void CinderCUDASampleApp::setup()
{
image_width = 640;
image_height = 480;
char argv[] = { "cinderCudaSample" };
findCudaGLDevice(1, NULL);
// init GL Buffers
ci::gl::Fbo::Format format;
format.setColorInternalFormat(GL_RGBA);
format.setWrapS(GL_CLAMP_TO_EDGE);
format.setWrapT(GL_CLAMP_TO_EDGE);
format.setMinFilter(GL_NEAREST);
format.setMagFilter(GL_NEAREST);
format.enableMipmapping(false);
format.enableDepthBuffer(false);
mFbo = ci::gl::Fbo(image_width, image_height, format);
cudaGraphicsGLRegisterImage(&cuda_tex_result_resource, mFbo.getId(), GL_TEXTURE_2D, cudaGraphicsMapFlagsWriteDiscard);
// init CUDA Buffers
num_texels = image_width * image_height;
num_values = num_texels * 4;
size_tex_data = sizeof(GLubyte) * num_values;
cudaMalloc((void **)&cuda_dest_resource, size_tex_data);
}
// General initialization call for CUDA Device
int chooseCudaDevice(int argc, char **argv, bool bUseOpenGL)
{
int result = 0;
if (bUseOpenGL)
{
result = findCudaGLDevice(argc, (const char **)argv);
}
else
{
result = findCudaDevice(argc, (const char **)argv);
}
return result;
}
int main(int argc, char **argv)
{
int devID;
cudaDeviceProp deviceProps;
printf("%s Starting...\n\n", sSDKname);
printf("[%s] - [OpenGL/CUDA simulation] starting...\n", sSDKname);
// 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 (false == initGL(&argc, argv))
{
exit(EXIT_SUCCESS);
}
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
devID = findCudaGLDevice(argc, (const char **)argv);
// get number of SMs on this GPU
checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));
printf("CUDA device [%s] has %d Multi-Processors\n",
deviceProps.name, deviceProps.multiProcessorCount);
// automated build testing harness
if (checkCmdLineFlag(argc, (const char **)argv, "file"))
{
getCmdLineArgumentString(argc, (const char **)argv, "file", &ref_file);
}
// Allocate and initialize host data
GLint bsize;
sdkCreateTimer(&timer);
sdkResetTimer(&timer);
hvfield = (cData *)malloc(sizeof(cData) * DS);
memset(hvfield, 0, sizeof(cData) * DS);
// Allocate and initialize device data
cudaMallocPitch((void **)&dvfield, &tPitch, sizeof(cData)*DIM, DIM);
cudaMemcpy(dvfield, hvfield, sizeof(cData) * DS,
cudaMemcpyHostToDevice);
// Temporary complex velocity field data
cudaMalloc((void **)&vxfield, sizeof(cData) * PDS);
cudaMalloc((void **)&vyfield, sizeof(cData) * PDS);
setupTexture(DIM, DIM);
bindTexture();
// Create particle array
particles = (cData *)malloc(sizeof(cData) * DS);
memset(particles, 0, sizeof(cData) * DS);
initParticles(particles, DIM, DIM);
// Create CUFFT transform plan configuration
cufftPlan2d(&planr2c, DIM, DIM, CUFFT_R2C);
cufftPlan2d(&planc2r, DIM, DIM, CUFFT_C2R);
// TODO: update kernels to use the new unpadded memory layout for perf
// rather than the old FFTW-compatible layout
cufftSetCompatibilityMode(planr2c, CUFFT_COMPATIBILITY_FFTW_PADDING);
cufftSetCompatibilityMode(planc2r, CUFFT_COMPATIBILITY_FFTW_PADDING);
glGenBuffersARB(1, &vbo);
glBindBufferARB(GL_ARRAY_BUFFER_ARB, vbo);
glBufferDataARB(GL_ARRAY_BUFFER_ARB, sizeof(cData) * DS,
particles, GL_DYNAMIC_DRAW_ARB);
glGetBufferParameterivARB(GL_ARRAY_BUFFER_ARB, GL_BUFFER_SIZE_ARB, &bsize);
if (bsize != (sizeof(cData) * DS))
goto EXTERR;
glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);
checkCudaErrors(cudaGraphicsGLRegisterBuffer(&cuda_vbo_resource, vbo, cudaGraphicsMapFlagsNone));
getLastCudaError("cudaGraphicsGLRegisterBuffer failed");
if (ref_file)
{
autoTest(argv);
cleanup();
cudaDeviceReset();
printf("[fluidsGL] - Test Results: %d Failures\n", g_TotalErrors);
exit(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
else
{
atexit(cleanup);
glutMainLoop();
}
cudaDeviceReset();
if (!ref_file)
{
exit(EXIT_SUCCESS);
}
return 0;
EXTERR:
printf("Failed to initialize GL extensions.\n");
cudaDeviceReset();
exit(EXIT_FAILURE);
}
//////////////////////////////////////////////////////////////////////////////
// Program main
//////////////////////////////////////////////////////////////////////////////
int
main(int argc, char **argv)
{
bool bTestResults = true;
if (checkCmdLineFlag(argc, (const char **)argv, "help"))
{
printf("\n> Command line options\n");
showHelp();
return 0;
}
printf("Run \"nbody -benchmark [-numbodies=<numBodies>]\" to measure perfomance.\n");
showHelp();
bFullscreen = (checkCmdLineFlag(argc, (const char **) argv, "fullscreen") != 0);
if (bFullscreen)
{
bShowSliders = false;
}
benchmark = (checkCmdLineFlag(argc, (const char **) argv, "benchmark") != 0);
compareToCPU = ((checkCmdLineFlag(argc, (const char **) argv, "compare") != 0) ||
(checkCmdLineFlag(argc, (const char **) argv, "qatest") != 0));
QATest = (checkCmdLineFlag(argc, (const char **) argv, "qatest") != 0);
useHostMem = (checkCmdLineFlag(argc, (const char **) argv, "hostmem") != 0);
fp64 = (checkCmdLineFlag(argc, (const char **) argv, "fp64") != 0);
flopsPerInteraction = fp64 ? 30 : 20;
useCpu = (checkCmdLineFlag(argc, (const char **) argv, "cpu") != 0);
if (checkCmdLineFlag(argc, (const char **)argv, "numdevices"))
{
numDevsRequested = getCmdLineArgumentInt(argc, (const char **) argv, "numdevices");
if (numDevsRequested < 1)
{
printf("Error: \"number of CUDA devices\" specified %d is invalid. Value should be >= 1\n", numDevsRequested);
exit(bTestResults ? EXIT_SUCCESS : EXIT_FAILURE);
}
else
{
printf("number of CUDA devices = %d\n", 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)
{
printf("Error: only %d Devices available, %d requested. Exiting.\n", numDevsAvailable, numDevsRequested);
exit(EXIT_SUCCESS);
}
printf("> %s mode\n", bFullscreen ? "Fullscreen" : "Windowed");
printf("> Simulation data stored in %s memory\n", useHostMem ? "system" : "video");
printf("> %s precision floating point simulation\n", fp64 ? "Double" : "Single");
printf("> %d Devices used for simulation\n", numDevsRequested);
int devID;
cudaDeviceProp props;
if (useCpu)
{
useHostMem = true;
compareToCPU = false;
bSupportDouble = true;
#ifdef OPENMP
printf("> Simulation with CPU using OpenMP\n");
#else
printf("> Simulation with CPU\n");
#endif
}
// Initialize GL and GLUT if necessary
if (!benchmark && !compareToCPU)
{
initGL(&argc, argv);
initParameters();
}
if(!useCpu)
{
// 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 (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
customGPU = true;
}
devID = findCudaDevice(argc, (const char **)argv);
}
else // or with GL interop:
{
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
customGPU = true;
}
devID = findCudaGLDevice(argc, (const char **)argv);
}
checkCudaErrors(cudaGetDevice(&devID));
checkCudaErrors(cudaGetDeviceProperties(&props, devID));
bSupportDouble = true;
#if CUDART_VERSION < 4000
if (numDevsRequested > 1)
{
printf("MultiGPU n-body requires CUDA 4.0 or later\n");
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
#endif
// Initialize devices
if (numDevsRequested > 1 && customGPU)
{
printf("You can't use --numdevices and --device at the same time.\n");
exit(EXIT_SUCCESS);
}
if (customGPU)
{
cudaDeviceProp props;
checkCudaErrors(cudaGetDeviceProperties(&props, devID));
printf("> Compute %d.%d CUDA device: [%s]\n", props.major, props.minor, props.name);
}
else
{
for (int i = 0; i < numDevsRequested; i++)
{
cudaDeviceProp props;
checkCudaErrors(cudaGetDeviceProperties(&props, i));
printf("> 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);
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
if (numDevsRequested > 1)
{
checkCudaErrors(cudaSetDevice(i));
}
checkCudaErrors(cudaSetDeviceFlags(cudaDeviceMapHost));
#else
fprintf(stderr, "This CUDART version does not support <cudaDeviceProp.canMapHostMemory> field\n");
cudaDeviceReset();
exit(EXIT_SUCCESS);
#endif
}
}
// CC 1.2 and earlier do not support double precision
if (props.major*10 + props.minor <= 12)
{
bSupportDouble = false;
}
}
//if(numDevsRequested > 1)
// checkCudaErrors(cudaSetDevice(devID));
if (fp64 && !bSupportDouble)
{
fprintf(stderr, "One or more of the requested devices does not support double precision floating-point\n");
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
}
numIterations = 0;
p = 0;
q = 1;
if (checkCmdLineFlag(argc, (const char **)argv, "i"))
{
numIterations = getCmdLineArgumentInt(argc, (const char **)argv, "i");
}
if (checkCmdLineFlag(argc, (const char **) argv, "p"))
{
p = getCmdLineArgumentInt(argc, (const char **)argv, "p");
}
if (checkCmdLineFlag(argc, (const char **) argv, "q"))
{
q = getCmdLineArgumentInt(argc, (const char **)argv, "q");
}
if (p == 0) // p not set on command line
{
p = 256;
if (q * p > 256)
{
p = 256 / q;
printf("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)
int main(int argc, char **argv)
{
int devID;
cudaDeviceProp deviceProps;
printf("%s Starting...\n\n", sSDKname);
printf("[%s] - [OpenGL/CUDA simulation] starting...\n", sSDKname);
// 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 (false == initGL(&argc, argv))
{
exit(EXIT_SUCCESS);
}
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
#ifndef OPTIMUS
devID = findCudaGLDevice(argc, (const char **)argv);
#else
devID = gpuGetMaxGflopsDeviceId();
#endif
// get number of SMs on this GPU
checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));
printf("CUDA device [%s] has %d Multi-Processors\n",
deviceProps.name, deviceProps.multiProcessorCount);
// automated build testing harness
if (checkCmdLineFlag(argc, (const char **)argv, "file"))
{
getCmdLineArgumentString(argc, (const char **)argv, "file", &ref_file);
}
// Allocate and initialize host data
GLint bsize;
sdkCreateTimer(&timer);
sdkResetTimer(&timer);
hvfield = (cData *)malloc(sizeof(cData) * DS);
memset(hvfield, 0, sizeof(cData) * DS);
// Allocate and initialize device data
cudaMallocPitch((void **)&dvfield, &tPitch, sizeof(cData)*DIM, DIM);
cudaMemcpy(dvfield, hvfield, sizeof(cData) * DS,
cudaMemcpyHostToDevice);
// Temporary complex velocity field data
cudaMalloc((void **)&vxfield, sizeof(cData) * PDS);
cudaMalloc((void **)&vyfield, sizeof(cData) * PDS);
setupTexture(DIM, DIM);
bindTexture();
// Create particle array in host memory
particles = (cData *)malloc(sizeof(cData) * DS);
memset(particles, 0, sizeof(cData) * DS);
#ifdef BROADCAST
int step = 1;
// Broadcasted visualization stepping.
if (argc > 3)
step = atoi(argv[3]);
// Create additional space to store particle packets
// for broadcasting.
wstep = step; hstep = step;
int npackets = sizeof(float) * (DIM / wstep) * (DIM / hstep) / UdpBroadcastServer::PacketSize;
if (sizeof(float) * (DIM / wstep) * (DIM / hstep) % UdpBroadcastServer::PacketSize)
npackets++;
packets = (char*)malloc(npackets *
(UdpBroadcastServer::PacketSize + sizeof(unsigned int)));
#endif
initParticles(particles, DIM, DIM);
#if defined(OPTIMUS) || defined(BROADCAST)
// Create particle array in device memory
cudaMalloc((void **)&particles_gpu, sizeof(cData) * DS);
cudaMemcpy(particles_gpu, particles, sizeof(cData) * DS, cudaMemcpyHostToDevice);
#endif
// Create CUFFT transform plan configuration
cufftPlan2d(&planr2c, DIM, DIM, CUFFT_R2C);
cufftPlan2d(&planc2r, DIM, DIM, CUFFT_C2R);
// TODO: update kernels to use the new unpadded memory layout for perf
// rather than the old FFTW-compatible layout
cufftSetCompatibilityMode(planr2c, CUFFT_COMPATIBILITY_FFTW_PADDING);
cufftSetCompatibilityMode(planc2r, CUFFT_COMPATIBILITY_FFTW_PADDING);
glGenBuffersARB(1, &vbo);
glBindBufferARB(GL_ARRAY_BUFFER_ARB, vbo);
glBufferDataARB(GL_ARRAY_BUFFER_ARB, sizeof(cData) * DS,
particles, GL_DYNAMIC_DRAW_ARB);
glGetBufferParameterivARB(GL_ARRAY_BUFFER_ARB, GL_BUFFER_SIZE_ARB, &bsize);
if (bsize != (sizeof(cData) * DS))
goto EXTERR;
glBindBufferARB(GL_ARRAY_BUFFER_ARB, 0);
#ifndef OPTIMUS
checkCudaErrors(cudaGraphicsGLRegisterBuffer(&cuda_vbo_resource, vbo, cudaGraphicsMapFlagsNone));
getLastCudaError("cudaGraphicsGLRegisterBuffer failed");
#endif
if (ref_file)
{
autoTest(argv);
cleanup();
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
printf("[fluidsGL] - Test Results: %d Failures\n", g_TotalErrors);
exit(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
else
{
#ifdef BROADCAST
const char *sv_addr = "127.0.0:9097";
const char *bc_addr = "127.255.255.2:9097";
// Server address
if (argc > 2)
sv_addr = argv[2];
// Broadcast address
if (argc > 1)
bc_addr = argv[1];
server.reset(new UdpBroadcastServer(sv_addr, bc_addr));
// Listen to clients' feedbacks in a separate thread.
{
pthread_t tid;
pthread_create(&tid, NULL, &feedback_listener, &step);
}
// Broadcast the particles state in a separate thread.
{
pthread_t tid;
pthread_create(&tid, NULL, &broadcaster, &step);
}
#endif
#if defined (__APPLE__) || defined(MACOSX)
atexit(cleanup);
#else
glutCloseFunc(cleanup);
#endif
glutMainLoop();
}
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
if (!ref_file)
{
exit(EXIT_SUCCESS);
}
return 0;
EXTERR:
printf("Failed to initialize GL extensions.\n");
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
exit(EXIT_FAILURE);
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main(int argc, char **argv)
{
pArgc = &argc;
pArgv = argv;
char *ref_file = NULL;
#if defined(__linux__)
setenv ("DISPLAY", ":0", 0);
#endif
printf("%s Starting...\n\n", sSDKsample);
printf("NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.\n\n");
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if (argc > 1)
{
if (checkCmdLineFlag(argc, (const char **)argv, "file"))
{
getCmdLineArgumentString(argc, (const char **)argv, "file", &ref_file);
fpsLimit = frameCheckNumber;
}
}
// Get the path of the filename
char *filename;
if (getCmdLineArgumentString(argc, (const char **) argv, "image", &filename))
{
image_filename = filename;
}
// load image
char *image_path = sdkFindFilePath(image_filename, argv[0]);
if (image_path == NULL)
{
fprintf(stderr, "Error unable to find and load image file: '%s'\n", image_filename);
exit(EXIT_FAILURE);
}
sdkLoadPPM4ub(image_path, (unsigned char **)&h_img, &width, &height);
if (!h_img)
{
printf("Error unable to load PPM file: '%s'\n", image_path);
exit(EXIT_FAILURE);
}
printf("Loaded '%s', %d x %d pixels\n", image_path, width, height);
if (checkCmdLineFlag(argc, (const char **)argv, "threads"))
{
nthreads = getCmdLineArgumentInt(argc, (const char **) argv, "threads");
}
if (checkCmdLineFlag(argc, (const char **)argv, "sigma"))
{
sigma = getCmdLineArgumentFloat(argc, (const char **) argv, "sigma");
}
runBenchmark = checkCmdLineFlag(argc, (const char **) argv, "benchmark");
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 = true;
runBenchmark = true;
}
// Benchmark or AutoTest mode detected, no OpenGL
if (runBenchmark == true || ref_file != NULL)
{
findCudaDevice(argc, (const char **)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);
findCudaGLDevice(argc, (const char **)argv);
}
initCudaBuffers();
if (ref_file)
{
printf("(Automated Testing)\n");
bool testPassed = runSingleTest(ref_file, argv[0]);
cleanup();
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
exit(testPassed ? EXIT_SUCCESS : EXIT_FAILURE);
}
if (runBenchmark)
{
printf("(Run Benchmark)\n");
benchmark(100);
cleanup();
// cudaDeviceReset causes the driver to clean up all state. While
// not mandatory in normal operation, it is good practice. It is also
// needed to ensure correct operation when the application is being
// profiled. Calling cudaDeviceReset causes all profile data to be
// flushed before the application exits
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
initGLBuffers();
glutMainLoop();
exit(EXIT_SUCCESS);
}