////////////////////////////////////////////////////////////////////////////////
//! Run a simple test for CUDA
////////////////////////////////////////////////////////////////////////////////
CUTBoolean runTest(int argc, char** argv)
{
if (!cutCheckCmdLineFlag(argc, (const char **)argv, "noqatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt"))
{
g_bQAReadback = true;
fpsLimit = frameCheckNumber;
}
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
if (CUTFalse == initGL(argc, argv)) {
return CUTFalse;
}
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") )
cutilGLDeviceInit(argc, argv);
else {
cudaGLSetGLDevice( cutGetMaxGflopsDeviceId() );
}
// Create the CUTIL timer
cutilCheckError( cutCreateTimer( &timer));
// register callbacks
glutDisplayFunc(display);
glutKeyboardFunc(keyboard);
glutMouseFunc(mouse);
glutMotionFunc(motion);
if (g_bQAReadback) {
g_CheckRender = new CheckBackBuffer(window_width, window_height, 4);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
}
// create VBO
createVBO(&vbo);
// run the cuda part
runCuda(vbo);
// check result of Cuda step
checkResultCuda(argc, argv, vbo);
atexit(cleanup);
// start rendering mainloop
glutMainLoop();
cudaThreadExit();
return CUTTrue;
}
bool initCUDA( int argc, char **argv)
{
return true;
if ( cutCheckCmdLineFlag(argc, (const char **)argv, "device"))
{
cutilGLDeviceInit(argc, argv);
}
else
{
cudaGLSetGLDevice (cutGetMaxGflopsDeviceId());
}
return true;
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main( int argc, char** argv)
{
//start logs
shrSetLogFileName ("volumeRender.txt");
shrLog("%s Starting...\n\n", argv[0]);
if (cutCheckCmdLineFlag(argc, (const char **)argv, "qatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt"))
{
g_bQAReadback = true;
fpsLimit = frameCheckNumber;
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "glverify"))
{
g_bQAGLVerify = true;
fpsLimit = frameCheckNumber;
}
if (g_bQAReadback) {
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilDeviceInit(argc, argv);
} else {
cudaSetDevice( cutGetMaxGflopsDeviceId() );
}
} else {
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
initGL( &argc, argv );
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilGLDeviceInit(argc, argv);
} else {
cudaGLSetGLDevice( cutGetMaxGflopsDeviceId() );
}
/*
int device;
struct cudaDeviceProp prop;
cudaGetDevice( &device );
cudaGetDeviceProperties( &prop, device );
if( !strncmp( "Tesla", prop.name, 5 ) ) {
shrLog("This sample needs a card capable of OpenGL and display.\n");
shrLog("Please choose a different device with the -device=x argument.\n");
cutilExit(argc, argv);
}
*/
}
// parse arguments
char *filename;
if (cutGetCmdLineArgumentstr( argc, (const char**) argv, "file", &filename)) {
volumeFilename = filename;
}
int n;
if (cutGetCmdLineArgumenti( argc, (const char**) argv, "size", &n)) {
volumeSize.width = volumeSize.height = volumeSize.depth = n;
}
if (cutGetCmdLineArgumenti( argc, (const char**) argv, "xsize", &n)) {
volumeSize.width = n;
}
if (cutGetCmdLineArgumenti( argc, (const char**) argv, "ysize", &n)) {
volumeSize.height = n;
}
if (cutGetCmdLineArgumenti( argc, (const char**) argv, "zsize", &n)) {
volumeSize.depth = n;
}
// load volume data
char* path = shrFindFilePath(volumeFilename, argv[0]);
if (path == 0) {
shrLog("Error finding file '%s'\n", volumeFilename);
exit(EXIT_FAILURE);
}
size_t size = volumeSize.width*volumeSize.height*volumeSize.depth*sizeof(VolumeType);
void *h_volume = loadRawFile(path, size);
initCuda(h_volume, volumeSize);
free(h_volume);
cutilCheckError( cutCreateTimer( &timer));
shrLog("Press '=' and '-' to change density\n"
" ']' and '[' to change brightness\n"
" ';' and ''' to modify transfer function offset\n"
" '.' and ',' to modify transfer function scale\n\n");
// calculate new grid size
gridSize = dim3(iDivUp(width, blockSize.x), iDivUp(height, blockSize.y));
if (g_bQAReadback) {
g_CheckRender = new CheckBackBuffer(width, height, 4, false);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
uint *d_output;
cutilSafeCall(cudaMalloc((void**)&d_output, width*height*sizeof(uint)));
cutilSafeCall(cudaMemset(d_output, 0, width*height*sizeof(uint)));
float modelView[16] =
{
1.0f, 0.0f, 0.0f, 0.0f,
0.0f, 1.0f, 0.0f, 0.0f,
0.0f, 0.0f, 1.0f, 0.0f,
0.0f, 0.0f, 4.0f, 1.0f
};
invViewMatrix[0] = modelView[0]; invViewMatrix[1] = modelView[4]; invViewMatrix[2] = modelView[8]; invViewMatrix[3] = modelView[12];
invViewMatrix[4] = modelView[1]; invViewMatrix[5] = modelView[5]; invViewMatrix[6] = modelView[9]; invViewMatrix[7] = modelView[13];
invViewMatrix[8] = modelView[2]; invViewMatrix[9] = modelView[6]; invViewMatrix[10] = modelView[10]; invViewMatrix[11] = modelView[14];
// call CUDA kernel, writing results to PBO
copyInvViewMatrix(invViewMatrix, sizeof(float4)*3);
// Start timer 0 and process n loops on the GPU
int nIter = 10;
for (int i = -1; i < nIter; i++)
{
if( i == 0 ) {
cudaThreadSynchronize();
cutStartTimer(timer);
}
render_kernel(gridSize, blockSize, d_output, width, height, density, brightness, transferOffset, transferScale);
}
cudaThreadSynchronize();
cutStopTimer(timer);
// Get elapsed time and throughput, then log to sample and master logs
double dAvgTime = cutGetTimerValue(timer)/(nIter * 1000.0);
shrLogEx(LOGBOTH | MASTER, 0, "volumeRender, Throughput = %.4f MTexels/s, Time = %.5f s, Size = %u Texels, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * width * height)/dAvgTime, dAvgTime, (width * height), 1, blockSize.x * blockSize.y);
cutilCheckMsg("Error: render_kernel() execution FAILED");
cutilSafeCall( cudaThreadSynchronize() );
cutilSafeCall( cudaMemcpy(g_CheckRender->imageData(), d_output, width*height*4, cudaMemcpyDeviceToHost) );
g_CheckRender->savePPM(sOriginal[g_Index], true, NULL);
if (!g_CheckRender->PPMvsPPM(sOriginal[g_Index], sReference[g_Index], MAX_EPSILON_ERROR, THRESHOLD)) {
shrLog("\nFAILED\n\n");
} else {
shrLog("\nPASSED\n\n");
}
cudaFree(d_output);
freeCudaBuffers();
if (g_CheckRender) {
delete g_CheckRender; g_CheckRender = NULL;
}
} else {
// This is the normal rendering path for VolumeRender
glutDisplayFunc(display);
glutKeyboardFunc(keyboard);
glutMouseFunc(mouse);
glutMotionFunc(motion);
glutReshapeFunc(reshape);
glutIdleFunc(idle);
initPixelBuffer();
if (g_bQAGLVerify) {
g_CheckRender = new CheckBackBuffer(width, height, 4);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
}
atexit(cleanup);
glutMainLoop();
}
cudaThreadExit();
shrEXIT(argc, (const char**)argv);
}
int main(int argc, char** argv)
{
printf("[%s]\n", sSDKsample);
if (argc > 1) {
if (cutCheckCmdLineFlag(argc, (const char **)argv, "help")) {
printHelp();
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "qatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt"))
{
g_bQAReadback = true;
fpsLimit = frameCheckNumber;
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "glverify"))
{
g_bOpenGLQA = true;
fpsLimit = frameCheckNumber;
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "fbo")) {
g_bFBODisplay = true;
fpsLimit = frameCheckNumber;
}
}
if (g_bQAReadback)
{
runAutoTest(argc, argv);
}
else
{
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
initGL( &argc, argv );
// use command-line specified CUDA device if possible, otherwise search for capable device
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilGLDeviceInit(argc, argv);
int device;
cudaGetDevice( &device );
if( checkCUDAProfile( device ) == false ) {
cudaThreadExit();
cutilExit(argc, argv);
}
} else {
//cudaGLSetGLDevice (cutGetMaxGflopsDeviceId() );
int dev = findCapableDevice(argc, argv);
if( dev != -1 )
cudaGLSetGLDevice( dev );
else {
cudaThreadExit();
cutilExit(argc, argv);
}
}
cutilCheckError(cutCreateTimer(&timer));
cutilCheckError(cutResetTimer(timer));
glutDisplayFunc(display);
glutKeyboardFunc(keyboard);
glutReshapeFunc(reshape);
glutIdleFunc(idle);
if (g_bOpenGLQA) {
loadDefaultImage( argc, argv );
}
if (argc > 1) {
char *filename;
if (cutGetCmdLineArgumentstr(argc, (const char **)argv, "file", &filename)) {
initializeData(filename, argc, argv);
}
} else {
loadDefaultImage( argc, argv );
}
// If code is not printing the USage, then we execute this path.
if (!bQuit) {
if (g_bOpenGLQA) {
g_CheckRender = new CheckBackBuffer(wWidth, wHeight, 4);
g_CheckRender->setPixelFormat(GL_BGRA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
}
printf("I: display image\n");
printf("T: display Sobel edge detection (computed with tex)\n");
printf("S: display Sobel edge detection (computed with tex+shared memory)\n");
printf("Use the '-' and '=' keys to change the brightness.\n");
printf("b: switch block filter operation (mean/Sobel)\n");
printf("p: swtich point filter operation (threshold on/off)\n");
fflush(stdout);
atexit(cleanup);
glutMainLoop();
}
}
cudaThreadExit();
cutilExit(argc, argv);
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main(int argc, char** argv)
{
int retVal = 0;
retVal = xnInit( argc, argv );
printf("[ %s ]\n", sSDKsample);
if (argc > 1) {
cutGetCmdLineArgumenti( argc, (const char**) argv, "n", (int *) &numParticles);
if (cutCheckCmdLineFlag(argc, (const char **)argv, "qatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt")
)
{
g_bQAReadback = true;
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "glverify"))
{
g_bQAGLVerify = true;
}
}
if (g_bQAReadback) {
// For Automated testing, we do not use OpenGL/CUDA interop
if ( cutCheckCmdLineFlag(argc, (const char **)argv, "device")) {
cutilDeviceInit (argc, argv);
} else {
cudaSetDevice (cutGetMaxGflopsDeviceId() );
}
g_CheckRender = new CheckBackBuffer(winWidth, winHeight, 4, false);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
// This code path is used for Automated Testing
initParticles(numParticles, false, false);
initParams();
if (emitterOn) {
runEmitter();
}
SimParams ¶ms = psystem->getParams();
params.cursorPos = make_float3(cursorPosLag.x, cursorPosLag.y, cursorPosLag.z);
psystem->step(timestep);
float4 *pos = NULL, *vel = NULL;
int g_TotalErrors = 0;
psystem->dumpBin(&pos, &vel);
g_CheckRender->dumpBin(pos, numParticles*sizeof(float4), "smokeParticles_pos.bin");
g_CheckRender->dumpBin(vel, numParticles*sizeof(float4), "smokeParticles_vel.bin");
if (!g_CheckRender->compareBin2BinFloat("smokeParticles_pos.bin", sRefBin[0], numParticles*sizeof(float4), MAX_EPSILON_ERROR, THRESHOLD))
g_TotalErrors++;
if (!g_CheckRender->compareBin2BinFloat("smokeParticles_vel.bin", sRefBin[1], numParticles*sizeof(float4), MAX_EPSILON_ERROR, THRESHOLD))
g_TotalErrors++;
delete psystem;
delete g_CheckRender;
printf("%s\n", (g_TotalErrors > 0) ? "FAILED" : "PASSED");
cudaThreadExit();
} else {
// Normal smokeParticles rendering path
// 1st initialize OpenGL context, so we can properly set the GL for CUDA.
// This is needed to achieve optimal performance with OpenGL/CUDA interop.
initGL( &argc, argv );
if ( cutCheckCmdLineFlag(argc, (const char **)argv, "device")) {
cutilGLDeviceInit (argc, argv);
} else {
cudaGLSetGLDevice (cutGetMaxGflopsDeviceId() );
}
if (g_bQAGLVerify) {
g_CheckRender = new CheckBackBuffer(winWidth, winHeight, 4);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
}
// This is the normal code path for SmokeParticles
initParticles(numParticles, true, true);
initParams();
initMenus();
glutDisplayFunc(display);
glutReshapeFunc(reshape);
glutMouseFunc(mouse);
glutMotionFunc(motion);
glutKeyboardFunc(key);
glutKeyboardUpFunc(keyUp);
glutSpecialFunc(special);
glutIdleFunc(idle);
glutMainLoop();
}
cutilExit(argc, argv);
return retVal;
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main( int argc, char** argv)
{
shrQAStart(argc, argv);
// start logs
shrSetLogFileName ("bilateralFilter.txt");
shrLog("%s Starting...\n\n", argv[0]);
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
cutGetCmdLineArgumenti( argc, (const char**) argv, "threads", &nthreads );
cutGetCmdLineArgumenti( argc, (const char**) argv, "radius", &filter_radius);
// load image to process
loadImageData(argc, argv);
if (cutCheckCmdLineFlag(argc, (const char **)argv, "qatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt"))
{
// Running CUDA kernel (bilateralFilter) without visualization (QA Testing/Verification)
runAutoTest(argc, argv);
shrQAFinishExit(argc, (const char **)argv, (g_TotalErrors == 0 ? QA_PASSED : QA_FAILED));
}
else if (cutCheckCmdLineFlag(argc, (const char **)argv, "benchmark"))
{
// Running CUDA kernel (bilateralFilter) in Benchmarking Mode
runBenchmark(argc, argv);
shrQAFinishExit(argc, (const char **)argv, (g_TotalErrors == 0 ? QA_PASSED : QA_FAILED));
}
else
{
// Running CUDA kernel (bilateralFilter) in CUDA + OpenGL Visualization Mode
if ( cutCheckCmdLineFlag(argc, (const char **)argv, "device")) {
printf("[%s]\n", argv[0]);
printf(" Does not explicitly support -device=n in OpenGL mode\n");
printf(" To use -device=n, the sample must be running w/o OpenGL\n\n");
printf(" > %s -device=n -qatest\n", argv[0]);
printf("exiting...\n");
exit(0);
}
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
initGL( argc, argv );
if ( cutCheckCmdLineFlag(argc, (const char **)argv, "device")) {
cutilGLDeviceInit(argc, argv);
} else {
cudaGLSetGLDevice (cutGetMaxGflopsDeviceId() );
}
initCuda();
initOpenGL();
}
atexit(cleanup);
printf("Running Standard Demonstration with GLUT loop...\n\n");
printf("Press '+' and '-' to change number of iterations\n"
"Press LEFT and RIGHT change euclidean delta\n"
"Press UP and DOWN to change gaussian delta\n"
"Press '1' to show original image\n"
"Press '2' to show result\n\n");
glutMainLoop();
cutilDeviceReset();
shrQAFinishExit(argc, (const char **)argv, (g_TotalErrors == 0 ? QA_PASSED : QA_FAILED));
}
//////////////////////////////////////////////////////////////////////////////
// Program main
//////////////////////////////////////////////////////////////////////////////
int
main( int argc, char** argv)
{
bool bTestResults = true;
shrQAStart(argc, argv);
if( cutCheckCmdLineFlag(argc, (const char**)argv, "help") ) {
showHelp();
return 0;
}
shrLog("Run \"nbody -benchmark [-n=<numBodies>]\" to measure perfomance.\n");
shrLog("\t-fullscreen (run n-body simulation in fullscreen mode)\n");
shrLog("\t-fp64 (use double precision floating point values for simulation)\n");
shrLog("\t-numdevices=N (use first N CUDA devices for simulation)\n");
// shrLog("\t-hostmem (stores simulation data in host memory)\n");
// shrLog("\t-cpu (performs simulation on the host)\n");
shrLog("\n");
bFullscreen = (cutCheckCmdLineFlag(argc, (const char**) argv, "fullscreen") != 0);
if (bFullscreen)
bShowSliders = false;
benchmark = (cutCheckCmdLineFlag(argc, (const char**) argv, "benchmark") != 0);
compareToCPU = ((cutCheckCmdLineFlag(argc, (const char**) argv, "compare") != 0) ||
(cutCheckCmdLineFlag(argc, (const char**) argv, "qatest") != 0));
QATest = (cutCheckCmdLineFlag(argc, (const char**) argv, "qatest") != 0);
useHostMem = (cutCheckCmdLineFlag(argc, (const char**) argv, "hostmem") != 0);
fp64 = (cutCheckCmdLineFlag(argc, (const char**) argv, "fp64") != 0);
flopsPerInteraction = fp64 ? 30 : 20;
useCpu = (cutCheckCmdLineFlag(argc, (const char**) argv, "cpu") != 0);
cutGetCmdLineArgumenti(argc, (const char**) argv, "numdevices", &numDevsRequested);
// for multi-device we currently require using host memory -- the devices share
// data via the host
if (numDevsRequested > 1)
useHostMem = true;
int numDevsAvailable = 0;
bool customGPU = false;
cudaGetDeviceCount(&numDevsAvailable);
if (numDevsAvailable < numDevsRequested) {
shrLog("Error: only %d Devices available, %d requested. Exiting.\n", numDevsAvailable, numDevsRequested);
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
}
shrLog("> %s mode\n", bFullscreen ? "Fullscreen" : "Windowed");
shrLog("> Simulation data stored in %s memory\n", useHostMem ? "system" : "video" );
shrLog("> %s precision floating point simulation\n", fp64 ? "Double" : "Single");
shrLog("> %d Devices used for simulation\n", numDevsRequested);
int devID;
cudaDeviceProp props;
// Initialize GL and GLUT if necessary
if (!benchmark && !compareToCPU) {
initGL(&argc, argv);
initParameters();
}
if (useCpu) {
useHostMem = true;
compareToCPU = false;
bSupportDouble = true;
#ifdef OPENMP
shrLog("> Simulation with CPU using OpenMP\n");
#else
shrLog("> Simulation with CPU\n");
#endif
}
else
{
// Now choose the CUDA Device
// Either without GL interop:
if (benchmark || compareToCPU || useHostMem)
{
// Note if we are using host memory for the body system, we
// don't use CUDA-GL interop.
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") )
{
devID = cutilDeviceInit(argc, argv);
if (devID < 0) {
printf("exiting...\n");
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
}
customGPU = true;
}
else
{
devID = cutGetMaxGflopsDeviceId();
cudaSetDevice( devID );
}
}
else // or with GL interop:
{
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilGLDeviceInit(argc, argv);
customGPU = true;
} else {
devID = cutGetMaxGflopsDeviceId();
cudaGLSetGLDevice( devID );
}
}
cutilSafeCall(cudaGetDevice(&devID));
cutilSafeCall(cudaGetDeviceProperties(&props, devID));
bSupportDouble = true;
#if CUDART_VERSION < 4000
if (numDevsRequested > 1)
{
shrLog("MultiGPU n-body requires CUDA 4.0 or later\n");
cutilDeviceReset();
shrQAFinishExit(argc, (const char**)argv, QA_PASSED);
}
#endif
// Initialize devices
if(numDevsRequested > 1 && customGPU)
{
printf("You can't use --numdevices and --device at the same time.\n");
shrQAFinishExit(argc, (const char**)argv, QA_PASSED);
}
if(customGPU) {
cudaDeviceProp props;
cutilSafeCall(cudaGetDeviceProperties(&props, devID));
shrLog("> Compute %d.%d CUDA device: [%s]\n", props.major, props.minor, props.name);
}
else
{
for (int i = 0; i < numDevsRequested; i++)
{
cudaDeviceProp props;
cutilSafeCall(cudaGetDeviceProperties(&props, i));
shrLog("> Compute %d.%d CUDA device: [%s]\n", props.major, props.minor, props.name);
if (useHostMem)
{
#if CUDART_VERSION >= 2020
if(!props.canMapHostMemory)
{
fprintf(stderr, "Device %d cannot map host memory!\n", devID);
cutilDeviceReset();
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
}
if (numDevsRequested > 1)
cutilSafeCall(cudaSetDevice(i));
cutilSafeCall(cudaSetDeviceFlags(cudaDeviceMapHost));
#else
fprintf(stderr, "This CUDART version does not support <cudaDeviceProp.canMapHostMemory> field\n");
cutilDeviceReset();
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
#endif
}
}
// CC 1.2 and earlier do not support double precision
if (props.major*10 + props.minor <= 12)
bSupportDouble = false;
}
//if(numDevsRequested > 1)
// cutilSafeCall(cudaSetDevice(devID));
if (fp64 && !bSupportDouble) {
fprintf(stderr, "One or more of the requested devices does not support double precision floating-point\n");
cutilDeviceReset();
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
}
}
numIterations = 0;
p = 0;
q = 1;
cutGetCmdLineArgumenti(argc, (const char**) argv, "i", &numIterations);
cutGetCmdLineArgumenti(argc, (const char**) argv, "p", &p);
cutGetCmdLineArgumenti(argc, (const char**) argv, "q", &q);
if (p == 0) // p not set on command line
{
p = 256;
if (q * p > 256)
{
p = 256 / q;
shrLog("Setting p=%d, q=%d to maintain %d threads per block\n", p, q, 256);
}
}
// default number of bodies is #SMs * 4 * CTA size
if (useCpu)
#ifdef OPENMP
numBodies = 8192;
#else
numBodies = 4096;
#endif
else if (numDevsRequested == 1)
//////////////////////////////////////////////////////////////////////////////
// Program main
//////////////////////////////////////////////////////////////////////////////
int
main( int argc, char** argv)
{
printf("Run \"nbody -benchmark [-n=<numBodies>]\" to measure perfomance.\n\n");
bool benchmark =
(cutCheckCmdLineFlag(argc, (const char**) argv, "benchmark") != 0);
bool compareToCPU =
((cutCheckCmdLineFlag(argc, (const char**) argv, "compare") != 0) ||
!(cutCheckCmdLineFlag(argc, (const char**) argv, "noqatest") != 0));
bool regression =
(cutCheckCmdLineFlag(argc, (const char**) argv, "regression") != 0);
int devID;
cudaDeviceProp props;
// nBody has a mode that allows it to be run without using GL interop
if (benchmark || compareToCPU || regression) {
/*
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilDeviceInit(argc, argv);
} else {
devID = cutGetMaxGflopsDeviceId();
cudaSetDevice( devID );
} */
}
else
{
// This mode shows the OpenGL results rendered
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
glutInit(&argc, argv);
glutInitDisplayMode(GLUT_RGB | GLUT_DEPTH | GLUT_DOUBLE);
glutInitWindowSize(720, 480);
glutCreateWindow("CUDA n-body system");
GLenum err = glewInit();
if (GLEW_OK != err)
{
printf("GLEW Error: %s\n", glewGetErrorString(err));
}
else
{
#if defined(WIN32)
wglSwapIntervalEXT(0);
#elif defined(LINUX)
glxSwapIntervalSGI(0);
#endif
}
initGL();
initParameters();
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilGLDeviceInit(argc, argv);
} else {
devID = cutGetMaxGflopsDeviceId();
cudaGLSetGLDevice( devID );
}
}
// get number of SMs on this GPU
cutilSafeCall(cudaGetDevice(&devID));
cutilSafeCall(cudaGetDeviceProperties(&props, devID));
numIterations = 0;
int p = 256;
int q = 1;
cutGetCmdLineArgumenti(argc, (const char**) argv, "i", &numIterations);
cutGetCmdLineArgumenti(argc, (const char**) argv, "p", &p);
cutGetCmdLineArgumenti(argc, (const char**) argv, "q", &q);
// default number of bodies is #SMs * 4 * CTA size
numBodies = compareToCPU ? 4096 : p*q*4*props.multiProcessorCount;
cutGetCmdLineArgumenti(argc, (const char**) argv, "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 8192:
activeParams.m_clusterScale = 1.98f;
activeParams.m_velocityScale = 2.9f;
break;
default:
case 16384:
activeParams.m_clusterScale = 1.54f;
activeParams.m_velocityScale = 8.f;
break;
case 32768:
activeParams.m_clusterScale = 1.44f;
activeParams.m_velocityScale = 11.f;
break;
}
if (q * p > 256)
{
p = 256 / q;
printf("Setting p=%d, q=%d to maintain %d threads per block\n", p, q, 256);
}
if (q == 1 && numBodies < p)
{
p = numBodies;
}
init(numBodies, p, q, !(benchmark || compareToCPU));
reset(nbody, numBodies, NBODY_CONFIG_SHELL, !(benchmark || compareToCPU));
if (benchmark)
{
if (numIterations <= 0)
numIterations = 100;
runBenchmark(numIterations);
}
else if (compareToCPU || regression)
{
compareResults(regression, numBodies);
}
else
{
glutDisplayFunc(display);
glutReshapeFunc(reshape);
glutMouseFunc(mouse);
glutMotionFunc(motion);
glutKeyboardFunc(key);
glutSpecialFunc(special);
glutIdleFunc(idle);
cutilSafeCall(cudaEventRecord(startEvent, 0));
glutMainLoop();
}
if (nbodyCPU)
delete nbodyCPU;
if (nbodyCUDA)
delete nbodyCUDA;
if (hPos)
delete [] hPos;
if (hVel)
delete [] hVel;
if (hColor)
delete [] hColor;
cutilSafeCall(cudaEventDestroy(startEvent));
cutilSafeCall(cudaEventDestroy(stopEvent));
cutilCheckError(cutDeleteTimer(demoTimer));
return 0;
}
int main(int argc, char** argv)
{
// EDISON //////////////////////////////////////////////////////////////////
sigmaS = 7.0f;
sigmaR = 6.5f;
edison.minRegion = 20.0f;
cutLoadPPMub("image.ppm", &edison.inputImage_, &width, &height);
edison.meanShift();
cutSavePPMub("segmimage.ppm", edison.segmImage_, width, height);
cutSavePPMub("filtimage.ppm", edison.filtImage_, width, height);
unsigned char data[height * width];
memset(data, 0, height * width * sizeof(unsigned char));
for(int i = 0; i < edison.numBoundaries_; i++) {
data[edison.boundaries_[i]] = 255;
}
cutSavePGMub("bndyimage.pgm", data, width, height);
//return 0;
// EDISON //////////////////////////////////////////////////////////////////
if (argc > 1) {
if (cutCheckCmdLineFlag(argc, (const char **)argv, "help")) {
printHelp();
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "qatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt"))
{
g_bQAReadback = true;
fpsLimit = frameCheckNumber;
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "glverify"))
{
g_bOpenGLQA = true;
fpsLimit = frameCheckNumber;
}
if (cutCheckCmdLineFlag(argc, (const char **)argv, "fbo")) {
g_bFBODisplay = true;
fpsLimit = frameCheckNumber;
}
}
if (g_bQAReadback) {
runAutoTest(argc, argv);
} else {
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
initGL( argc, argv );
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if( cutCheckCmdLineFlag(argc, (const char**)argv, "device") ) {
cutilGLDeviceInit(argc, argv);
} else {
cudaGLSetGLDevice (cutGetMaxGflopsDeviceId() );
}
int device;
struct cudaDeviceProp prop;
cudaGetDevice( &device );
cudaGetDeviceProperties( &prop, device );
if(!strncmp( "Tesla", prop.name, 5 )) {
printf("This sample needs a card capable of OpenGL and display.\n");
printf("Please choose a different device with the -device=x argument.\n");
cudaThreadExit();
cutilExit(argc, argv);
}
cutilCheckError(cutCreateTimer(&timer));
cutilCheckError(cutResetTimer(timer));
glutDisplayFunc(display);
glutKeyboardFunc(keyboard);
glutReshapeFunc(reshape);
glutIdleFunc(idle);
if (g_bOpenGLQA) {
loadDefaultImage( argv[0] );
}
if (argc > 1) {
char *filename;
if (cutGetCmdLineArgumentstr(argc, (const char **)argv, "file", &filename)) {
initializeData(filename);
}
} else {
loadDefaultImage( argv[0]);
}
// If code is not printing the USage, then we execute this path.
if (!bQuit) {
if (g_bOpenGLQA) {
g_CheckRender = new CheckBackBuffer(wWidth, wHeight, 4);
g_CheckRender->setPixelFormat(GL_BGRA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
}
printf("I: display image\n");
printf("T: display Sobel edge detection (computed with tex)\n");
printf("S: display Sobel edge detection (computed with tex+shared memory)\n");
printf("Use the '-' and '=' keys to change the brightness.\n");
fflush(stdout);
atexit(cleanup);
glutMainLoop();
}
}
cudaThreadExit();
cutilExit(argc, argv);
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main( int argc, char** argv)
{
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if (!cutCheckCmdLineFlag(argc, (const char **)argv, "noqatest") ||
cutCheckCmdLineFlag(argc, (const char **)argv, "noprompt"))
{
g_bQAReadback = true;
fpsLimit = frameCheckNumber;
}
if (argc > 1) {
if (cutCheckCmdLineFlag(argc, (const char **)argv, "glverify")) {
g_bOpenGLQA = true;
fpsLimit = frameCheckNumber;
}
}
printf("[%s] ", sSDKsample);
if (g_bQAReadback) printf("(Automated Testing)\n");
if (g_bOpenGLQA) printf("(OpenGL Readback)\n");
// Get the path of the filename
char *filename;
if (cutGetCmdLineArgumentstr(argc, (const char**) argv, "image", &filename)) {
image_filename = filename;
}
// load image
char* image_path = cutFindFilePath(image_filename, argv[0]);
if (image_path == 0) {
fprintf(stderr, "Error finding image file '%s'\n", image_filename);
cudaThreadExit();
exit(EXIT_FAILURE);
}
cutilCheckError( cutLoadPPM4ub(image_path, (unsigned char **) &h_img, &width, &height));
if (!h_img) {
printf("Error opening file '%s'\n", image_path);
cudaThreadExit();
exit(-1);
}
printf("Loaded '%s', %d x %d pixels\n", image_path, width, height);
cutGetCmdLineArgumenti(argc, (const char**) argv, "threads", &nthreads);
cutGetCmdLineArgumentf(argc, (const char**) argv, "sigma", &sigma);
runBenchmark = cutCheckCmdLineFlag(argc, (const char**) argv, "bench");
int device;
struct cudaDeviceProp prop;
cudaGetDevice( &device );
cudaGetDeviceProperties( &prop, device );
if( !strncmp( "Tesla", prop.name, 5 ) ) {
printf("Tesla card detected, running the test in benchmark mode (no OpenGL display)\n");
// runBenchmark = CUTTrue;
g_bQAReadback = true;
}
// Benchmark or AutoTest mode detected, no OpenGL
if (runBenchmark == CUTTrue || g_bQAReadback) {
if( cutCheckCmdLineFlag( argc, (const char **)argv, "device" ) )
cutilDeviceInit( argc, argv );
else
cudaSetDevice( cutGetMaxGflopsDeviceId() );
} else {
// First initialize OpenGL context, so we can properly set the GL for CUDA.
// This is necessary in order to achieve optimal performance with OpenGL/CUDA interop.
initGL(argc, argv);
if( cutCheckCmdLineFlag( argc, (const char **)argv, "device" ) )
cutilGLDeviceInit( argc, argv );
else
cudaGLSetGLDevice( cutGetMaxGflopsDeviceId() );
}
initCudaBuffers();
if (g_bOpenGLQA) {
g_CheckRender = new CheckBackBuffer(width, height, 4);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
}
if (g_bQAReadback) {
// This is the automated testing path
g_CheckRender = new CheckBackBuffer(width, height, 4, false);
g_CheckRender->setPixelFormat(GL_RGBA);
g_CheckRender->setExecPath(argv[0]);
g_CheckRender->EnableQAReadback(true);
runAutoTest(argc, argv);
cleanup();
cudaThreadExit();
cutilExit(argc, argv);
}
if (runBenchmark) {
benchmark(100);
cleanup();
cudaThreadExit();
exit(0);
}
initGLBuffers();
atexit(cleanup);
glutMainLoop();
cudaThreadExit();
cutilExit(argc, argv);
}
