// Main program
//*****************************************************************************
int main(int argc, char** argv)
{
numParticles = NUM_PARTICLES;
uint gridDim = GRID_SIZE;
pArgc = &argc;
pArgv = argv;
shrQAStart(argc, argv);
// Start logs and timers
cExecutableName = argv[0];
shrSetLogFileName ("oclParticles.txt");
shrLog("%s Starting...\n\n", argv[0]);
// check command line flags and parameters
if (argc > 1)
{
shrGetCmdLineArgumenti(argc, (const char**)argv, "n", (int*)&numParticles);
shrGetCmdLineArgumenti(argc, (const char**)argv, "grid", (int*)&gridDim);
bQATest = shrCheckCmdLineFlag(argc, (const char**)argv, "qatest");
bNoPrompt = shrCheckCmdLineFlag(argc, (const char**)argv, "noprompt");
}
// Set and log grid size and particle count, after checking optional command-line inputs
gridSize.x = gridSize.y = gridSize.z = gridDim;
shrLog(" grid: %d x %d x %d = %d cells\n", gridSize.x, gridSize.y, gridSize.z, gridSize.x * gridSize.y * gridSize.z);
shrLog(" particles: %d\n\n", numParticles);
// initialize GLUT and GLEW
if(!bQATest)
{
InitGL(&argc, argv);
}
// initialize OpenCL
startupOpenCL(argc,(const char**)argv);
// init simulation parameters and objects
initParticleSystem(numParticles, gridSize);
initParams();
// Init timers
shrDeltaT(0); // timer 0 is for processing time measurements
shrDeltaT(1); // timer 1 is for fps measurement
// Start main GLUT rendering loop for processing and rendering,
// or otherwise run No-GL Q/A test sequence
if(!bQATest)
{
glutMainLoop();
}
else
{
TestNoGL();
}
// Normally unused return path
shrQAFinish(argc, (const char **)argv, QA_PASSED);
Cleanup(EXIT_SUCCESS);
}
////////////////////////////////////////////////////////////////////////////////
//! Run a simple test for
////////////////////////////////////////////////////////////////////////////////
int runTest(int argc, const char** argv)
{
cl_platform_id cpPlatform = NULL;
cl_uint ciDeviceCount = 0;
cl_device_id *cdDevices = NULL;
cl_int ciErrNum = CL_SUCCESS;
//Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create OpenCL context!\n");
return ciErrNum;
}
//Get the devices
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 0, NULL, &ciDeviceCount);
cdDevices = (cl_device_id *)malloc(ciDeviceCount * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, ciDeviceCount, cdDevices, NULL);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create OpenCL context!\n");
return ciErrNum;
}
//Create the context
cxGPUContext = clCreateContext(0, ciDeviceCount, cdDevices, NULL, NULL, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create OpenCL context!\n");
return ciErrNum;
}
if(shrCheckCmdLineFlag(argc, (const char**)argv, "device"))
{
// User specified GPUs
char* deviceList;
char* deviceStr;
char* next_token;
shrGetCmdLineArgumentstr(argc, (const char**)argv, "device", &deviceList);
#ifdef WIN32
deviceStr = strtok_s (deviceList," ,.-", &next_token);
#else
deviceStr = strtok (deviceList," ,.-");
#endif
ciDeviceCount = 0;
while(deviceStr != NULL)
{
// get and print the device for this queue
cl_device_id device = oclGetDev(cxGPUContext, atoi(deviceStr));
if( device == (cl_device_id) -1 ) {
shrLog(" Device %s does not exist!\n", deviceStr);
return -1;
}
shrLog("Device %s: ", deviceStr);
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create command queue
commandQueue[ciDeviceCount] = clCreateCommandQueue(cxGPUContext, device, CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clCreateCommandQueue call !!!\n\n", ciErrNum);
return ciErrNum;
}
++ciDeviceCount;
#ifdef WIN32
deviceStr = strtok_s (NULL," ,.-", &next_token);
#else
deviceStr = strtok (NULL," ,.-");
#endif
}
free(deviceList);
}
else
{
// Find out how many GPU's to compute on all available GPUs
size_t nDeviceBytes;
ciErrNum |= clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &nDeviceBytes);
ciDeviceCount = (cl_uint)nDeviceBytes/sizeof(cl_device_id);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clGetDeviceIDs call !!!\n\n", ciErrNum);
return ciErrNum;
}
else if (ciDeviceCount == 0)
{
shrLog(" There are no devices supporting OpenCL (return code %i)\n\n", ciErrNum);
return -1;
}
// create command-queues
for(unsigned int i = 0; i < ciDeviceCount; ++i)
{
// get and print the device for this queue
cl_device_id device = oclGetDev(cxGPUContext, i);
shrLog("Device %d: ", i);
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create command queue
commandQueue[i] = clCreateCommandQueue(cxGPUContext, device, CL_QUEUE_PROFILING_ENABLE, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clCreateCommandQueue call !!!\n\n", ciErrNum);
return ciErrNum;
}
}
}
// Optional Command-line multiplier for matrix sizes
shrGetCmdLineArgumenti(argc, (const char**)argv, "sizemult", &iSizeMultiple);
iSizeMultiple = CLAMP(iSizeMultiple, 1, 10);
uiWA = WA * iSizeMultiple;
uiHA = HA * iSizeMultiple;
uiWB = WB * iSizeMultiple;
uiHB = HB * iSizeMultiple;
uiWC = WC * iSizeMultiple;
uiHC = HC * iSizeMultiple;
shrLog("\nUsing Matrix Sizes: A(%u x %u), B(%u x %u), C(%u x %u)\n",
uiWA, uiHA, uiWB, uiHB, uiWC, uiHC);
// allocate host memory for matrices A and B
unsigned int size_A = uiWA * uiHA;
unsigned int mem_size_A = sizeof(float) * size_A;
float* h_A_data = (float*)malloc(mem_size_A);
unsigned int size_B = uiWB * uiHB;
unsigned int mem_size_B = sizeof(float) * size_B;
float* h_B_data = (float*)malloc(mem_size_B);
// initialize host memory
srand(2006);
shrFillArray(h_A_data, size_A);
shrFillArray(h_B_data, size_B);
// allocate host memory for result
unsigned int size_C = uiWC * uiHC;
unsigned int mem_size_C = sizeof(float) * size_C;
float* h_C = (float*) malloc(mem_size_C);
// create OpenCL buffer pointing to the host memory
cl_mem h_A = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR,
mem_size_A, h_A_data, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: clCreateBuffer\n");
return ciErrNum;
}
// Program Setup
size_t program_length;
const char* header_path = shrFindFilePath("matrixMul.h", argv[0]);
oclCheckError(header_path != NULL, shrTRUE);
char* header = oclLoadProgSource(header_path, "", &program_length);
if(!header)
{
shrLog("Error: Failed to load the header %s!\n", header_path);
return -1000;
}
const char* source_path = shrFindFilePath("matrixMul.cl", argv[0]);
oclCheckError(source_path != NULL, shrTRUE);
char *source = oclLoadProgSource(source_path, header, &program_length);
if(!source)
{
shrLog("Error: Failed to load compute program %s!\n", source_path);
return -2000;
}
// create the program
cl_program cpProgram = clCreateProgramWithSource(cxGPUContext, 1, (const char **)&source,
&program_length, &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create program\n");
return ciErrNum;
}
free(header);
free(source);
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-fast-relaxed-math", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then return error
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclMatrixMul.ptx");
return ciErrNum;
}
// write out PTX if requested on the command line
if(shrCheckCmdLineFlag(argc, argv, "dump-ptx") )
{
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclMatrixMul.ptx");
}
// Create Kernel
for(unsigned int i = 0; i < ciDeviceCount; ++i) {
multiplicationKernel[i] = clCreateKernel(cpProgram, "matrixMul", &ciErrNum);
if (ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failed to create kernel\n");
return ciErrNum;
}
}
// Run multiplication on 1..deviceCount GPUs to compare improvement
shrLog("\nRunning Computations on 1 - %d GPU's...\n\n", ciDeviceCount);
for(unsigned int k = 1; k <= ciDeviceCount; ++k)
{
matrixMulGPU(k, h_A, h_B_data, mem_size_B, h_C);
}
// compute reference solution
shrLog("Comparing results with CPU computation... \n\n");
float* reference = (float*) malloc(mem_size_C);
computeGold(reference, h_A_data, h_B_data, uiHA, uiWA, uiWB);
// check result
shrBOOL res = shrCompareL2fe(reference, h_C, size_C, 1.0e-6f);
if (res != shrTRUE)
{
printDiff(reference, h_C, uiWC, uiHC, 100, 1.0e-5f);
}
// clean up OCL resources
ciErrNum = clReleaseMemObject(h_A);
for(unsigned int k = 0; k < ciDeviceCount; ++k)
{
ciErrNum |= clReleaseKernel( multiplicationKernel[k] );
ciErrNum |= clReleaseCommandQueue( commandQueue[k] );
}
ciErrNum |= clReleaseProgram(cpProgram);
ciErrNum |= clReleaseContext(cxGPUContext);
if(ciErrNum != CL_SUCCESS)
{
shrLog("Error: Failure releasing OpenCL resources: %d\n", ciErrNum);
return ciErrNum;
}
// clean up memory
free(h_A_data);
free(h_B_data);
free(h_C);
free(reference);
return ((shrTRUE == res) ? CL_SUCCESS : -3000);
}
bool fdtdGPU(float *output, const float *input, const float *coeff, const int dimx, const int dimy, const int dimz, const int radius, const int timesteps, const int argc, const char **argv)
{
bool ok = true;
const int outerDimx = dimx + 2 * radius;
const int outerDimy = dimy + 2 * radius;
const int outerDimz = dimz + 2 * radius;
const size_t volumeSize = outerDimx * outerDimy * outerDimz;
cl_context context = 0;
cl_platform_id platform = 0;
cl_device_id *devices = 0;
cl_command_queue commandQueue = 0;
cl_mem bufferOut = 0;
cl_mem bufferIn = 0;
cl_mem bufferCoeff = 0;
cl_program program = 0;
cl_kernel kernel = 0;
cl_event *kernelEvents = 0;
#ifdef GPU_PROFILING
cl_ulong kernelEventStart;
cl_ulong kernelEventEnd;
#endif
double hostElapsedTimeS;
char *cPathAndName = 0;
char *cSourceCL = 0;
size_t szKernelLength;
size_t globalWorkSize[2];
size_t localWorkSize[2];
cl_uint deviceCount = 0;
cl_uint targetDevice = 0;
cl_int errnum = 0;
char buildOptions[128];
// Ensure that the inner data starts on a 128B boundary
const int padding = (128 / sizeof(float)) - radius;
const size_t paddedVolumeSize = volumeSize + padding;
#ifdef GPU_PROFILING
const int profileTimesteps = timesteps - 1;
if (ok)
{
if (profileTimesteps < 1)
{
shrLog(" cannot profile with fewer than two timesteps (timesteps=%d), profiling is disabled.\n", timesteps);
}
}
#endif
// Get the NVIDIA platform
if (ok)
{
shrLog(" oclGetPlatformID...\n");
errnum = oclGetPlatformID(&platform);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("oclGetPlatformID (returned %d).\n", errnum);
ok = false;
}
}
// Get the list of GPU devices associated with the platform
if (ok)
{
shrLog(" clGetDeviceIDs");
errnum = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &deviceCount);
devices = (cl_device_id *)malloc(deviceCount * sizeof(cl_device_id) );
errnum = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, deviceCount, devices, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clGetDeviceIDs (returned %d).\n", errnum);
ok = false;
}
}
// Create the OpenCL context
if (ok)
{
shrLog(" clCreateContext...\n");
context = clCreateContext(0, deviceCount, devices, NULL, NULL, &errnum);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateContext (returned %d).\n", errnum);
ok = false;
}
}
// Select target device (device 0 by default)
if (ok)
{
char *device = 0;
if (shrGetCmdLineArgumentstr(argc, argv, "device", &device))
{
targetDevice = (cl_uint)atoi(device);
if (targetDevice >= deviceCount)
{
shrLogEx(LOGBOTH | ERRORMSG, -2001, STDERROR);
shrLog("invalid target device specified on command line (device %d does not exist).\n", targetDevice);
ok = false;
}
}
else
{
targetDevice = 0;
}
if (device)
{
free(device);
}
}
// Create a command-queue
if (ok)
{
shrLog(" clCreateCommandQueue\n");
commandQueue = clCreateCommandQueue(context, devices[targetDevice], CL_QUEUE_PROFILING_ENABLE, &errnum);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateCommandQueue (returned %d).\n", errnum);
ok = false;
}
}
// Create memory buffer objects
if (ok)
{
shrLog(" clCreateBuffer bufferOut\n");
bufferOut = clCreateBuffer(context, CL_MEM_READ_WRITE, paddedVolumeSize * sizeof(float), NULL, &errnum);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateBuffer (returned %d).\n", errnum);
ok = false;
}
}
if (ok)
{
shrLog(" clCreateBuffer bufferIn\n");
bufferIn = clCreateBuffer(context, CL_MEM_READ_WRITE, paddedVolumeSize * sizeof(float), NULL, &errnum);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateBuffer (returned %d).\n", errnum);
ok = false;
}
}
if (ok)
{
shrLog(" clCreateBuffer bufferCoeff\n");
bufferCoeff = clCreateBuffer(context, CL_MEM_READ_ONLY, (radius + 1) * sizeof(float), NULL, &errnum);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateBuffer (returned %d).\n", errnum);
ok = false;
}
}
// Load the kernel from file
if (ok)
{
shrLog(" shrFindFilePath\n");
cPathAndName = shrFindFilePath(clSourceFile, argv[0]);
if (cPathAndName == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -2002, STDERROR);
shrLog("shrFindFilePath returned null.\n");
ok = false;
}
}
if (ok)
{
shrLog(" oclLoadProgSource\n");
cSourceCL = oclLoadProgSource(cPathAndName, "// Preamble\n", &szKernelLength);
if (cSourceCL == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -2003, STDERROR);
shrLog("oclLoadProgSource returned null.\n");
ok = false;
}
}
// Create the program
if (ok)
{
shrLog(" clCreateProgramWithSource\n");
program = clCreateProgramWithSource(context, 1, (const char **)&cSourceCL, &szKernelLength, &errnum);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateProgramWithSource (returned %d).\n", errnum);
ok = false;
}
}
// Check for a command-line specified work group size
size_t userWorkSize;
int localWorkMaxY;
if (ok)
{
int userWorkSizeInt;
if (shrGetCmdLineArgumenti(argc, argv, "work-group-size", &userWorkSizeInt))
{
// We can't clamp to CL_KERNEL_WORK_GROUP_SIZE yet since that is
// dependent on the build.
if (userWorkSizeInt < k_localWorkMin || userWorkSizeInt > k_localWorkMax)
{
shrLogEx(LOGBOTH | ERRORMSG, -2004, STDERROR);
shrLog("invalid work group size specified on command line (must be between %d and %d).\n", k_localWorkMin, k_localWorkMax);
ok = false;
}
// Constrain to a multiple of k_localWorkX
userWorkSize = (userWorkSizeInt / k_localWorkX * k_localWorkX);
}
else
{
userWorkSize = k_localWorkY * k_localWorkX;
}
// Divide by k_localWorkX (integer division to clamp)
localWorkMaxY = userWorkSize / k_localWorkX;
}
// Build the program
if (ok)
{
#ifdef WIN32
if (sprintf_s(buildOptions, sizeof(buildOptions), "-DRADIUS=%d -DMAXWORKX=%d -DMAXWORKY=%d -cl-fast-relaxed-math", radius, k_localWorkX, localWorkMaxY) < 0)
{
shrLogEx(LOGBOTH | ERRORMSG, -2005, STDERROR);
shrLog("sprintf_s (failed).\n");
ok = false;
}
#else
if (snprintf(buildOptions, sizeof(buildOptions), "-DRADIUS=%d -DMAXWORKX=%d -DMAXWORKY=%d -cl-fast-relaxed-math", radius, k_localWorkX, localWorkMaxY) < 0)
{
shrLogEx(LOGBOTH | ERRORMSG, -2005, STDERROR);
shrLog("snprintf (failed).\n");
ok = false;
}
#endif
}
if (ok)
{
shrLog(" clBuildProgram (%s)\n", buildOptions);
errnum = clBuildProgram(program, 0, NULL, buildOptions, NULL, NULL);
if (errnum != CL_SUCCESS)
{
char buildLog[10240];
clGetProgramBuildInfo(program, devices[0], CL_PROGRAM_BUILD_LOG, sizeof(buildLog), buildLog, NULL);
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clBuildProgram (returned %d).\n", errnum);
shrLog("Log:\n%s\n", buildLog);
ok = false;
}
}
// Create the kernel
if (ok)
{
shrLog(" clCreateKernel\n");
kernel = clCreateKernel(program, "FiniteDifferences", &errnum);
if (kernel == (cl_kernel)NULL || errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clCreateKernel (returned %d).\n", errnum);
ok = false;
}
}
// Get the maximum work group size
size_t maxWorkSize;
if (ok)
{
shrLog(" clGetKernelWorkGroupInfo\n");
errnum = clGetKernelWorkGroupInfo(kernel, devices[targetDevice], CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &maxWorkSize, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clGetKernelWorkGroupInfo (returned %d).\n", errnum);
ok = false;
}
}
// Set the work group size
if (ok)
{
userWorkSize = CLAMP(userWorkSize, k_localWorkMin, maxWorkSize);
localWorkSize[0] = k_localWorkX;
localWorkSize[1] = userWorkSize / k_localWorkX;
globalWorkSize[0] = localWorkSize[0] * (unsigned int)ceil((float)dimx / localWorkSize[0]);
globalWorkSize[1] = localWorkSize[1] * (unsigned int)ceil((float)dimy / localWorkSize[1]);
shrLog(" set local work group size to %dx%d\n", localWorkSize[0], localWorkSize[1]);
shrLog(" set total work size to %dx%d\n", globalWorkSize[0], globalWorkSize[1]);
}
// Copy the input to the device input buffer
if (ok)
{
shrLog(" clEnqueueWriteBuffer bufferIn\n");
errnum = clEnqueueWriteBuffer(commandQueue, bufferIn, CL_TRUE, padding * sizeof(float), volumeSize * sizeof(float), input, 0, NULL, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clEnqueueWriteBuffer bufferIn (returned %d).\n", errnum);
ok = false;
}
}
// Copy the input to the device output buffer (actually only need the halo)
if (ok)
{
shrLog(" clEnqueueWriteBuffer bufferOut\n");
errnum = clEnqueueWriteBuffer(commandQueue, bufferOut, CL_TRUE, padding * sizeof(float), volumeSize * sizeof(float), input, 0, NULL, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clEnqueueWriteBuffer bufferOut (returned %d).\n", errnum);
ok = false;
}
}
// Copy the coefficients to the device coefficient buffer
if (ok)
{
shrLog(" clEnqueueWriteBuffer bufferCoeff\n");
errnum = clEnqueueWriteBuffer(commandQueue, bufferCoeff, CL_TRUE, 0, (radius + 1) * sizeof(float), coeff, 0, NULL, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clEnqueueWriteBuffer bufferCoeff (returned %d).\n", errnum);
ok = false;
}
}
// Allocate the events
if (ok)
{
shrLog(" calloc events\n");
if ((kernelEvents = (cl_event *)calloc(timesteps, sizeof(cl_event))) == NULL)
{
shrLogEx(LOGBOTH | ERRORMSG, -2006, STDERROR);
shrLog("Insufficient memory for events calloc, please try a smaller volume (use --help for syntax).\n");
ok = false;
}
}
// Start the clock
shrDeltaT(0);
// Set the constant arguments
if (ok)
{
shrLog(" clSetKernelArg 2-6\n");
errnum = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&bufferCoeff);
errnum |= clSetKernelArg(kernel, 3, sizeof(int), &dimx);
errnum |= clSetKernelArg(kernel, 4, sizeof(int), &dimy);
errnum |= clSetKernelArg(kernel, 5, sizeof(int), &dimz);
errnum |= clSetKernelArg(kernel, 6, sizeof(int), &padding);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clSetKernelArg 2-6 (returned %d).\n", errnum);
ok = false;
}
}
// Execute the FDTD
cl_mem bufferSrc = bufferIn;
cl_mem bufferDst = bufferOut;
if (ok)
{
shrLog(" GPU FDTD loop\n");
}
for (int it = 0 ; ok && it < timesteps ; it++)
{
shrLog("\tt = %d ", it);
// Set the dynamic arguments
if (ok)
{
shrLog(" clSetKernelArg 0-1,");
errnum = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&bufferDst);
errnum |= clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&bufferSrc);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clSetKernelArg 0-1 (returned %d).\n", errnum);
ok = false;
}
}
// Launch the kernel
if (ok)
{
shrLog(" clEnqueueNDRangeKernel\n");
errnum = clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL, globalWorkSize, localWorkSize, 0, NULL, &kernelEvents[it]);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clEnqueueNDRangeKernel (returned %d).\n", errnum);
ok = false;
}
}
// Toggle the buffers
cl_mem tmp = bufferSrc;
bufferSrc = bufferDst;
bufferDst = tmp;
}
if (ok)
shrLog("\n");
// Wait for the kernel to complete
if (ok)
{
shrLog(" clWaitForEvents\n");
errnum = clWaitForEvents(1, &kernelEvents[timesteps-1]);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clWaitForEvents (returned %d).\n", errnum);
ok = false;
}
}
// Stop the clock
hostElapsedTimeS = shrDeltaT(0);
// Read the result back, result is in bufferSrc (after final toggle)
if (ok)
{
shrLog(" clEnqueueReadBuffer\n");
errnum = clEnqueueReadBuffer(commandQueue, bufferSrc, CL_TRUE, padding * sizeof(float), volumeSize * sizeof(float), output, 0, NULL, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clEnqueueReadBuffer bufferSrc (returned %d).\n", errnum);
ok = false;
}
}
// Report time
#ifdef GPU_PROFILING
double elapsedTime = 0.0;
if (ok && profileTimesteps > 0)
shrLog(" Collect profile information\n");
for (int it = 1 ; ok && it <= profileTimesteps ; it++)
{
shrLog("\tt = %d ", it);
shrLog(" clGetEventProfilingInfo,", it);
errnum = clGetEventProfilingInfo(kernelEvents[it], CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &kernelEventStart, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clGetEventProfilingInfo (returned %d).\n", errnum);
ok = false;
}
shrLog(" clGetEventProfilingInfo\n", it);
errnum = clGetEventProfilingInfo(kernelEvents[it], CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &kernelEventEnd, NULL);
if (errnum != CL_SUCCESS)
{
shrLogEx(LOGBOTH | ERRORMSG, errnum, STDERROR);
shrLog("clGetEventProfilingInfo (returned %d).\n", errnum);
ok = false;
}
elapsedTime += (double)kernelEventEnd - (double)kernelEventStart;
}
if (ok && profileTimesteps > 0)
{
shrLog("\n");
// Convert nanoseconds to seconds
elapsedTime *= 1.0e-9;
double avgElapsedTime = elapsedTime / (double)profileTimesteps;
// Determine number of computations per timestep
size_t pointsComputed = dimx * dimy * dimz;
// Determine throughput
double throughputM = 1.0e-6 * (double)pointsComputed / avgElapsedTime;
shrLogEx(LOGBOTH | MASTER, 0, "oclFDTD3d, Throughput = %.4f MPoints/s, Time = %.5f s, Size = %u Points, NumDevsUsed = %i, Workgroup = %u\n",
throughputM, avgElapsedTime, pointsComputed, 1, localWorkSize[0] * localWorkSize[1]);
}
#endif
// Cleanup
if (kernelEvents)
{
for (int it = 0 ; it < timesteps ; it++)
{
if (kernelEvents[it])
clReleaseEvent(kernelEvents[it]);
}
free(kernelEvents);
}
if (kernel)
clReleaseKernel(kernel);
if (program)
clReleaseProgram(program);
if (cSourceCL)
free(cSourceCL);
if (cPathAndName)
free(cPathAndName);
if (bufferCoeff)
clReleaseMemObject(bufferCoeff);
if (bufferIn)
clReleaseMemObject(bufferIn);
if (bufferOut)
clReleaseMemObject(bufferOut);
if (commandQueue)
clReleaseCommandQueue(commandQueue);
if (devices)
free(devices);
if (context)
clReleaseContext(context);
return 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 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, char** argv)
{
shrQAStart(argc, argv);
int use_gpu = 0;
for(int i = 0; i < argc && argv; i++)
{
if(!argv[i])
continue;
if(strstr(argv[i], "cpu"))
use_gpu = 0;
else if(strstr(argv[i], "gpu"))
use_gpu = 1;
}
// start logs
shrSetLogFileName ("oclDXTCompression.txt");
shrLog("%s Starting...\n\n", argv[0]);
cl_platform_id cpPlatform = NULL;
cl_uint uiNumDevices = 0;
cl_device_id *cdDevices = NULL;
cl_context cxGPUContext;
cl_command_queue cqCommandQueue;
cl_program cpProgram;
cl_kernel ckKernel;
cl_mem cmMemObjs[3];
cl_mem cmAlphaTable4, cmProds4;
cl_mem cmAlphaTable3, cmProds3;
size_t szGlobalWorkSize[1];
size_t szLocalWorkSize[1];
cl_int ciErrNum;
// Get the path of the filename
char *filename;
if (shrGetCmdLineArgumentstr(argc, (const char **)argv, "image", &filename)) {
image_filename = filename;
}
// load image
const char* image_path = shrFindFilePath(image_filename, argv[0]);
oclCheckError(image_path != NULL, shrTRUE);
shrLoadPPM4ub(image_path, (unsigned char **)&h_img, &width, &height);
oclCheckError(h_img != NULL, shrTRUE);
shrLog("Loaded '%s', %d x %d pixels\n\n", image_path, width, height);
// Convert linear image to block linear.
const uint memSize = width * height * sizeof(cl_uint);
uint* block_image = (uint*)malloc(memSize);
// Convert linear image to block linear.
for(uint by = 0; by < height/4; by++) {
for(uint bx = 0; bx < width/4; bx++) {
for (int i = 0; i < 16; i++) {
const int x = i & 3;
const int y = i / 4;
block_image[(by * width/4 + bx) * 16 + i] =
((uint *)h_img)[(by * 4 + y) * 4 * (width/4) + bx * 4 + x];
}
}
}
// Get the NVIDIA platform
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckError(ciErrNum, CL_SUCCESS);
// Get the platform's GPU devices
ciErrNum = clGetDeviceIDs(cpPlatform, use_gpu?CL_DEVICE_TYPE_GPU:CL_DEVICE_TYPE_CPU, 0, NULL, &uiNumDevices);
oclCheckError(ciErrNum, CL_SUCCESS);
cdDevices = (cl_device_id *)malloc(uiNumDevices * sizeof(cl_device_id) );
ciErrNum = clGetDeviceIDs(cpPlatform, use_gpu?CL_DEVICE_TYPE_GPU:CL_DEVICE_TYPE_CPU, uiNumDevices, cdDevices, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
// Create the context
cxGPUContext = clCreateContext(0, uiNumDevices, cdDevices, NULL, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// get and log device
cl_device_id device;
if( shrCheckCmdLineFlag(argc, (const char **)argv, "device") ) {
int device_nr = 0;
shrGetCmdLineArgumenti(argc, (const char **)argv, "device", &device_nr);
device = oclGetDev(cxGPUContext, device_nr);
if( device == (cl_device_id)-1 ) {
shrLog(" Invalid GPU Device: devID=%d. %d valid GPU devices detected\n\n", device_nr, uiNumDevices);
shrLog(" exiting...\n");
return -1;
}
} else {
device = oclGetMaxFlopsDev(cxGPUContext);
}
oclPrintDevName(LOGBOTH, device);
shrLog("\n");
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, device, 0, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Memory Setup
// Constants
cmAlphaTable4 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_float), (void*)&alphaTable4[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
cmProds4 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_int), (void*)&prods4[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
cmAlphaTable3 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_float), (void*)&alphaTable3[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
cmProds3 = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 4 * sizeof(cl_int), (void*)&prods3[0], &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Compute permutations.
cl_uint permutations[1024];
computePermutations(permutations);
// Upload permutations.
cmMemObjs[0] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(cl_uint) * 1024, permutations, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Image
cmMemObjs[1] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY, memSize, NULL, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// Result
const uint compressedSize = (width / 4) * (height / 4) * 8;
cmMemObjs[2] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY, compressedSize, NULL , &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
unsigned int * h_result = (uint*)malloc(compressedSize);
// Program Setup
size_t program_length;
const char* source_path = shrFindFilePath("DXTCompression.cl", argv[0]);
oclCheckError(source_path != NULL, shrTRUE);
char *source = oclLoadProgSource(source_path, "", &program_length);
oclCheckError(source != NULL, shrTRUE);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,
(const char **) &source, &program_length, &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-fast-relaxed-math", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLogEx(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclDXTCompression.ptx");
oclCheckError(ciErrNum, CL_SUCCESS);
}
// create the kernel
ckKernel = clCreateKernel(cpProgram, "compress", &ciErrNum);
oclCheckError(ciErrNum, CL_SUCCESS);
// set the args values
ciErrNum = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void *) &cmMemObjs[0]);
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void *) &cmMemObjs[1]);
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void *) &cmMemObjs[2]);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(cl_mem), (void*)&cmAlphaTable4);
ciErrNum |= clSetKernelArg(ckKernel, 4, sizeof(cl_mem), (void*)&cmProds4);
ciErrNum |= clSetKernelArg(ckKernel, 5, sizeof(cl_mem), (void*)&cmAlphaTable3);
ciErrNum |= clSetKernelArg(ckKernel, 6, sizeof(cl_mem), (void*)&cmProds3);
oclCheckError(ciErrNum, CL_SUCCESS);
// Copy input data host to device
clEnqueueWriteBuffer(cqCommandQueue, cmMemObjs[1], CL_FALSE, 0, sizeof(cl_uint) * width * height, block_image, 0,0,0);
// Determine launch configuration and run timed computation numIterations times
int blocks = ((width + 3) / 4) * ((height + 3) / 4); // rounds up by 1 block in each dim if %4 != 0
// Restrict the numbers of blocks to launch on low end GPUs to avoid kernel timeout
cl_uint compute_units;
clGetDeviceInfo(device, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(compute_units), &compute_units, NULL);
int blocksPerLaunch = MIN(blocks, 768 * (int)compute_units);
// set work-item dimensions
szGlobalWorkSize[0] = blocksPerLaunch * NUM_THREADS;
szLocalWorkSize[0]= NUM_THREADS;
#ifdef GPU_PROFILING
shrLog("\nRunning DXT Compression on %u x %u image...\n", width, height);
shrLog("\n%u Workgroups, %u Work Items per Workgroup, %u Work Items in NDRange...\n\n",
blocks, NUM_THREADS, blocks * NUM_THREADS);
int numIterations = 50;
for (int i = -1; i < numIterations; ++i) {
if (i == 0) { // start timing only after the first warmup iteration
clFinish(cqCommandQueue); // flush command queue
shrDeltaT(0); // start timer
}
#endif
// execute kernel
for( int j=0; j<blocks; j+= blocksPerLaunch ) {
clSetKernelArg(ckKernel, 7, sizeof(int), &j);
szGlobalWorkSize[0] = MIN( blocksPerLaunch, blocks-j ) * NUM_THREADS;
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL,
szGlobalWorkSize, szLocalWorkSize,
0, NULL, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
}
#ifdef GPU_PROFILING
}
clFinish(cqCommandQueue);
double dAvgTime = shrDeltaT(0) / (double)numIterations;
shrLogEx(LOGBOTH | MASTER, 0, "oclDXTCompression, Throughput = %.4f MPixels/s, Time = %.5f s, Size = %u Pixels, NumDevsUsed = %i, Workgroup = %d\n",
(1.0e-6 * (double)(width * height)/ dAvgTime), dAvgTime, (width * height), 1, szLocalWorkSize[0]);
#endif
// blocking read output
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmMemObjs[2], CL_TRUE, 0,
compressedSize, h_result, 0, NULL, NULL);
oclCheckError(ciErrNum, CL_SUCCESS);
// Write DDS file.
FILE* fp = NULL;
char output_filename[1024];
#ifdef WIN32
strcpy_s(output_filename, 1024, image_path);
strcpy_s(output_filename + strlen(image_path) - 3, 1024 - strlen(image_path) + 3, "dds");
fopen_s(&fp, output_filename, "wb");
#else
strcpy(output_filename, image_path);
strcpy(output_filename + strlen(image_path) - 3, "dds");
fp = fopen(output_filename, "wb");
#endif
oclCheckError(fp != NULL, shrTRUE);
DDSHeader header;
header.fourcc = FOURCC_DDS;
header.size = 124;
header.flags = (DDSD_WIDTH|DDSD_HEIGHT|DDSD_CAPS|DDSD_PIXELFORMAT|DDSD_LINEARSIZE);
header.height = height;
header.width = width;
header.pitch = compressedSize;
header.depth = 0;
header.mipmapcount = 0;
memset(header.reserved, 0, sizeof(header.reserved));
header.pf.size = 32;
header.pf.flags = DDPF_FOURCC;
header.pf.fourcc = FOURCC_DXT1;
header.pf.bitcount = 0;
header.pf.rmask = 0;
header.pf.gmask = 0;
header.pf.bmask = 0;
header.pf.amask = 0;
header.caps.caps1 = DDSCAPS_TEXTURE;
header.caps.caps2 = 0;
header.caps.caps3 = 0;
header.caps.caps4 = 0;
header.notused = 0;
fwrite(&header, sizeof(DDSHeader), 1, fp);
fwrite(h_result, compressedSize, 1, fp);
fclose(fp);
// Make sure the generated image matches the reference image (regression check)
shrLog("\nComparing against Host/C++ computation...\n");
const char* reference_image_path = shrFindFilePath(refimage_filename, argv[0]);
oclCheckError(reference_image_path != NULL, shrTRUE);
// read in the reference image from file
#ifdef WIN32
fopen_s(&fp, reference_image_path, "rb");
#else
fp = fopen(reference_image_path, "rb");
#endif
oclCheckError(fp != NULL, shrTRUE);
fseek(fp, sizeof(DDSHeader), SEEK_SET);
uint referenceSize = (width / 4) * (height / 4) * 8;
uint * reference = (uint *)malloc(referenceSize);
fread(reference, referenceSize, 1, fp);
fclose(fp);
// compare the reference image data to the sample/generated image
float rms = 0;
for (uint y = 0; y < height; y += 4)
{
for (uint x = 0; x < width; x += 4)
{
// binary comparison of data
uint referenceBlockIdx = ((y/4) * (width/4) + (x/4));
uint resultBlockIdx = ((y/4) * (width/4) + (x/4));
int cmp = compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
// log deviations, if any
if (cmp != 0.0f)
{
compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
shrLog("Deviation at (%d, %d):\t%f rms\n", x/4, y/4, float(cmp)/16/3);
}
rms += cmp;
}
}
rms /= width * height * 3;
shrLog("RMS(reference, result) = %f\n\n", rms);
// Free OpenCL resources
oclDeleteMemObjs(cmMemObjs, 3);
clReleaseMemObject(cmAlphaTable4);
clReleaseMemObject(cmProds4);
clReleaseMemObject(cmAlphaTable3);
clReleaseMemObject(cmProds3);
clReleaseKernel(ckKernel);
clReleaseProgram(cpProgram);
clReleaseCommandQueue(cqCommandQueue);
clReleaseContext(cxGPUContext);
// Free host memory
free(source);
free(h_img);
// finish
shrQAFinishExit(argc, (const char **)argv, (rms <= ERROR_THRESHOLD) ? QA_PASSED : QA_FAILED);
}
// main function
//*****************************************************************************
int main(int argc, 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);
}
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 cqCommandQueue[MAX_GPU_COUNT]; // OpenCL command que
cl_int ciErrNum;
shrSetLogFileName ("oclRadixSort.txt");
shrLog("%s starting...\n\n", argv[0]);
shrLog("clGetPlatformID...\n");
ciErrNum = oclGetPlatformID(&cpPlatform);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
shrLog("clGetDeviceIDs...\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("Create command queue...\n\n");
int id_device;
if(shrGetCmdLineArgumenti(argc, argv, "device", &id_device)) // Set up command queue(s) for GPU specified on the command line
{
// get & log device index # and name
cl_device_id cdDevice = cdDevices[id_device];
// create a command que
cqCommandQueue[0] = clCreateCommandQueue(cxGPUContext, cdDevice, 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
oclPrintDevInfo(LOGBOTH, cdDevice);
nDevice = 1;
}
else
{
// create command queues for all available devices
for (cl_uint i = 0; i < nDevice; i++)
{
cqCommandQueue[i] = clCreateCommandQueue(cxGPUContext, cdDevices[i], 0, &ciErrNum);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
}
for (cl_uint i = 0; i < nDevice; i++) oclPrintDevInfo(LOGBOTH, cdDevices[i]);
}
int ctaSize;
if (!shrGetCmdLineArgumenti(argc, argv, "work-group-size", &ctaSize))
{
ctaSize = 128;
}
shrLog("Running Radix Sort on %d GPU(s) ...\n\n", nDevice);
unsigned int numElements = 1048576;//128*128*128*2;
// Alloc and init some data on the host, then alloc and init GPU buffer
unsigned int **h_keys = (unsigned int**)malloc(nDevice * sizeof(unsigned int*));
unsigned int **h_keysSorted = (unsigned int**)malloc(nDevice * sizeof(unsigned int*));
cl_mem *d_keys = (cl_mem* )malloc(nDevice * sizeof(cl_mem));
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
h_keys[iDevice] = (unsigned int*)malloc(numElements * sizeof(unsigned int));
h_keysSorted[iDevice] = (unsigned int*)malloc(numElements * sizeof(unsigned int));
makeRandomUintVector(h_keys[iDevice], numElements, keybits);
d_keys[iDevice] = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE,
sizeof(unsigned int) * numElements, NULL, &ciErrNum);
ciErrNum |= clEnqueueWriteBuffer(cqCommandQueue[iDevice], d_keys[iDevice], CL_TRUE, 0,
sizeof(unsigned int) * numElements, h_keys[iDevice], 0, NULL, NULL);
oclCheckErrorEX(ciErrNum, CL_SUCCESS, NULL);
}
// instantiate RadixSort objects
RadixSort **radixSort = (RadixSort**)malloc(nDevice * sizeof(RadixSort*));
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
radixSort[iDevice] = new RadixSort(cxGPUContext, cqCommandQueue[iDevice], numElements, argv[0], ctaSize, true);
}
#ifdef GPU_PROFILING
int numIterations = 30;
for (int i = -1; i < numIterations; i++)
{
if (i == 0)
{
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clFinish(cqCommandQueue[iDevice]);
}
shrDeltaT(1);
}
#endif
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
radixSort[iDevice]->sort(d_keys[iDevice], 0, numElements, keybits);
}
#ifdef GPU_PROFILING
}
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clFinish(cqCommandQueue[iDevice]);
}
double gpuTime = shrDeltaT(1)/(double)numIterations;
shrLogEx(LOGBOTH | MASTER, 0, "oclRadixSort, Throughput = %.4f MElements/s, Time = %.5f s, Size = %u elements, NumDevsUsed = %d, Workgroup = %d\n",
(1.0e-6 * (double)(nDevice * numElements)/gpuTime), gpuTime, nDevice * numElements, nDevice, ctaSize);
#endif
// copy sorted keys to CPU
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clEnqueueReadBuffer(cqCommandQueue[iDevice], d_keys[iDevice], CL_TRUE, 0, sizeof(unsigned int) * numElements,
h_keysSorted[iDevice], 0, NULL, NULL);
}
// Check results
bool passed = true;
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
passed &= verifySortUint(h_keysSorted[iDevice], NULL, h_keys[iDevice], numElements);
}
shrLog("\n%s\n\n", passed ? "PASSED" : "FAILED");
// cleanup allocs
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clReleaseMemObject(d_keys[iDevice]);
free(h_keys[iDevice]);
free(h_keysSorted[iDevice]);
delete radixSort[iDevice];
}
free(radixSort);
free(h_keys);
free(h_keysSorted);
// remaining cleanup and exit
free(cdDevices);
for (cl_uint iDevice = 0; iDevice < nDevice; iDevice++)
{
clReleaseCommandQueue(cqCommandQueue[iDevice]);
}
clReleaseContext(cxGPUContext);
shrEXIT(argc, argv);
}
// main function
//*****************************************************************************
int main(int argc, const char **argv)
{
cl_context cxGPUContext; // OpenCL context
cl_command_queue cqCommandQue[MAX_GPU_COUNT]; // OpenCL command que
cl_device_id* cdDevices; // OpenCL device list
cl_int err = 1; // Error code var
shrSetLogFileName ("oclHiddenMarkovModel.txt");
shrLog(LOGBOTH, 0, "%s Starting...\n\n", argv[0]);
shrLog(LOGBOTH, 0, "Create context\n");
cxGPUContext = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, NULL, NULL, &err);
shrCheckErrorEX(err, CL_SUCCESS, NULL);
shrLog(LOGBOTH, 0, "Get device info...\n");
int nDevice = 0;
size_t nDeviceBytes;
err |= clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, 0, NULL, &nDeviceBytes);
cdDevices = (cl_device_id*)malloc(nDeviceBytes);
err |= clGetContextInfo(cxGPUContext, CL_CONTEXT_DEVICES, nDeviceBytes, cdDevices, NULL);
nDevice = (int)(nDeviceBytes/sizeof(cl_device_id));
shrLog(LOGBOTH, 0, "clCreateCommandQueue\n");
int id_device;
if(shrGetCmdLineArgumenti(argc, argv, "device", &id_device)) // Set up command queue(s) for GPU specified on the command line
{
// get & log device index # and name
cl_device_id cdDevice = cdDevices[id_device];
// create a command que
cqCommandQue[0] = clCreateCommandQueue(cxGPUContext, cdDevice, 0, &err);
shrCheckErrorEX(err, CL_SUCCESS, NULL);
oclPrintDevInfo(LOGBOTH, cdDevice);
nDevice = 1;
}
else
{ // create command queues for all available devices
for (int i = 0; i < nDevice; i++)
{
cqCommandQue[i] = clCreateCommandQueue(cxGPUContext, cdDevices[i], 0, &err);
shrCheckErrorEX(err, CL_SUCCESS, NULL);
}
for (int i = 0; i < nDevice; i++) oclPrintDevInfo(LOGBOTH, cdDevices[i]);
}
shrLog(LOGBOTH, 0, "\nUsing %d GPU(s)...\n\n", nDevice);
int wgSize;
if (!shrGetCmdLineArgumenti(argc, argv, "work-group-size", &wgSize))
{
wgSize = 256;
}
shrLog(LOGBOTH, 0, "Init Hidden Markov Model parameters\n");
int nState = 256*16; // number of states
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 = 500; // 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 (int 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(LOGBOTH, 0, "# of states = %d\n# of possible observations = %d \nSize of observational sequence = %d\n\n",
nState, nEmit, nObs);
shrLog(LOGBOTH, 0, "Compute Viterbi path on GPU\n\n");
HMM **oHmm = (HMM**)malloc(nDevice*sizeof(HMM*));
for (int 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 (int iDevice = 0; iDevice < nDevice; iDevice++)
{
vProb[iDevice] = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(float), NULL, &err);
vPath[iDevice] = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, sizeof(int)*nObs, NULL, &err);
}
#ifdef GPU_PROFILING
for (int iDevice = 0; iDevice < nDevice; iDevice++)
{
clFinish(cqCommandQue[iDevice]);;
}
shrDeltaT(1);
#endif
size_t szWorkGroup;
for (int iDevice = 0; iDevice < nDevice; iDevice++)
{
szWorkGroup = oHmm[iDevice]->ViterbiSearch(vProb[iDevice], vPath[iDevice], obs[iDevice]);
}
for (int iDevice = 0; iDevice < nDevice; iDevice++)
{
clFinish(cqCommandQue[iDevice]);
}
#ifdef GPU_PROFILING
double dElapsedTime = shrDeltaT(1);
shrLog(LOGBOTH | MASTER, 0, "oclHiddenMarkovModel, Throughput = %.4f, Time = %.5f, Size = %u, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * nDevice * nState * nObs)/dElapsedTime, dElapsedTime, (nDevice * nState * nObs), nDevice, szWorkGroup);
#endif
for (int iDevice = 0; iDevice < nDevice; iDevice++)
{
err = clEnqueueReadBuffer(cqCommandQue[iDevice], vPath[iDevice], CL_TRUE, 0, sizeof(int)*nObs, viterbiPathGPU[iDevice], 0, NULL, NULL);
}
shrLog(LOGBOTH, 0, "\nCompute Viterbi path on CPU\n");
for (int iDevice = 0; iDevice < nDevice; iDevice++)
{
err = ViterbiCPU(viterbiProbCPU[iDevice], viterbiPathCPU[iDevice], obs[iDevice], nObs, initProb, mtState, nState, mtEmit);
}
if (!err)
{
shrEXIT(argc, argv);
}
bool pass = true;
for (int iDevice = 0; iDevice < nDevice; iDevice++)
{
for (int i = 0; i < nObs; i++)
{
if (viterbiPathCPU[iDevice][i] != viterbiPathGPU[iDevice][i])
{
pass = false;
break;
}
}
}
shrLog(LOGBOTH, 0, "\nTEST %s\n\n", (pass) ? "PASSED" : "FAILED !!!");
// NOTE: Most properly this should be done at any of the exit points above, but it is omitted elsewhere for clarity.
shrLog(LOGBOTH, 0, "Release CPU buffers and OpenCL objects...\n");
free(initProb);
free(mtState);
free(mtEmit);
for (int 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);
shrEXIT(argc, argv);
}
// Main function
// *********************************************************************
int main(const int argc, const char** argv)
{
// start logs
shrSetLogFileName ("oclDXTCompression.txt");
shrLog(LOGBOTH, 0, "%s Starting...\n\n", argv[0]);
cl_context cxGPUContext;
cl_command_queue cqCommandQueue;
cl_program cpProgram;
cl_kernel ckKernel;
cl_mem cmMemObjs[3];
size_t szGlobalWorkSize[1];
size_t szLocalWorkSize[1];
cl_int ciErrNum;
// Get the path of the filename
char *filename;
if (shrGetCmdLineArgumentstr(argc, argv, "image", &filename)) {
image_filename = filename;
}
// load image
const char* image_path = shrFindFilePath(image_filename, argv[0]);
shrCheckError(image_path != NULL, shrTRUE);
shrLoadPPM4ub(image_path, (unsigned char **)&h_img, &width, &height);
shrCheckError(h_img != NULL, shrTRUE);
shrLog(LOGBOTH, 0, "Loaded '%s', %d x %d pixels\n", image_path, width, height);
// Convert linear image to block linear.
uint * block_image = (uint *) malloc(width * height * 4);
// Convert linear image to block linear.
for(uint by = 0; by < height/4; by++) {
for(uint bx = 0; bx < width/4; bx++) {
for (int i = 0; i < 16; i++) {
const int x = i & 3;
const int y = i / 4;
block_image[(by * width/4 + bx) * 16 + i] =
((uint *)h_img)[(by * 4 + y) * 4 * (width/4) + bx * 4 + x];
}
}
}
// create the OpenCL context on a GPU device
cxGPUContext = clCreateContextFromType(0, CL_DEVICE_TYPE_GPU, NULL, NULL, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// get and log device
cl_device_id device;
if( shrCheckCmdLineFlag(argc, argv, "device") ) {
int device_nr = 0;
shrGetCmdLineArgumenti(argc, argv, "device", &device_nr);
device = oclGetDev(cxGPUContext, device_nr);
} else {
device = oclGetMaxFlopsDev(cxGPUContext);
}
oclPrintDevInfo(LOGBOTH, device);
// create a command-queue
cqCommandQueue = clCreateCommandQueue(cxGPUContext, device, 0, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// Memory Setup
// Compute permutations.
cl_uint permutations[1024];
computePermutations(permutations);
// Upload permutations.
cmMemObjs[0] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
sizeof(cl_uint) * 1024, permutations, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// Image
cmMemObjs[1] = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY ,
sizeof(cl_uint) * width * height, NULL, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// Result
const uint compressedSize = (width / 4) * (height / 4) * 8;
cmMemObjs[2] = clCreateBuffer(cxGPUContext, CL_MEM_WRITE_ONLY,
compressedSize, NULL , &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
unsigned int * h_result = (uint *)malloc(compressedSize);
// Program Setup
size_t program_length;
const char* source_path = shrFindFilePath("DXTCompression.cl", argv[0]);
shrCheckError(source_path != NULL, shrTRUE);
char *source = oclLoadProgSource(source_path, "", &program_length);
shrCheckError(source != NULL, shrTRUE);
// create the program
cpProgram = clCreateProgramWithSource(cxGPUContext, 1,
(const char **) &source, &program_length, &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// build the program
ciErrNum = clBuildProgram(cpProgram, 0, NULL, "-cl-mad-enable", NULL, NULL);
if (ciErrNum != CL_SUCCESS)
{
// write out standard error, Build Log and PTX, then cleanup and exit
shrLog(LOGBOTH | ERRORMSG, ciErrNum, STDERROR);
oclLogBuildInfo(cpProgram, oclGetFirstDev(cxGPUContext));
oclLogPtx(cpProgram, oclGetFirstDev(cxGPUContext), "oclDXTCompression.ptx");
shrCheckError(ciErrNum, CL_SUCCESS);
}
// create the kernel
ckKernel = clCreateKernel(cpProgram, "compress", &ciErrNum);
shrCheckError(ciErrNum, CL_SUCCESS);
// set the args values
ciErrNum = clSetKernelArg(ckKernel, 0, sizeof(cl_mem), (void *) &cmMemObjs[0]);
ciErrNum |= clSetKernelArg(ckKernel, 1, sizeof(cl_mem), (void *) &cmMemObjs[1]);
ciErrNum |= clSetKernelArg(ckKernel, 2, sizeof(cl_mem), (void *) &cmMemObjs[2]);
ciErrNum |= clSetKernelArg(ckKernel, 3, sizeof(float) * 4 * 16, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 4, sizeof(float) * 4 * 16, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 5, sizeof(int) * 64, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 6, sizeof(float) * 16 * 6, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 7, sizeof(unsigned int) * 160, NULL);
ciErrNum |= clSetKernelArg(ckKernel, 8, sizeof(int) * 16, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
shrLog(LOGBOTH, 0, "Running DXT Compression on %u x %u image...\n\n", width, height);
// Upload the image
clEnqueueWriteBuffer(cqCommandQueue, cmMemObjs[1], CL_FALSE, 0, sizeof(cl_uint) * width * height, block_image, 0,0,0);
// set work-item dimensions
szGlobalWorkSize[0] = width * height * (NUM_THREADS/16);
szLocalWorkSize[0]= NUM_THREADS;
#ifdef GPU_PROFILING
int numIterations = 100;
for (int i = -1; i < numIterations; ++i) {
if (i == 0) { // start timing only after the first warmup iteration
clFinish(cqCommandQueue); // flush command queue
shrDeltaT(0); // start timer
}
#endif
// execute kernel
ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL,
szGlobalWorkSize, szLocalWorkSize,
0, NULL, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
#ifdef GPU_PROFILING
}
clFinish(cqCommandQueue);
double dAvgTime = shrDeltaT(0) / (double)numIterations;
shrLog(LOGBOTH | MASTER, 0, "oclDXTCompression, Throughput = %.4f, Time = %.5f, Size = %u, NumDevsUsed = %i\n",
(1.0e-6 * (double)(width * height)/ dAvgTime), dAvgTime, (width * height), 1);
#endif
// blocking read output
ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmMemObjs[2], CL_TRUE, 0,
compressedSize, h_result, 0, NULL, NULL);
shrCheckError(ciErrNum, CL_SUCCESS);
// Write DDS file.
FILE* fp = NULL;
char output_filename[1024];
#ifdef WIN32
strcpy_s(output_filename, 1024, image_path);
strcpy_s(output_filename + strlen(image_path) - 3, 1024 - strlen(image_path) + 3, "dds");
fopen_s(&fp, output_filename, "wb");
#else
strcpy(output_filename, image_path);
strcpy(output_filename + strlen(image_path) - 3, "dds");
fp = fopen(output_filename, "wb");
#endif
shrCheckError(fp != NULL, shrTRUE);
DDSHeader header;
header.fourcc = FOURCC_DDS;
header.size = 124;
header.flags = (DDSD_WIDTH|DDSD_HEIGHT|DDSD_CAPS|DDSD_PIXELFORMAT|DDSD_LINEARSIZE);
header.height = height;
header.width = width;
header.pitch = compressedSize;
header.depth = 0;
header.mipmapcount = 0;
memset(header.reserved, 0, sizeof(header.reserved));
header.pf.size = 32;
header.pf.flags = DDPF_FOURCC;
header.pf.fourcc = FOURCC_DXT1;
header.pf.bitcount = 0;
header.pf.rmask = 0;
header.pf.gmask = 0;
header.pf.bmask = 0;
header.pf.amask = 0;
header.caps.caps1 = DDSCAPS_TEXTURE;
header.caps.caps2 = 0;
header.caps.caps3 = 0;
header.caps.caps4 = 0;
header.notused = 0;
fwrite(&header, sizeof(DDSHeader), 1, fp);
fwrite(h_result, compressedSize, 1, fp);
fclose(fp);
// Make sure the generated image matches the reference image (regression check)
shrLog(LOGBOTH, 0, "\nComparing against Host/C++ computation...\n");
const char* reference_image_path = shrFindFilePath(refimage_filename, argv[0]);
shrCheckError(reference_image_path != NULL, shrTRUE);
// read in the reference image from file
#ifdef WIN32
fopen_s(&fp, reference_image_path, "rb");
#else
fp = fopen(reference_image_path, "rb");
#endif
shrCheckError(fp != NULL, shrTRUE);
fseek(fp, sizeof(DDSHeader), SEEK_SET);
uint referenceSize = (width / 4) * (height / 4) * 8;
uint * reference = (uint *)malloc(referenceSize);
fread(reference, referenceSize, 1, fp);
fclose(fp);
// compare the reference image data to the sample/generated image
float rms = 0;
for (uint y = 0; y < height; y += 4)
{
for (uint x = 0; x < width; x += 4)
{
// binary comparison of data
uint referenceBlockIdx = ((y/4) * (width/4) + (x/4));
uint resultBlockIdx = ((y/4) * (width/4) + (x/4));
int cmp = compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
// log deviations, if any
if (cmp != 0.0f)
{
compareBlock(((BlockDXT1 *)h_result) + resultBlockIdx, ((BlockDXT1 *)reference) + referenceBlockIdx);
shrLog(LOGBOTH, 0, "Deviation at (%d, %d):\t%f rms\n", x/4, y/4, float(cmp)/16/3);
}
rms += cmp;
}
}
rms /= width * height * 3;
shrLog(LOGBOTH, 0, "RMS(reference, result) = %f\n\n", rms);
shrLog(LOGBOTH, 0, "TEST %s\n\n", (rms <= ERROR_THRESHOLD) ? "PASSED" : "FAILED !!!");
// Free OpenCL resources
oclDeleteMemObjs(cmMemObjs, 3);
clReleaseKernel(ckKernel);
clReleaseProgram(cpProgram);
clReleaseCommandQueue(cqCommandQueue);
clReleaseContext(cxGPUContext);
// Free host memory
free(source);
free(h_img);
// finish
shrEXIT(argc, argv);
}
///////////////////////////////////////////////////////////////////////////////
//Parse args, run the appropriate tests
///////////////////////////////////////////////////////////////////////////////
int runTest(const int argc, const char **argv)
{
int start = DEFAULT_SIZE;
int end = DEFAULT_SIZE;
int startDevice = 0;
int endDevice = 0;
int increment = DEFAULT_INCREMENT;
testMode mode = QUICK_MODE;
bool htod = false;
bool dtoh = false;
bool dtod = false;
char *modeStr;
char *device = NULL;
printMode printmode = USER_READABLE;
char *memModeStr = NULL;
memoryMode memMode = PAGEABLE;
accessMode accMode = DIRECT;
//process command line args
if(shrCheckCmdLineFlag( argc, argv, "help"))
{
printHelp();
return 0;
}
if(shrCheckCmdLineFlag( argc, argv, "csv"))
{
printmode = CSV;
}
// Get host memory mode type from command line
if(shrGetCmdLineArgumentstr(argc, argv, "memory", &memModeStr))
{
if(strcmp(memModeStr, "pageable") == 0 )
{
memMode = PAGEABLE;
}
else if(strcmp(memModeStr, "pinned") == 0)
{
memMode = PINNED;
}
else
{
shrLog("Invalid memory mode - valid modes are pageable or pinned\n");
shrLog("See --help for more information\n");
return -1000;
}
}
else
{
//default - pageable memory
memMode = PAGEABLE;
}
// Access type from command line
if(shrGetCmdLineArgumentstr(argc, argv, "access", &memModeStr))
{
if(strcmp(memModeStr, "direct") == 0)
{
accMode = DIRECT;
}
else if(strcmp(memModeStr, "mapped") == 0)
{
accMode = MAPPED;
}
else
{
shrLog("Invalid access mode - valid modes are direct or mapped\n");
shrLog("See --help for more information\n");
return -2000;
}
}
else
{
//default - direct
accMode = DIRECT;
}
// Get OpenCL platform ID for NVIDIA if available, otherwise default
cl_platform_id clSelectedPlatformID = NULL;
cl_int ciErrNum = oclGetPlatformID (&clSelectedPlatformID);
oclCheckError(ciErrNum, CL_SUCCESS);
// Find out how many devices there are
cl_uint ciDeviceCount;
ciErrNum = clGetDeviceIDs (clSelectedPlatformID, CL_DEVICE_TYPE_GPU, 0, NULL, &ciDeviceCount);
if (ciErrNum != CL_SUCCESS)
{
shrLog(" Error %i in clGetDeviceIDs call !!!\n\n", ciErrNum);
return ciErrNum;
}
else if (ciDeviceCount == 0)
{
shrLog(" There are no devices supporting OpenCL (return code %i)\n\n", ciErrNum);
return ciErrNum;
}
// Get command line device options and config accordingly
if(shrGetCmdLineArgumentstr(argc, argv, "device", &device))
{
if(strcmp (device, "all") == 0)
{
shrLog("\n!!!Cumulative Bandwidth to be computed from all the devices !!!\n\n");
startDevice = 0;
endDevice = (int)(ciDeviceCount-1);
}
else
{
startDevice = endDevice = atoi(device);
if(startDevice < 0 || ((size_t)startDevice) >= ciDeviceCount)
{
shrLog("\n!!!Invalid GPU number %d given hence default gpu %d will be used !!!\n", startDevice,0);
startDevice = endDevice = 0;
}
}
}
// Get and log the device info
shrLog("Running on...\n\n");
devices = (cl_device_id*) malloc(sizeof(cl_device_id) * ciDeviceCount);
ciErrNum = clGetDeviceIDs (clSelectedPlatformID, CL_DEVICE_TYPE_GPU, ciDeviceCount, devices, &ciDeviceCount);
for(int currentDevice = startDevice; currentDevice <= endDevice; currentDevice++)
{
oclPrintDevName(LOGBOTH, devices[currentDevice]);
shrLog("\n");
}
shrLog("\n");
// Get command line mode(s) and config accordingly
if(shrGetCmdLineArgumentstr(argc, argv, "mode", &modeStr))
{
//figure out the mode
if(strcmp(modeStr, "quick") == 0)
{
shrLog("Quick Mode\n\n");
mode = QUICK_MODE;
}
else if(strcmp(modeStr, "shmoo") == 0)
{
shrLog("Shmoo Mode\n\n");
mode = SHMOO_MODE;
}
else if(strcmp(modeStr, "range") == 0)
{
shrLog("Range Mode\n\n");
mode = RANGE_MODE;
}
else
{
shrLog("Invalid mode - valid modes are quick, range, or shmoo\n");
shrLog("See --help for more information\n\n");
return -3000;
}
}
else
{
//default mode - quick
shrLog("Quick Mode\n\n");
mode = QUICK_MODE;
}
if(shrCheckCmdLineFlag(argc, argv, "htod"))
htod = true;
if(shrCheckCmdLineFlag(argc, argv, "dtoh"))
dtoh = true;
if(shrCheckCmdLineFlag(argc, argv, "dtod"))
dtod = true;
if(!htod && !dtoh && !dtod)
{
//default: All
htod = true;
dtoh = true;
dtod = true;
}
if(RANGE_MODE == mode)
{
if(shrGetCmdLineArgumenti( argc, argv, "start", &start))
{
if( start <= 0 )
{
shrLog("Illegal argument - start must be greater than zero\n");
return -4000;
}
}
else
{
shrLog("Must specify a starting size in range mode\n");
shrLog("See --help for more information\n");
return -5000;
}
if(shrGetCmdLineArgumenti( argc, argv, "end", &end))
{
if(end <= 0)
{
shrLog("Illegal argument - end must be greater than zero\n");
return -6000;
}
if(start > end)
{
shrLog("Illegal argument - start is greater than end\n");
return -7000;
}
}
else
{
shrLog("Must specify an end size in range mode.\n");
shrLog("See --help for more information\n");
return -8000;
}
if(shrGetCmdLineArgumenti( argc, argv, "increment", &increment))
{
if(increment <= 0)
{
shrLog("Illegal argument - increment must be greater than zero\n");
return -9000;
}
}
else
{
shrLog("Must specify an increment in user mode\n");
shrLog("See --help for more information\n");
return -10000;
}
}
// Create the OpenCL context
cxGPUContext = clCreateContext(0, ciDeviceCount, devices, NULL, NULL, NULL);
if (cxGPUContext == (cl_context)0)
{
shrLog("Failed to create OpenCL context!\n");
return -11000;
}
// Run tests
if(htod)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, HOST_TO_DEVICE, printmode, accMode, memMode, startDevice, endDevice);
}
if(dtoh)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, DEVICE_TO_HOST, printmode, accMode, memMode, startDevice, endDevice);
}
if(dtod)
{
testBandwidth((unsigned int)start, (unsigned int)end, (unsigned int)increment,
mode, DEVICE_TO_DEVICE, printmode, accMode, memMode, startDevice, endDevice);
}
// Clean up
free(memModeStr);
if(cqCommandQueue)clReleaseCommandQueue(cqCommandQueue);
if(cxGPUContext)clReleaseContext(cxGPUContext);
if(devices)free(devices);
return 0;
}
// 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);
}
bool
runTest( int argc, const char** argv, ReduceType datatype)
{
int size = 1<<24; // number of elements to reduce
int maxThreads;
cl_kernel reductionKernel = getReductionKernel(datatype, 0, 64, 1);
clReleaseKernel(reductionKernel);
if (smallBlock)
maxThreads = 64; // number of threads per block
else
maxThreads = 128;
int whichKernel = 6;
int maxBlocks = 64;
bool cpuFinalReduction = false;
int cpuFinalThreshold = 1;
shrGetCmdLineArgumenti( argc, (const char**) argv, "n", &size);
shrGetCmdLineArgumenti( argc, (const char**) argv, "threads", &maxThreads);
shrGetCmdLineArgumenti( argc, (const char**) argv, "kernel", &whichKernel);
shrGetCmdLineArgumenti( argc, (const char**) argv, "maxblocks", &maxBlocks);
shrLog(" %d elements\n", size);
shrLog(" %d threads (max)\n", maxThreads);
cpuFinalReduction = (shrCheckCmdLineFlag( argc, (const char**) argv, "cpufinal") == shrTRUE);
shrGetCmdLineArgumenti( argc, (const char**) argv, "cputhresh", &cpuFinalThreshold);
bool runShmoo = (shrCheckCmdLineFlag(argc, (const char**) argv, "shmoo") == shrTRUE);
#ifdef GPU_PROFILING
if (runShmoo)
{
shmoo<T>(1, 33554432, maxThreads, maxBlocks, datatype);
return true;
}
else
#endif
{
// create random input data on CPU
unsigned int bytes = size * sizeof(T);
T* h_idata = (T*)malloc(bytes);
for(int i=0; i<size; i++)
{
// Keep the numbers small so we don't get truncation error in the sum
if (datatype == REDUCE_INT)
h_idata[i] = (T)(rand() & 0xFF);
else
h_idata[i] = (rand() & 0xFF) / (T)RAND_MAX;
}
int numBlocks = 0;
int numThreads = 0;
getNumBlocksAndThreads(whichKernel, size, maxBlocks, maxThreads, numBlocks, numThreads);
if (numBlocks == 1) cpuFinalThreshold = 1;
shrLog(" %d blocks\n\n", numBlocks);
// allocate mem for the result on host side
T* h_odata = (T*)malloc(numBlocks * sizeof(T));
// allocate device memory and data
cl_mem d_idata = clCreateBuffer(cxGPUContext, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, bytes, h_idata, NULL);
cl_mem d_odata = clCreateBuffer(cxGPUContext, CL_MEM_READ_WRITE, numBlocks * sizeof(T), NULL, NULL);
int testIterations = 100;
double dTotalTime = 0.0;
T gpu_result = 0;
gpu_result = profileReduce<T>(datatype, size, numThreads, numBlocks, maxThreads, maxBlocks,
whichKernel, testIterations, cpuFinalReduction,
cpuFinalThreshold, &dTotalTime,
h_odata, d_idata, d_odata);
#ifdef GPU_PROFILING
double reduceTime = dTotalTime/(double)testIterations;
shrLogEx(LOGBOTH | MASTER, 0, "oclReduction, Throughput = %.4f GB/s, Time = %.5f s, Size = %u Elements, NumDevsUsed = %d, Workgroup = %u\n",
1.0e-9 * ((double)bytes)/reduceTime, reduceTime, size, 1, numThreads);
#endif
// compute reference solution
shrLog("\nComparing against Host/C++ computation...\n");
T cpu_result = reduceCPU<T>(h_idata, size);
if (datatype == REDUCE_INT)
{
shrLog(" GPU result = %d\n", gpu_result);
shrLog(" CPU result = %d\n\n", cpu_result);
shrLog("%s\n\n", (gpu_result == cpu_result) ? "PASSED" : "FAILED");
}
else
{
shrLog(" GPU result = %.9f\n", gpu_result);
shrLog(" CPU result = %.9f\n\n", cpu_result);
double threshold = (datatype == REDUCE_FLOAT) ? 1e-8 * size : 1e-12;
double diff = abs((double)gpu_result - (double)cpu_result);
shrLog("%s\n\n", (diff < threshold) ? "PASSED" : "FAILED");
}
// cleanup
free(h_idata);
free(h_odata);
clReleaseMemObject(d_idata);
clReleaseMemObject(d_odata);
return (gpu_result == cpu_result);
}
}
