GPUImageProcessor::GPUImageProcessor(int width,int height,int nChannels)
{
//cout << "gpu computing konstr" << endl;
GPUError = oclGetPlatformID(&cpPlatform);
CheckError(GPUError);
cl_uint uiNumAllDevs = 0;
// Get the number of GPU devices available to the platform
GPUError = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumAllDevs);
CheckError(GPUError);
uiDevCount = uiNumAllDevs;
// Create the device list
cdDevices = new cl_device_id [uiDevCount];
GPUError = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiDevCount, cdDevices, NULL);
CheckError(GPUError);
// Create the OpenCL context on a GPU device
GPUContext = clCreateContext(0, uiNumAllDevs, cdDevices, NULL, NULL, &GPUError);
CheckError(GPUError);
//The command-queue can be used to queue a set of operations (referred to as commands) in order.
GPUCommandQueue = clCreateCommandQueue(GPUContext, cdDevices[0], 0, &GPUError);
CheckError(GPUError);
Transfer = new GPUTransferManager(GPUContext,GPUCommandQueue,width,height,nChannels);
oclPrintDevName(LOGBOTH, cdDevices[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);
}
//-----------------------------------------------------------------------------
// Name: InitCL()
// Desc: Get platform and devices and create context and queues
//-----------------------------------------------------------------------------
HRESULT InitCL(int argc, const char** argv)
{
cl_platform_id cpPlatform;
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
//
// Initialize extension functions for D3D9
//
INITPFN(clGetDeviceIDsFromD3D9NV);
INITPFN(clCreateFromD3D9VertexBufferNV);
INITPFN(clCreateFromD3D9IndexBufferNV);
INITPFN(clCreateFromD3D9SurfaceNV);
INITPFN(clCreateFromD3D9TextureNV);
INITPFN(clCreateFromD3D9CubeTextureNV);
INITPFN(clCreateFromD3D9VolumeTextureNV);
INITPFN(clEnqueueAcquireD3D9ObjectsNV);
INITPFN(clEnqueueReleaseD3D9ObjectsNV);
INITPFN(clGetDeviceIDsFromD3D9NV);
// Query the OpenCL device that would be good for the current D3D device
// We need to take the one that is on the same Gfx card.
// Get the device ids for the adapter
cl_device_id cdDevice;
cl_uint num_devices = 0;
ciErrNum = clGetDeviceIDsFromD3D9NV(
cpPlatform,
CL_D3D9_DEVICE_NV,//CL_D3D9_ADAPTER_NAME_NV,
g_pD3DDevice,//adapterName,
CL_PREFERRED_DEVICES_FOR_D3D9_NV, //CL_ALL_DEVICES_FOR_D3D9_NV,
1,
&cdDevice,
&num_devices);
if (ciErrNum == -1) {
shrLog("No OpenCL device available that supports D3D9, exiting...\n");
Cleanup (EXIT_SUCCESS);
} else {
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
}
cl_context_properties props[] =
{
CL_CONTEXT_D3D9_DEVICE_NV, (cl_context_properties)g_pD3DDevice,
CL_CONTEXT_PLATFORM, (cl_context_properties)cpPlatform,
0
};
cxGPUContext = clCreateContext(props, 1, &cdDevice, NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Log device used
shrLog("Device: ");
oclPrintDevName(LOGBOTH, cdDevice);
shrLog("\n");
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevice, 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
CreateKernelProgram(argv[0], "texture_2d.cl", "texture_2d.ptx", "cl_kernel_texture_2d", cpProgram_tex2d, ckKernel_tex2d);
CreateKernelProgram(argv[0], "texture_cube.cl", "texture_cube.ptx", "cl_kernel_texture_cube", cpProgram_texcube, ckKernel_texcube);
CreateKernelProgram(argv[0], "texture_volume.cl", "texture_volume.ptx", "cl_kernel_texture_volume", cpProgram_texvolume, ckKernel_texvolume);
return S_OK;
}
// Main program
//*****************************************************************************
int main(int argc, char** argv)
{
// Locals used with command line args
int p = 256; // workgroup X dimension
int q = 1; // workgroup Y dimension
pArgc = &argc;
pArgv = argv;
shrQAStart(argc, argv);
// latch the executable path for other funcs to use
cExecutablePath = argv[0];
// start logs and show command line help
shrSetLogFileName ("oclNbody.txt");
shrLog("%s Starting...\n\n", cExecutablePath);
shrLog("Command line switches:\n");
shrLog(" --qatest\t\tCheck correctness of GPU execution and measure performance)\n");
shrLog(" --noprompt\t\tQuit simulation automatically after a brief period\n");
shrLog(" --n=<numbodies>\tSpecify # of bodies to simulate (default = %d)\n", numBodies);
shrLog(" --double\t\tUse double precision floating point values for simulation\n");
shrLog(" --p=<workgroup X dim>\tSpecify X dimension of workgroup (default = %d)\n", p);
shrLog(" --q=<workgroup Y dim>\tSpecify Y dimension of workgroup (default = %d)\n\n", q);
// Get command line arguments if there are any and set vars accordingly
if (argc > 0)
{
shrGetCmdLineArgumenti(argc, (const char**)argv, "p", &p);
shrGetCmdLineArgumenti(argc, (const char**)argv, "q", &q);
shrGetCmdLineArgumenti(argc, (const char**)argv, "n", &numBodies);
bDouble = (shrTRUE == shrCheckCmdLineFlag(argc, (const char**)argv, "double"));
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
bQATest = shrCheckCmdLineFlag(argc, (const char**)argv, "qatest");
}
//Get the NVIDIA platform
cl_int ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
shrLog("clGetPlatformID...\n\n");
if (bDouble)
{
shrLog("Double precision execution...\n\n");
}
else
{
shrLog("Single precision execution...\n\n");
}
flopsPerInteraction = bDouble ? 30 : 20;
//Get all the devices
shrLog("Get the Device info and select Device...\n");
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Set target device and Query number of compute units on uiTargetDevice
shrLog(" # of Devices Available = %u\n", uiNumDevices);
if(shrGetCmdLineArgumentu(argc, (const char**)argv, "device", &uiTargetDevice)== shrTRUE)
{
uiTargetDevice = CLAMP(uiTargetDevice, 0, (uiNumDevices - 1));
}
shrLog(" Using Device %u, ", uiTargetDevice);
oclPrintDevName(LOGBOTH, cdDevices[uiTargetDevice]);
cl_uint uiNumComputeUnits;
clGetDeviceInfo(cdDevices[uiTargetDevice], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(uiNumComputeUnits), &uiNumComputeUnits, NULL);
shrLog(" # of Compute Units = %u\n", uiNumComputeUnits);
//Create the context
shrLog("clCreateContext...\n");
cxContext = clCreateContext(0, uiNumDevsUsed, &cdDevices[uiTargetDevice], NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create a command-queue
shrLog("clCreateCommandQueue...\n\n");
cqCommandQueue = clCreateCommandQueue(cxContext, cdDevices[uiTargetDevice], CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Log and config for number of bodies
shrLog("Number of Bodies = %d\n", numBodies);
switch (numBodies)
{
case 1024:
activeParams.m_clusterScale = 1.52f;
activeParams.m_velocityScale = 2.f;
break;
case 2048:
activeParams.m_clusterScale = 1.56f;
activeParams.m_velocityScale = 2.64f;
break;
case 4096:
activeParams.m_clusterScale = 1.68f;
activeParams.m_velocityScale = 2.98f;
break;
case 7680:
case 8192:
activeParams.m_clusterScale = 1.98f;
activeParams.m_velocityScale = 2.9f;
break;
default:
case 15360:
case 16384:
activeParams.m_clusterScale = 1.54f;
activeParams.m_velocityScale = 8.f;
break;
case 30720:
case 32768:
activeParams.m_clusterScale = 1.44f;
activeParams.m_velocityScale = 11.f;
break;
}
if ((q * p) > 256)
{
p = 256 / q;
shrLog("Setting p=%d to maintain %d threads per block\n", p, 256);
}
if ((q == 1) && (numBodies < p))
{
p = numBodies;
shrLog("Setting p=%d because # of bodies < p\n", p);
}
shrLog("Workgroup Dims = (%d x %d)\n\n", p, q);
// Initialize OpenGL items if using GL
if (bQATest == shrFALSE)
{
assert(0);
/*
shrLog("Calling InitGL...\n");
InitGL(&argc, argv);
*/
}
else
{
shrLog("Skipping InitGL...\n");
}
// CL/GL interop disabled
bUsePBO = (false && (bQATest == shrFALSE));
InitNbody(cdDevices[uiTargetDevice], cxContext, cqCommandQueue, numBodies, p, q, bUsePBO, bDouble);
ResetSim(nbody, numBodies, NBODY_CONFIG_SHELL, bUsePBO);
// init timers
shrDeltaT(DEMOTIME); // timer 0 is for timing demo periods
shrDeltaT(FUNCTIME); // timer 1 is for logging function delta t's
shrDeltaT(FPSTIME); // timer 2 is for fps measurement
// Standard simulation
if (bQATest == shrFALSE)
{
assert(0);
/*
shrLog("Running standard oclNbody simulation...\n\n");
glutDisplayFunc(DisplayGL);
glutReshapeFunc(ReshapeGL);
glutMouseFunc(MouseGL);
glutMotionFunc(MotionGL);
glutKeyboardFunc(KeyboardGL);
glutSpecialFunc(SpecialGL);
glutIdleFunc(IdleGL);
glutMainLoop();
*/
}
// Compare to host, profile and write out file for regression analysis
if (bQATest == shrTRUE) {
bool bTestResults = false;
shrLog("Running oclNbody Results Comparison...\n\n");
bTestResults = CompareResults(numBodies);
//shrLog("Profiling oclNbody...\n\n");
//RunProfiling(100, (unsigned int)(p * q)); // 100 iterations
shrQAFinish(argc, (const char **)argv, bTestResults ? QA_PASSED : QA_FAILED);
} else {
// Cleanup/exit
bNoPrompt = shrTRUE;
shrQAFinish2(false, *pArgc, (const char **)pArgv, QA_PASSED);
}
Cleanup(EXIT_SUCCESS);
}
// Main function
// *********************************************************************
int main(int argc, char **argv)
{
gp_argc = &argc;
gp_argv = &argv;
shrQAStart(argc, argv);
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clGetPlatformID...\n");
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clGetPlatformID...\n");
//Get all the devices
cl_uint uiNumDevices = 0; // Number of devices available
cl_uint uiTargetDevice = 0; // Default Device to compute on
cl_uint uiNumComputeUnits; // Number of compute units (SM's on NV GPU)
shrLog("Get the Device info and select Device...\n");
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
// Get command line device options and config accordingly
shrLog(" # of Devices Available = %u\n", uiNumDevices);
if(shrGetCmdLineArgumentu(argc, (const char**)argv, "device", &uiTargetDevice)== shrTRUE)
{
uiTargetDevice = CLAMP(uiTargetDevice, 0, (uiNumDevices - 1));
}
shrLog(" Using Device %u: ", uiTargetDevice);
oclPrintDevName(LOGBOTH, cdDevices[uiTargetDevice]);
ciErrNum = clGetDeviceInfo(cdDevices[uiTargetDevice], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(uiNumComputeUnits), &uiNumComputeUnits, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("\n # of Compute Units = %u\n", uiNumComputeUnits);
// get command line arg for quick test, if provided
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
// start logs
cExecutableName = argv[0];
shrSetLogFileName ("oclDotProduct.txt");
shrLog("%s Starting...\n\n# of float elements per Array \t= %u\n", argv[0], iNumElements);
// set and log Global and Local work size dimensions
szLocalWorkSize = 256;
szGlobalWorkSize = shrRoundUp((int)szLocalWorkSize, iNumElements); // rounded up to the nearest multiple of the LocalWorkSize
shrLog("Global Work Size \t\t= %u\nLocal Work Size \t\t= %u\n# of Work Groups \t\t= %u\n\n",
szGlobalWorkSize, szLocalWorkSize, (szGlobalWorkSize % szLocalWorkSize + szGlobalWorkSize/szLocalWorkSize));
// Allocate and initialize host arrays
shrLog( "Allocate and Init Host Mem...\n");
srcA = (void *)malloc(sizeof(cl_float4) * szGlobalWorkSize);
srcB = (void *)malloc(sizeof(cl_float4) * szGlobalWorkSize);
dst = (void *)malloc(sizeof(cl_float) * szGlobalWorkSize);
Golden = (void *)malloc(sizeof(cl_float) * iNumElements);
shrFillArray((float*)srcA, 4 * iNumElements);
shrFillArray((float*)srcB, 4 * iNumElements);
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Get a GPU device
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &cdDevices[uiTargetDevice], NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create the context
cxGPUContext = clCreateContext(0, 1, &cdDevices[uiTargetDevice], NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Create a command-queue
shrLog("clCreateCommandQueue...\n");
cqCommandQueue = clCreateCommandQueue(cxGPUContext, cdDevices[uiTargetDevice], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Allocate the OpenCL buffer memory objects for source and result on the device GMEM
shrLog("clCreateBuffer (SrcA, SrcB and Dst in Device GMEM)...\n");
cmDevSrcA = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize * 4, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmDevSrcB = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, sizeof(cl_float) * szGlobalWorkSize * 4, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
cmDevDst = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, sizeof(cl_float) * szGlobalWorkSize, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Read the OpenCL kernel in from source file
shrLog("oclLoadProgSource (%s)...\n", cSourceFile);
cPathAndName = shrFindFilePath(cSourceFile, argv[0]);
oclCheckErrorEX(cPathAndName != NULL, shrTRUE, pCleanup);
cSourceCL = oclLoadProgSource(cPathAndName, "", &szKernelLength);
oclCheckErrorEX(cSourceCL != NULL, shrTRUE, pCleanup);
// Create the program
shrLog("clCreateProgramWithSource...\n");
cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&cSourceCL, &szKernelLength, &ciErrNum);
// Build the program with 'mad' Optimization option
#ifdef MAC
char* flags = "-cl-fast-relaxed-math -DMAC";
#else
char* flags = "-cl-fast-relaxed-math";
#endif
shrLog("clBuildProgram...\n");
ciErrNum = clBuildProgram(cpProgram, 0, NULL, NULL, NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclDotProduct.ptx");
Cleanup(EXIT_FAILURE);
}
// Create the kernel
shrLog("clCreateKernel (DotProduct)...\n");
ckKernel = clCreateKernel(cpProgram, "DotProduct", &ciErrNum);
// Set the Argument values
shrLog("clSetKernelArg 0 - 3...\n\n");
ciErrNum = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void*)&cmDevSrcA);
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void*)&cmDevSrcB);
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void*)&cmDevDst);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(cl_int), (void*)&iNumElements);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// --------------------------------------------------------
// Core sequence... copy input data to GPU, compute, copy results back
// Asynchronous write of data to GPU device
shrLog("clEnqueueWriteBuffer (SrcA and SrcB)...\n");
ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcA, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize * 4, srcA, 0, NULL, NULL);
ciErrNum |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcB, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize * 4, srcB, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Launch kernel
shrLog("clEnqueueNDRangeKernel (DotProduct)...\n");
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Read back results and check accumulated errors
shrLog("clEnqueueReadBuffer (Dst)...\n\n");
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmDevDst, CL_TRUE, 0, sizeof(cl_float) * szGlobalWorkSize, dst, 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, pCleanup);
// Compute and compare results for golden-host and report errors and pass/fail
shrLog("Comparing against Host/C++ computation...\n\n");
DotProductHost ((const float*)srcA, (const float*)srcB, (float*)Golden, iNumElements);
shrBOOL bMatch = shrComparefet((const float*)Golden, (const float*)dst, (unsigned int)iNumElements, 0.0f, 0);
// Cleanup and leave
Cleanup (EXIT_SUCCESS);
}
// 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);
}
// main function
//*****************************************************************************
int main(int argc, const char **argv)
{
cl_platform_id cpPlatform; // OpenCL platform
cl_uint nDevice; // OpenCL device count
cl_device_id* cdDevices; // OpenCL device list
cl_context cxGPUContext; // OpenCL context
cl_command_queue cqCommandQue[MAX_GPU_COUNT]; // OpenCL command que
cl_int ciErrNum = 1; // Error code var
shrQAStart(argc, (char **)argv);
// 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);
shrSetLogFileName ("oclHiddenMarkovModel.txt");
shrLog("%s Starting...\n\n", argv[0]);
shrLog("Get platform...\n");
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("Get devices...\n");
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &nDevice);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
cdDevices = (cl_device_id *)malloc(nDevice * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, nDevice, cdDevices, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clCreateContext\n");
cxGPUContext = clCreateContext(0, nDevice, cdDevices, NULL, NULL, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clCreateCommandQueue\n");
int id_device;
if(shrGetCmdLineArgumenti(argc, argv, "device", &id_device)) // Set up command queue(s) for GPU specified on the command line
{
// create a command que
cqCommandQue[0] = clCreateCommandQueue(cxGPUContext, cdDevices[id_device], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
oclPrintDevInfo(LOGBOTH, cdDevices[id_device]);
nDevice = 1;
}
else
{ // create command queues for all available devices
for (cl_uint i = 0; i < nDevice; i++)
{
cqCommandQue[i] = clCreateCommandQueue(cxGPUContext, cdDevices[i], CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
}
for (cl_uint i = 0; i < nDevice; i++) oclPrintDevInfo(LOGBOTH, cdDevices[i]);
}
shrLog("\nUsing %d GPU(s)...\n\n", nDevice);
int wgSize;
if (!shrGetCmdLineArgumenti(argc, argv, "work-group-size", &wgSize))
{
wgSize = 256;
}
shrLog("Init Hidden Markov Model parameters\n");
int nState = 256*16; // number of states, must be a multiple of 256
int nEmit = 128; // number of possible observations
float *initProb = (float*)malloc(sizeof(float)*nState); // initial probability
float *mtState = (float*)malloc(sizeof(float)*nState*nState); // state transition matrix
float *mtEmit = (float*)malloc(sizeof(float)*nEmit*nState); // emission matrix
initHMM(initProb, mtState, mtEmit, nState, nEmit);
// define observational sequence
int nObs = 100; // size of observational sequence
int **obs = (int**)malloc(nDevice*sizeof(int*));
int **viterbiPathCPU = (int**)malloc(nDevice*sizeof(int*));
int **viterbiPathGPU = (int**)malloc(nDevice*sizeof(int*));
float *viterbiProbCPU = (float*)malloc(nDevice*sizeof(float));
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
obs[iDevice] = (int*)malloc(sizeof(int)*nObs);
for (int i = 0; i < nObs; i++)
obs[iDevice][i] = i % 15;
viterbiPathCPU[iDevice] = (int*)malloc(sizeof(int)*nObs);
viterbiPathGPU[iDevice] = (int*)malloc(sizeof(int)*nObs);
}
shrLog("# of states = %d\n# of possible observations = %d \nSize of observational sequence = %d\n\n",
nState, nEmit, nObs);
shrLog("Compute Viterbi path on GPU\n\n");
HMM **oHmm = (HMM**)malloc(nDevice*sizeof(HMM*));
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
oHmm[iDevice] = new HMM(cxGPUContext, cqCommandQue[iDevice], initProb, mtState, mtEmit, nState, nEmit, nObs, argv[0], wgSize);
}
cl_mem *vProb = (cl_mem*)malloc(sizeof(cl_mem)*nDevice);
cl_mem *vPath = (cl_mem*)malloc(sizeof(cl_mem)*nDevice);
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
vProb[iDevice] = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(float), NULL, &ciErrNum);
vPath[iDevice] = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(int)*nObs, NULL, &ciErrNum);
}
#ifdef GPU_PROFILING
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clFinish(cqCommandQue[iDevice]);;
}
shrDeltaT(1);
#endif
size_t szWorkGroup;
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
szWorkGroup = oHmm[iDevice]->ViterbiSearch(vProb[iDevice], vPath[iDevice], obs[iDevice]);
}
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clFinish(cqCommandQue[iDevice]);
}
#ifdef GPU_PROFILING
double dElapsedTime = shrDeltaT(1);
shrLogEx(LOGBOTH | MASTER, 0, "oclHiddenMarkovModel, Throughput = %.4f GB/s, Time = %.5f s, Size = %u items, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-9 * 2.0 * sizeof(float) * nDevice * nState * nState * (nObs-1))/dElapsedTime, dElapsedTime, (nDevice * nState * nObs), nDevice, szWorkGroup);
#endif
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
ciErrNum = clEnqueueReadBuffer(cqCommandQue[iDevice], vPath[iDevice], CL_TRUE, 0, sizeof(int)*nObs, viterbiPathGPU[iDevice], 0, NULL, NULL);
}
shrLog("\nCompute Viterbi path on CPU\n");
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
ciErrNum = ViterbiCPU(viterbiProbCPU[iDevice], viterbiPathCPU[iDevice], obs[iDevice], nObs, initProb, mtState, nState, mtEmit);
}
if (!ciErrNum)
{
shrEXIT(argc, argv);
}
bool pass = true;
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
for (int i = 0; i < nObs; i++)
{
if (viterbiPathCPU[iDevice][i] != viterbiPathGPU[iDevice][i])
{
pass = false;
break;
}
}
}
// NOTE: Most properly this should be done at any of the exit points above, but it is omitted elsewhere for clarity.
shrLog("Release CPU buffers and OpenCL objects...\n");
free(initProb);
free(mtState);
free(mtEmit);
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
free(obs[iDevice]);
free(viterbiPathCPU[iDevice]);
free(viterbiPathGPU[iDevice]);
delete oHmm[iDevice];
clReleaseCommandQueue(cqCommandQue[iDevice]);
}
free(obs);
free(viterbiPathCPU);
free(viterbiPathGPU);
free(viterbiProbCPU);
free(cdDevices);
free(oHmm);
clReleaseContext(cxGPUContext);
// finish
shrQAFinishExit(argc, (const char **)argv, pass ? QA_PASSED : QA_FAILED);
shrEXIT(argc, argv);
}
// 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;
}
///////////////////////////////////////////////////////////////////////////////
//Parse args, run the appropriate tests
///////////////////////////////////////////////////////////////////////////////
int runTest(const int argc, const char **argv)
{
int start = DEFAULT_SIZE;
int end = DEFAULT_SIZE;
int startDevice = 0;
int endDevice = 0;
int increment = DEFAULT_INCREMENT;
testMode mode = QUICK_MODE;
bool htod = false;
bool dtoh = false;
bool dtod = false;
char *modeStr;
char *device = NULL;
printMode printmode = USER_READABLE;
char *memModeStr = NULL;
memoryMode memMode = PAGEABLE;
accessMode accMode = DIRECT;
//process command line args
if(shrCheckCmdLineFlag( argc, argv, "help"))
{
printHelp();
return 0;
}
if(shrCheckCmdLineFlag( argc, argv, "csv"))
{
printmode = CSV;
}
// Get host memory mode type from command line
if(shrGetCmdLineArgumentstr(argc, argv, "memory", &memModeStr))
{
if(strcmp(memModeStr, "pageable") == 0 )
{
memMode = PAGEABLE;
}
else if(strcmp(memModeStr, "pinned") == 0)
{
memMode = PINNED;
}
else
{
shrLog("Invalid memory mode - valid modes are pageable or pinned\n");
shrLog("See --help for more information\n");
return -1000;
}
}
else
{
//default - pageable memory
memMode = PAGEABLE;
}
// Access type from command line
if(shrGetCmdLineArgumentstr(argc, argv, "access", &memModeStr))
{
if(strcmp(memModeStr, "direct") == 0)
{
accMode = DIRECT;
}
else if(strcmp(memModeStr, "mapped") == 0)
{
accMode = MAPPED;
}
else
{
shrLog("Invalid access mode - valid modes are direct or mapped\n");
shrLog("See --help for more information\n");
return -2000;
}
}
else
{
//default - direct
accMode = DIRECT;
}
// Get OpenCL platform ID for NVIDIA if available, otherwise default
cl_platform_id clSelectedPlatformID = NULL;
cl_int ciErrNum = oclGetPlatformID (&clSelectedPlatformID);
oclCheckError(ciErrNum, CL_SUCCESS);
// Find out how many devices there are
cl_uint ciDeviceCount;
ciErrNum = clGetDeviceIDs (clSelectedPlatformID, CL_DEVICE_TYPE_GPU, 0, NULL, &ciDeviceCount);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clGetDeviceIDs call !!!\n\n", ciErrNum);
return ciErrNum;
}
else if (ciDeviceCount == 0)
{
shrLog(" There are no devices supporting OpenCL (return code %i)\n\n", ciErrNum);
return ciErrNum;
}
// Get command line device options and config accordingly
if(shrGetCmdLineArgumentstr(argc, argv, "device", &device))
{
if(strcmp (device, "all") == 0)
{
shrLog("\n!!!Cumulative Bandwidth to be computed from all the devices !!!\n\n");
startDevice = 0;
endDevice = (int)(ciDeviceCount-1);
}
else
{
startDevice = endDevice = atoi(device);
if(startDevice < 0 || ((size_t)startDevice) >= ciDeviceCount)
{
shrLog("\n!!!Invalid GPU number %d given hence default gpu %d will be used !!!\n", startDevice,0);
startDevice = endDevice = 0;
}
}
}
// Get and log the device info
shrLog("Running on...\n\n");
devices = (cl_device_id*) malloc(sizeof(cl_device_id) * ciDeviceCount);
ciErrNum = clGetDeviceIDs (clSelectedPlatformID, CL_DEVICE_TYPE_GPU, ciDeviceCount, devices, &ciDeviceCount);
for(int currentDevice = startDevice; currentDevice <= endDevice; currentDevice++)
{
oclPrintDevName(LOGBOTH, devices[currentDevice]);
shrLog("\n");
}
shrLog("\n");
// Get command line mode(s) and config accordingly
if(shrGetCmdLineArgumentstr(argc, argv, "mode", &modeStr))
{
//figure out the mode
if(strcmp(modeStr, "quick") == 0)
{
shrLog("Quick Mode\n\n");
mode = QUICK_MODE;
}
else if(strcmp(modeStr, "shmoo") == 0)
{
shrLog("Shmoo Mode\n\n");
mode = SHMOO_MODE;
}
else if(strcmp(modeStr, "range") == 0)
{
shrLog("Range Mode\n\n");
mode = RANGE_MODE;
}
else
{
shrLog("Invalid mode - valid modes are quick, range, or shmoo\n");
shrLog("See --help for more information\n\n");
return -3000;
}
}
else
{
//default mode - quick
shrLog("Quick Mode\n\n");
mode = QUICK_MODE;
}
if(shrCheckCmdLineFlag(argc, argv, "htod"))
htod = true;
if(shrCheckCmdLineFlag(argc, argv, "dtoh"))
dtoh = true;
if(shrCheckCmdLineFlag(argc, argv, "dtod"))
dtod = true;
if(!htod && !dtoh && !dtod)
{
//default: All
htod = true;
dtoh = true;
dtod = true;
}
if(RANGE_MODE == mode)
{
if(shrGetCmdLineArgumenti( argc, argv, "start", &start))
{
if( start <= 0 )
{
shrLog("Illegal argument - start must be greater than zero\n");
return -4000;
}
}
else
{
shrLog("Must specify a starting size in range mode\n");
shrLog("See --help for more information\n");
return -5000;
}
if(shrGetCmdLineArgumenti( argc, argv, "end", &end))
{
if(end <= 0)
{
shrLog("Illegal argument - end must be greater than zero\n");
return -6000;
}
if(start > end)
{
shrLog("Illegal argument - start is greater than end\n");
return -7000;
}
}
else
{
shrLog("Must specify an end size in range mode.\n");
shrLog("See --help for more information\n");
return -8000;
}
if(shrGetCmdLineArgumenti( argc, argv, "increment", &increment))
{
if(increment <= 0)
{
shrLog("Illegal argument - increment must be greater than zero\n");
return -9000;
}
}
else
{
shrLog("Must specify an increment in user mode\n");
shrLog("See --help for more information\n");
return -10000;
}
}
// Create the OpenCL context
cxGPUContext = clCreateContext(0, ciDeviceCount, devices, NULL, NULL, NULL);
if (cxGPUContext == (cl_context)0)
{
shrLog("Failed to create OpenCL context!\n");
return -11000;
}
// Run tests
if(htod)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, HOST_TO_DEVICE, printmode, accMode, memMode, startDevice, endDevice);
}
if(dtoh)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, DEVICE_TO_HOST, printmode, accMode, memMode, startDevice, endDevice);
}
if(dtod)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, DEVICE_TO_DEVICE, printmode, accMode, memMode, startDevice, endDevice);
}
// Clean up
free(memModeStr);
if(cqCommandQueue)clReleaseCommandQueue(cqCommandQueue);
if(cxGPUContext)clReleaseContext(cxGPUContext);
if(devices)free(devices);
return 0;
}
GPUBase::GPUBase(char* source, char* KernelName)
{
kernelFuncName = KernelName;
size_t szKernelLength;
size_t szKernelLengthFilter;
size_t szKernelLengthSum;
char* SourceOpenCLShared;
char* SourceOpenCL;
iBlockDimX = 16;
iBlockDimY = 16;
GPUError = oclGetPlatformID(&cpPlatform);
CheckError(GPUError);
cl_uint uiNumAllDevs = 0;
// Get the number of GPU devices available to the platform
GPUError = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumAllDevs);
CheckError(GPUError);
uiDevCount = uiNumAllDevs;
// Create the device list
cdDevices = new cl_device_id [uiDevCount];
GPUError = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiDevCount, cdDevices, NULL);
CheckError(GPUError);
// Create the OpenCL context on a GPU device
GPUContext = clCreateContext(0, uiNumAllDevs, cdDevices, NULL, NULL, &GPUError);
CheckError(GPUError);
//The command-queue can be used to queue a set of operations (referred to as commands) in order.
GPUCommandQueue = clCreateCommandQueue(GPUContext, cdDevices[0], 0, &GPUError);
CheckError(GPUError);
oclPrintDevName(LOGBOTH, cdDevices[0]);
// Load OpenCL kernel
SourceOpenCLShared = oclLoadProgSource("C:\\Dropbox\\MGR\\GPUFeatureExtraction\\GPU\\OpenCL\\GPUCode.cl", "// My comment\n", &szKernelLength);
SourceOpenCL = oclLoadProgSource(source, "// My comment\n", &szKernelLengthFilter);
szKernelLengthSum = szKernelLength + szKernelLengthFilter;
char* sourceCL = new char[szKernelLengthSum];
strcpy(sourceCL,SourceOpenCLShared);
strcat (sourceCL, SourceOpenCL);
GPUProgram = clCreateProgramWithSource( GPUContext , 1, (const char **)&sourceCL, &szKernelLengthSum, &GPUError);
CheckError(GPUError);
// Build the program with 'mad' Optimization option
char *flags = "-cl-unsafe-math-optimizations -cl-fast-relaxed-math -cl-mad-enable";
GPUError = clBuildProgram(GPUProgram, 0, NULL, flags, NULL, NULL);
//error checking code
if(!GPUError)
{
//print kernel compilation error
char programLog[1024];
clGetProgramBuildInfo(GPUProgram, cdDevices[0], CL_PROGRAM_BUILD_LOG, 1024, programLog, 0);
cout<<programLog<<endl;
}
cout << kernelFuncName << endl;
GPUKernel = clCreateKernel(GPUProgram, kernelFuncName, &GPUError);
CheckError(GPUError);
}
// Main function
// *********************************************************************
int main( int argc, const char** argv)
{
shrQAStart(argc, (char **)argv);
// start logs
shrSetLogFileName ("oclReduction.txt");
shrLog("%s Starting...\n\n", argv[0]);
char *typeChoice;
shrGetCmdLineArgumentstr(argc, argv, "type", &typeChoice);
// determine type of array from command line args
if (0 == typeChoice)
{
typeChoice = (char*)malloc(7 * sizeof(char));
#ifdef WIN32
strcpy_s(typeChoice, 7 * sizeof(char) + 1, "int");
#else
strcpy(typeChoice, "int");
#endif
}
ReduceType datatype = REDUCE_INT;
#ifdef WIN32
if (!_strcmpi(typeChoice, "float"))
datatype = REDUCE_FLOAT;
else if (!_strcmpi(typeChoice, "double"))
datatype = REDUCE_DOUBLE;
else
datatype = REDUCE_INT;
#else
if (!strcmp(typeChoice, "float"))
datatype = REDUCE_FLOAT;
else if (!strcmp(typeChoice, "double"))
datatype = REDUCE_DOUBLE;
else
datatype = REDUCE_INT;
#endif
shrLog("Reducing array of type %s.\n", typeChoice);
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckError(ciErrNum, CL_SUCCESS);
//Get the devices
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &uiNumDevices);
oclCheckError(ciErrNum, CL_SUCCESS);
cl_device_id *cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, uiNumDevices, cdDevices, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
//Create the context
cxGPUContext = clCreateContext(0, uiNumDevices, cdDevices, NULL, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// get and log the device info
if( shrCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
int device_nr = 0;
shrGetCmdLineArgumenti(argc, (const char**)argv, "device", &device_nr);
if( device_nr < uiNumDevices ) {
device = oclGetDev(cxGPUContext, device_nr);
} else {
shrLog("Invalid Device %d Requested.\n", device_nr);
shrExitEX(argc, argv, EXIT_FAILURE);
}
} else {
device = oclGetMaxFlopsDev(cxGPUContext);
}
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, device, 0, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
source_path = shrFindFilePath("oclReduction_kernel.cl", argv[0]);
bool bSuccess = false;
switch (datatype)
{
default:
case REDUCE_INT:
bSuccess = runTest<int>( argc, argv, datatype);
break;
case REDUCE_FLOAT:
bSuccess = runTest<float>( argc, argv, datatype);
break;
}
// finish
shrQAFinishExit(argc, (const char **)argv, bSuccess ? QA_PASSED : QA_FAILED);
}
