// 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);
}
////////////////////////////////////////////////////////////////////////////////
//! 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);
}
////////////////////////////////////////////////////////////////////////////////
// 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, 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)
{
//////////////////////////////////////////////////////////////////////////
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;
//////////////////////////////////////////////////////////////////////////
}
int main(int argc, char *argv[])
{
cl_platform_id platform;
cl_device_id device;
cl_context context;
cl_command_queue queue;
cl_program program;
cl_kernel kernel;
cl_mem buff_A, buff_B, buff_C;
int mult = 1;
uint32_t uiWA, uiHA, uiWB, uiHB, uiWC, uiHC;
uiWA = WA * mult;
uiHA = HA * mult;
uiWB = WB * mult;
uiHB = HB * mult;
uiWC = WC * mult;
uiHC = HC * mult;
printf("sizes WA %u HA %u WB %u HB %u WC %u HC %u\n",
uiWA, uiHA, uiWB, uiHB, uiWC, uiHC);
uint32_t size_A = uiWA * uiHA;
uint32_t size_B = uiWB * uiHB;
uint32_t size_C = uiWC * uiHC;
size_t mem_size_A = size_A * sizeof(float);
size_t mem_size_B = size_B * sizeof(float);
size_t mem_size_C = size_C * sizeof(float);
float *data_A = (float *)malloc(mem_size_A);
float *data_B = (float *)malloc(mem_size_B);
float *data_C = (float *)malloc(mem_size_C);
srand(2012);
shrFillArray(data_A, size_A);
shrFillArray(data_B, size_B);
size_t global_work_size[2];
size_t local_work_size[] = { BLOCK_SIZE, BLOCK_SIZE };
global_work_size[0] = shrRoundUp(BLOCK_SIZE, uiWC);
global_work_size[1] = shrRoundUp(BLOCK_SIZE, uiHA);
const char *source = load_program_source("MatrixMul.cl");
size_t source_len = strlen(source);;
cl_uint err = 0;
char *flags = "-cl-fast-relaxed-math";
clGetPlatformIDs(1, &platform, NULL);
printf("platform %p err %d\n", platform, err);
clGetDeviceIDs(platform, CL_DEVICE_TYPE_CPU, 1, &device, &err);
printf("device %p err %d\n", device, err);
context = clCreateContext(0, 1, &device, NULL, NULL, &err);
printf("context %p err %d\n", context, err);
queue = clCreateCommandQueue(context, device, 0, &err);
printf("queue %p err %d\n", queue, err);
program = clCreateProgramWithSource(context, 1, &source, &source_len, &err);
printf("program %p err %d\n", program, err);
err = clBuildProgram(program, 0, NULL, flags, NULL, NULL);
printf("err %d\n", err);
kernel = clCreateKernel(program, "matrixMul", &err);
printf("kernel %p err %d\n", kernel, err);
buff_A = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
mem_size_A, data_A, NULL);
printf("buff_A %p\n", buff_A);
buff_B = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
mem_size_B, data_B, NULL);
printf("buff_B %p\n", buff_B);
buff_C = clCreateBuffer(context, CL_MEM_WRITE_ONLY, mem_size_C, NULL, NULL);
printf("buff_C %p\n", buff_C);
err = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void*)&buff_C);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&buff_A);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&buff_B);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 3, sizeof(float) * BLOCK_SIZE * BLOCK_SIZE, NULL);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 4, sizeof(float) * BLOCK_SIZE * BLOCK_SIZE, NULL);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 5, sizeof(cl_int), (void*)&uiWA);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 6, sizeof(cl_int), (void*)&uiWB);
printf("err %d\n", err);
err = clSetKernelArg(kernel, 7, sizeof(cl_int), (void*)&uiHA);
printf("err %d\n", err);
err = clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size,
local_work_size, 0, NULL, NULL);
printf("err %d\n", err);
err = clFlush(queue);
printf("err %d\n", err);
err = clFinish(queue);
printf("err %d\n", err);
err = clEnqueueReadBuffer(queue, buff_C, CL_TRUE, 0, mem_size_C, data_C, 0,
NULL, NULL);
printf("err %d\n", err);
int i;
for (i = 0; i < size_C; i++) {
printf("%d %f\n", i, data_C[i]);
}
clReleaseMemObject(buff_A);
clReleaseMemObject(buff_B);
clReleaseMemObject(buff_C);
clReleaseProgram(program);
clReleaseKernel(kernel);
clReleaseCommandQueue(queue);
clReleaseProgram(program);
}
