extern "C" void initScan(cl_context cxGPUContext, cl_command_queue cqParamCommandQue, const char **argv) {
cl_int ciErrNum;
size_t kernelLength;
shrLog(" ...loading Scan.cl\n");
char *cScan = oclLoadProgSource(shrFindFilePath("Scan.cl", argv[0]), "// My comment\n", &kernelLength);
oclCheckError(cScan != NULL, shrTRUE);
shrLog(" ...creating scan program\n");
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cScan, &kernelLength, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog(" ...building scan program\n");
ciErrNum = clBuildProgram(cpProgram, 0, NULL, compileOptions, 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), "oclScan.ptx");
oclCheckError(ciErrNum, CL_SUCCESS);
}
shrLog(" ...creating scan kernels\n");
ckScanExclusiveLocal1 = clCreateKernel(cpProgram, "scanExclusiveLocal1", &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
ckScanExclusiveLocal2 = clCreateKernel(cpProgram, "scanExclusiveLocal2", &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
ckUniformUpdate = clCreateKernel(cpProgram, "uniformUpdate", &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog( " ...checking minimum supported workgroup size\n");
//Check for work group size
cl_device_id device;
size_t szScanExclusiveLocal1, szScanExclusiveLocal2, szUniformUpdate;
ciErrNum = clGetCommandQueueInfo(cqParamCommandQue, CL_QUEUE_DEVICE, sizeof(cl_device_id), &device, NULL);
ciErrNum |= clGetKernelWorkGroupInfo(ckScanExclusiveLocal1, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &szScanExclusiveLocal1, NULL);
ciErrNum |= clGetKernelWorkGroupInfo(ckScanExclusiveLocal2, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &szScanExclusiveLocal2, NULL);
ciErrNum |= clGetKernelWorkGroupInfo(ckUniformUpdate, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &szUniformUpdate, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
if( (szScanExclusiveLocal1 < WORKGROUP_SIZE) || (szScanExclusiveLocal2 < WORKGROUP_SIZE) || (szUniformUpdate < WORKGROUP_SIZE) ) {
shrLog("ERROR: Minimum work-group size %u required by this application is not supported on this device.\n", WORKGROUP_SIZE);
exit(0);
}
shrLog(" ...allocating internal buffers\n");
d_Buffer = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, (MAX_BATCH_ELEMENTS / (4 * WORKGROUP_SIZE)) * sizeof(uint), NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
//Discard temp storage
free(cScan);
}
extern "C" void initHistogram64(cl_context cxGPUContext, cl_command_queue cqParamCommandQue, const char **argv){
cl_int ciErrNum;
size_t kernelLength;
shrLog("...loading Histogram64.cl from file\n");
char *cHistogram64 = oclLoadProgSource(shrFindFilePath("Histogram64.cl", argv[0]), "// My comment\n", &kernelLength);
shrCheckError(cHistogram64 != NULL, shrTRUE);
shrLog("...creating histogram64 program\n");
cpHistogram64 = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cHistogram64, &kernelLength, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog("...building histogram64 program\n");
ciErrNum = clBuildProgram(cpHistogram64, 0, NULL, compileOptions, NULL, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog("...creating histogram64 kernels\n");
ckHistogram64 = clCreateKernel(cpHistogram64, "histogram64", &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
ckMergeHistogram64 = clCreateKernel(cpHistogram64, "mergeHistogram64", &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog("...allocating internal histogram64 buffer\n");
d_PartialHistograms = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, MAX_PARTIAL_HISTOGRAM64_COUNT * HISTOGRAM64_BIN_COUNT * sizeof(uint), NULL, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
//Save default command queue
cqDefaultCommandQue = cqParamCommandQue;
//Discard temp storage
free(cHistogram64);
//Save ptx code to separate file
oclLogPtx(cpHistogram64, oclGetFirstDev(cxGPUContext), "Histogram64.ptx");
}
//-----------------------------------------------------------------------------
// Name: CreateKernelProgram()
// Desc: Creates OpenCL program and kernel instances
//-----------------------------------------------------------------------------
HRESULT CreateKernelProgram(
const char *exepath, const char *clName, const char *clPtx, const char *kernelEntryPoint,
cl_program &cpProgram,
cl_kernel &ckKernel )
{
// Program Setup
size_t program_length;
const char* source_path = shrFindFilePath(clName, exepath);
char *source = oclLoadProgSource(source_path, "", &program_length);
oclCheckErrorEX(source != NULL, shrTRUE, pCleanup);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,(const char **) &source, &program_length, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
free(source);
// build the program
#ifdef USE_STAGING_BUFFER
static char *opts = "-cl-fast-relaxed-math -DUSE_STAGING_BUFFER";
#else
static char *opts = "-cl-fast-relaxed-math";
#endif
ciErrNum = clBuildProgram(cpProgram, 0, NULL, opts, 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), clPtx);
Cleanup(EXIT_FAILURE);
}
// create the kernel
ckKernel = clCreateKernel(cpProgram, kernelEntryPoint, &ciErrNum);
if (!ckKernel)
{
Cleanup(EXIT_FAILURE);
}
// set the args values
return ciErrNum ? E_FAIL : S_OK;
}
// ****************************************************************************
// Method: dumpPTXCode
//
// Purpose:
//
//
// Arguments:
// ctx context
// prg program
// name file name prefix to output to
//
// Programmer: Gabriel Marin
// Creation: July 14, 2009
//
// ****************************************************************************
inline bool
dumpPTXCode (cl_context ctx, cl_program prog, const char *name)
{
std::cout << "Dumping the PTX code" << std::endl;
size_t ptx_length;
char* ptx_code;
char buf[64];
oclGetProgBinary (prog, oclGetFirstDev(ctx), &ptx_code, &ptx_length);
FILE* ptxFile = NULL;
sprintf (buf, "%.59s.ptx", name);
#ifdef WIN32
fopen_s (&ptxFile, buf, "w");
#else
ptxFile = fopen (buf,"w");
#endif
if (ptxFile)
{
fwrite (ptx_code, ptx_length, 1, ptxFile);
fclose (ptxFile);
}
free (ptx_code);
return (ptx_code!=0);
}
// 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);
}
extern "C" void initBlackScholes(cl_context cxGPUContext, cl_command_queue cqParamCommandQueue, const char **argv){
cl_int ciErrNum;
size_t kernelLength;
shrLog("...loading BlackScholes.cl\n");
char *cPathAndName = shrFindFilePath("BlackScholes.cl", argv[0]);
shrCheckError(cPathAndName != NULL, shrTRUE);
char *cBlackScholes = oclLoadProgSource(cPathAndName, "// My comment\n", &kernelLength);
shrCheckError(cBlackScholes != NULL, shrTRUE);
shrLog("...creating BlackScholes program\n");
cpBlackScholes = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cBlackScholes, &kernelLength, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog("...building BlackScholes program\n");
ciErrNum = clBuildProgram(cpBlackScholes, 0, NULL, "-cl-fast-relaxed-math -Werror", NULL, NULL);
if(ciErrNum != CL_BUILD_SUCCESS){
shrLog("*** Compilation failure ***\n");
size_t deviceNum;
cl_device_id *cdDevices;
ciErrNum = clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &deviceNum);
shrCheckError(ciErrNum, CL_SUCCESS);
cdDevices = (cl_device_id *)malloc(deviceNum * sizeof(cl_device_id));
shrCheckError(cdDevices != NULL, shrTRUE);
ciErrNum = clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, deviceNum * sizeof(cl_device_id), cdDevices, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
size_t logSize;
char *logTxt;
ciErrNum = clGetProgramBuildInfo(cpBlackScholes, cdDevices[0], CL_PROGRAM_BUILD_LOG, 0, NULL, &logSize);
shrCheckError(ciErrNum, CL_SUCCESS);
logTxt = (char *)malloc(logSize);
shrCheckError(logTxt != NULL, shrTRUE);
ciErrNum = clGetProgramBuildInfo(cpBlackScholes, cdDevices[0], CL_PROGRAM_BUILD_LOG, logSize, logTxt, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog("%s\n", logTxt);
shrLog("*** Exiting ***\n");
free(logTxt);
free(cdDevices);
exit(666);
}
//Save ptx code to separate file
oclLogPtx(cpBlackScholes, oclGetFirstDev(cxGPUContext), "BlackScholes.ptx");
shrLog("...creating BlackScholes kernels\n");
ckBlackScholes = clCreateKernel(cpBlackScholes, "BlackScholes", &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
cqDefaultCommandQueue = cqParamCommandQueue;
free(cBlackScholes);
free(cPathAndName);
}
// 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 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 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);
}
/**
* Helper function that will load the OpenCL source program, build and return a handle to that OpenCL kernel
* @param context - the OpenCL context
* @param device - the device to compile for
* @param program - the program that is being built (in/out)
* @param sourcePath - full path to the source file
* @param options - build options, e.g. define flags("-D name=value")
* @return an error code on failure, 0 on success
*/
cl_program CLState::compileOCLProgram(const char* sourcePath,
const std::string& options) {
cl_int errNum;
size_t program_length;
oclCheckError(sourcePath != NULL, shrTRUE);
char *source = oclLoadProgSource(sourcePath, "", &program_length);
if (!source) {
shrLog("Error: Failed to load compute program %s!\n", sourcePath);
BOOST_THROW_EXCEPTION(
runtime_error()
<< error_message(
(std::string(
"Error: Failed to load cl program source from ")
+ sourcePath).c_str()));
}
// create the simple increment OpenCL program
cl_program program = clCreateProgramWithSource(context, 1,
(const char **) &source, &program_length, &errNum);
if (errNum != CL_SUCCESS) {
shrLog("Error: Failed to create program\n");
BOOST_THROW_EXCEPTION(runtime_error() << cl_error_code(errNum));
} else {
shrLog("clCreateProgramWithSource <%s> succeeded, program_length=%d\n",
sourcePath, program_length);
}
free(source);
// build the program
cl_build_status build_status = CL_SUCCESS;
errNum =
clBuildProgram(program, 0, NULL,
std::string(
"-cl-fast-relaxed-math -cl-nv-verbose" + options).c_str(),
NULL, NULL);
if (errNum != CL_SUCCESS) {
// write out standard error, Build Log and PTX, then return error
shrLogEx(LOGBOTH | ERRORMSG, errNum, STDERROR);
oclLogBuildInfo(program, oclGetFirstDev(context));
oclLogPtx(program, oclGetFirstDev(context), "build_error.ptx");
BOOST_THROW_EXCEPTION(runtime_error() << cl_error_code(errNum));
} else {
shrLog("clBuildProgram <%s> succeeded\n", sourcePath);
if (this->execDevices != NULL) {
for (uint iDevice = 0;
iDevice < this->execDeviceCount
&& build_status == CL_SUCCESS
&& errNum == CL_SUCCESS; iDevice++) {
cl_device_id device = this->execDevices[iDevice];
errNum = clGetProgramBuildInfo(program, device,
CL_PROGRAM_BUILD_STATUS, sizeof(cl_build_status), &build_status,
NULL);
shrLog("clGetProgramBuildInfo returned: ");
if (build_status == CL_SUCCESS) {
shrLog("CL_SUCCESS\n");
} else {
shrLog("CLErrorNumber = %d\n", errNum);
}
// print out the build log, note in the case where there is nothing shown, some OpenCL PTX->SASS caching has happened
{
char *build_log;
size_t ret_val_size;
errNum = clGetProgramBuildInfo(program, device,
CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
if (errNum != CL_SUCCESS) {
shrLog(
"clGetProgramBuildInfo device %d, failed to get the log size at line %d\n",
device, __LINE__);
}
build_log = (char *) malloc(ret_val_size + 1);
errNum = clGetProgramBuildInfo(program, device,
CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
if (errNum != CL_SUCCESS) {
shrLog(
"clGetProgramBuildInfo device %d, failed to get the build log at line %d\n",
device, __LINE__);
}
// to be carefully, terminate with \0
// there's no information in the reference whether the string is 0 terminated or not
build_log[ret_val_size] = '\0';
shrLog("%s\n", build_log);
}
}
}
}
this->programs.push_back(program);
return program;
}
// Init OpenCL
//*****************************************************************************
int initCL(int argc, const char** argv)
{
cl_platform_id cpPlatform;
cl_uint uiDevCount;
cl_device_id *cdDevices;
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Get the number of GPU devices available to the platform
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiDevCount);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create the device list
cdDevices = new cl_device_id [uiDevCount];
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiDevCount, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Get device requested on command line, if any
unsigned int uiDeviceUsed = 0;
unsigned int uiEndDev = uiDevCount - 1;
if(shrGetCmdLineArgumentu(argc, argv, "device", &uiDeviceUsed))
{
uiDeviceUsed = CLAMP(uiDeviceUsed, 0, uiEndDev);
uiEndDev = uiDeviceUsed;
}
// Check if the requested device (or any of the devices if none requested) supports context sharing with OpenGL
if(bGLinterop && !bQATest)
{
bool bSharingSupported = false;
for(unsigned int i = uiDeviceUsed; (!bSharingSupported && (i <= uiEndDev)); ++i)
{
size_t extensionSize;
ciErrNum = clGetDeviceInfo(cdDevices[i], CL_DEVICE_EXTENSIONS, 0, NULL, &extensionSize );
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
if(extensionSize > 0)
{
char* extensions = (char*)malloc(extensionSize);
ciErrNum = clGetDeviceInfo(cdDevices[i], CL_DEVICE_EXTENSIONS, extensionSize, extensions, &extensionSize);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
std::string stdDevString(extensions);
free(extensions);
size_t szOldPos = 0;
size_t szSpacePos = stdDevString.find(' ', szOldPos); // extensions string is space delimited
while (szSpacePos != stdDevString.npos)
{
if( strcmp(GL_SHARING_EXTENSION, stdDevString.substr(szOldPos, szSpacePos - szOldPos).c_str()) == 0 )
{
// Device supports context sharing with OpenGL
uiDeviceUsed = i;
bSharingSupported = true;
break;
}
do
{
szOldPos = szSpacePos + 1;
szSpacePos = stdDevString.find(' ', szOldPos);
}
while (szSpacePos == szOldPos);
}
}
}
shrLog("%s...\n\n", bSharingSupported ? "Using CL-GL Interop" : "No device found that supports CL/GL context sharing");
oclCheckErrorEX(bSharingSupported, true, pCleanup);
// Define OS-specific context properties and create the OpenCL context
#if defined (__APPLE__) || defined (MACOSX)
CGLContextObj kCGLContext = CGLGetCurrentContext();
CGLShareGroupObj kCGLShareGroup = CGLGetShareGroup(kCGLContext);
cl_context_properties props[] =
{
CL_CONTEXT_PROPERTY_USE_CGL_SHAREGROUP_APPLE, (cl_context_properties)kCGLShareGroup,
0
};
cxGPUContext = clCreateContext(props, 0,0, NULL, NULL, &ciErrNum);
#else
#ifdef UNIX
cl_context_properties props[] =
{
CL_GL_CONTEXT_KHR, (cl_context_properties)glXGetCurrentContext(),
CL_GLX_DISPLAY_KHR, (cl_context_properties)glXGetCurrentDisplay(),
CL_CONTEXT_PLATFORM, (cl_context_properties)cpPlatform,
0
};
cxGPUContext = clCreateContext(props, 1, &cdDevices[uiDeviceUsed], NULL, NULL, &ciErrNum);
#else // Win32
cl_context_properties props[] =
{
CL_GL_CONTEXT_KHR, (cl_context_properties)wglGetCurrentContext(),
CL_WGL_HDC_KHR, (cl_context_properties)wglGetCurrentDC(),
CL_CONTEXT_PLATFORM, (cl_context_properties)cpPlatform,
0
};
cxGPUContext = clCreateContext(props, 1, &cdDevices[uiDeviceUsed], NULL, NULL, &ciErrNum);
#endif
#endif
}
else
{
// No GL interop
cl_context_properties props[] = {CL_CONTEXT_PLATFORM, (cl_context_properties)cpPlatform, 0};
cxGPUContext = clCreateContext(props, 1, &cdDevices[uiDeviceUsed], NULL, NULL, &ciErrNum);
bGLinterop = shrFALSE;
}
shrCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Log device used
shrLog("Device # %u, ", uiDeviceUsed);
oclPrintDevName(LOGBOTH, cdDevices[uiDeviceUsed]);
shrLog("\n");
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevices[uiDeviceUsed], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Memory Setup
if( bGLinterop ) {
cl_pbos[0] = clCreateFromGLBuffer(cxGPUContext, CL_MEM_READ_ONLY, pbo_source, &ciErrNum);
cl_pbos[1] = clCreateFromGLBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, pbo_dest, &ciErrNum);
} else {
cl_pbos[0] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, 4 * image_width * image_height, NULL, &ciErrNum);
cl_pbos[1] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, 4 * image_width * image_height, NULL, &ciErrNum);
}
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Program Setup
size_t program_length;
const char* source_path = shrFindFilePath(clSourcefile, argv[0]);
char *source = oclLoadProgSource(source_path, "", &program_length);
oclCheckErrorEX(source != NULL, shrTRUE, pCleanup);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,(const char **) &source, &program_length, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
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 cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclPostProcessGL.ptx");
Cleanup(EXIT_FAILURE);
}
// create the kernel
ckKernel = clCreateKernel(cpProgram, "postprocess", &ciErrNum);
// set the args values
ciErrNum |= clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void *) &(cl_pbos[0]));
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void *) &(cl_pbos[1]));
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(image_width), &image_width);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(image_width), &image_height);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
return 0;
}
int InitOpenCLContext()
{
// start logs
shrSetLogFileName ("oclVolumeRender.txt");
// get command line arg for quick test, if provided
// process command line arguments
// First initialize OpenGL context, so we can properly setup the OpenGL / OpenCL interop.
// glewInit();
// GLboolean bGLEW = glewIsSupported("GL_VERSION_2_0 GL_ARB_pixel_buffer_object");
// oclCheckErrorEX(bGLEW, shrTRUE, pCleanup);
g_glInterop = true;
// Create OpenCL context, get device info, select device, select options for image/texture and CL-GL interop
createCLContext();
// Print device info
clGetDeviceInfo(cdDevices[uiDeviceUsed], CL_DEVICE_IMAGE_SUPPORT, sizeof(g_bImageSupport), &g_bImageSupport, NULL);
//shrLog("%s...\n\n", g_bImageSupport ? "Using Image (Texture)" : "No Image (Texuture) Support");
// shrLog("Detailed Device info:\n\n");
oclPrintDevInfo(LOGBOTH, cdDevices[uiDeviceUsed]);
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevices[uiDeviceUsed], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Program Setup
size_t program_length;
cPathAndName = shrFindFilePath("Transform.cl", ".");
oclCheckErrorEX(cPathAndName != NULL, shrTRUE, pCleanup);
cSourceCL = oclLoadProgSource(cPathAndName, "", &program_length);
oclCheckErrorEX(cSourceCL != NULL, shrTRUE, pCleanup);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,
(const char **)&cSourceCL, &program_length, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// build the program
std::string buildOpts = "-cl-single-precision_constant";
// buildOpts += g_bImageSupport ? " -DIMAGE_SUPPORT" : "";
// ciErrNum = clBuildProgram(cpProgram, 1, &cdDevices[uiDeviceUsed],"-cl-fast-relaxed-math", NULL, NULL);
ciErrNum = clBuildProgram(cpProgram, 1, &cdDevices[uiDeviceUsed],NULL, NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and return error
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclVolumeRender.ptx");
Cleanup(EXIT_FAILURE);
}
// create the kernel
ScalseKernel = clCreateKernel(cpProgram, "d_render", &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
TransformKernel = clCreateKernel(cpProgram, "angle", &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
LongToShortKernel = clCreateKernel(cpProgram, "transfer", &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
return TRUE;
}
// 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);
}
// Function to read in kernel from uncompiled source, create the OCL program and build the OCL program
// **************************************************************************************************
int CreateProgramAndKernel(cl_context cxGPUContext, cl_device_id* cdDevices, const char *kernel_name, cl_kernel *kernel, bool bDouble)
{
cl_program cpProgram;
size_t szSourceLen;
cl_int ciErrNum = CL_SUCCESS;
// Read the kernel in from file
shrLog("\nLoading Uncompiled kernel from .cl file, using %s\n", clSourcefile);
char* cPathAndFile = shrFindFilePath(clSourcefile, cExecutablePath);
oclCheckError(cPathAndFile != NULL, shrTRUE);
char* pcSource = oclLoadProgSource(cPathAndFile, "", &szSourceLen);
oclCheckError(pcSource != NULL, shrTRUE);
// Check OpenCL version -> vec3 types are supported only from version 1.1 and above
char cOCLVersion[32];
clGetDeviceInfo(cdDevices[0], CL_DEVICE_VERSION, sizeof(cOCLVersion), &cOCLVersion, 0);
int iVec3Length = 3;
if( strncmp("OpenCL 1.0", cOCLVersion, 10) == 0 ) {
iVec3Length = 4;
}
//for double precision
char *pcSourceForDouble;
std::stringstream header;
if (bDouble)
{
header << "#define REAL double";
header << std::endl;
header << "#define REAL4 double4";
header << std::endl;
header << "#define REAL3 double" << iVec3Length;
header << std::endl;
header << "#define ZERO3 {0.0, 0.0, 0.0" << ((iVec3Length == 4) ? ", 0.0}" : "}");
header << std::endl;
}
else
{
header << "#define REAL float";
header << std::endl;
header << "#define REAL4 float4";
header << std::endl;
header << "#define REAL3 float" << iVec3Length;
header << std::endl;
header << "#define ZERO3 {0.0f, 0.0f, 0.0f" << ((iVec3Length == 4) ? ", 0.0f}" : "}");
header << std::endl;
}
header << pcSource;
pcSourceForDouble = (char *)malloc(header.str().size() + 1);
szSourceLen = header.str().size();
#ifdef WIN32
strcpy_s(pcSourceForDouble, szSourceLen + 1, header.str().c_str());
#else
strcpy(pcSourceForDouble, header.str().c_str());
#endif
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&pcSourceForDouble, &szSourceLen, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
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)
{
// 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), "oclNbody.ptx");
oclCheckError(ciErrNum, CL_SUCCESS);
}
shrLog("clBuildProgram\n");
// create the kernel
*kernel = clCreateKernel(cpProgram, kernel_name, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
shrLog("clCreateKernel\n");
size_t wgSize;
ciErrNum = clGetKernelWorkGroupInfo(*kernel, cdDevices[0], CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &wgSize, NULL);
if (wgSize == 64) {
shrLog(
"ERROR: Minimum work-group size 256 required by this application is not supported on this device.\n");
exit(0);
}
free(pcSourceForDouble);
return 0;
}
////////////////////////////////////////////////////////////////////////////////
//! 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);
}
// Helper function to create and build program and kernel
// *********************************************************************
cl_kernel getReductionKernel(ReduceType datatype, int whichKernel, int blockSize, int isPowOf2)
{
// compile cl program
size_t program_length;
char *source;
std::ostringstream preamble;
// create the program
// with type specification depending on datatype argument
switch (datatype)
{
default:
case REDUCE_INT:
preamble << "#define T int" << std::endl;
break;
case REDUCE_FLOAT:
preamble << "#define T float" << std::endl;
break;
}
// set blockSize at compile time
preamble << "#define blockSize " << blockSize << std::endl;
// set isPow2 at compile time
preamble << "#define nIsPow2 " << isPowOf2 << std::endl;
// Load the source code and prepend the preamble
source = oclLoadProgSource(source_path, preamble.str().c_str(), &program_length);
oclCheckError(source != NULL, shrTRUE);
cl_program cpProgram = clCreateProgramWithSource(cxGPUContext, 1,(const char **) &source,
&program_length, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
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 cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclReduction.ptx");
oclCheckError(ciErrNum, CL_SUCCESS);
}
// create Kernel
std::ostringstream kernelName;
kernelName << "reduce" << whichKernel;
cl_kernel ckKernel = clCreateKernel(cpProgram, kernelName.str().c_str(), &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
size_t wgSize;
ciErrNum = clGetKernelWorkGroupInfo(ckKernel, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &wgSize, NULL);
if (wgSize == 64)
smallBlock = true;
else smallBlock = false;
// NOTE: the program will get deleted when the kernel is also released
clReleaseProgram(cpProgram);
return ckKernel;
}
