueue, cmDevDst_label_ptr, CL_TRUE, 0, sizeof(cl_float) * szGlobalWorkSize, label_ptr, 0, NULL, NULL);
//////////////////////////////////////////////////////////////////////////
shrLog("clEnqueueReadBuffer (Dst)...\n\n");
if (ciErr1 != CL_SUCCESS)
{
shrLog("Error in clEnqueueReadBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(EXIT_FAILURE);
}
//--------------------------------------------------------
// Compute and compare results for golden-host and report errors and pass/fail
shrLog("Comparing against Host/C++ computation...\n\n");
VectorAddHost ((const float*)srcA, (const float*)srcB, (float*)Golden, iNumElements);
shrBOOL bMatch = shrComparefet((const float*)Golden, (const float*)dst, (unsigned int)iNumElements, 0.0f, 0);
shrLog("%s\n\n", (bMatch == shrTRUE) ? "PASSED" : "FAILED");
//////////////////////////////////////////////////////////////////////////
//float *a = (float *)srcA;
//float *b = (float *)srcB;
//float *c = (float *)dst;
//float *d = (float *)Golden;
//for (int i=0; i<iNumElements; i++)
//{
// printf("%f+%f=%f=%f\t", a[i], b[i], c[i], a[i]+b[i]);
// printf("%s\n", (a[i]+b[i]==c[i]?"equal":"not equal"));
//}
//for (int i=0; i<iNumElements; i++)
//{
// printf("%f\n", ((float *)dst)[i]);
//}
//////////////////////////////////////////////////////////////////////////
// Cleanup and leave
Cleanup (EXIT_SUCCESS);
//////////////////////////////////////////////////////////////////////////
delete [] scalar_value;
delete [] gradient_magnitude;
delete [] second_derivative_magnitude;
delete [] label_ptr;
//////////////////////////////////////////////////////////////////////////
}
// Main program
//*****************************************************************************
int main(int argc, char** argv)
{
pArgc = &argc;
pArgv = argv;
shrQAStart(argc, argv);
// Start logs
cExecutableName = argv[0];
shrSetLogFileName ("oclSobelFilter.txt");
shrLog("%s Starting (Using %s)...\n\n", argv[0], clSourcefile);
// Get command line args for quick test or QA test, if provided
bNoPrompt = (bool)shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
bQATest = (bool)shrCheckCmdLineFlag(argc, (const char**)argv, "qatest");
// Menu items
if (!(bQATest))
{
ShowMenuItems();
}
// Find the path from the exe to the image file
cPathAndName = shrFindFilePath(cImageFile, argv[0]);
oclCheckErrorEX(cPathAndName != NULL, shrTRUE, pCleanup);
shrLog("Image File\t = %s\nImage Dimensions = %u w x %u h x %u bpp\n\n", cPathAndName, uiImageWidth, uiImageHeight, sizeof(unsigned int)<<3);
// Initialize OpenGL items (if not No-GL QA test)
shrLog("%sInitGL...\n\n", bQATest ? "Skipping " : "Calling ");
if (!(bQATest))
{
InitGL(&argc, argv);
}
//Get the NVIDIA platform if available, otherwise use default
char cBuffer[1024];
bool bNV = false;
shrLog("Get Platform ID... ");
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
ciErrNum = clGetPlatformInfo (cpPlatform, CL_PLATFORM_NAME, sizeof(cBuffer), cBuffer, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("%s\n\n", cBuffer);
bNV = (strstr(cBuffer, "NVIDIA") != NULL);
//Get the devices
shrLog("Get Device Info...\n");
cl_uint uiNumAllDevs = 0;
GpuDevMngr = new DeviceManager(cpPlatform, &uiNumAllDevs, pCleanup);
// Get selected device if specified, otherwise examine avaiable ones and choose by perf
cl_int iSelectedDevice = 0;
if((shrGetCmdLineArgumenti(argc, (const char**)argv, "device", &iSelectedDevice)) || (uiNumAllDevs == 1))
{
// Use 1 selected device
GpuDevMngr->uiUsefulDevCt = 1;
iSelectedDevice = CLAMP((cl_uint)iSelectedDevice, 0, (uiNumAllDevs - 1));
GpuDevMngr->uiUsefulDevs[0] = iSelectedDevice;
GpuDevMngr->fLoadProportions[0] = 1.0f;
shrLog(" Using 1 Selected Device for Sobel Filter Computation...\n");
}
else
{
// Use available useful devices and Compute the device load proportions
ciErrNum = GpuDevMngr->GetDevLoadProportions(bNV);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
if (GpuDevMngr->uiUsefulDevCt == 1)
{
iSelectedDevice = GpuDevMngr->uiUsefulDevs[0];
}
shrLog(" Using %u Device(s) for Sobel Filter Computation\n", GpuDevMngr->uiUsefulDevCt);
}
//Create the context
shrLog("\nclCreateContext...\n\n");
cxGPUContext = clCreateContext(0, uiNumAllDevs, GpuDevMngr->cdDevices, NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Allocate per-device OpenCL objects for useful devices
cqCommandQueue = new cl_command_queue[GpuDevMngr->uiUsefulDevCt];
ckSobel = new cl_kernel[GpuDevMngr->uiUsefulDevCt];
cmDevBufIn = new cl_mem[GpuDevMngr->uiUsefulDevCt];
cmDevBufOut = new cl_mem[GpuDevMngr->uiUsefulDevCt];
szAllocDevBytes = new size_t[GpuDevMngr->uiUsefulDevCt];
uiInHostPixOffsets = new cl_uint[GpuDevMngr->uiUsefulDevCt];
uiOutHostPixOffsets = new cl_uint[GpuDevMngr->uiUsefulDevCt];
uiDevImageHeight = new cl_uint[GpuDevMngr->uiUsefulDevCt];
// Create command queue(s) for device(s)
shrLog("clCreateCommandQueue...\n");
for (cl_uint i = 0; i < GpuDevMngr->uiUsefulDevCt; i++)
{
cqCommandQueue[i] = clCreateCommandQueue(cxGPUContext, GpuDevMngr->cdDevices[GpuDevMngr->uiUsefulDevs[i]], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog(" CommandQueue %u, Device %u, Device Load Proportion = %.2f, ", i, GpuDevMngr->uiUsefulDevs[i], GpuDevMngr->fLoadProportions[i]);
oclPrintDevName(LOGBOTH, GpuDevMngr->cdDevices[GpuDevMngr->uiUsefulDevs[i]]);
shrLog("\n");
}
// Allocate pinned input and output host image buffers: mem copy operations to/from pinned memory is much faster than paged memory
szBuffBytes = uiImageWidth * uiImageHeight * sizeof (unsigned int);
cmPinnedBufIn = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, szBuffBytes, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmPinnedBufOut = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, szBuffBytes, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("\nclCreateBuffer (Input and Output Pinned Host buffers)...\n");
// Get mapped pointers for writing to pinned input and output host image pointers
uiInput = (cl_uint*)clEnqueueMapBuffer(cqCommandQueue[0], cmPinnedBufIn, CL_TRUE, CL_MAP_WRITE, 0, szBuffBytes, 0, NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
uiOutput = (cl_uint*)clEnqueueMapBuffer(cqCommandQueue[0], cmPinnedBufOut, CL_TRUE, CL_MAP_READ, 0, szBuffBytes, 0, NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clEnqueueMapBuffer (Pointer to Input and Output pinned host buffers)...\n");
// Load image data from file to pinned input host buffer
ciErrNum = shrLoadPPM4ub(cPathAndName, (unsigned char **)&uiInput, &uiImageWidth, &uiImageHeight);
oclCheckErrorEX(ciErrNum, shrTRUE, pCleanup);
shrLog("Load Input Image to Input pinned host buffer...\n");
// Read the kernel in from file
free(cPathAndName);
cPathAndName = shrFindFilePath(clSourcefile, argv[0]);
oclCheckErrorEX(cPathAndName != NULL, shrTRUE, pCleanup);
cSourceCL = oclLoadProgSource(cPathAndName, "// My comment\n", &szKernelLength);
oclCheckErrorEX(cSourceCL != NULL, shrTRUE, pCleanup);
shrLog("Load OpenCL Prog Source from File...\n");
// Create the program object
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clCreateProgramWithSource...\n");
// Build the program with 'mad' Optimization option
#ifdef MAC
char *flags = "-cl-fast-relaxed-math -DMAC";
#else
char *flags = "-cl-fast-relaxed-math";
#endif
ciErrNum = clBuildProgram(cpProgram, 0, NULL, flags, NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// On error: write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclSobelFilter.ptx");
Cleanup(EXIT_FAILURE);
}
shrLog("clBuildProgram...\n\n");
// Determine, the size/shape of the image portions for each dev and create the device buffers
unsigned uiSumHeight = 0;
for (cl_uint i = 0; i < GpuDevMngr->uiUsefulDevCt; i++)
{
// Create kernel instance
ckSobel[i] = clCreateKernel(cpProgram, "ckSobel", &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clCreateKernel (ckSobel), Device %u...\n", i);
// Allocations and offsets for the portion of the image worked on by each device
if (GpuDevMngr->uiUsefulDevCt == 1)
{
// One device processes the whole image with no offset
uiDevImageHeight[i] = uiImageHeight;
uiInHostPixOffsets[i] = 0;
uiOutHostPixOffsets[i] = 0;
szAllocDevBytes[i] = uiDevImageHeight[i] * uiImageWidth * sizeof(cl_uint);
}
else if (i == 0)
{
// Multiple devices, top stripe zone including topmost row of image:
// Over-allocate on device by 1 row
// Set offset and size to copy extra 1 padding row H2D (below bottom of stripe)
// Won't return the last row (dark/garbage row) D2H
uiInHostPixOffsets[i] = 0;
uiOutHostPixOffsets[i] = 0;
uiDevImageHeight[i] = (cl_uint)(GpuDevMngr->fLoadProportions[GpuDevMngr->uiUsefulDevs[i]] * (float)uiImageHeight); // height is proportional to dev perf
uiSumHeight += uiDevImageHeight[i];
uiDevImageHeight[i] += 1;
szAllocDevBytes[i] = uiDevImageHeight[i] * uiImageWidth * sizeof(cl_uint);
}
else if (i < (GpuDevMngr->uiUsefulDevCt - 1))
{
// Multiple devices, middle stripe zone:
// Over-allocate on device by 2 rows
// Set offset and size to copy extra 2 padding rows H2D (above top and below bottom of stripe)
// Won't return the first and last rows (dark/garbage rows) D2H
uiInHostPixOffsets[i] = (uiSumHeight - 1) * uiImageWidth;
uiOutHostPixOffsets[i] = uiInHostPixOffsets[i] + uiImageWidth;
uiDevImageHeight[i] = (cl_uint)(GpuDevMngr->fLoadProportions[GpuDevMngr->uiUsefulDevs[i]] * (float)uiImageHeight); // height is proportional to dev perf
uiSumHeight += uiDevImageHeight[i];
uiDevImageHeight[i] += 2;
szAllocDevBytes[i] = uiDevImageHeight[i] * uiImageWidth * sizeof(cl_uint);
}
else
{
// Multiple devices, last boundary tile:
// Over-allocate on device by 1 row
// Set offset and size to copy extra 1 padding row H2D (above top of stripe)
// Won't return the first row (dark/garbage rows D2H
uiInHostPixOffsets[i] = (uiSumHeight - 1) * uiImageWidth;
uiOutHostPixOffsets[i] = uiInHostPixOffsets[i] + uiImageWidth;
uiDevImageHeight[i] = uiImageHeight - uiSumHeight; // "leftover" rows
uiSumHeight += uiDevImageHeight[i];
uiDevImageHeight[i] += 1;
szAllocDevBytes[i] = uiDevImageHeight[i] * uiImageWidth * sizeof(cl_uint);
}
shrLog("Image Height (rows) for Device %u = %u...\n", i, uiDevImageHeight[i]);
// Create the device buffers in GMEM on each device
cmDevBufIn[i] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, szAllocDevBytes[i], NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmDevBufOut[i] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, szAllocDevBytes[i], NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clCreateBuffer (Input and Output GMEM buffers, Device %u)...\n", i);
// Set the common argument values for the Median kernel instance for each device
int iLocalPixPitch = iBlockDimX + 2;
ciErrNum = clSetKernelArg(ckSobel[i], 0, sizeof(cl_mem), (void*)&cmDevBufIn[i]);
ciErrNum |= clSetKernelArg(ckSobel[i], 1, sizeof(cl_mem), (void*)&cmDevBufOut[i]);
ciErrNum |= clSetKernelArg(ckSobel[i], 2, (iLocalPixPitch * (iBlockDimY + 2) * sizeof(cl_uchar4)), NULL);
ciErrNum |= clSetKernelArg(ckSobel[i], 3, sizeof(cl_int), (void*)&iLocalPixPitch);
ciErrNum |= clSetKernelArg(ckSobel[i], 4, sizeof(cl_uint), (void*)&uiImageWidth);
ciErrNum |= clSetKernelArg(ckSobel[i], 6, sizeof(cl_float), (void*)&fThresh);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clSetKernelArg (0-4), Device %u...\n\n", i);
}
// Set common global and local work sizes for Median kernel
szLocalWorkSize[0] = iBlockDimX;
szLocalWorkSize[1] = iBlockDimY;
szGlobalWorkSize[0] = shrRoundUp((int)szLocalWorkSize[0], uiImageWidth);
// init running timers
shrDeltaT(0); // timer 0 used for computation timing
shrDeltaT(1); // timer 1 used for fps computation
// Start main GLUT rendering loop for processing and rendering,
// or otherwise run No-GL Q/A test sequence
if (!(bQATest))
{
glutMainLoop();
}
else
{
TestNoGL();
}
Cleanup(EXIT_SUCCESS);
}
// Main Program
//*****************************************************************************
int main(int argc, char** argv)
{
pArgc = &argc;
pArgv = argv;
shrQAStart(argc, argv);
// start logs
cExecutableName = argv[0];
shrSetLogFileName ("oclPostProcessGL.txt");
shrLog("%s Starting...\n\n", argv[0]);
// process command line arguments
if (argc > 1)
{
bQATest = shrCheckCmdLineFlag(argc, (const char**)argv, "qatest");
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
}
shrLog(" Image Width = %d, Image Height = %d, Blur Radius = %d\n\n", image_width, image_height, blur_radius);
// init GL
if(!bQATest)
{
InitGL(&argc, argv);
// create pbo
createPBO(&pbo_source);
createPBO(&pbo_dest);
// create texture for blitting onto the screen
createTexture(&tex_screen, image_width, image_height);
bGLinterop = shrTRUE;
}
// init CL
if( initCL(argc, (const char**)argv) != 0 )
{
return -1;
}
// init fps timer
shrDeltaT (1);
// Create buffers and textures,
// and then start main GLUT rendering loop for processing and rendering,
// or otherwise run No-GL Q/A test sequence
shrLog("\n%s...\n", bQATest ? "No-GL test sequence" : "Standard GL Loop");
printf("\n"
"\tControls\n"
"\t(right click mouse button for Menu)\n"
"\t[ ] : Toggle Post-Processing (blur filter) ON/OFF\n"
"\t[ p ] : Toggle Processing (between GPU or CPU)\n"
"\t[ a ] : Toggle OpenGL Animation (rotation) ON/OFF\n"
"\t[+/=] : Increase Blur Radius\n"
"\t[-/_] : Decrease Blur Radius\n"
"\t[Esc] - Quit\n\n"
);
if(!bQATest)
{
glutMainLoop();
}
else
{
TestNoGL();
}
Cleanup(EXIT_SUCCESS);
}
// Main function
// *********************************************************************
int main(const int argc, const char** argv)
{
// start logs
shrSetLogFileName ("oclDXTCompression.txt");
shrLog(LOGBOTH, 0, "%s Starting...\n\n", argv[0]);
cl_context cxGPUContext;
cl_command_queue cqCommandQueue;
cl_program cpProgram;
cl_kernel ckKernel;
cl_mem cmMemObjs[3];
size_t szGlobalWorkSize[1];
size_t szLocalWorkSize[1];
cl_int ciErrNum;
// Get the path of the filename
char *filename;
if (shrGetCmdLineArgumentstr(argc, argv, "image", &filename)) {
image_filename = filename;
}
// load image
const char* image_path = shrFindFilePath(image_filename, argv[0]);
shrCheckError(image_path != NULL, shrTRUE);
shrLoadPPM4ub(image_path, (unsigned char **)&h_img, &width, &height);
shrCheckError(h_img != NULL, shrTRUE);
shrLog(LOGBOTH, 0, "Loaded '%s', %d x %d pixels\n", image_path, width, height);
// Convert linear image to block linear.
uint * block_image = (uint *) malloc(width * height * 4);
// Convert linear image to block linear.
for(uint by = 0; by < height/4; by++) {
for(uint bx = 0; bx < width/4; bx++) {
for (int i = 0; i < 16; i++) {
const int x = i & 3;
const int y = i / 4;
block_image[(by * width/4 + bx) * 16 + i] =
((uint *)h_img)[(by * 4 + y) * 4 * (width/4) + bx * 4 + x];
}
}
}
// create the OpenCL context on a GPU device
cxGPUContext = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, NULL, NULL, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// get and log device
cl_device_id device;
if( shrCheckCmdLineFlag(argc, argv, "device") ) {
int device_nr = 0;
shrGetCmdLineArgumenti(argc, argv, "device", &device_nr);
device = oclGetDev(cxGPUContext, device_nr);
} else {
device = oclGetMaxFlopsDev(cxGPUContext);
}
oclPrintDevInfo(LOGBOTH, device);
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, device, 0, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// Memory Setup
// Compute permutations.
cl_uint permutations[1024];
computePermutations(permutations);
// Upload permutations.
cmMemObjs[0] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(cl_uint) * 1024, permutations, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// Image
cmMemObjs[1] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY ,
sizeof(cl_uint) * width * height, NULL, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// Result
const uint compressedSize = (width / 4) * (height / 4) * 8;
cmMemObjs[2] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY,
compressedSize, NULL , &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
unsigned int * h_result = (uint *)malloc(compressedSize);
// Program Setup
size_t program_length;
const char* source_path = shrFindFilePath("DXTCompression.cl", argv[0]);
shrCheckError(source_path != NULL, shrTRUE);
char *source = oclLoadProgSource(source_path, "", &program_length);
shrCheckError(source != NULL, shrTRUE);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,
(const char **) &source, &program_length, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-mad-enable", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLog(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclDXTCompression.ptx");
shrCheckError(ciErrNum, CL_SUCCESS);
}
// create the kernel
ckKernel = clCreateKernel(cpProgram, "compress", &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// set the args values
ciErrNum = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void *) &cmMemObjs[0]);
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void *) &cmMemObjs[1]);
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void *) &cmMemObjs[2]);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(float) * 4 * 16, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 4, sizeof(float) * 4 * 16, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 5, sizeof(int) * 64, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 6, sizeof(float) * 16 * 6, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 7, sizeof(unsigned int) * 160, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 8, sizeof(int) * 16, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog(LOGBOTH, 0, "Running DXT Compression on %u x %u image...\n\n", width, height);
// Upload the image
clEnqueueWriteBuffer(cqCommandQueue, cmMemObjs[1], CL_FALSE, 0, sizeof(cl_uint) * width * height, block_image, 0,0,0);
// set work-item dimensions
szGlobalWorkSize[0] = width * height * (NUM_THREADS/16);
szLocalWorkSize[0]= NUM_THREADS;
#ifdef GPU_PROFILING
int numIterations = 100;
for (int i = -1; i < numIterations; ++i) {
if (i == 0) { // start timing only after the first warmup iteration
clFinish(cqCommandQueue); // flush command queue
shrDeltaT(0); // start timer
}
#endif
// execute kernel
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL,
szGlobalWorkSize, szLocalWorkSize,
0, NULL, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
#ifdef GPU_PROFILING
}
clFinish(cqCommandQueue);
double dAvgTime = shrDeltaT(0) / (double)numIterations;
shrLog(LOGBOTH | MASTER, 0, "oclDXTCompression, Throughput = %.4f, Time = %.5f, Size = %u, NumDevsUsed = %i\n",
(1.0e-6 * (double)(width * height)/ dAvgTime), dAvgTime, (width * height), 1);
#endif
// blocking read output
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmMemObjs[2], CL_TRUE, 0,
compressedSize, h_result, 0, NULL, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
// Write DDS file.
FILE* fp = NULL;
char output_filename[1024];
#ifdef WIN32
strcpy_s(output_filename, 1024, image_path);
strcpy_s(output_filename + strlen(image_path) - 3, 1024 - strlen(image_path) + 3, "dds");
fopen_s(&fp, output_filename, "wb");
#else
strcpy(output_filename, image_path);
strcpy(output_filename + strlen(image_path) - 3, "dds");
fp = fopen(output_filename, "wb");
#endif
shrCheckError(fp != NULL, shrTRUE);
DDSHeader header;
header.fourcc = FOURCC_DDS;
header.size = 124;
header.flags = (DDSD_WIDTH|DDSD_HEIGHT|DDSD_CAPS|DDSD_PIXELFORMAT|DDSD_LINEARSIZE);
header.height = height;
header.width = width;
header.pitch = compressedSize;
header.depth = 0;
header.mipmapcount = 0;
memset(header.reserved, 0, sizeof(header.reserved));
header.pf.size = 32;
header.pf.flags = DDPF_FOURCC;
header.pf.fourcc = FOURCC_DXT1;
header.pf.bitcount = 0;
header.pf.rmask = 0;
header.pf.gmask = 0;
header.pf.bmask = 0;
header.pf.amask = 0;
header.caps.caps1 = DDSCAPS_TEXTURE;
header.caps.caps2 = 0;
header.caps.caps3 = 0;
header.caps.caps4 = 0;
header.notused = 0;
fwrite(&header, sizeof(DDSHeader), 1, fp);
fwrite(h_result, compressedSize, 1, fp);
fclose(fp);
// Make sure the generated image matches the reference image (regression check)
shrLog(LOGBOTH, 0, "\nComparing against Host/C++ computation...\n");
const char* reference_image_path = shrFindFilePath(refimage_filename, argv[0]);
shrCheckError(reference_image_path != NULL, shrTRUE);
// read in the reference image from file
#ifdef WIN32
fopen_s(&fp, reference_image_path, "rb");
#else
fp = fopen(reference_image_path, "rb");
#endif
shrCheckError(fp != NULL, shrTRUE);
fseek(fp, sizeof(DDSHeader), SEEK_SET);
uint referenceSize = (width / 4) * (height / 4) * 8;
uint * reference = (uint *)malloc(referenceSize);
fread(reference, referenceSize, 1, fp);
fclose(fp);
// compare the reference image data to the sample/generated image
float rms = 0;
for (uint y = 0; y < height; y += 4)
{
for (uint x = 0; x < width; x += 4)
{
// binary comparison of data
uint referenceBlockIdx = ((y/4) * (width/4) + (x/4));
uint resultBlockIdx = ((y/4) * (width/4) + (x/4));
int cmp = compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
// log deviations, if any
if (cmp != 0.0f)
{
compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
shrLog(LOGBOTH, 0, "Deviation at (%d, %d):\t%f rms\n", x/4, y/4, float(cmp)/16/3);
}
rms += cmp;
}
}
rms /= width * height * 3;
shrLog(LOGBOTH, 0, "RMS(reference, result) = %f\n\n", rms);
shrLog(LOGBOTH, 0, "TEST %s\n\n", (rms <= ERROR_THRESHOLD) ? "PASSED" : "FAILED !!!");
// Free OpenCL resources
oclDeleteMemObjs(cmMemObjs, 3);
clReleaseKernel(ckKernel);
clReleaseProgram(cpProgram);
clReleaseCommandQueue(cqCommandQueue);
clReleaseContext(cxGPUContext);
// Free host memory
free(source);
free(h_img);
// finish
shrEXIT(argc, argv);
}
///////////////////////////////////////////////////////////////////////////////
//Parse args, run the appropriate tests
///////////////////////////////////////////////////////////////////////////////
int runTest(const int argc, const char **argv)
{
int start = DEFAULT_SIZE;
int end = DEFAULT_SIZE;
int startDevice = 0;
int endDevice = 0;
int increment = DEFAULT_INCREMENT;
testMode mode = QUICK_MODE;
bool htod = false;
bool dtoh = false;
bool dtod = false;
char *modeStr;
char *device = NULL;
printMode printmode = USER_READABLE;
char *memModeStr = NULL;
memoryMode memMode = PAGEABLE;
accessMode accMode = DIRECT;
//process command line args
if(shrCheckCmdLineFlag( argc, argv, "help"))
{
printHelp();
return 0;
}
if(shrCheckCmdLineFlag( argc, argv, "csv"))
{
printmode = CSV;
}
// Get host memory mode type from command line
if(shrGetCmdLineArgumentstr(argc, argv, "memory", &memModeStr))
{
if(strcmp(memModeStr, "pageable") == 0 )
{
memMode = PAGEABLE;
}
else if(strcmp(memModeStr, "pinned") == 0)
{
memMode = PINNED;
}
else
{
shrLog("Invalid memory mode - valid modes are pageable or pinned\n");
shrLog("See --help for more information\n");
return -1000;
}
}
else
{
//default - pageable memory
memMode = PAGEABLE;
}
// Access type from command line
if(shrGetCmdLineArgumentstr(argc, argv, "access", &memModeStr))
{
if(strcmp(memModeStr, "direct") == 0)
{
accMode = DIRECT;
}
else if(strcmp(memModeStr, "mapped") == 0)
{
accMode = MAPPED;
}
else
{
shrLog("Invalid access mode - valid modes are direct or mapped\n");
shrLog("See --help for more information\n");
return -2000;
}
}
else
{
//default - direct
accMode = DIRECT;
}
// Get OpenCL platform ID for NVIDIA if available, otherwise default
cl_platform_id clSelectedPlatformID = NULL;
cl_int ciErrNum = oclGetPlatformID (&clSelectedPlatformID);
oclCheckError(ciErrNum, CL_SUCCESS);
// Find out how many devices there are
cl_uint ciDeviceCount;
ciErrNum = clGetDeviceIDs (clSelectedPlatformID, CL_DEVICE_TYPE_GPU, 0, NULL, &ciDeviceCount);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clGetDeviceIDs call !!!\n\n", ciErrNum);
return ciErrNum;
}
else if (ciDeviceCount == 0)
{
shrLog(" There are no devices supporting OpenCL (return code %i)\n\n", ciErrNum);
return ciErrNum;
}
// Get command line device options and config accordingly
if(shrGetCmdLineArgumentstr(argc, argv, "device", &device))
{
if(strcmp (device, "all") == 0)
{
shrLog("\n!!!Cumulative Bandwidth to be computed from all the devices !!!\n\n");
startDevice = 0;
endDevice = (int)(ciDeviceCount-1);
}
else
{
startDevice = endDevice = atoi(device);
if(startDevice < 0 || ((size_t)startDevice) >= ciDeviceCount)
{
shrLog("\n!!!Invalid GPU number %d given hence default gpu %d will be used !!!\n", startDevice,0);
startDevice = endDevice = 0;
}
}
}
// Get and log the device info
shrLog("Running on...\n\n");
devices = (cl_device_id*) malloc(sizeof(cl_device_id) * ciDeviceCount);
ciErrNum = clGetDeviceIDs (clSelectedPlatformID, CL_DEVICE_TYPE_GPU, ciDeviceCount, devices, &ciDeviceCount);
for(int currentDevice = startDevice; currentDevice <= endDevice; currentDevice++)
{
oclPrintDevName(LOGBOTH, devices[currentDevice]);
shrLog("\n");
}
shrLog("\n");
// Get command line mode(s) and config accordingly
if(shrGetCmdLineArgumentstr(argc, argv, "mode", &modeStr))
{
//figure out the mode
if(strcmp(modeStr, "quick") == 0)
{
shrLog("Quick Mode\n\n");
mode = QUICK_MODE;
}
else if(strcmp(modeStr, "shmoo") == 0)
{
shrLog("Shmoo Mode\n\n");
mode = SHMOO_MODE;
}
else if(strcmp(modeStr, "range") == 0)
{
shrLog("Range Mode\n\n");
mode = RANGE_MODE;
}
else
{
shrLog("Invalid mode - valid modes are quick, range, or shmoo\n");
shrLog("See --help for more information\n\n");
return -3000;
}
}
else
{
//default mode - quick
shrLog("Quick Mode\n\n");
mode = QUICK_MODE;
}
if(shrCheckCmdLineFlag(argc, argv, "htod"))
htod = true;
if(shrCheckCmdLineFlag(argc, argv, "dtoh"))
dtoh = true;
if(shrCheckCmdLineFlag(argc, argv, "dtod"))
dtod = true;
if(!htod && !dtoh && !dtod)
{
//default: All
htod = true;
dtoh = true;
dtod = true;
}
if(RANGE_MODE == mode)
{
if(shrGetCmdLineArgumenti( argc, argv, "start", &start))
{
if( start <= 0 )
{
shrLog("Illegal argument - start must be greater than zero\n");
return -4000;
}
}
else
{
shrLog("Must specify a starting size in range mode\n");
shrLog("See --help for more information\n");
return -5000;
}
if(shrGetCmdLineArgumenti( argc, argv, "end", &end))
{
if(end <= 0)
{
shrLog("Illegal argument - end must be greater than zero\n");
return -6000;
}
if(start > end)
{
shrLog("Illegal argument - start is greater than end\n");
return -7000;
}
}
else
{
shrLog("Must specify an end size in range mode.\n");
shrLog("See --help for more information\n");
return -8000;
}
if(shrGetCmdLineArgumenti( argc, argv, "increment", &increment))
{
if(increment <= 0)
{
shrLog("Illegal argument - increment must be greater than zero\n");
return -9000;
}
}
else
{
shrLog("Must specify an increment in user mode\n");
shrLog("See --help for more information\n");
return -10000;
}
}
// Create the OpenCL context
cxGPUContext = clCreateContext(0, ciDeviceCount, devices, NULL, NULL, NULL);
if (cxGPUContext == (cl_context)0)
{
shrLog("Failed to create OpenCL context!\n");
return -11000;
}
// Run tests
if(htod)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, HOST_TO_DEVICE, printmode, accMode, memMode, startDevice, endDevice);
}
if(dtoh)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, DEVICE_TO_HOST, printmode, accMode, memMode, startDevice, endDevice);
}
if(dtod)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, DEVICE_TO_DEVICE, printmode, accMode, memMode, startDevice, endDevice);
}
// Clean up
free(memModeStr);
if(cqCommandQueue)clReleaseCommandQueue(cqCommandQueue);
if(cxGPUContext)clReleaseContext(cxGPUContext);
if(devices)free(devices);
return 0;
}
//-----------------------------------------------------------------------------
// Program main
//-----------------------------------------------------------------------------
int main(int argc, char** argv)
{
pArgc = &argc;
pArgv = argv;
// start logs
shrQAStart(argc, argv);
shrSetLogFileName ("oclSimpleD3D9Texture.txt");
shrLog("%s Starting...\n\n", argv[0]);
// process command line arguments
if (argc > 1)
{
bQATest = shrCheckCmdLineFlag(argc, (const char **)argv, "qatest");
bNoPrompt = shrCheckCmdLineFlag(argc, (const char **)argv, "noprompt");
}
//
// create window
//
// Register the window class
WNDCLASSEX wc = { sizeof(WNDCLASSEX), CS_CLASSDC, MsgProc, 0L, 0L,
GetModuleHandle(NULL), NULL, NULL, NULL, NULL,
"OpenCL/D3D9 Texture InterOP", NULL };
RegisterClassEx( &wc );
int xBorder = ::GetSystemMetrics(SM_CXSIZEFRAME);
int yMenu = ::GetSystemMetrics(SM_CYMENU);
int yBorder = ::GetSystemMetrics(SM_CYSIZEFRAME);
// Create the application's window (padding by window border for uniform BB sizes across OSs)
HWND hWnd = CreateWindow( wc.lpszClassName, "OpenCL/D3D9 Texture InterOP",
WS_OVERLAPPEDWINDOW, 0, 0, g_WindowWidth + 2*xBorder, g_WindowHeight+ 2*yBorder+yMenu,
NULL, NULL, wc.hInstance, NULL );
ShowWindow(hWnd, SW_SHOWDEFAULT);
UpdateWindow(hWnd);
// init fps timer
shrDeltaT (1);
// Initialize Direct3D
if( SUCCEEDED( InitD3D9(hWnd) ) &&
SUCCEEDED( InitCL(argc, (const char **)argv) ) &&
SUCCEEDED( InitTextures() ) )
{
if (!g_bDeviceLost)
{
RegisterD3D9ResourceWithCL();
}
}
//
// the main loop
//
while(false == g_bDone)
{
RunCL();
DrawScene();
//
// handle I/O
//
MSG msg;
ZeroMemory( &msg, sizeof(msg) );
while( msg.message!=WM_QUIT )
{
if( PeekMessage( &msg, NULL, 0U, 0U, PM_REMOVE ) )
{
TranslateMessage( &msg );
DispatchMessage( &msg );
}
else
{
RunCL();
DrawScene();
if(bQATest)
{
for(int count=0;count<g_iFrameToCompare;count++)
{
RunCL();
DrawScene();
}
const char *ref_image_path = "ref_oclSimpleD3D9Texture.ppm";
const char *cur_image_path = "oclSimpleD3D9Texture.ppm";
// Save a reference of our current test run image
CheckRenderD3D9::BackbufferToPPM(g_pD3DDevice,cur_image_path);
// compare to offical reference image, printing PASS or FAIL.
g_bPassed = CheckRenderD3D9::PPMvsPPM(cur_image_path,ref_image_path,argv[0],MAX_EPSILON, 0.15f);
PostQuitMessage(0);
g_bDone = true;
}
}
}
};
// Unregister windows class
UnregisterClass( wc.lpszClassName, wc.hInstance );
// Cleanup and leave
Cleanup (g_bPassed ? EXIT_SUCCESS : EXIT_FAILURE);
}
// Main function
// *********************************************************************
int main(int argc, char **argv)
{
//shrQAStart(argc, argv);
// get command line arg for quick test, if provided
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
// start logs
cExecutableName = argv[0];
shrSetLogFileName ("Barrier.txt");
printf("%s Starting...\n\n# of THREADS \t= %i\n", argv[0], iNumElements);
// set and log Global and Local work size dimensions
szLocalWorkSize = NUM_THREADS ;
szGlobalWorkSize = shrRoundUp((int)szLocalWorkSize, iNumElements); // rounded up to the nearest multiple of the LocalWorkSize
printf("Global Work Size \t\t= %u\nLocal Work Size \t\t= %u\n# of Work Groups \t\t= %u\n\n",
szGlobalWorkSize, szLocalWorkSize, (szGlobalWorkSize % szLocalWorkSize + szGlobalWorkSize/szLocalWorkSize));
//Get an OpenCL platform
ciErr1 = clGetPlatformIDs(1, &cpPlatform, NULL);
printf("clGetPlatformID...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clGetPlatformID, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
//Get the devices
ciErr1 = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &cdDevice, NULL);
printf("clGetDeviceIDs...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clGetDeviceIDs, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
//Create the context
cxGPUContext = clCreateContext(0, 1, &cdDevice, NULL, NULL, &ciErr1);
printf("clCreateContext...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clCreateContext, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevice, CL_QUEUE_PROFILING_ENABLE, &ciErr1);
printf("clCreateCommandQueue...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clCreateCommandQueue, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Read the OpenCL kernel in from source file
printf("oclLoadProgSource (%s)...\n", cSourceFile);
cPathAndName = shrFindFilePath(cSourceFile, argv[0]);
cSourceCL = oclLoadProgSource(cPathAndName, "", &szKernelLength);
// Create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErr1);
printf("clCreateProgramWithSource...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clCreateProgramWithSource, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Build the program with 'mad' Optimization option
#ifdef MAC
char* flags = "-cl-fast-relaxed-math -DMAC";
#else
char* flags = "-cl-fast-relaxed-math";
#endif
ciErr1 = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL);
printf("clBuildProgram...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clBuildProgram, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Create the kernel
ckKernel = clCreateKernel(cpProgram, "Barrier", &ciErr1);
printf("clCreateKernel (Barrier)...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clCreateKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Allocate and initialize host arrays
printf( "Allocate and Init Host Mem...\n");
input = (int *)malloc(sizeof(int) * NUM_BLOCKS);
for(int i =0; i<=NUM_BLOCKS; i++)
{
input[i]=0;
}
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
array_in = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(int)* NUM_BLOCKS, NULL, &ciErr1);
array_out = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(int)* NUM_BLOCKS, NULL, &ciErr1);
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clCreateBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Set the Argument values
ciErr1 = clSetKernelArg(ckKernel, 0, sizeof(cl_int), (void*)&goal_val);
ciErr1 |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void*)&array_in);
ciErr1 |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void*)&array_out);
// ciErr1 |= clSetKernelArg(ckKernel, 1, sizeof(cl_int), (void*)&iNumElements);
printf("clSetKernelArg 0 - 2...\n\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clSetKernelArg, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// --------------------------------------------------------
// Start Core sequence... copy input data to GPU, compute, copy results back
ciErr1 = clEnqueueWriteBuffer(cqCommandQueue, array_in, CL_FALSE, 0, sizeof(int) * NUM_BLOCKS,(void*) input, 0, NULL, NULL);
printf("clEnqueueWriteBuffer (SrcA and SrcB)...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clEnqueueWriteBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
// Launch kernel
ciErr1 = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, &ceEvent);
printf("clEnqueueNDRangeKernel (Barrier)...\n");
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clEnqueueNDRangeKernel, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}
/*ciErr1 = clEnqueueReadBuffer(cqCommandQueue, global_mutex, CL_TRUE, 0, sizeof(cl_int), &original_goal, 0, NULL, NULL);
printf("clEnqueueReadBuffer (Dst)...%d \n\n", original_goal);
if (ciErr1 != CL_SUCCESS)
{
printf("Error in clEnqueueReadBuffer, Line %u in file %s !!!\n\n", __LINE__, __FILE__);
Cleanup(argc, argv, EXIT_FAILURE);
}*/
//GPU_PROFILING
ciErr1=clWaitForEvents(1, &ceEvent);
if (ciErr1 != CL_SUCCESS)
{
printf("Error 1 !\n\n");
Cleanup(argc, argv, EXIT_FAILURE);
}
cl_ulong start, end;
ciErr1 = clGetEventProfilingInfo(ceEvent, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL);
ciErr1 |= clGetEventProfilingInfo(ceEvent, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, NULL);
//oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
if (ciErr1 != CL_SUCCESS)
{
printf("Error 2 !\n\n");
Cleanup(argc, argv, EXIT_FAILURE);
}
double dSeconds = 1.0e-9 * (double)(end - start);
printf("Done! time taken %llu \n",end - start );
// printf("Done! Kernel execution time: %.5f s\n\n", dSeconds);
// Release event
clReleaseEvent(ceEvent);
ceEvent = 0;
Cleanup (argc, argv, EXIT_SUCCESS);
}
// Main function
// *********************************************************************
int main(int argc, char** argv)
{
shrQAStart(argc, argv);
// get command line arg for quick test, if provided
bNoPrompt = shrCheckCmdLineFlag(argc, (const char **)argv, "noprompt");
// start logs
cExecutableName = argv[0];
shrSetLogFileName ("oclMatVecMul.txt");
shrLog("%s Starting...\n\n", argv[0]);
// calculate matrix height given GPU memory
shrLog("Determining Matrix height from available GPU mem...\n");
memsize_t memsize;
getTargetDeviceGlobalMemSize(&memsize, argc, (const char **)argv);
height = memsize/width/16;
if (height > MAX_HEIGHT)
height = MAX_HEIGHT;
shrLog(" Matrix width\t= %u\n Matrix height\t= %u\n\n", width, height);
// Allocate and initialize host arrays
shrLog("Allocate and Init Host Mem...\n\n");
unsigned int size = width * height;
unsigned int mem_size_M = size * sizeof(float);
M = (float*)malloc(mem_size_M);
unsigned int mem_size_V = width * sizeof(float);
V = (float*)malloc(mem_size_V);
unsigned int mem_size_W = height * sizeof(float);
W = (float*)malloc(mem_size_W);
shrFillArray(M, size);
shrFillArray(V, width);
Golden = (float*)malloc(mem_size_W);
MatVecMulHost(M, V, width, height, Golden);
//Get the NVIDIA platform
shrLog("Get the Platform ID...\n\n");
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
//Get all the devices
shrLog("Get the Device info and select Device...\n");
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Set target device and Query number of compute units on targetDevice
shrLog(" # of Devices Available = %u\n", uiNumDevices);
if(shrGetCmdLineArgumentu(argc, (const char **)argv, "device", &targetDevice)== shrTRUE)
{
targetDevice = CLAMP(targetDevice, 0, (uiNumDevices - 1));
}
shrLog(" Using Device %u: ", targetDevice);
oclPrintDevName(LOGBOTH, cdDevices[targetDevice]);
cl_uint num_compute_units;
clGetDeviceInfo(cdDevices[targetDevice], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(num_compute_units), &num_compute_units, NULL);
shrLog("\n # of Compute Units = %u\n\n", num_compute_units);
//Create the context
shrLog("clCreateContext...\n");
cxGPUContext = clCreateContext(0, uiNumDevsUsed, &cdDevices[targetDevice], NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create a command-queue
shrLog("clCreateCommandQueue...\n");
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevices[targetDevice], CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
shrLog("clCreateBuffer (M, V and W in device global memory, mem_size_m = %u)...\n", mem_size_M);
cmM = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, mem_size_M, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmV = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, mem_size_V, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmW = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, mem_size_W, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Read the OpenCL kernel in from source file
shrLog("oclLoadProgSource (%s)...\n", cSourceFile);
cPathAndName = shrFindFilePath(cSourceFile, argv[0]);
oclCheckErrorEX(cPathAndName != NULL, shrTRUE, pCleanup);
cSourceCL = oclLoadProgSource(cPathAndName, "", &szKernelLength);
oclCheckErrorEX(cSourceCL != NULL, shrTRUE, pCleanup);
// Create the program
shrLog("clCreateProgramWithSource...\n");
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErrNum);
// Build the program
shrLog("clBuildProgram...\n");
ciErrNum = clBuildProgram(cpProgram, uiNumDevsUsed, &cdDevices[targetDevice], "-cl-fast-relaxed-math", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclMatVecMul.ptx");
shrQAFinish(argc, (const char **)argv, QA_FAILED);
Cleanup(EXIT_FAILURE);
}
// --------------------------------------------------------
// Core sequence... copy input data to GPU, compute, copy results back
// Asynchronous write of data to GPU device
shrLog("clEnqueueWriteBuffer (M and V)...\n\n");
ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, cmM, CL_FALSE, 0, mem_size_M, M, 0, NULL, NULL);
ciErrNum |= clEnqueueWriteBuffer(cqCommandQueue, cmV, CL_FALSE, 0, mem_size_V, V, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Kernels
const char* kernels[] = {
"MatVecMulUncoalesced0",
"MatVecMulUncoalesced1",
"MatVecMulCoalesced0",
"MatVecMulCoalesced1",
"MatVecMulCoalesced2",
"MatVecMulCoalesced3" };
for (int k = 0; k < (int)(sizeof(kernels)/sizeof(char*)); ++k) {
shrLog("Running with Kernel %s...\n\n", kernels[k]);
// Clear result
shrLog(" Clear result with clEnqueueWriteBuffer (W)...\n");
memset(W, 0, mem_size_W);
ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, cmW, CL_FALSE, 0, mem_size_W, W, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create the kernel
shrLog(" clCreateKernel...\n");
if (ckKernel) {
clReleaseKernel(ckKernel);
ckKernel = 0;
}
ckKernel = clCreateKernel(cpProgram, kernels[k], &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Set and log Global and Local work size dimensions
szLocalWorkSize = 256;
if (k == 0)
szGlobalWorkSize = shrRoundUp((int)szLocalWorkSize, height); // rounded up to the nearest multiple of the LocalWorkSize
else
// Some experiments should be done here for determining the best global work size for a given device
// We will assume here that we can run 2 work-groups per compute unit
szGlobalWorkSize = 2 * num_compute_units * szLocalWorkSize;
shrLog(" Global Work Size \t\t= %u\n Local Work Size \t\t= %u\n # of Work Groups \t\t= %u\n",
szGlobalWorkSize, szLocalWorkSize, (szGlobalWorkSize % szLocalWorkSize + szGlobalWorkSize/szLocalWorkSize));
// Set the Argument values
shrLog(" clSetKernelArg...\n\n");
int n = 0;
ciErrNum = clSetKernelArg(ckKernel, n++, sizeof(cl_mem), (void*)&cmM);
ciErrNum |= clSetKernelArg(ckKernel, n++, sizeof(cl_mem), (void*)&cmV);
ciErrNum |= clSetKernelArg(ckKernel, n++, sizeof(cl_int), (void*)&width);
ciErrNum |= clSetKernelArg(ckKernel, n++, sizeof(cl_int), (void*)&height);
ciErrNum |= clSetKernelArg(ckKernel, n++, sizeof(cl_mem), (void*)&cmW);
if (k > 1)
ciErrNum |= clSetKernelArg(ckKernel, n++, szLocalWorkSize * sizeof(float), 0);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Launch kernel
shrLog(" clEnqueueNDRangeKernel (%s)...\n", kernels[k]);
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, &ceEvent);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Read back results and check accumulated errors
shrLog(" clEnqueueReadBuffer (W)...\n");
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmW, CL_TRUE, 0, mem_size_W, W, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
#ifdef GPU_PROFILING
// Execution time
ciErrNum = clWaitForEvents(1, &ceEvent);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cl_ulong start, end;
ciErrNum = clGetEventProfilingInfo(ceEvent, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL);
ciErrNum |= clGetEventProfilingInfo(ceEvent, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
double dSeconds = 1.0e-9 * (double)(end - start);
shrLog(" Kernel execution time: %.5f s\n\n", dSeconds);
#endif
// Compare results for golden-host and report errors and pass/fail
shrLog(" Comparing against Host/C++ computation...\n\n");
shrBOOL res = shrCompareL2fe(Golden, W, height, 1e-6f);
shrLog(" GPU Result %s CPU Result within allowable tolerance\n\n", (res == shrTRUE) ? "MATCHES" : "DOESN'T MATCH");
bPassFlag &= (res == shrTRUE);
// Release event
ciErrNum = clReleaseEvent(ceEvent);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
ceEvent = 0;
}
// Master status Pass/Fail (all tests)
shrQAFinish(argc, (const char **)argv, (bPassFlag ? QA_PASSED : QA_FAILED) );
// Cleanup and leave
Cleanup (EXIT_SUCCESS);
}
// Main program
//*****************************************************************************
int main(int argc, char** argv)
{
numParticles = NUM_PARTICLES;
uint gridDim = GRID_SIZE;
pArgc = &argc;
pArgv = argv;
shrQAStart(argc, argv);
// Start logs and timers
cExecutableName = argv[0];
shrSetLogFileName ("oclParticles.txt");
shrLog("%s Starting...\n\n", argv[0]);
// check command line flags and parameters
if (argc > 1)
{
shrGetCmdLineArgumenti(argc, (const char**)argv, "n", (int*)&numParticles);
shrGetCmdLineArgumenti(argc, (const char**)argv, "grid", (int*)&gridDim);
bQATest = shrCheckCmdLineFlag(argc, (const char**)argv, "qatest");
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
}
// Set and log grid size and particle count, after checking optional command-line inputs
gridSize.x = gridSize.y = gridSize.z = gridDim;
shrLog(" grid: %d x %d x %d = %d cells\n", gridSize.x, gridSize.y, gridSize.z, gridSize.x * gridSize.y * gridSize.z);
shrLog(" particles: %d\n\n", numParticles);
// initialize GLUT and GLEW
if(!bQATest)
{
InitGL(&argc, argv);
}
// initialize OpenCL
startupOpenCL(argc,(const char**)argv);
// init simulation parameters and objects
initParticleSystem(numParticles, gridSize);
initParams();
// Init timers
shrDeltaT(0); // timer 0 is for processing time measurements
shrDeltaT(1); // timer 1 is for fps measurement
// Start main GLUT rendering loop for processing and rendering,
// or otherwise run No-GL Q/A test sequence
if(!bQATest)
{
glutMainLoop();
}
else
{
TestNoGL();
}
// Normally unused return path
shrQAFinish(argc, (const char **)argv, QA_PASSED);
Cleanup(EXIT_SUCCESS);
}
int main(int argc, char **argv)
{
uchar *h_Data;
uint *h_HistogramCPU, *h_HistogramGPU;
uchar *d_Data;
uint *d_Histogram;
uint hTimer;
int PassFailFlag = 1;
uint byteCount = 64 * 1048576;
uint uiSizeMult = 1;
cudaDeviceProp deviceProp;
deviceProp.major = 0;
deviceProp.minor = 0;
int dev;
shrQAStart(argc, argv);
// set logfile name and start logs
shrSetLogFileName ("histogram.txt");
//Use command-line specified CUDA device, otherwise use device with highest Gflops/s
if( shrCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
dev = cutilDeviceInit(argc, argv);
if (dev < 0) {
printf("No CUDA Capable Devices found, exiting...\n");
shrQAFinishExit(argc, (const char **)argv, QA_WAIVED);
}
} else {
cudaSetDevice( dev = cutGetMaxGflopsDeviceId() );
cutilSafeCall( cudaChooseDevice(&dev, &deviceProp) );
}
cutilSafeCall( cudaGetDeviceProperties(&deviceProp, dev) );
printf("CUDA device [%s] has %d Multi-Processors, Compute %d.%d\n",
deviceProp.name, deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
int version = deviceProp.major * 0x10 + deviceProp.minor;
if(version < 0x11)
{
printf("There is no device supporting a minimum of CUDA compute capability 1.1 for this SDK sample\n");
cutilDeviceReset();
shrQAFinishExit(argc, (const char **)argv, QA_WAIVED);
}
cutilCheckError(cutCreateTimer(&hTimer));
// Optional Command-line multiplier to increase size of array to histogram
if (shrGetCmdLineArgumentu(argc, (const char**)argv, "sizemult", &uiSizeMult))
{
uiSizeMult = CLAMP(uiSizeMult, 1, 10);
byteCount *= uiSizeMult;
}
shrLog("Initializing data...\n");
shrLog("...allocating CPU memory.\n");
h_Data = (uchar *)malloc(byteCount);
h_HistogramCPU = (uint *)malloc(HISTOGRAM256_BIN_COUNT * sizeof(uint));
h_HistogramGPU = (uint *)malloc(HISTOGRAM256_BIN_COUNT * sizeof(uint));
shrLog("...generating input data\n");
srand(2009);
for(uint i = 0; i < byteCount; i++)
h_Data[i] = rand() % 256;
shrLog("...allocating GPU memory and copying input data\n\n");
cutilSafeCall( cudaMalloc((void **)&d_Data, byteCount ) );
cutilSafeCall( cudaMalloc((void **)&d_Histogram, HISTOGRAM256_BIN_COUNT * sizeof(uint) ) );
cutilSafeCall( cudaMemcpy(d_Data, h_Data, byteCount, cudaMemcpyHostToDevice) );
{
shrLog("Starting up 64-bin histogram...\n\n");
initHistogram64();
shrLog("Running 64-bin GPU histogram for %u bytes (%u runs)...\n\n", byteCount, numRuns);
for(int iter = -1; iter < numRuns; iter++){
//iter == -1 -- warmup iteration
if(iter == 0){
cutilSafeCall( cutilDeviceSynchronize() );
cutilCheckError( cutResetTimer(hTimer) );
cutilCheckError( cutStartTimer(hTimer) );
}
histogram64(d_Histogram, d_Data, byteCount);
}
cutilSafeCall( cutilDeviceSynchronize() );
cutilCheckError( cutStopTimer(hTimer));
double dAvgSecs = 1.0e-3 * (double)cutGetTimerValue(hTimer) / (double)numRuns;
shrLog("histogram64() time (average) : %.5f sec, %.4f MB/sec\n\n", dAvgSecs, ((double)byteCount * 1.0e-6) / dAvgSecs);
shrLogEx(LOGBOTH | MASTER, 0, "histogram64, Throughput = %.4f MB/s, Time = %.5f s, Size = %u Bytes, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * (double)byteCount / dAvgSecs), dAvgSecs, byteCount, 1, HISTOGRAM64_THREADBLOCK_SIZE);
shrLog("\nValidating GPU results...\n");
shrLog(" ...reading back GPU results\n");
cutilSafeCall( cudaMemcpy(h_HistogramGPU, d_Histogram, HISTOGRAM64_BIN_COUNT * sizeof(uint), cudaMemcpyDeviceToHost) );
shrLog(" ...histogram64CPU()\n");
histogram64CPU(
h_HistogramCPU,
h_Data,
byteCount
);
shrLog(" ...comparing the results...\n");
for(uint i = 0; i < HISTOGRAM64_BIN_COUNT; i++)
if(h_HistogramGPU[i] != h_HistogramCPU[i]) PassFailFlag = 0;
shrLog(PassFailFlag ? " ...64-bin histograms match\n\n" : " ***64-bin histograms do not match!!!***\n\n" );
shrLog("Shutting down 64-bin histogram...\n\n\n");
closeHistogram64();
}
{
shrLog("Initializing 256-bin histogram...\n");
initHistogram256();
shrLog("Running 256-bin GPU histogram for %u bytes (%u runs)...\n\n", byteCount, numRuns);
for(int iter = -1; iter < numRuns; iter++){
//iter == -1 -- warmup iteration
if(iter == 0){
cutilSafeCall( cutilDeviceSynchronize() );
cutilCheckError( cutResetTimer(hTimer) );
cutilCheckError( cutStartTimer(hTimer) );
}
histogram256(d_Histogram, d_Data, byteCount);
}
cutilSafeCall( cutilDeviceSynchronize() );
cutilCheckError( cutStopTimer(hTimer));
double dAvgSecs = 1.0e-3 * (double)cutGetTimerValue(hTimer) / (double)numRuns;
shrLog("histogram256() time (average) : %.5f sec, %.4f MB/sec\n\n", dAvgSecs, ((double)byteCount * 1.0e-6) / dAvgSecs);
shrLogEx(LOGBOTH | MASTER, 0, "histogram256, Throughput = %.4f MB/s, Time = %.5f s, Size = %u Bytes, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * (double)byteCount / dAvgSecs), dAvgSecs, byteCount, 1, HISTOGRAM256_THREADBLOCK_SIZE);
shrLog("\nValidating GPU results...\n");
shrLog(" ...reading back GPU results\n");
cutilSafeCall( cudaMemcpy(h_HistogramGPU, d_Histogram, HISTOGRAM256_BIN_COUNT * sizeof(uint), cudaMemcpyDeviceToHost) );
shrLog(" ...histogram256CPU()\n");
histogram256CPU(
h_HistogramCPU,
h_Data,
byteCount
);
shrLog(" ...comparing the results\n");
for(uint i = 0; i < HISTOGRAM256_BIN_COUNT; i++)
if(h_HistogramGPU[i] != h_HistogramCPU[i]) PassFailFlag = 0;
shrLog(PassFailFlag ? " ...256-bin histograms match\n\n" : " ***256-bin histograms do not match!!!***\n\n" );
shrLog("Shutting down 256-bin histogram...\n\n\n");
closeHistogram256();
}
shrLog("Shutting down...\n");
cutilCheckError(cutDeleteTimer(hTimer));
cutilSafeCall( cudaFree(d_Histogram) );
cutilSafeCall( cudaFree(d_Data) );
free(h_HistogramGPU);
free(h_HistogramCPU);
free(h_Data);
cutilDeviceReset();
shrLog("%s - Test Summary\n", sSDKsample);
// pass or fail (for both 64 bit and 256 bit histograms)
shrQAFinishExit(argc, (const char **)argv, (PassFailFlag ? QA_PASSED : QA_FAILED));
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test for
////////////////////////////////////////////////////////////////////////////////
int runTest(int argc, const char** argv)
{
cl_platform_id cpPlatform = NULL;
cl_uint ciDeviceCount = 0;
cl_device_id *cdDevices = NULL;
cl_int ciErrNum = CL_SUCCESS;
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create OpenCL context!\n");
return ciErrNum;
}
//Get the devices
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &ciDeviceCount);
cdDevices = (cl_device_id *)malloc(ciDeviceCount * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, ciDeviceCount, cdDevices, NULL);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create OpenCL context!\n");
return ciErrNum;
}
//Create the context
cxGPUContext = clCreateContext(0, ciDeviceCount, cdDevices, NULL, NULL, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create OpenCL context!\n");
return ciErrNum;
}
if(shrCheckCmdLineFlag(argc, (const char**)argv, "device"))
{
// User specified GPUs
char* deviceList;
char* deviceStr;
char* next_token;
shrGetCmdLineArgumentstr(argc, (const char**)argv, "device", &deviceList);
#ifdef WIN32
deviceStr = strtok_s (deviceList," ,.-", &next_token);
#else
deviceStr = strtok (deviceList," ,.-");
#endif
ciDeviceCount = 0;
while(deviceStr != NULL)
{
// get and print the device for this queue
cl_device_id device = oclGetDev(cxGPUContext, atoi(deviceStr));
if( device == (cl_device_id) -1 ) {
shrLog(" Device %s does not exist!\n", deviceStr);
return -1;
}
shrLog("Device %s: ", deviceStr);
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create command queue
commandQueue[ciDeviceCount] = clCreateCommandQueue(cxGPUContext, device, CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clCreateCommandQueue call !!!\n\n", ciErrNum);
return ciErrNum;
}
++ciDeviceCount;
#ifdef WIN32
deviceStr = strtok_s (NULL," ,.-", &next_token);
#else
deviceStr = strtok (NULL," ,.-");
#endif
}
free(deviceList);
}
else
{
// Find out how many GPU's to compute on all available GPUs
size_t nDeviceBytes;
ciErrNum |= clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &nDeviceBytes);
ciDeviceCount = (cl_uint)nDeviceBytes/sizeof(cl_device_id);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clGetDeviceIDs call !!!\n\n", ciErrNum);
return ciErrNum;
}
else if (ciDeviceCount == 0)
{
shrLog(" There are no devices supporting OpenCL (return code %i)\n\n", ciErrNum);
return -1;
}
// create command-queues
for(unsigned int i = 0; i < ciDeviceCount; ++i)
{
// get and print the device for this queue
cl_device_id device = oclGetDev(cxGPUContext, i);
shrLog("Device %d: ", i);
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create command queue
commandQueue[i] = clCreateCommandQueue(cxGPUContext, device, CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clCreateCommandQueue call !!!\n\n", ciErrNum);
return ciErrNum;
}
}
}
// Optional Command-line multiplier for matrix sizes
shrGetCmdLineArgumenti(argc, (const char**)argv, "sizemult", &iSizeMultiple);
iSizeMultiple = CLAMP(iSizeMultiple, 1, 10);
uiWA = WA * iSizeMultiple;
uiHA = HA * iSizeMultiple;
uiWB = WB * iSizeMultiple;
uiHB = HB * iSizeMultiple;
uiWC = WC * iSizeMultiple;
uiHC = HC * iSizeMultiple;
shrLog("\nUsing Matrix Sizes: A(%u x %u), B(%u x %u), C(%u x %u)\n",
uiWA, uiHA, uiWB, uiHB, uiWC, uiHC);
// allocate host memory for matrices A and B
unsigned int size_A = uiWA * uiHA;
unsigned int mem_size_A = sizeof(float) * size_A;
float* h_A_data = (float*)malloc(mem_size_A);
unsigned int size_B = uiWB * uiHB;
unsigned int mem_size_B = sizeof(float) * size_B;
float* h_B_data = (float*)malloc(mem_size_B);
// initialize host memory
srand(2006);
shrFillArray(h_A_data, size_A);
shrFillArray(h_B_data, size_B);
// allocate host memory for result
unsigned int size_C = uiWC * uiHC;
unsigned int mem_size_C = sizeof(float) * size_C;
float* h_C = (float*) malloc(mem_size_C);
// create OpenCL buffer pointing to the host memory
cl_mem h_A = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR,
mem_size_A, h_A_data, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: clCreateBuffer\n");
return ciErrNum;
}
// Program Setup
size_t program_length;
const char* header_path = shrFindFilePath("matrixMul.h", argv[0]);
oclCheckError(header_path != NULL, shrTRUE);
char* header = oclLoadProgSource(header_path, "", &program_length);
if(!header)
{
shrLog("Error: Failed to load the header %s!\n", header_path);
return -1000;
}
const char* source_path = shrFindFilePath("matrixMul.cl", argv[0]);
oclCheckError(source_path != NULL, shrTRUE);
char *source = oclLoadProgSource(source_path, header, &program_length);
if(!source)
{
shrLog("Error: Failed to load compute program %s!\n", source_path);
return -2000;
}
// create the program
cl_program cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&source,
&program_length, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create program\n");
return ciErrNum;
}
free(header);
free(source);
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-fast-relaxed-math", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then return error
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclMatrixMul.ptx");
return ciErrNum;
}
// write out PTX if requested on the command line
if(shrCheckCmdLineFlag(argc, argv, "dump-ptx") )
{
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclMatrixMul.ptx");
}
// Create Kernel
for(unsigned int i = 0; i < ciDeviceCount; ++i) {
multiplicationKernel[i] = clCreateKernel(cpProgram, "matrixMul", &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create kernel\n");
return ciErrNum;
}
}
// Run multiplication on 1..deviceCount GPUs to compare improvement
shrLog("\nRunning Computations on 1 - %d GPU's...\n\n", ciDeviceCount);
for(unsigned int k = 1; k <= ciDeviceCount; ++k)
{
matrixMulGPU(k, h_A, h_B_data, mem_size_B, h_C);
}
// compute reference solution
shrLog("Comparing results with CPU computation... \n\n");
float* reference = (float*) malloc(mem_size_C);
computeGold(reference, h_A_data, h_B_data, uiHA, uiWA, uiWB);
// check result
shrBOOL res = shrCompareL2fe(reference, h_C, size_C, 1.0e-6f);
if (res != shrTRUE)
{
printDiff(reference, h_C, uiWC, uiHC, 100, 1.0e-5f);
}
// clean up OCL resources
ciErrNum = clReleaseMemObject(h_A);
for(unsigned int k = 0; k < ciDeviceCount; ++k)
{
ciErrNum |= clReleaseKernel( multiplicationKernel[k] );
ciErrNum |= clReleaseCommandQueue( commandQueue[k] );
}
ciErrNum |= clReleaseProgram(cpProgram);
ciErrNum |= clReleaseContext(cxGPUContext);
if(ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failure releasing OpenCL resources: %d\n", ciErrNum);
return ciErrNum;
}
// clean up memory
free(h_A_data);
free(h_B_data);
free(h_C);
free(reference);
return ((shrTRUE == res) ? CL_SUCCESS : -3000);
}
bool
runTest( int argc, const char** argv, ReduceType datatype)
{
int size = 1<<24; // number of elements to reduce
int maxThreads;
cl_kernel reductionKernel = getReductionKernel(datatype, 0, 64, 1);
clReleaseKernel(reductionKernel);
if (smallBlock)
maxThreads = 64; // number of threads per block
else
maxThreads = 128;
int whichKernel = 6;
int maxBlocks = 64;
bool cpuFinalReduction = false;
int cpuFinalThreshold = 1;
shrGetCmdLineArgumenti( argc, (const char**) argv, "n", &size);
shrGetCmdLineArgumenti( argc, (const char**) argv, "threads", &maxThreads);
shrGetCmdLineArgumenti( argc, (const char**) argv, "kernel", &whichKernel);
shrGetCmdLineArgumenti( argc, (const char**) argv, "maxblocks", &maxBlocks);
shrLog(" %d elements\n", size);
shrLog(" %d threads (max)\n", maxThreads);
cpuFinalReduction = (shrCheckCmdLineFlag( argc, (const char**) argv, "cpufinal") == shrTRUE);
shrGetCmdLineArgumenti( argc, (const char**) argv, "cputhresh", &cpuFinalThreshold);
bool runShmoo = (shrCheckCmdLineFlag(argc, (const char**) argv, "shmoo") == shrTRUE);
#ifdef GPU_PROFILING
if (runShmoo)
{
shmoo<T>(1, 33554432, maxThreads, maxBlocks, datatype);
return true;
}
else
#endif
{
// create random input data on CPU
unsigned int bytes = size * sizeof(T);
T* h_idata = (T*)malloc(bytes);
for(int i=0; i<size; i++)
{
// Keep the numbers small so we don't get truncation error in the sum
if (datatype == REDUCE_INT)
h_idata[i] = (T)(rand() & 0xFF);
else
h_idata[i] = (rand() & 0xFF) / (T)RAND_MAX;
}
int numBlocks = 0;
int numThreads = 0;
getNumBlocksAndThreads(whichKernel, size, maxBlocks, maxThreads, numBlocks, numThreads);
if (numBlocks == 1) cpuFinalThreshold = 1;
shrLog(" %d blocks\n\n", numBlocks);
// allocate mem for the result on host side
T* h_odata = (T*)malloc(numBlocks * sizeof(T));
// allocate device memory and data
cl_mem d_idata = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, bytes, h_idata, NULL);
cl_mem d_odata = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, numBlocks * sizeof(T), NULL, NULL);
int testIterations = 100;
double dTotalTime = 0.0;
T gpu_result = 0;
gpu_result = profileReduce<T>(datatype, size, numThreads, numBlocks, maxThreads, maxBlocks,
whichKernel, testIterations, cpuFinalReduction,
cpuFinalThreshold, &dTotalTime,
h_odata, d_idata, d_odata);
#ifdef GPU_PROFILING
double reduceTime = dTotalTime/(double)testIterations;
shrLogEx(LOGBOTH | MASTER, 0, "oclReduction, Throughput = %.4f GB/s, Time = %.5f s, Size = %u Elements, NumDevsUsed = %d, Workgroup = %u\n",
1.0e-9 * ((double)bytes)/reduceTime, reduceTime, size, 1, numThreads);
#endif
// compute reference solution
shrLog("\nComparing against Host/C++ computation...\n");
T cpu_result = reduceCPU<T>(h_idata, size);
if (datatype == REDUCE_INT)
{
shrLog(" GPU result = %d\n", gpu_result);
shrLog(" CPU result = %d\n\n", cpu_result);
shrLog("%s\n\n", (gpu_result == cpu_result) ? "PASSED" : "FAILED");
}
else
{
shrLog(" GPU result = %.9f\n", gpu_result);
shrLog(" CPU result = %.9f\n\n", cpu_result);
double threshold = (datatype == REDUCE_FLOAT) ? 1e-8 * size : 1e-12;
double diff = abs((double)gpu_result - (double)cpu_result);
shrLog("%s\n\n", (diff < threshold) ? "PASSED" : "FAILED");
}
// cleanup
free(h_idata);
free(h_odata);
clReleaseMemObject(d_idata);
clReleaseMemObject(d_odata);
return (gpu_result == cpu_result);
}
}
