//--------------------------------------------------------------------------------
int main(int argc, char** argv)
{
std::cout << "We are alive!" << std::endl;
// set device
cudaDeviceProp device_prop;
int dev_id = findCudaDevice(argc, (const char **) argv);
checkCudaErrors(cudaGetDeviceProperties(&device_prop, dev_id));
std::cout << "set cuda device id: " << dev_id << std::endl;
cv::Mat im1 = cv::imread("/home/mwerlberger/data/std/Growth_of_cubic_bacteria_25x16.jpg");
cv::Mat scaled;
cv::resize(im1, scaled, cv::Size(), 0.5, 0.5);
std::cout << "im1 size: " << im1.cols << "x" << im1.rows << std::endl;
// copy im1 to global memory
uchar* d_im1;
const size_t size = im1.rows * im1.cols * sizeof(float);
cv::imshow("input (scaled)", scaled);
// //cv::imshow("result", im1);
cv::waitKey();
return EXIT_SUCCESS;
}
void runAutoTest(int argc, char **argv, const char *filename, int kernel_param)
{
printf("[%s] - (automated testing w/ readback)\n", sSDKsample);
int devID = findCudaDevice(argc, (const char **)argv);
// First load the image, so we know what the size of the image (imageW and imageH)
printf("Allocating host and CUDA memory and loading image file...\n");
const char *image_path = sdkFindFilePath("portrait_noise.bmp", argv[0]);
if (image_path == NULL)
{
printf("imageDenoisingGL was unable to find and load image file <portrait_noise.bmp>.\nExiting...\n");
exit(EXIT_FAILURE);
}
LoadBMPFile(&h_Src, &imageW, &imageH, image_path);
printf("Data init done.\n");
checkCudaErrors(CUDA_MallocArray(&h_Src, imageW, imageH));
TColor *d_dst = NULL;
unsigned char *h_dst = NULL;
checkCudaErrors(cudaMalloc((void **)&d_dst, imageW*imageH*sizeof(TColor)));
h_dst = (unsigned char *)malloc(imageH*imageW*4);
{
g_Kernel = kernel_param;
printf("[AutoTest]: %s <%s>\n", sSDKsample, filterMode[g_Kernel]);
checkCudaErrors(CUDA_Bind2TextureArray());
runImageFilters(d_dst);
checkCudaErrors(CUDA_UnbindTexture());
checkCudaErrors(cudaDeviceSynchronize());
checkCudaErrors(cudaMemcpy(h_dst, d_dst, imageW*imageH*sizeof(TColor), cudaMemcpyDeviceToHost));
sdkSavePPM4ub(filename, h_dst, imageW, imageH);
}
checkCudaErrors(CUDA_FreeArray());
free(h_Src);
checkCudaErrors(cudaFree(d_dst));
free(h_dst);
printf("\n[%s] -> Kernel %d, Saved: %s\n", sSDKsample, kernel_param, filename);
// 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(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
// 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;
}
void runAutoTest(int argc, char **argv, const char *dump_filename, eFilterMode filter_mode)
{
cudaDeviceProp deviceProps;
int devID = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&deviceProps, devID));
printf("[%s] (automated testing w/ readback)\n", sSDKsample);
printf("CUDA device [%s] has %d Multi-Processors\n", deviceProps.name, deviceProps.multiProcessorCount);
loadImageData(argc, argv);
uchar4 *d_output;
checkCudaErrors(cudaMalloc((void **)&d_output, imageWidth*imageHeight*4));
unsigned int *h_result = (unsigned int *)malloc(width * height * sizeof(unsigned int));
printf("AutoTest: %s Filter Mode: <%s>\n", sSDKsample, sFilterMode[g_FilterMode]);
render(imageWidth, imageHeight,
tx, ty, scale, cx, cy,
blockSize, gridSize, filter_mode, d_output);
// check if kernel execution generated an error
getLastCudaError("Error: render (bicubicTexture) Kernel execution FAILED");
checkCudaErrors(cudaDeviceSynchronize());
cudaMemcpy(h_result, d_output, imageWidth*imageHeight*4, cudaMemcpyDeviceToHost);
sdkSavePPM4ub(dump_filename, (unsigned char *)h_result, imageWidth, imageHeight);
checkCudaErrors(cudaFree(d_output));
free(h_result);
// 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();
}
inline int cudaDeviceInit(int argc, const char **argv)
{
int deviceCount;
checkCudaErrors(cudaGetDeviceCount(&deviceCount));
if (deviceCount == 0)
{
std::cerr << "CUDA error: no devices supporting CUDA." << std::endl;
exit(EXIT_FAILURE);
}
int dev = findCudaDevice(argc, argv);
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, dev);
std::cerr << "cudaSetDevice GPU" << dev << " = " << deviceProp.name << std::endl;
checkCudaErrors(cudaSetDevice(dev));
return dev;
}
int main(int argc, char **argv)
{
#ifdef GPU
int dev = findCudaDevice(argc, (const char **) argv);
if (dev == -1)
return 0;
if (cublasCreate(&handle) != CUBLAS_STATUS_SUCCESS)
{
fprintf(stdout, "CUBLAS initialization failed!\n");
cudaDeviceReset();
exit(EXIT_FAILURE);
}
#endif // GPU
char input_file_name[1024];
char model_file_name[1024];
const char *error_msg;
parse_command_line(argc, argv, input_file_name, model_file_name);
time_t t1 = clock();
read_problem(input_file_name);
time_t t2 = clock();
printf("reading the input file took %f seconds.\n", float(t2-t1)/CLOCKS_PER_SEC);
error_msg = check_parameter(&prob,¶m);
if(error_msg)
{
fprintf(stderr,"ERROR: %s\n",error_msg);
exit(1);
}
if(flag_cross_validation)
{
do_cross_validation();
}
else
{
model_=train(&prob, ¶m);
if(save_model(model_file_name, model_))
{
fprintf(stderr,"can't save model to file %s\n",model_file_name);
exit(1);
}
free_and_destroy_model(&model_);
}
destroy_param(¶m);
free(prob.y);
free(prob.x);
free(x_space);
free(line);
#ifdef GPU
cublasDestroy(handle);
cudaDeviceReset();
#endif // GPU
printf("reading the input file took %f seconds.\n", float(t2-t1)/CLOCKS_PER_SEC);
return 0;
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
// start logs
int devID;
char *ref_file = NULL;
printf("%s Starting...\n\n", argv[0]);
#if defined(__linux__)
setenv ("DISPLAY", ":0", 0);
#endif
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if (argc > 1)
{
if (checkCmdLineFlag(argc, (const char **)argv, "radius"))
{
filter_radius = getCmdLineArgumentInt(argc, (const char **) argv, "radius");
}
if (checkCmdLineFlag(argc, (const char **)argv, "passes"))
{
iterations = getCmdLineArgumentInt(argc, (const char **)argv, "passes");
}
if (checkCmdLineFlag(argc, (const char **)argv, "file"))
{
getCmdLineArgumentString(argc, (const char **)argv, "file", (char **)&ref_file);
}
}
// load image to process
loadImageData(argc, argv);
if (checkCmdLineFlag(argc, (const char **)argv, "benchmark"))
{
// This is a separate mode of the sample, where we are benchmark the kernels for performance
devID = findCudaDevice(argc, (const char **)argv);
// Running CUDA kernels (bilateralfilter) in Benchmarking mode
g_TotalErrors += runBenchmark(argc, argv);
// 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(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
else if (checkCmdLineFlag(argc, (const char **)argv, "radius") ||
checkCmdLineFlag(argc, (const char **)argv, "passes"))
{
// This overrides the default mode. Users can specify the radius used by the filter kernel
devID = findCudaDevice(argc, (const char **)argv);
g_TotalErrors += runSingleTest(ref_file, argv[0]);
// 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(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
else
{
// Default mode running with OpenGL visualization and in automatic mode
// the output automatically changes animation
printf("\n");
// 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, (char **)argv);
int dev = findCapableDevice(argc, argv);
if (dev != -1)
{
dev = gpuGLDeviceInit(argc, (const char **)argv);
if (dev == -1)
{
exit(EXIT_FAILURE);
}
}
else
{
// 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);
}
// Now we can create a CUDA context and bind it to the OpenGL context
initCuda();
initGLResources();
// sets the callback function so it will call cleanup upon exit
#if defined (__APPLE__) || defined(MACOSX)
atexit(cleanup);
#else
glutCloseFunc(cleanup);
#endif
printf("Running Standard Demonstration with GLUT loop...\n\n");
printf("Press '+' and '-' to change filter width\n"
"Press ']' and '[' to change number of iterations\n"
"Press 'e' and 'E' to change Euclidean delta\n"
"Press 'g' and 'G' to changle Gaussian delta\n"
"Press 'a' or 'A' to change Animation mode ON/OFF\n\n");
// Main OpenGL loop that will run visualization for every vsync
glutMainLoop();
}
}
int
main(int argc, char **argv)
{
pArgc = &argc;
pArgv = argv;
char *ref_file = NULL;
printf("%s Starting...\n\n", sSDKsample);
//start logs
if (checkCmdLineFlag(argc, (const char **)argv, "help"))
{
printHelp();
exit(EXIT_SUCCESS);
}
if (checkCmdLineFlag(argc, (const char **)argv, "file"))
{
fpsLimit = frameCheckNumber;
getCmdLineArgumentString(argc, (const char **)argv, "file", &ref_file);
}
if (ref_file)
{
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
int device = findCudaDevice(argc, (const char **)argv);
if (device < 0)
{
printf("No CUDA Capable devices found, exiting...\n");
exit(EXIT_SUCCESS);
}
checkDeviceMeetComputeSpec(argc, argv);
}
else
{
int dev = findCapableDevice(argc, argv);
if (dev != -1)
{
cudaSetDevice(dev);
}
else
{
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
}
}
else
{
if (checkCmdLineFlag(argc, (const char **)argv, "device"))
{
printf(" This SDK does not explicitly support -device=n when running with OpenGL.\n");
printf(" When specifying -device=n (n=0,1,2,....) the sample must not use OpenGL.\n");
printf(" See details below to run without OpenGL:\n\n");
printf(" > %s -device=n -file=output.bin\n\n", argv[0]);
printf("exiting...\n");
exit(EXIT_SUCCESS);
}
// 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);
int dev = findCapableDevice(argc, argv);
if (dev != -1)
{
cudaGLSetGLDevice(dev);
}
else
{
exit(EXIT_SUCCESS);
}
}
// load volume data
initData(argc, argv);
printf(
"Press \n"
" 'SPACE' to toggle animation\n"
" 'p' to toggle pre-integrated transfer function\n"
" '+' and '-' to change density (0.01 increments)\n"
" ']' and '[' to change brightness\n"
" ';' and ''' to modify transfer function offset\n"
" '.' and ',' to modify transfer function scale\n\n");
if (ref_file)
{
runSingleTest(ref_file, argv[0]);
}
else
{
// This is the normal rendering path for VolumeRender
glutDisplayFunc(display);
glutKeyboardFunc(keyboard);
glutMouseFunc(mouse);
glutMotionFunc(motion);
glutReshapeFunc(reshape);
glutIdleFunc(idle);
initPixelBuffer();
atexit(cleanup);
glutMainLoop();
}
cudaDeviceReset();
}
int main(int argc, char **argv)
{
int M = 0, N = 0, nz = 0, *I = NULL, *J = NULL;
float *val = NULL;
const float tol = 1e-5f;
const int max_iter = 10000;
float *x;
float *rhs;
float a, b, na, r0, r1;
int *d_col, *d_row;
float *d_val, *d_x, dot;
float *d_r, *d_p, *d_Ax;
int k;
float alpha, beta, alpham1;
shrQAStart(argc, argv);
// This will pick the best possible CUDA capable device
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
if (devID < 0) {
printf("exiting...\n");
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
exit(0);
}
checkCudaErrors( cudaGetDeviceProperties(&deviceProp, devID) );
// Statistics about the GPU device
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
int version = (deviceProp.major * 0x10 + deviceProp.minor);
if(version < 0x11)
{
printf("%s: requires a minimum CUDA compute 1.1 capability\n", sSDKname);
cudaDeviceReset();
shrQAFinishExit(argc, (const char **)argv, QA_PASSED);
}
/* Generate a random tridiagonal symmetric matrix in CSR format */
M = N = 1048576;
nz = (N-2)*3 + 4;
I = (int*)malloc(sizeof(int)*(N+1));
J = (int*)malloc(sizeof(int)*nz);
val = (float*)malloc(sizeof(float)*nz);
genTridiag(I, J, val, N, nz);
x = (float*)malloc(sizeof(float)*N);
rhs = (float*)malloc(sizeof(float)*N);
for (int i = 0; i < N; i++) {
rhs[i] = 1.0;
x[i] = 0.0;
}
/* Get handle to the CUBLAS context */
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
if ( checkCublasStatus (cublasStatus, "!!!! CUBLAS initialization error\n") ) return EXIT_FAILURE;
/* Get handle to the CUSPARSE context */
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
if ( checkCusparseStatus (cusparseStatus, "!!!! CUSPARSE initialization error\n") ) return EXIT_FAILURE;
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
if ( checkCusparseStatus (cusparseStatus, "!!!! CUSPARSE cusparseCreateMatDescr error\n") ) return EXIT_FAILURE;
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
checkCudaErrors( cudaMalloc((void**)&d_col, nz*sizeof(int)) );
checkCudaErrors( cudaMalloc((void**)&d_row, (N+1)*sizeof(int)) );
checkCudaErrors( cudaMalloc((void**)&d_val, nz*sizeof(float)) );
checkCudaErrors( cudaMalloc((void**)&d_x, N*sizeof(float)) );
checkCudaErrors( cudaMalloc((void**)&d_r, N*sizeof(float)) );
checkCudaErrors( cudaMalloc((void**)&d_p, N*sizeof(float)) );
checkCudaErrors( cudaMalloc((void**)&d_Ax, N*sizeof(float)) );
cudaMemcpy(d_col, J, nz*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_row, I, (N+1)*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_val, val, nz*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_r, rhs, N*sizeof(float), cudaMemcpyHostToDevice);
alpha = 1.0;
alpham1 = -1.0;
beta = 0.0;
r0 = 0.;
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha, descr, d_val, d_row, d_col, d_x, &beta, d_Ax);
cublasSaxpy(cublasHandle, N, &alpham1, d_Ax, 1, d_r, 1);
cublasStatus = cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
k = 1;
while (r1 > tol*tol && k <= max_iter) {
if (k > 1) {
b = r1 / r0;
cublasStatus = cublasSscal(cublasHandle, N, &b, d_p, 1);
cublasStatus = cublasSaxpy(cublasHandle, N, &alpha, d_r, 1, d_p, 1);
} else {
cublasStatus = cublasScopy(cublasHandle, N, d_r, 1, d_p, 1);
}
cusparseScsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha, descr, d_val, d_row, d_col, d_p, &beta, d_Ax);
cublasStatus = cublasSdot(cublasHandle, N, d_p, 1, d_Ax, 1, &dot);
a = r1 / dot;
cublasStatus = cublasSaxpy(cublasHandle, N, &a, d_p, 1, d_x, 1);
na = -a;
cublasStatus = cublasSaxpy(cublasHandle, N, &na, d_Ax, 1, d_r, 1);
r0 = r1;
cublasStatus = cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
cudaThreadSynchronize();
printf("iteration = %3d, residual = %e\n", k, sqrt(r1));
k++;
}
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
float rsum, diff, err = 0.0;
for (int i = 0; i < N; i++) {
rsum = 0.0;
for (int j = I[i]; j < I[i+1]; j++) {
rsum += val[j]*x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err) err = diff;
}
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
free(I);
free(J);
free(val);
free(x);
free(rhs);
cudaFree(d_col);
cudaFree(d_row);
cudaFree(d_val);
cudaFree(d_x);
cudaFree(d_r);
cudaFree(d_p);
cudaFree(d_Ax);
cudaDeviceReset();
printf("Test Summary: Error amount = %f\n", err);
shrQAFinishExit(argc, (const char **)argv, (k <= max_iter) ? QA_PASSED : QA_FAILED );
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main(int argc, char **argv)
{
int devID = 0;
char *ref_file = NULL;
pArgc = &argc;
pArgv = argv;
// start logs
printf("%s Starting...\n\n", argv[0]);
if (checkCmdLineFlag(argc, (const char **)argv, "help"))
{
printHelp();
exit(EXIT_SUCCESS);
}
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
if (argc > 1)
{
if (checkCmdLineFlag(argc, (const char **)argv, "threads"))
{
nthreads = getCmdLineArgumentInt(argc, (const char **) argv, "threads");
}
if (checkCmdLineFlag(argc, (const char **)argv, "radius"))
{
filter_radius = getCmdLineArgumentInt(argc, (const char **) argv, "radius");
}
if (checkCmdLineFlag(argc, (const char **)argv, "passes"))
{
iterations = getCmdLineArgumentInt(argc, (const char **) argv, "passes");
}
if (checkCmdLineFlag(argc, (const char **)argv, "file"))
{
getCmdLineArgumentString(argc, (const char **)argv, "file", (char **)&ref_file);
}
}
// load image to process
loadImageData(argc, argv);
if (checkCmdLineFlag(argc, (const char **)argv, "benchmark"))
{
// This is a separate mode of the sample, where we are benchmark the kernels for performance
devID = findCudaDevice(argc, (const char **)argv);
// Running CUDA kernels (boxfilter) in Benchmarking mode
g_TotalErrors += runBenchmark();
exit(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
else if (checkCmdLineFlag(argc, (const char **)argv, "radius") ||
checkCmdLineFlag(argc, (const char **)argv, "passes"))
{
// This overrides the default mode. Users can specify the radius used by the filter kernel
devID = findCudaDevice(argc, (const char **)argv);
g_TotalErrors += runSingleTest(ref_file, argv[0]);
exit(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
else
{
// Default mode running with OpenGL visualization and in automatic mode
// the output automatically changes animation
printf("\n");
// 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);
int dev = findCapableDevice(argc, argv);
if (dev != -1)
{
cudaGLSetGLDevice(dev);
}
else
{
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
// Now we can create a CUDA context and bind it to the OpenGL context
initCuda();
initGLResources();
// sets the callback function so it will call cleanup upon exit
atexit(cleanup);
printf("Running Standard Demonstration with GLUT loop...\n\n");
printf("Press '+' and '-' to change filter width\n"
"Press ']' and '[' to change number of iterations\n"
"Press 'a' or 'A' to change animation ON/OFF\n\n");
// Main OpenGL loop that will run visualization for every vsync
glutMainLoop();
cudaDeviceReset();
exit(g_TotalErrors == 0 ? EXIT_SUCCESS : EXIT_FAILURE);
}
}
////////////////////////////////////////////////////////////////////////////////
// Program main
////////////////////////////////////////////////////////////////////////////////
int
main(int argc, char **argv)
{
printf("%s Starting...\n\n", argv[0]);
char *typeInput = 0;
getCmdLineArgumentString(argc, (const char **) argv, "type", &typeInput);
ReduceType datatype = REDUCE_INT;
if (0 != typeInput)
{
if (!strcasecmp(typeInput, "float"))
{
datatype = REDUCE_FLOAT;
}
else if (!strcasecmp(typeInput, "double"))
{
datatype = REDUCE_DOUBLE;
}
else if (strcasecmp(typeInput, "int"))
{
printf("Type %s is not recognized. Using default type int.\n\n", typeInput);
}
}
cudaDeviceProp deviceProp;
deviceProp.major = 1;
deviceProp.minor = 0;
int minimumComputeVersion = 10;
if (datatype == REDUCE_DOUBLE)
{
deviceProp.minor = 3;
minimumComputeVersion = 13;
}
int dev;
dev = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
if ((deviceProp.major * 10 + deviceProp.minor) >= minimumComputeVersion)
{
printf("Using Device %d: %s\n\n", dev, deviceProp.name);
checkCudaErrors(cudaSetDevice(dev));
}
else
{
printf("Error: the selected device does not support the minimum compute capability of %d.%d.\n\n",
minimumComputeVersion / 10, minimumComputeVersion % 10);
// 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);
}
printf("Reducing array of type %s\n\n", getReduceTypeString(datatype));
bool bResult = false;
switch (datatype)
{
default:
case REDUCE_INT:
bResult = runTest<int>(argc, argv, datatype);
break;
case REDUCE_FLOAT:
bResult = runTest<float>(argc, argv, datatype);
break;
case REDUCE_DOUBLE:
bResult = runTest<double>(argc, argv, datatype);
break;
}
// 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(bResult ? "Test passed\n" : "Test failed!\n");
}
int main(int argc, char **argv)
{
int N = 0, nz = 0, *I = NULL, *J = NULL;
float *val = NULL;
const float tol = 1e-5f;
const int max_iter = 10000;
float *x;
float *rhs;
float a, b, na, r0, r1;
float dot;
float *r, *p, *Ax;
int k;
float alpha, beta, alpham1;
printf("Starting [%s]...\n", sSDKname);
// This will pick the best possible CUDA capable device
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
#if defined(__APPLE__) || defined(MACOSX)
fprintf(stderr, "Unified Memory not currently supported on OS X\n");
cudaDeviceReset();
exit(EXIT_WAIVED);
#endif
if (sizeof(void *) != 8)
{
fprintf(stderr, "Unified Memory requires compiling for a 64-bit system.\n");
cudaDeviceReset();
exit(EXIT_WAIVED);
}
if (((deviceProp.major << 4) + deviceProp.minor) < 0x30)
{
fprintf(stderr, "%s requires Compute Capability of SM 3.0 or higher to run.\nexiting...\n", argv[0]);
cudaDeviceReset();
exit(EXIT_WAIVED);
}
// Statistics about the GPU device
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
/* Generate a random tridiagonal symmetric matrix in CSR format */
N = 1048576;
nz = (N-2)*3 + 4;
cudaMallocManaged((void **)&I, sizeof(int)*(N+1));
cudaMallocManaged((void **)&J, sizeof(int)*nz);
cudaMallocManaged((void **)&val, sizeof(float)*nz);
genTridiag(I, J, val, N, nz);
cudaMallocManaged((void **)&x, sizeof(float)*N);
cudaMallocManaged((void **)&rhs, sizeof(float)*N);
for (int i = 0; i < N; i++)
{
rhs[i] = 1.0;
x[i] = 0.0;
}
/* Get handle to the CUBLAS context */
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
checkCudaErrors(cublasStatus);
/* Get handle to the CUSPARSE context */
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
checkCudaErrors(cusparseStatus);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
// temp memory for CG
checkCudaErrors(cudaMallocManaged((void **)&r, N*sizeof(float)));
checkCudaErrors(cudaMallocManaged((void **)&p, N*sizeof(float)));
checkCudaErrors(cudaMallocManaged((void **)&Ax, N*sizeof(float)));
cudaDeviceSynchronize();
for (int i=0; i < N; i++)
{
r[i] = rhs[i];
}
alpha = 1.0;
alpham1 = -1.0;
beta = 0.0;
r0 = 0.;
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha, descr, val, I, J, x, &beta, Ax);
cublasSaxpy(cublasHandle, N, &alpham1, Ax, 1, r, 1);
cublasStatus = cublasSdot(cublasHandle, N, r, 1, r, 1, &r1);
k = 1;
while (r1 > tol*tol && k <= max_iter)
{
if (k > 1)
{
b = r1 / r0;
cublasStatus = cublasSscal(cublasHandle, N, &b, p, 1);
cublasStatus = cublasSaxpy(cublasHandle, N, &alpha, r, 1, p, 1);
}
else
{
cublasStatus = cublasScopy(cublasHandle, N, r, 1, p, 1);
}
cusparseScsrmv(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &alpha, descr, val, I, J, p, &beta, Ax);
cublasStatus = cublasSdot(cublasHandle, N, p, 1, Ax, 1, &dot);
a = r1 / dot;
cublasStatus = cublasSaxpy(cublasHandle, N, &a, p, 1, x, 1);
na = -a;
cublasStatus = cublasSaxpy(cublasHandle, N, &na, Ax, 1, r, 1);
r0 = r1;
cublasStatus = cublasSdot(cublasHandle, N, r, 1, r, 1, &r1);
cudaThreadSynchronize();
printf("iteration = %3d, residual = %e\n", k, sqrt(r1));
k++;
}
printf("Final residual: %e\n",sqrt(r1));
fprintf(stdout,"&&&& uvm_cg test %s\n", (sqrt(r1) < tol) ? "PASSED" : "FAILED");
float rsum, diff, err = 0.0;
for (int i = 0; i < N; i++)
{
rsum = 0.0;
for (int j = I[i]; j < I[i+1]; j++)
{
rsum += val[j]*x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err)
{
err = diff;
}
}
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
cudaFree(I);
cudaFree(J);
cudaFree(val);
cudaFree(x);
cudaFree(r);
cudaFree(p);
cudaFree(Ax);
cudaDeviceReset();
printf("Test Summary: Error amount = %f, result = %s\n", err, (k <= max_iter) ? "SUCCESS" : "FAILURE");
exit((k <= max_iter) ? EXIT_SUCCESS : EXIT_FAILURE);
}
void runAutoTest(int argc, char *argv[])
{
printf("[%s] (automated testing w/ readback)\n", sSDKsample);
int devID = findCudaDevice(argc, (const char **)argv);
loadDefaultImage(argv[0]);
Pixel *d_result;
checkCudaErrors(cudaMalloc((void **)&d_result, imWidth*imHeight*sizeof(Pixel)));
char *ref_file = NULL;
char dump_file[256];
int mode = 0;
mode = getCmdLineArgumentInt(argc, (const char **)argv, "mode");
getCmdLineArgumentString(argc, (const char **)argv, "file", &ref_file);
switch (mode)
{
case 0:
g_SobelDisplayMode = SOBELDISPLAY_IMAGE;
sprintf(dump_file, "lena_orig.pgm");
break;
case 1:
g_SobelDisplayMode = SOBELDISPLAY_SOBELTEX;
sprintf(dump_file, "lena_tex.pgm");
break;
case 2:
g_SobelDisplayMode = SOBELDISPLAY_SOBELSHARED;
sprintf(dump_file, "lena_shared.pgm");
break;
default:
printf("Invalid Filter Mode File\n");
exit(EXIT_FAILURE);
break;
}
printf("AutoTest: %s <%s>\n", sSDKsample, filterMode[g_SobelDisplayMode]);
sobelFilter(d_result, imWidth, imHeight, g_SobelDisplayMode, imageScale);
checkCudaErrors(cudaDeviceSynchronize());
unsigned char *h_result = (unsigned char *)malloc(imWidth*imHeight*sizeof(Pixel));
checkCudaErrors(cudaMemcpy(h_result, d_result, imWidth*imHeight*sizeof(Pixel), cudaMemcpyDeviceToHost));
sdkSavePGM(dump_file, h_result, imWidth, imHeight);
if (!sdkComparePGM(dump_file, sdkFindFilePath(ref_file, argv[0]), MAX_EPSILON_ERROR, 0.15f, false))
{
g_TotalErrors++;
}
checkCudaErrors(cudaFree(d_result));
free(h_result);
if (g_TotalErrors != 0)
{
printf("Test failed!\n");
exit(EXIT_FAILURE);
}
printf("Test passed!\n");
exit(EXIT_SUCCESS);
}
///////////////////////////////////////////////////////////////////////////////
// Main program
///////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
// Start logs
shrQAStart(argc, argv);
// initialize the GPU, either identified by --device
// or by picking the device with highest flop rate.
int devID = findCudaDevice(argc, (const char **)argv);
// parsing the number of random numbers to generate
int rand_n = DEFAULT_RAND_N;
if( checkCmdLineFlag(argc, (const char**) argv, "count") )
{
rand_n = getCmdLineArgumentInt(argc, (const char**) argv, "count");
}
printf("Allocating data for %i samples...\n", rand_n);
// parsing the seed
int seed = DEFAULT_SEED;
if( checkCmdLineFlag(argc, (const char**) argv, "seed") )
{
seed = getCmdLineArgumentInt(argc, (const char**) argv, "seed");
}
printf("Seeding with %i ...\n", seed);
float *d_Rand;
checkCudaErrors( cudaMalloc((void **)&d_Rand, rand_n * sizeof(float)) );
curandGenerator_t prngGPU;
checkCurandErrors( curandCreateGenerator(&prngGPU, CURAND_RNG_PSEUDO_MTGP32) );
checkCurandErrors( curandSetPseudoRandomGeneratorSeed(prngGPU, seed) );
curandGenerator_t prngCPU;
checkCurandErrors( curandCreateGeneratorHost(&prngCPU, CURAND_RNG_PSEUDO_MTGP32) );
checkCurandErrors( curandSetPseudoRandomGeneratorSeed(prngCPU, seed) );
//
// Example 1: Compare random numbers generated on GPU and CPU
float *h_RandGPU = (float *)malloc(rand_n * sizeof(float));
printf("Generating random numbers on GPU...\n\n");
checkCurandErrors( curandGenerateUniform(prngGPU, (float*) d_Rand, rand_n) );
printf("\nReading back the results...\n");
checkCudaErrors( cudaMemcpy(h_RandGPU, d_Rand, rand_n * sizeof(float), cudaMemcpyDeviceToHost) );
float *h_RandCPU = (float *)malloc(rand_n * sizeof(float));
printf("Generating random numbers on CPU...\n\n");
checkCurandErrors( curandGenerateUniform(prngCPU, (float*) h_RandCPU, rand_n) );
printf("Comparing CPU/GPU random numbers...\n\n");
float L1norm = compareResults(rand_n, h_RandGPU, h_RandCPU);
//
// Example 2: Timing of random number generation on GPU
const int numIterations = 10;
int i;
StopWatchInterface *hTimer;
checkCudaErrors( cudaDeviceSynchronize() );
sdkCreateTimer(&hTimer);
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
for (i = 0; i < numIterations; i++)
{
checkCurandErrors( curandGenerateUniform(prngGPU, (float*) d_Rand, rand_n) );
}
checkCudaErrors( cudaDeviceSynchronize() );
sdkStopTimer(&hTimer);
double gpuTime = 1.0e-3 * sdkGetTimerValue(&hTimer)/(double)numIterations;
printf("MersenneTwister, Throughput = %.4f GNumbers/s, Time = %.5f s, Size = %u Numbers\n",
1.0e-9 * rand_n / gpuTime, gpuTime, rand_n);
printf("Shutting down...\n");
checkCurandErrors( curandDestroyGenerator(prngGPU) );
checkCurandErrors( curandDestroyGenerator(prngCPU) );
checkCudaErrors( cudaFree(d_Rand) );
sdkDeleteTimer( &hTimer);
free(h_RandGPU);
free(h_RandCPU);
cudaDeviceReset();
shrQAFinishExit(argc, (const char**)argv, (L1norm < 1e-6) ? QA_PASSED : QA_FAILED);
}
////////////////////////////////////////////////////////////////////////////////
// Test driver
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
cudaError_t error;
printf("%s Starting...\n\n", argv[0]);
printf("Starting up CUDA context...\n");
int dev = findCudaDevice(argc, (const char **)argv);
uint *h_InputKey, *h_InputVal, *h_OutputKeyGPU, *h_OutputValGPU;
uint *d_InputKey, *d_InputVal, *d_OutputKey, *d_OutputVal;
StopWatchInterface *hTimer = NULL;
const uint N = 1048576;
const uint DIR = 0;
const uint numValues = 65536;
const uint numIterations = 1;
printf("Allocating and initializing host arrays...\n\n");
sdkCreateTimer(&hTimer);
h_InputKey = (uint *)malloc(N * sizeof(uint));
h_InputVal = (uint *)malloc(N * sizeof(uint));
h_OutputKeyGPU = (uint *)malloc(N * sizeof(uint));
h_OutputValGPU = (uint *)malloc(N * sizeof(uint));
srand(2001);
for (uint i = 0; i < N; i++)
{
h_InputKey[i] = rand() % numValues;
h_InputVal[i] = i;
}
printf("Allocating and initializing CUDA arrays...\n\n");
error = cudaMalloc((void **)&d_InputKey, N * sizeof(uint));
checkCudaErrors(error);
error = cudaMalloc((void **)&d_InputVal, N * sizeof(uint));
checkCudaErrors(error);
error = cudaMalloc((void **)&d_OutputKey, N * sizeof(uint));
checkCudaErrors(error);
error = cudaMalloc((void **)&d_OutputVal, N * sizeof(uint));
checkCudaErrors(error);
error = cudaMemcpy(d_InputKey, h_InputKey, N * sizeof(uint), cudaMemcpyHostToDevice);
checkCudaErrors(error);
error = cudaMemcpy(d_InputVal, h_InputVal, N * sizeof(uint), cudaMemcpyHostToDevice);
checkCudaErrors(error);
int flag = 1;
printf("Running GPU bitonic sort (%u identical iterations)...\n\n", numIterations);
for (uint arrayLength = 64; arrayLength <= N; arrayLength *= 2)
{
printf("Testing array length %u (%u arrays per batch)...\n", arrayLength, N / arrayLength);
error = cudaDeviceSynchronize();
checkCudaErrors(error);
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
uint threadCount = 0;
for (uint i = 0; i < numIterations; i++)
threadCount = bitonicSort(
d_OutputKey,
d_OutputVal,
d_InputKey,
d_InputVal,
N / arrayLength,
arrayLength,
DIR
);
error = cudaDeviceSynchronize();
checkCudaErrors(error);
sdkStopTimer(&hTimer);
printf("Average time: %f ms\n\n", sdkGetTimerValue(&hTimer) / numIterations);
if (arrayLength == N)
{
double dTimeSecs = 1.0e-3 * sdkGetTimerValue(&hTimer) / numIterations;
printf("sortingNetworks-bitonic, Throughput = %.4f MElements/s, Time = %.5f s, Size = %u elements, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * (double)arrayLength/dTimeSecs), dTimeSecs, arrayLength, 1, threadCount);
}
printf("\nValidating the results...\n");
printf("...reading back GPU results\n");
error = cudaMemcpy(h_OutputKeyGPU, d_OutputKey, N * sizeof(uint), cudaMemcpyDeviceToHost);
checkCudaErrors(error);
error = cudaMemcpy(h_OutputValGPU, d_OutputVal, N * sizeof(uint), cudaMemcpyDeviceToHost);
checkCudaErrors(error);
int keysFlag = validateSortedKeys(h_OutputKeyGPU, h_InputKey, N / arrayLength, arrayLength, numValues, DIR);
int valuesFlag = validateValues(h_OutputKeyGPU, h_OutputValGPU, h_InputKey, N / arrayLength, arrayLength);
flag = flag && keysFlag && valuesFlag;
printf("\n");
}
printf("Shutting down...\n");
sdkDeleteTimer(&hTimer);
cudaFree(d_OutputVal);
cudaFree(d_OutputKey);
cudaFree(d_InputVal);
cudaFree(d_InputKey);
free(h_OutputValGPU);
free(h_OutputKeyGPU);
free(h_InputVal);
free(h_InputKey);
cudaDeviceReset();
exit(flag ? EXIT_SUCCESS : EXIT_FAILURE);
}
int main(int argc, char **argv)
{
// Start logs
printf("%s Starting...\n\n", argv[0]);
unsigned int tableCPU[QRNG_DIMENSIONS][QRNG_RESOLUTION];
float *h_OutputGPU, *d_Output;
int dim, pos;
double delta, ref, sumDelta, sumRef, L1norm, gpuTime;
StopWatchInterface *hTimer = NULL;
if (sizeof(INT64) != 8)
{
printf("sizeof(INT64) != 8\n");
return 0;
}
cudaDeviceProp deviceProp;
int dev = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, dev));
if (((deviceProp.major << 4) + deviceProp.minor) < 0x20)
{
fprintf(stderr, "quasirandomGenerator requires Compute Capability of SM 2.0 or higher to run.\n");
cudaDeviceReset();
exit(EXIT_WAIVED);
}
sdkCreateTimer(&hTimer);
printf("Allocating GPU memory...\n");
checkCudaErrors(cudaMalloc((void **)&d_Output, QRNG_DIMENSIONS * N * sizeof(float)));
printf("Allocating CPU memory...\n");
h_OutputGPU = (float *)malloc(QRNG_DIMENSIONS * N * sizeof(float));
printf("Initializing QRNG tables...\n\n");
initQuasirandomGenerator(tableCPU);
initTableGPU(tableCPU);
printf("Testing QRNG...\n\n");
checkCudaErrors(cudaMemset(d_Output, 0, QRNG_DIMENSIONS * N * sizeof(float)));
int numIterations = 20;
for (int i = -1; i < numIterations; i++)
{
if (i == 0)
{
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
}
quasirandomGeneratorGPU(d_Output, 0, N);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer)/(double)numIterations*1e-3;
printf("quasirandomGenerator, Throughput = %.4f GNumbers/s, Time = %.5f s, Size = %u Numbers, NumDevsUsed = %u, Workgroup = %u\n",
(double)QRNG_DIMENSIONS * (double)N * 1.0E-9 / gpuTime, gpuTime, QRNG_DIMENSIONS*N, 1, 128*QRNG_DIMENSIONS);
printf("\nReading GPU results...\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, QRNG_DIMENSIONS * N * sizeof(float), cudaMemcpyDeviceToHost));
printf("Comparing to the CPU results...\n\n");
sumDelta = 0;
sumRef = 0;
for (dim = 0; dim < QRNG_DIMENSIONS; dim++)
for (pos = 0; pos < N; pos++)
{
ref = getQuasirandomValue63(pos, dim);
delta = (double)h_OutputGPU[dim * N + pos] - ref;
sumDelta += fabs(delta);
sumRef += fabs(ref);
}
printf("L1 norm: %E\n", sumDelta / sumRef);
printf("\nTesting inverseCNDgpu()...\n\n");
checkCudaErrors(cudaMemset(d_Output, 0, QRNG_DIMENSIONS * N * sizeof(float)));
for (int i = -1; i < numIterations; i++)
{
if (i == 0)
{
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
}
inverseCNDgpu(d_Output, NULL, QRNG_DIMENSIONS * N);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer)/(double)numIterations*1e-3;
printf("quasirandomGenerator-inverse, Throughput = %.4f GNumbers/s, Time = %.5f s, Size = %u Numbers, NumDevsUsed = %u, Workgroup = %u\n",
(double)QRNG_DIMENSIONS * (double)N * 1E-9 / gpuTime, gpuTime, QRNG_DIMENSIONS*N, 1, 128);
printf("Reading GPU results...\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, QRNG_DIMENSIONS * N * sizeof(float), cudaMemcpyDeviceToHost));
printf("\nComparing to the CPU results...\n");
sumDelta = 0;
sumRef = 0;
unsigned int distance = ((unsigned int)-1) / (QRNG_DIMENSIONS * N + 1);
for (pos = 0; pos < QRNG_DIMENSIONS * N; pos++)
{
unsigned int d = (pos + 1) * distance;
ref = MoroInvCNDcpu(d);
delta = (double)h_OutputGPU[pos] - ref;
sumDelta += fabs(delta);
sumRef += fabs(ref);
}
printf("L1 norm: %E\n\n", L1norm = sumDelta / sumRef);
printf("Shutting down...\n");
sdkDeleteTimer(&hTimer);
free(h_OutputGPU);
checkCudaErrors(cudaFree(d_Output));
// 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(L1norm < 1e-6 ? EXIT_SUCCESS : EXIT_FAILURE);
}
/* Main */
int main(int argc, char **argv)
{
cublasStatus_t status;
float *h_A;
float *h_B;
float *h_C;
float *h_C_rnd;
float *h_C_ref;
float *d_A = 0;
float *d_B = 0;
float *d_C = 0;
float alpha = 1.0f;
float beta = 0.0f;
int n2 = N * N;
int i;
cublasHandle_t handle;
int dev_id;
cudaDeviceProp device_prop;
bool do_device_api_test = false;
float host_api_test_ratio, device_api_test_ratio;
/* Initialize CUBLAS */
printf("simpleCUBLAS test running...\n");
dev_id = findCudaDevice(argc, (const char **) argv);
checkCudaErrors(cudaGetDeviceProperties(&device_prop, dev_id));
if ((device_prop.major << 4) + device_prop.minor >= 0x35)
{
printf("Host and device APIs will be tested.\n");
do_device_api_test = true;
}
/* else if ((device_prop.major << 4) + device_prop.minor >= 0x20)
{
printf("Host API will be tested.\n");
do_device_api_test = false;
}
*/
else
{
fprintf(stderr, "simpleDevLibCUBLAS examples requires Compute Capability of SM 3.5 or higher\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();
return EXIT_SUCCESS;
}
status = cublasCreate(&handle);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! CUBLAS initialization error\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();
return EXIT_FAILURE;
}
/* Allocate host memory for the matrices */
h_A = (float *)malloc(n2 * sizeof(h_A[0]));
if (h_A == 0)
{
fprintf(stderr, "!!!! host memory allocation error (A)\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();
return EXIT_FAILURE;
}
h_B = (float *)malloc(n2 * sizeof(h_B[0]));
if (h_B == 0)
{
fprintf(stderr, "!!!! host memory allocation error (B)\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();
return EXIT_FAILURE;
}
h_C_rnd = (float *)malloc(n2 * sizeof(h_C_rnd[0]));
if (h_C_rnd == 0)
{
fprintf(stderr, "!!!! host memory allocation error (C_rnd)\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();
return EXIT_FAILURE;
}
h_C = (float *)malloc(n2 * sizeof(h_C_ref[0]));
if (h_C == 0)
{
fprintf(stderr, "!!!! host memory allocation error (C)\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();
return EXIT_FAILURE;
}
/* Fill the matrices with test data */
for (i = 0; i < n2; i++)
{
h_A[i] = rand() / (float)RAND_MAX;
h_B[i] = rand() / (float)RAND_MAX;
h_C_rnd[i] = rand() / (float)RAND_MAX;
h_C[i] = h_C_rnd[i];
}
/* Allocate device memory for the matrices */
if (cudaMalloc((void **)&d_A, n2 * sizeof(d_A[0])) != cudaSuccess)
{
fprintf(stderr, "!!!! device memory allocation error (allocate A)\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();
return EXIT_FAILURE;
}
if (cudaMalloc((void **)&d_B, n2 * sizeof(d_B[0])) != cudaSuccess)
{
fprintf(stderr, "!!!! device memory allocation error (allocate B)\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();
return EXIT_FAILURE;
}
if (cudaMalloc((void **)&d_C, n2 * sizeof(d_C[0])) != cudaSuccess)
{
fprintf(stderr, "!!!! device memory allocation error (allocate C)\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();
return EXIT_FAILURE;
}
/* Initialize the device matrices with the host matrices */
status = cublasSetVector(n2, sizeof(h_A[0]), h_A, 1, d_A, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write A)\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();
return EXIT_FAILURE;
}
status = cublasSetVector(n2, sizeof(h_B[0]), h_B, 1, d_B, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write B)\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();
return EXIT_FAILURE;
}
status = cublasSetVector(n2, sizeof(h_C_rnd[0]), h_C_rnd, 1, d_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write C)\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();
return EXIT_FAILURE;
}
/*
* Performs operation using plain C code
*/
simple_sgemm(N, alpha, h_A, h_B, beta, h_C);
h_C_ref = h_C;
/*
* Performs operation using cublas
*/
status = cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, N, N, N, &alpha, d_A, N, d_B, N, &beta, d_C, N);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! kernel execution error\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();
return EXIT_FAILURE;
}
/* Allocate host memory for reading back the result from device memory */
h_C = (float *)malloc(n2 * sizeof(h_C[0]));
if (h_C == 0)
{
fprintf(stderr, "!!!! host memory allocation error (C)\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();
return EXIT_FAILURE;
}
/* Read the result back */
status = cublasGetVector(n2, sizeof(h_C[0]), d_C, 1, h_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (read C)\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();
return EXIT_FAILURE;
}
/* Check result against reference */
host_api_test_ratio = check_result(h_C, h_C_ref, n2);
if (do_device_api_test)
{
/* Reset device resident C matrix */
status = cublasSetVector(n2, sizeof(h_C_rnd[0]), h_C_rnd, 1, d_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write C)\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();
return EXIT_FAILURE;
}
/*
* Performs operation using the device API of CUBLAS library
*/
device_cublas_sgemm(N, alpha, d_A, d_B, beta, d_C);
/* Read the result back */
status = cublasGetVector(n2, sizeof(h_C[0]), d_C, 1, h_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (read C)\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();
return EXIT_FAILURE;
}
/* Check result against reference */
device_api_test_ratio = check_result(h_C, h_C_ref, n2);
}
/* Memory clean up */
free(h_A);
free(h_B);
free(h_C);
free(h_C_rnd);
free(h_C_ref);
if (cudaFree(d_A) != cudaSuccess)
{
fprintf(stderr, "!!!! memory free error (A)\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();
return EXIT_FAILURE;
}
if (cudaFree(d_B) != cudaSuccess)
{
fprintf(stderr, "!!!! memory free error (B)\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();
return EXIT_FAILURE;
}
if (cudaFree(d_C) != cudaSuccess)
{
fprintf(stderr, "!!!! memory free error (C)\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();
return EXIT_FAILURE;
}
/* Shutdown */
status = cublasDestroy(handle);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! shutdown error (A)\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();
return EXIT_FAILURE;
}
// 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();
bool test_result = do_device_api_test ?
host_api_test_ratio < 1e-6 &&
device_api_test_ratio < 1e-6 :
host_api_test_ratio < 1e-6;
printf("simpleCUBLAS completed, returned %s\n",
test_result ? "OK" : "ERROR!");
exit(test_result ? EXIT_SUCCESS : 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);
}
int main(int argc, char **argv)
{
// Start logs
printf("%s Starting...\n\n", argv[0]);
unsigned int useDoublePrecision;
char *precisionChoice;
getCmdLineArgumentString(argc, (const char **)argv, "type", &precisionChoice);
if (precisionChoice == NULL)
{
useDoublePrecision = 0;
}
else
{
if (!STRCASECMP(precisionChoice, "double"))
{
useDoublePrecision = 1;
}
else
{
useDoublePrecision = 0;
}
}
unsigned int tableCPU[QRNG_DIMENSIONS][QRNG_RESOLUTION];
float *h_OutputGPU, *d_Output;
int dim, pos;
double delta, ref, sumDelta, sumRef, L1norm, gpuTime;
StopWatchInterface *hTimer = NULL;
if (sizeof(INT64) != 8)
{
printf("sizeof(INT64) != 8\n");
return 0;
}
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
int dev = findCudaDevice(argc, (const char **)argv);
sdkCreateTimer(&hTimer);
int deviceIndex;
checkCudaErrors(cudaGetDevice(&deviceIndex));
cudaDeviceProp deviceProp;
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, deviceIndex));
int version = deviceProp.major * 10 + deviceProp.minor;
if (useDoublePrecision && version < 13)
{
printf("Double precision not supported.\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();
return 0;
}
printf("Allocating GPU memory...\n");
checkCudaErrors(cudaMalloc((void **)&d_Output, QRNG_DIMENSIONS * N * sizeof(float)));
printf("Allocating CPU memory...\n");
h_OutputGPU = (float *)malloc(QRNG_DIMENSIONS * N * sizeof(float));
printf("Initializing QRNG tables...\n\n");
initQuasirandomGenerator(tableCPU);
if (useDoublePrecision)
{
initTable_SM13(tableCPU);
}
else
{
initTable_SM10(tableCPU);
}
printf("Testing QRNG...\n\n");
checkCudaErrors(cudaMemset(d_Output, 0, QRNG_DIMENSIONS * N * sizeof(float)));
int numIterations = 20;
for (int i = -1; i < numIterations; i++)
{
if (i == 0)
{
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
}
if (useDoublePrecision)
{
quasirandomGenerator_SM13(d_Output, 0, N);
}
else
{
quasirandomGenerator_SM10(d_Output, 0, N);
}
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer)/(double)numIterations*1e-3;
printf("quasirandomGenerator, Throughput = %.4f GNumbers/s, Time = %.5f s, Size = %u Numbers, NumDevsUsed = %u, Workgroup = %u\n",
(double)QRNG_DIMENSIONS * (double)N * 1.0E-9 / gpuTime, gpuTime, QRNG_DIMENSIONS*N, 1, 128*QRNG_DIMENSIONS);
printf("\nReading GPU results...\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, QRNG_DIMENSIONS * N * sizeof(float), cudaMemcpyDeviceToHost));
printf("Comparing to the CPU results...\n\n");
sumDelta = 0;
sumRef = 0;
for (dim = 0; dim < QRNG_DIMENSIONS; dim++)
for (pos = 0; pos < N; pos++)
{
ref = getQuasirandomValue63(pos, dim);
delta = (double)h_OutputGPU[dim * N + pos] - ref;
sumDelta += fabs(delta);
sumRef += fabs(ref);
}
printf("L1 norm: %E\n", sumDelta / sumRef);
printf("\nTesting inverseCNDgpu()...\n\n");
checkCudaErrors(cudaMemset(d_Output, 0, QRNG_DIMENSIONS * N * sizeof(float)));
for (int i = -1; i < numIterations; i++)
{
if (i == 0)
{
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
}
if (useDoublePrecision)
{
inverseCND_SM13(d_Output, NULL, QRNG_DIMENSIONS * N);
}
else
{
inverseCND_SM10(d_Output, NULL, QRNG_DIMENSIONS * N);
}
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer)/(double)numIterations*1e-3;
printf("quasirandomGenerator-inverse, Throughput = %.4f GNumbers/s, Time = %.5f s, Size = %u Numbers, NumDevsUsed = %u, Workgroup = %u\n",
(double)QRNG_DIMENSIONS * (double)N * 1E-9 / gpuTime, gpuTime, QRNG_DIMENSIONS*N, 1, 128);
printf("Reading GPU results...\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, QRNG_DIMENSIONS * N * sizeof(float), cudaMemcpyDeviceToHost));
printf("\nComparing to the CPU results...\n");
sumDelta = 0;
sumRef = 0;
unsigned int distance = ((unsigned int)-1) / (QRNG_DIMENSIONS * N + 1);
for (pos = 0; pos < QRNG_DIMENSIONS * N; pos++)
{
unsigned int d = (pos + 1) * distance;
ref = MoroInvCNDcpu(d);
delta = (double)h_OutputGPU[pos] - ref;
sumDelta += fabs(delta);
sumRef += fabs(ref);
}
printf("L1 norm: %E\n\n", L1norm = sumDelta / sumRef);
printf("Shutting down...\n");
sdkDeleteTimer(&hTimer);
free(h_OutputGPU);
checkCudaErrors(cudaFree(d_Output));
// 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(L1norm < 1e-6 ? EXIT_SUCCESS : EXIT_FAILURE);
}
int _tmain(int argc, _TCHAR* argv[])
{
int M = 0, N = 0, nz = 0, *I = NULL, *J = NULL;
cuDoubleComplex *val = NULL;
cuDoubleComplex *x, *y;
cuDoubleComplex *d_x, *d_y;
double duration, duration_setup;
std::clock_t setup_clock;
setup_clock = std::clock();
// This will pick the best possible CUDA capable device
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
if (devID < 0)
{
printf("no devices found...\n");
exit(EXIT_SUCCESS);
}
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
// Statistics about the GPU device
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
int version = (deviceProp.major * 0x10 + deviceProp.minor);
if (version < 0x11)
{
printf("Requires a minimum CUDA compute 1.1 capability\n");
cudaDeviceReset();
exit(EXIT_SUCCESS);
}
M = N = 8388608; //2 ^ 23
//M = N = 4194304; //2 ^ 22
//M = N = 2097152; //2 ^ 21
//M = N = 1048576; //2 ^ 20
//M = N = 524288; //2 ^ 19
nz = N * 8;
I = (int *)malloc(sizeof(int)*(N + 1));
J = (int *)malloc(sizeof(int)*nz);
val = (cuDoubleComplex *)malloc(sizeof(cuDoubleComplex)*nz);
genTridiag(I, J, val, N, nz);
x = (cuDoubleComplex*)malloc(sizeof(cuDoubleComplex)* N);
y = (cuDoubleComplex*)malloc(sizeof(cuDoubleComplex)* N);
//create an array for the answer array (Y) and set all of the answers to 0 for the test (could do random)
for (int i = 0; i < N; i++)
{
y[i] = make_cuDoubleComplex(0.0, 0.0);
}
//Get handle to the CUBLAS context
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
checkCudaErrors(cublasStatus);
//Get handle to the CUSPARSE context
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
checkCudaErrors(cusparseStatus);
//Get handle to a CUSPARSE matrix descriptor
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
//Get handle to a matrix_solve_info object
cusparseSolveAnalysisInfo_t info = 0;
cusparseStatus = cusparseCreateSolveAnalysisInfo(&info);
checkCudaErrors(cusparseStatus);
cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr, CUSPARSE_INDEX_BASE_ZERO);
duration_setup = (std::clock() - setup_clock) / (double)CLOCKS_PER_SEC;
printf("setup_time: %f\r\n", duration_setup);
std::clock_t start;
start = std::clock();
checkCudaErrors(cudaMalloc((void **)&d_x, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_y, N*sizeof(float)));
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice);
//Analyze the matrix. The info variable is needed to perform additional operations on the matrix
cusparseStatus = cusparseZcsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, nz, descr, val, J, I, info);
//Uses infor gathered from the matrix to solve the matrix.
cusparseStatus = cusparseZcsrsv_solve(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, 0, descr, val, J, I, info, d_x, d_y);
//Get the result back from the device
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost);
duration = (std::clock() - start) / (double)CLOCKS_PER_SEC;
printf("time ellapsed: %f", duration);
//free up memory
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
free(I);
free(J);
free(val);
free(x);
cudaFree(d_x);
cudaDeviceReset();
//Wait for user input so they can see the results
char* s = (char*)malloc(sizeof(char) * 8);
scanf(s);
exit(0);
}
////////////////////////////////////////////////////////////////////////////////
// Main program
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
printf("[%s] - Starting...\n", argv[0]);
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
if (((deviceProp.major << 4) + deviceProp.minor) < 0x20)
{
fprintf(stderr, "binomialOptions requires Compute Capability of SM 2.0 or higher to run.\n");
cudaDeviceReset();
exit(EXIT_WAIVED);
}
const int OPT_N = MAX_OPTIONS;
TOptionData optionData[MAX_OPTIONS];
real
callValueBS[MAX_OPTIONS],
callValueGPU[MAX_OPTIONS],
callValueCPU[MAX_OPTIONS];
real
sumDelta, sumRef, gpuTime, errorVal;
StopWatchInterface *hTimer = NULL;
int i;
sdkCreateTimer(&hTimer);
printf("Generating input data...\n");
//Generate options set
srand(123);
for (i = 0; i < OPT_N; i++)
{
optionData[i].S = randData(5.0f, 30.0f);
optionData[i].X = randData(1.0f, 100.0f);
optionData[i].T = randData(0.25f, 10.0f);
optionData[i].R = 0.06f;
optionData[i].V = 0.10f;
BlackScholesCall(callValueBS[i], optionData[i]);
}
printf("Running GPU binomial tree...\n");
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
binomialOptionsGPU(callValueGPU, optionData, OPT_N);
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer);
printf("Options count : %i \n", OPT_N);
printf("Time steps : %i \n", NUM_STEPS);
printf("binomialOptionsGPU() time: %f msec\n", gpuTime);
printf("Options per second : %f \n", OPT_N / (gpuTime * 0.001));
printf("Running CPU binomial tree...\n");
for (i = 0; i < OPT_N; i++)
{
binomialOptionsCPU(callValueCPU[i], optionData[i]);
}
printf("Comparing the results...\n");
sumDelta = 0;
sumRef = 0;
printf("GPU binomial vs. Black-Scholes\n");
for (i = 0; i < OPT_N; i++)
{
sumDelta += fabs(callValueBS[i] - callValueGPU[i]);
sumRef += fabs(callValueBS[i]);
}
if (sumRef >1E-5)
{
printf("L1 norm: %E\n", (double)(sumDelta / sumRef));
}
else
{
printf("Avg. diff: %E\n", (double)(sumDelta / (real)OPT_N));
}
printf("CPU binomial vs. Black-Scholes\n");
sumDelta = 0;
sumRef = 0;
for (i = 0; i < OPT_N; i++)
{
sumDelta += fabs(callValueBS[i]- callValueCPU[i]);
sumRef += fabs(callValueBS[i]);
}
if (sumRef >1E-5)
{
printf("L1 norm: %E\n", sumDelta / sumRef);
}
else
{
printf("Avg. diff: %E\n", (double)(sumDelta / (real)OPT_N));
}
printf("CPU binomial vs. GPU binomial\n");
sumDelta = 0;
sumRef = 0;
for (i = 0; i < OPT_N; i++)
{
sumDelta += fabs(callValueGPU[i] - callValueCPU[i]);
sumRef += callValueCPU[i];
}
if (sumRef > 1E-5)
{
printf("L1 norm: %E\n", errorVal = sumDelta / sumRef);
}
else
{
printf("Avg. diff: %E\n", (double)(sumDelta / (real)OPT_N));
}
printf("Shutting down...\n");
sdkDeleteTimer(&hTimer);
// 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("\nNOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.\n\n");
if (errorVal > 5e-4)
{
printf("Test failed!\n");
exit(EXIT_FAILURE);
}
printf("Test passed\n");
exit(EXIT_SUCCESS);
}
/* Main */
int main(int argc, char **argv)
{
cublasStatus_t status;
float *h_A;
float *h_B;
float *h_C;
float *h_C_ref;
float *d_A = 0;
float *d_B = 0;
float *d_C = 0;
float alpha = 1.0f;
float beta = 0.0f;
int n2 = N * N;
int i;
float error_norm;
float ref_norm;
float diff;
cublasHandle_t handle;
int dev = findCudaDevice(argc, (const char **) argv);
if (dev == -1)
{
return EXIT_FAILURE;
}
/* Initialize CUBLAS */
printf("simpleCUBLAS test running..\n");
status = cublasCreate(&handle);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! CUBLAS initialization error\n");
return EXIT_FAILURE;
}
/* Allocate host memory for the matrices */
h_A = (float *)malloc(n2 * sizeof(h_A[0]));
if (h_A == 0)
{
fprintf(stderr, "!!!! host memory allocation error (A)\n");
return EXIT_FAILURE;
}
h_B = (float *)malloc(n2 * sizeof(h_B[0]));
if (h_B == 0)
{
fprintf(stderr, "!!!! host memory allocation error (B)\n");
return EXIT_FAILURE;
}
h_C = (float *)malloc(n2 * sizeof(h_C[0]));
if (h_C == 0)
{
fprintf(stderr, "!!!! host memory allocation error (C)\n");
return EXIT_FAILURE;
}
/* Fill the matrices with test data */
for (i = 0; i < n2; i++)
{
h_A[i] = rand() / (float)RAND_MAX;
h_B[i] = rand() / (float)RAND_MAX;
h_C[i] = rand() / (float)RAND_MAX;
}
/* Allocate device memory for the matrices */
if (cudaMalloc((void **)&d_A, n2 * sizeof(d_A[0])) != cudaSuccess)
{
fprintf(stderr, "!!!! device memory allocation error (allocate A)\n");
return EXIT_FAILURE;
}
if (cudaMalloc((void **)&d_B, n2 * sizeof(d_B[0])) != cudaSuccess)
{
fprintf(stderr, "!!!! device memory allocation error (allocate B)\n");
return EXIT_FAILURE;
}
if (cudaMalloc((void **)&d_C, n2 * sizeof(d_C[0])) != cudaSuccess)
{
fprintf(stderr, "!!!! device memory allocation error (allocate C)\n");
return EXIT_FAILURE;
}
/* Initialize the device matrices with the host matrices */
status = cublasSetVector(n2, sizeof(h_A[0]), h_A, 1, d_A, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write A)\n");
return EXIT_FAILURE;
}
status = cublasSetVector(n2, sizeof(h_B[0]), h_B, 1, d_B, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write B)\n");
return EXIT_FAILURE;
}
status = cublasSetVector(n2, sizeof(h_C[0]), h_C, 1, d_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (write C)\n");
return EXIT_FAILURE;
}
/* Performs operation using plain C code */
simple_sgemm(N, alpha, h_A, h_B, beta, h_C);
h_C_ref = h_C;
/* Performs operation using cublas */
status = cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, N, N, N, &alpha, d_A, N, d_B, N, &beta, d_C, N);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! kernel execution error.\n");
return EXIT_FAILURE;
}
/* Allocate host memory for reading back the result from device memory */
h_C = (float *)malloc(n2 * sizeof(h_C[0]));
if (h_C == 0)
{
fprintf(stderr, "!!!! host memory allocation error (C)\n");
return EXIT_FAILURE;
}
/* Read the result back */
status = cublasGetVector(n2, sizeof(h_C[0]), d_C, 1, h_C, 1);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! device access error (read C)\n");
return EXIT_FAILURE;
}
/* Check result against reference */
error_norm = 0;
ref_norm = 0;
for (i = 0; i < n2; ++i)
{
diff = h_C_ref[i] - h_C[i];
error_norm += diff * diff;
ref_norm += h_C_ref[i] * h_C_ref[i];
}
error_norm = (float)sqrt((double)error_norm);
ref_norm = (float)sqrt((double)ref_norm);
if (fabs(ref_norm) < 1e-7)
{
fprintf(stderr, "!!!! reference norm is 0\n");
return EXIT_FAILURE;
}
/* Memory clean up */
free(h_A);
free(h_B);
free(h_C);
free(h_C_ref);
if (cudaFree(d_A) != cudaSuccess)
{
fprintf(stderr, "!!!! memory free error (A)\n");
return EXIT_FAILURE;
}
if (cudaFree(d_B) != cudaSuccess)
{
fprintf(stderr, "!!!! memory free error (B)\n");
return EXIT_FAILURE;
}
if (cudaFree(d_C) != cudaSuccess)
{
fprintf(stderr, "!!!! memory free error (C)\n");
return EXIT_FAILURE;
}
/* Shutdown */
status = cublasDestroy(handle);
if (status != CUBLAS_STATUS_SUCCESS)
{
fprintf(stderr, "!!!! shutdown error (A)\n");
return EXIT_FAILURE;
}
printf("CUBLAS program finished\n");
exit(error_norm / ref_norm < 1e-6f ? EXIT_SUCCESS : EXIT_FAILURE);
}
/* Solve Ax=b using the conjugate gradient method a) without any preconditioning, b) using an Incomplete Cholesky preconditioner and c) using an ILU0 preconditioner. */
int main(int argc, char **argv)
{
const int max_iter = 1000;
int k, M = 0, N = 0, nz = 0, *I = NULL, *J = NULL;
int *d_col, *d_row;
int qatest = 0;
const float tol = 1e-12f;
float *x, *rhs;
float r0, r1, alpha, beta;
float *d_val, *d_x;
float *d_zm1, *d_zm2, *d_rm2;
float *d_r, *d_p, *d_omega, *d_y;
float *val = NULL;
float *d_valsILU0;
float *valsILU0;
float rsum, diff, err = 0.0;
float qaerr1, qaerr2 = 0.0;
float dot, numerator, denominator, nalpha;
const float floatone = 1.0;
const float floatzero = 0.0;
int nErrors = 0;
printf("conjugateGradientPrecond starting...\n");
/* QA testing mode */
if (checkCmdLineFlag(argc, (const char **)argv, "qatest"))
{
qatest = 1;
}
/* This will pick the best possible CUDA capable device */
cudaDeviceProp deviceProp;
int devID = findCudaDevice(argc, (const char **)argv);
printf("GPU selected Device ID = %d \n", devID);
if (devID < 0)
{
printf("Invalid GPU device %d selected, exiting...\n", devID);
exit(EXIT_SUCCESS);
}
checkCudaErrors(cudaGetDeviceProperties(&deviceProp, devID));
/* Statistics about the GPU device */
printf("> GPU device has %d Multi-Processors, SM %d.%d compute capabilities\n\n",
deviceProp.multiProcessorCount, deviceProp.major, deviceProp.minor);
int version = (deviceProp.major * 0x10 + deviceProp.minor);
if (version < 0x11)
{
printf("%s: requires a minimum CUDA compute 1.1 capability\n", sSDKname);
// 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);
}
/* Generate a random tridiagonal symmetric matrix in CSR (Compressed Sparse Row) format */
M = N = 16384;
nz = 5*N-4*(int)sqrt((double)N);
I = (int *)malloc(sizeof(int)*(N+1)); // csr row pointers for matrix A
J = (int *)malloc(sizeof(int)*nz); // csr column indices for matrix A
val = (float *)malloc(sizeof(float)*nz); // csr values for matrix A
x = (float *)malloc(sizeof(float)*N);
rhs = (float *)malloc(sizeof(float)*N);
for (int i = 0; i < N; i++)
{
rhs[i] = 0.0; // Initialize RHS
x[i] = 0.0; // Initial approximation of solution
}
genLaplace(I, J, val, M, N, nz, rhs);
/* Create CUBLAS context */
cublasHandle_t cublasHandle = 0;
cublasStatus_t cublasStatus;
cublasStatus = cublasCreate(&cublasHandle);
checkCudaErrors(cublasStatus);
/* Create CUSPARSE context */
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
checkCudaErrors(cusparseStatus);
/* Description of the A matrix*/
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
checkCudaErrors(cusparseStatus);
/* Define the properties of the matrix */
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
/* Allocate required memory */
checkCudaErrors(cudaMalloc((void **)&d_col, nz*sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_row, (N+1)*sizeof(int)));
checkCudaErrors(cudaMalloc((void **)&d_val, nz*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_x, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_y, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_r, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_p, N*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_omega, N*sizeof(float)));
cudaMemcpy(d_col, J, nz*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_row, I, (N+1)*sizeof(int), cudaMemcpyHostToDevice);
cudaMemcpy(d_val, val, nz*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_r, rhs, N*sizeof(float), cudaMemcpyHostToDevice);
/* Conjugate gradient without preconditioning.
------------------------------------------
Follows the description by Golub & Van Loan, "Matrix Computations 3rd ed.", Section 10.2.6 */
printf("Convergence of conjugate gradient without preconditioning: \n");
k = 0;
r0 = 0;
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
while (r1 > tol*tol && k <= max_iter)
{
k++;
if (k == 1)
{
cublasScopy(cublasHandle, N, d_r, 1, d_p, 1);
}
else
{
beta = r1/r0;
cublasSscal(cublasHandle, N, &beta, d_p, 1);
cublasSaxpy(cublasHandle, N, &floatone, d_r, 1, d_p, 1) ;
}
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nz, &floatone, descr, d_val, d_row, d_col, d_p, &floatzero, d_omega);
cublasSdot(cublasHandle, N, d_p, 1, d_omega, 1, &dot);
alpha = r1/dot;
cublasSaxpy(cublasHandle, N, &alpha, d_p, 1, d_x, 1);
nalpha = -alpha;
cublasSaxpy(cublasHandle, N, &nalpha, d_omega, 1, d_r, 1);
r0 = r1;
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
}
printf(" iteration = %3d, residual = %e \n", k, sqrt(r1));
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
/* check result */
err = 0.0;
for (int i = 0; i < N; i++)
{
rsum = 0.0;
for (int j = I[i]; j < I[i+1]; j++)
{
rsum += val[j]*x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err)
{
err = diff;
}
}
printf(" Convergence Test: %s \n", (k <= max_iter) ? "OK" : "FAIL");
nErrors += (k > max_iter) ? 1 : 0;
qaerr1 = err;
if (0)
{
// output result in matlab-style array
int n=(int)sqrt((double)N);
printf("a = [ ");
for (int iy=0; iy<n; iy++)
{
for (int ix=0; ix<n; ix++)
{
printf(" %f ", x[iy*n+ix]);
}
if (iy == n-1)
{
printf(" ]");
}
printf("\n");
}
}
/* Preconditioned Conjugate Gradient using ILU.
--------------------------------------------
Follows the description by Golub & Van Loan, "Matrix Computations 3rd ed.", Algorithm 10.3.1 */
printf("\nConvergence of conjugate gradient using incomplete LU preconditioning: \n");
int nzILU0 = 2*N-1;
valsILU0 = (float *) malloc(nz*sizeof(float));
checkCudaErrors(cudaMalloc((void **)&d_valsILU0, nz*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_zm1, (N)*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_zm2, (N)*sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_rm2, (N)*sizeof(float)));
/* create the analysis info object for the A matrix */
cusparseSolveAnalysisInfo_t infoA = 0;
cusparseStatus = cusparseCreateSolveAnalysisInfo(&infoA);
checkCudaErrors(cusparseStatus);
/* Perform the analysis for the Non-Transpose case */
cusparseStatus = cusparseScsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE,
N, nz, descr, d_val, d_row, d_col, infoA);
checkCudaErrors(cusparseStatus);
/* Copy A data to ILU0 vals as input*/
cudaMemcpy(d_valsILU0, d_val, nz*sizeof(float), cudaMemcpyDeviceToDevice);
/* generate the Incomplete LU factor H for the matrix A using cudsparseScsrilu0 */
cusparseStatus = cusparseScsrilu0(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, descr, d_valsILU0, d_row, d_col, infoA);
checkCudaErrors(cusparseStatus);
/* Create info objects for the ILU0 preconditioner */
cusparseSolveAnalysisInfo_t info_u;
cusparseCreateSolveAnalysisInfo(&info_u);
cusparseMatDescr_t descrL = 0;
cusparseStatus = cusparseCreateMatDescr(&descrL);
cusparseSetMatType(descrL,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descrL,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatFillMode(descrL, CUSPARSE_FILL_MODE_LOWER);
cusparseSetMatDiagType(descrL, CUSPARSE_DIAG_TYPE_UNIT);
cusparseMatDescr_t descrU = 0;
cusparseStatus = cusparseCreateMatDescr(&descrU);
cusparseSetMatType(descrU,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descrU,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatFillMode(descrU, CUSPARSE_FILL_MODE_UPPER);
cusparseSetMatDiagType(descrU, CUSPARSE_DIAG_TYPE_NON_UNIT);
cusparseStatus = cusparseScsrsv_analysis(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, nz, descrU, d_val, d_row, d_col, info_u);
/* reset the initial guess of the solution to zero */
for (int i = 0; i < N; i++)
{
x[i] = 0.0;
}
checkCudaErrors(cudaMemcpy(d_r, rhs, N*sizeof(float), cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice));
k = 0;
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
while (r1 > tol*tol && k <= max_iter)
{
// Forward Solve, we can re-use infoA since the sparsity pattern of A matches that of L
cusparseStatus = cusparseScsrsv_solve(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, &floatone, descrL,
d_valsILU0, d_row, d_col, infoA, d_r, d_y);
checkCudaErrors(cusparseStatus);
// Back Substitution
cusparseStatus = cusparseScsrsv_solve(cusparseHandle, CUSPARSE_OPERATION_NON_TRANSPOSE, N, &floatone, descrU,
d_valsILU0, d_row, d_col, info_u, d_y, d_zm1);
checkCudaErrors(cusparseStatus);
k++;
if (k == 1)
{
cublasScopy(cublasHandle, N, d_zm1, 1, d_p, 1);
}
else
{
cublasSdot(cublasHandle, N, d_r, 1, d_zm1, 1, &numerator);
cublasSdot(cublasHandle, N, d_rm2, 1, d_zm2, 1, &denominator);
beta = numerator/denominator;
cublasSscal(cublasHandle, N, &beta, d_p, 1);
cublasSaxpy(cublasHandle, N, &floatone, d_zm1, 1, d_p, 1) ;
}
cusparseScsrmv(cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE, N, N, nzILU0, &floatone, descrU, d_val, d_row, d_col, d_p, &floatzero, d_omega);
cublasSdot(cublasHandle, N, d_r, 1, d_zm1, 1, &numerator);
cublasSdot(cublasHandle, N, d_p, 1, d_omega, 1, &denominator);
alpha = numerator / denominator;
cublasSaxpy(cublasHandle, N, &alpha, d_p, 1, d_x, 1);
cublasScopy(cublasHandle, N, d_r, 1, d_rm2, 1);
cublasScopy(cublasHandle, N, d_zm1, 1, d_zm2, 1);
nalpha = -alpha;
cublasSaxpy(cublasHandle, N, &nalpha, d_omega, 1, d_r, 1);
cublasSdot(cublasHandle, N, d_r, 1, d_r, 1, &r1);
}
printf(" iteration = %3d, residual = %e \n", k, sqrt(r1));
cudaMemcpy(x, d_x, N*sizeof(float), cudaMemcpyDeviceToHost);
/* check result */
err = 0.0;
for (int i = 0; i < N; i++)
{
rsum = 0.0;
for (int j = I[i]; j < I[i+1]; j++)
{
rsum += val[j]*x[J[j]];
}
diff = fabs(rsum - rhs[i]);
if (diff > err)
{
err = diff;
}
}
printf(" Convergence Test: %s \n", (k <= max_iter) ? "OK" : "FAIL");
nErrors += (k > max_iter) ? 1 : 0;
qaerr2 = err;
/* Destroy parameters */
cusparseDestroySolveAnalysisInfo(infoA);
cusparseDestroySolveAnalysisInfo(info_u);
/* Destroy contexts */
cusparseDestroy(cusparseHandle);
cublasDestroy(cublasHandle);
/* Free device memory */
free(I);
free(J);
free(val);
free(x);
free(rhs);
free(valsILU0);
cudaFree(d_col);
cudaFree(d_row);
cudaFree(d_val);
cudaFree(d_x);
cudaFree(d_y);
cudaFree(d_r);
cudaFree(d_p);
cudaFree(d_omega);
cudaFree(d_valsILU0);
cudaFree(d_zm1);
cudaFree(d_zm2);
cudaFree(d_rm2);
// 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(" Test Summary:\n");
printf(" Counted total of %d errors\n", nErrors);
printf(" qaerr1 = %f qaerr2 = %f\n\n", fabs(qaerr1), fabs(qaerr2));
exit((nErrors == 0 &&fabs(qaerr1)<1e-5 && fabs(qaerr2) < 1e-5 ? EXIT_SUCCESS : EXIT_FAILURE));
}
////////////////////////////////////////////////////////////////////////////////
// Main program
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
float
*h_Kernel,
*h_Input,
*h_Buffer,
*h_OutputCPU,
*h_OutputGPU;
cudaArray
*a_Src;
cudaChannelFormatDesc floatTex = cudaCreateChannelDesc<float>();
float
*d_Output;
float
gpuTime;
StopWatchInterface *hTimer = NULL;
const int imageW = 3072;
const int imageH = 3072 / 2;
const unsigned int iterations = 10;
printf("[%s] - Starting...\n", argv[0]);
// use command-line specified CUDA device, otherwise use device with highest Gflops/s
findCudaDevice(argc, (const char **)argv);
sdkCreateTimer(&hTimer);
printf("Initializing data...\n");
h_Kernel = (float *)malloc(KERNEL_LENGTH * sizeof(float));
h_Input = (float *)malloc(imageW * imageH * sizeof(float));
h_Buffer = (float *)malloc(imageW * imageH * sizeof(float));
h_OutputCPU = (float *)malloc(imageW * imageH * sizeof(float));
h_OutputGPU = (float *)malloc(imageW * imageH * sizeof(float));
checkCudaErrors(cudaMallocArray(&a_Src, &floatTex, imageW, imageH));
checkCudaErrors(cudaMalloc((void **)&d_Output, imageW * imageH * sizeof(float)));
srand(2009);
for (unsigned int i = 0; i < KERNEL_LENGTH; i++)
{
h_Kernel[i] = (float)(rand() % 16);
}
for (unsigned int i = 0; i < imageW * imageH; i++)
{
h_Input[i] = (float)(rand() % 16);
}
setConvolutionKernel(h_Kernel);
checkCudaErrors(cudaMemcpyToArray(a_Src, 0, 0, h_Input, imageW * imageH * sizeof(float), cudaMemcpyHostToDevice));
printf("Running GPU rows convolution (%u identical iterations)...\n", iterations);
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
for (unsigned int i = 0; i < iterations; i++)
{
convolutionRowsGPU(
d_Output,
a_Src,
imageW,
imageH
);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer) / (float)iterations;
printf("Average convolutionRowsGPU() time: %f msecs; //%f Mpix/s\n", gpuTime, imageW * imageH * 1e-6 / (0.001 * gpuTime));
//While CUDA kernels can't write to textures directly, this copy is inevitable
printf("Copying convolutionRowGPU() output back to the texture...\n");
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
checkCudaErrors(cudaMemcpyToArray(a_Src, 0, 0, d_Output, imageW * imageH * sizeof(float), cudaMemcpyDeviceToDevice));
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer);
printf("cudaMemcpyToArray() time: %f msecs; //%f Mpix/s\n", gpuTime, imageW * imageH * 1e-6 / (0.001 * gpuTime));
printf("Running GPU columns convolution (%i iterations)\n", iterations);
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
for (int i = 0; i < iterations; i++)
{
convolutionColumnsGPU(
d_Output,
a_Src,
imageW,
imageH
);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
gpuTime = sdkGetTimerValue(&hTimer) / (float)iterations;
printf("Average convolutionColumnsGPU() time: %f msecs; //%f Mpix/s\n", gpuTime, imageW * imageH * 1e-6 / (0.001 * gpuTime));
printf("Reading back GPU results...\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, imageW * imageH * sizeof(float), cudaMemcpyDeviceToHost));
printf("Checking the results...\n");
printf("...running convolutionRowsCPU()\n");
convolutionRowsCPU(
h_Buffer,
h_Input,
h_Kernel,
imageW,
imageH,
KERNEL_RADIUS
);
printf("...running convolutionColumnsCPU()\n");
convolutionColumnsCPU(
h_OutputCPU,
h_Buffer,
h_Kernel,
imageW,
imageH,
KERNEL_RADIUS
);
double delta = 0;
double sum = 0;
for (unsigned int i = 0; i < imageW * imageH; i++)
{
sum += h_OutputCPU[i] * h_OutputCPU[i];
delta += (h_OutputGPU[i] - h_OutputCPU[i]) * (h_OutputGPU[i] - h_OutputCPU[i]);
}
double L2norm = sqrt(delta / sum);
printf("Relative L2 norm: %E\n", L2norm);
printf("Shutting down...\n");
checkCudaErrors(cudaFree(d_Output));
checkCudaErrors(cudaFreeArray(a_Src));
free(h_OutputGPU);
free(h_Buffer);
free(h_Input);
free(h_Kernel);
sdkDeleteTimer(&hTimer);
// 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 (L2norm > 1e-6)
{
printf("Test failed!\n");
exit(EXIT_FAILURE);
}
printf("Test passed\n");
exit(EXIT_SUCCESS);
}
////////////////////////////////////////////////////////////////////////////////
// Main program
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
// start logs
printf("[%s] - Starting...\n", argv[0]);
float
*h_Kernel,
*h_Input,
*h_Buffer,
*h_OutputCPU,
*h_OutputGPU;
float
*d_Input,
*d_Output,
*d_Buffer;
const int imageW = 3072;
const int imageH = 3072;
const int iterations = 16;
StopWatchInterface *hTimer = NULL;
//Use command-line specified CUDA device, otherwise use device with highest Gflops/s
findCudaDevice(argc, (const char **)argv);
sdkCreateTimer(&hTimer);
printf("Image Width x Height = %i x %i\n\n", imageW, imageH);
printf("Allocating and initializing host arrays...\n");
h_Kernel = (float *)malloc(KERNEL_LENGTH * sizeof(float));
h_Input = (float *)malloc(imageW * imageH * sizeof(float));
h_Buffer = (float *)malloc(imageW * imageH * sizeof(float));
h_OutputCPU = (float *)malloc(imageW * imageH * sizeof(float));
h_OutputGPU = (float *)malloc(imageW * imageH * sizeof(float));
srand(200);
for (unsigned int i = 0; i < KERNEL_LENGTH; i++)
{
h_Kernel[i] = (float)(rand() % 16);
}
for (unsigned i = 0; i < imageW * imageH; i++)
{
h_Input[i] = (float)(rand() % 16);
}
printf("Allocating and initializing CUDA arrays...\n");
checkCudaErrors(cudaMalloc((void **)&d_Input, imageW * imageH * sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_Output, imageW * imageH * sizeof(float)));
checkCudaErrors(cudaMalloc((void **)&d_Buffer , imageW * imageH * sizeof(float)));
setConvolutionKernel(h_Kernel);
checkCudaErrors(cudaMemcpy(d_Input, h_Input, imageW * imageH * sizeof(float), cudaMemcpyHostToDevice));
printf("Running GPU convolution (%u identical iterations)...\n\n", iterations);
for (int i = -1; i < iterations; i++)
{
//i == -1 -- warmup iteration
if (i == 0)
{
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
}
convolutionRowsGPU(
d_Buffer,
d_Input,
imageW,
imageH
);
convolutionColumnsGPU(
d_Output,
d_Buffer,
imageW,
imageH
);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
double gpuTime = 0.001 * sdkGetTimerValue(&hTimer) / (double)iterations;
printf("convolutionSeparable, Throughput = %.4f MPixels/sec, Time = %.5f s, Size = %u Pixels, NumDevsUsed = %i, Workgroup = %u\n",
(1.0e-6 * (double)(imageW * imageH)/ gpuTime), gpuTime, (imageW * imageH), 1, 0);
printf("\nReading back GPU results...\n\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, imageW * imageH * sizeof(float), cudaMemcpyDeviceToHost));
printf("Checking the results...\n");
printf(" ...running convolutionRowCPU()\n");
convolutionRowCPU(
h_Buffer,
h_Input,
h_Kernel,
imageW,
imageH,
KERNEL_RADIUS
);
printf(" ...running convolutionColumnCPU()\n");
convolutionColumnCPU(
h_OutputCPU,
h_Buffer,
h_Kernel,
imageW,
imageH,
KERNEL_RADIUS
);
printf(" ...comparing the results\n");
double sum = 0, delta = 0;
for (unsigned i = 0; i < imageW * imageH; i++)
{
delta += (h_OutputGPU[i] - h_OutputCPU[i]) * (h_OutputGPU[i] - h_OutputCPU[i]);
sum += h_OutputCPU[i] * h_OutputCPU[i];
}
double L2norm = sqrt(delta / sum);
printf(" ...Relative L2 norm: %E\n\n", L2norm);
printf("Shutting down...\n");
checkCudaErrors(cudaFree(d_Buffer));
checkCudaErrors(cudaFree(d_Output));
checkCudaErrors(cudaFree(d_Input));
free(h_OutputGPU);
free(h_OutputCPU);
free(h_Buffer);
free(h_Input);
free(h_Kernel);
sdkDeleteTimer(&hTimer);
// 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 (L2norm > 1e-6)
{
printf("Test failed!\n");
exit(EXIT_FAILURE);
}
printf("Test passed\n");
exit(EXIT_SUCCESS);
}
////////////////////////////////////////////////////////////////////////////////
// Test driver
////////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
uint *h_SrcKey, *h_SrcVal, *h_DstKey, *h_DstVal;
uint *d_SrcKey, *d_SrcVal, *d_BufKey, *d_BufVal, *d_DstKey, *d_DstVal;
StopWatchInterface *hTimer = NULL;
const uint N = 4 * 1048576;
const uint DIR = 1;
const uint numValues = 65536;
printf("%s Starting...\n\n", argv[0]);
int dev = findCudaDevice(argc, (const char **) argv);
if (dev == -1)
{
return EXIT_FAILURE;
}
printf("Allocating and initializing host arrays...\n\n");
sdkCreateTimer(&hTimer);
h_SrcKey = (uint *)malloc(N * sizeof(uint));
h_SrcVal = (uint *)malloc(N * sizeof(uint));
h_DstKey = (uint *)malloc(N * sizeof(uint));
h_DstVal = (uint *)malloc(N * sizeof(uint));
srand(2009);
for (uint i = 0; i < N; i++)
{
h_SrcKey[i] = rand() % numValues;
}
fillValues(h_SrcVal, N);
printf("Allocating and initializing CUDA arrays...\n\n");
checkCudaErrors(cudaMalloc((void **)&d_DstKey, N * sizeof(uint)));
checkCudaErrors(cudaMalloc((void **)&d_DstVal, N * sizeof(uint)));
checkCudaErrors(cudaMalloc((void **)&d_BufKey, N * sizeof(uint)));
checkCudaErrors(cudaMalloc((void **)&d_BufVal, N * sizeof(uint)));
checkCudaErrors(cudaMalloc((void **)&d_SrcKey, N * sizeof(uint)));
checkCudaErrors(cudaMalloc((void **)&d_SrcVal, N * sizeof(uint)));
checkCudaErrors(cudaMemcpy(d_SrcKey, h_SrcKey, N * sizeof(uint), cudaMemcpyHostToDevice));
checkCudaErrors(cudaMemcpy(d_SrcVal, h_SrcVal, N * sizeof(uint), cudaMemcpyHostToDevice));
printf("Initializing GPU merge sort...\n");
initMergeSort();
printf("Running GPU merge sort...\n");
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
mergeSort(
d_DstKey,
d_DstVal,
d_BufKey,
d_BufVal,
d_SrcKey,
d_SrcVal,
N,
DIR
);
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
printf("Time: %f ms\n", sdkGetTimerValue(&hTimer));
printf("Reading back GPU merge sort results...\n");
checkCudaErrors(cudaMemcpy(h_DstKey, d_DstKey, N * sizeof(uint), cudaMemcpyDeviceToHost));
checkCudaErrors(cudaMemcpy(h_DstVal, d_DstVal, N * sizeof(uint), cudaMemcpyDeviceToHost));
printf("Inspecting the results...\n");
uint keysFlag = validateSortedKeys(
h_DstKey,
h_SrcKey,
1,
N,
numValues,
DIR
);
uint valuesFlag = validateSortedValues(
h_DstKey,
h_DstVal,
h_SrcKey,
1,
N
);
printf("Shutting down...\n");
closeMergeSort();
sdkDeleteTimer(&hTimer);
checkCudaErrors(cudaFree(d_SrcVal));
checkCudaErrors(cudaFree(d_SrcKey));
checkCudaErrors(cudaFree(d_BufVal));
checkCudaErrors(cudaFree(d_BufKey));
checkCudaErrors(cudaFree(d_DstVal));
checkCudaErrors(cudaFree(d_DstKey));
free(h_DstVal);
free(h_DstKey);
free(h_SrcVal);
free(h_SrcKey);
// 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((keysFlag && valuesFlag) ? EXIT_SUCCESS : EXIT_FAILURE);
}
int main(int argc, char **argv)
{
printf("%s Starting...\n\n", argv[0]);
//Use command-line specified CUDA device, otherwise use device with highest Gflops/s
findCudaDevice(argc, (const char **)argv);
uint *d_Input, *d_Output;
uint *h_Input, *h_OutputCPU, *h_OutputGPU;
StopWatchInterface *hTimer = NULL;
const uint N = 13 * 1048576 / 2;
printf("Allocating and initializing host arrays...\n");
sdkCreateTimer(&hTimer);
h_Input = (uint *)malloc(N * sizeof(uint));
h_OutputCPU = (uint *)malloc(N * sizeof(uint));
h_OutputGPU = (uint *)malloc(N * sizeof(uint));
srand(2009);
for (uint i = 0; i < N; i++)
{
h_Input[i] = rand();
}
printf("Allocating and initializing CUDA arrays...\n");
checkCudaErrors(cudaMalloc((void **)&d_Input, N * sizeof(uint)));
checkCudaErrors(cudaMalloc((void **)&d_Output, N * sizeof(uint)));
checkCudaErrors(cudaMemcpy(d_Input, h_Input, N * sizeof(uint), cudaMemcpyHostToDevice));
printf("Initializing CUDA-C scan...\n\n");
initScan();
int globalFlag = 1;
size_t szWorkgroup;
const int iCycles = 100;
printf("*** Running GPU scan for short arrays (%d identical iterations)...\n\n", iCycles);
for (uint arrayLength = MIN_SHORT_ARRAY_SIZE; arrayLength <= MAX_SHORT_ARRAY_SIZE; arrayLength <<= 1)
{
printf("Running scan for %u elements (%u arrays)...\n", arrayLength, N / arrayLength);
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
for (int i = 0; i < iCycles; i++)
{
szWorkgroup = scanExclusiveShort(d_Output, d_Input, N / arrayLength, arrayLength);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
double timerValue = 1.0e-3 * sdkGetTimerValue(&hTimer) / iCycles;
printf("Validating the results...\n");
printf("...reading back GPU results\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, N * sizeof(uint), cudaMemcpyDeviceToHost));
printf(" ...scanExclusiveHost()\n");
scanExclusiveHost(h_OutputCPU, h_Input, N / arrayLength, arrayLength);
// Compare GPU results with CPU results and accumulate error for this test
printf(" ...comparing the results\n");
int localFlag = 1;
for (uint i = 0; i < N; i++)
{
if (h_OutputCPU[i] != h_OutputGPU[i])
{
localFlag = 0;
break;
}
}
// Log message on individual test result, then accumulate to global flag
printf(" ...Results %s\n\n", (localFlag == 1) ? "Match" : "DON'T Match !!!");
globalFlag = globalFlag && localFlag;
// Data log
if (arrayLength == MAX_SHORT_ARRAY_SIZE)
{
printf("\n");
printf("scan-Short, Throughput = %.4f MElements/s, Time = %.5f s, Size = %u Elements, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * (double)arrayLength/timerValue), timerValue, (unsigned int)arrayLength, 1, (unsigned int)szWorkgroup);
printf("\n");
}
}
printf("***Running GPU scan for large arrays (%u identical iterations)...\n\n", iCycles);
for (uint arrayLength = MIN_LARGE_ARRAY_SIZE; arrayLength <= MAX_LARGE_ARRAY_SIZE; arrayLength <<= 1)
{
printf("Running scan for %u elements (%u arrays)...\n", arrayLength, N / arrayLength);
checkCudaErrors(cudaDeviceSynchronize());
sdkResetTimer(&hTimer);
sdkStartTimer(&hTimer);
for (int i = 0; i < iCycles; i++)
{
szWorkgroup = scanExclusiveLarge(d_Output, d_Input, N / arrayLength, arrayLength);
}
checkCudaErrors(cudaDeviceSynchronize());
sdkStopTimer(&hTimer);
double timerValue = 1.0e-3 * sdkGetTimerValue(&hTimer) / iCycles;
printf("Validating the results...\n");
printf("...reading back GPU results\n");
checkCudaErrors(cudaMemcpy(h_OutputGPU, d_Output, N * sizeof(uint), cudaMemcpyDeviceToHost));
printf("...scanExclusiveHost()\n");
scanExclusiveHost(h_OutputCPU, h_Input, N / arrayLength, arrayLength);
// Compare GPU results with CPU results and accumulate error for this test
printf(" ...comparing the results\n");
int localFlag = 1;
for (uint i = 0; i < N; i++)
{
if (h_OutputCPU[i] != h_OutputGPU[i])
{
localFlag = 0;
break;
}
}
// Log message on individual test result, then accumulate to global flag
printf(" ...Results %s\n\n", (localFlag == 1) ? "Match" : "DON'T Match !!!");
globalFlag = globalFlag && localFlag;
// Data log
if (arrayLength == MAX_LARGE_ARRAY_SIZE)
{
printf("\n");
printf("scan-Large, Throughput = %.4f MElements/s, Time = %.5f s, Size = %u Elements, NumDevsUsed = %u, Workgroup = %u\n",
(1.0e-6 * (double)arrayLength/timerValue), timerValue, (unsigned int)arrayLength, 1, (unsigned int)szWorkgroup);
printf("\n");
}
}
printf("Shutting down...\n");
closeScan();
checkCudaErrors(cudaFree(d_Output));
checkCudaErrors(cudaFree(d_Input));
sdkDeleteTimer(&hTimer);
// 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();
// pass or fail (cumulative... all tests in the loop)
exit(globalFlag ? EXIT_SUCCESS : 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)
///////////////////////////////////////////////////////////////////////////////
/// application entry point
///////////////////////////////////////////////////////////////////////////////
int main(int argc, char **argv)
{
// welcome message
printf("%s Starting...\n\n", sSDKsample);
// pick GPU
findCudaDevice(argc, (const char **)argv);
// find images
const char *const sourceFrameName = "frame10.ppm";
const char *const targetFrameName = "frame11.ppm";
// image dimensions
int width;
int height;
// row access stride
int stride;
// flow is computed from source image to target image
float *h_source; // source image, host memory
float *h_target; // target image, host memory
// load image from file
if (!LoadImageAsFP32(h_source, width, height, stride, sourceFrameName, argv[0]))
{
exit(EXIT_FAILURE);
}
if (!LoadImageAsFP32(h_target, width, height, stride, targetFrameName, argv[0]))
{
exit(EXIT_FAILURE);
}
// allocate host memory for CPU results
float *h_uGold = new float [stride * height];
float *h_vGold = new float [stride * height];
// allocate host memory for GPU results
float *h_u = new float [stride * height];
float *h_v = new float [stride * height];
// smoothness
// if image brightness is not within [0,1]
// this paramter should be scaled appropriately
const float alpha = 0.2f;
// number of pyramid levels
const int nLevels = 5;
// number of solver iterations on each level
const int nSolverIters = 500;
// number of warping iterations
const int nWarpIters = 3;
ComputeFlowGold(h_source, h_target, width, height, stride, alpha,
nLevels, nWarpIters, nSolverIters, h_uGold, h_vGold);
ComputeFlowCUDA(h_source, h_target, width, height, stride, alpha,
nLevels, nWarpIters, nSolverIters, h_u, h_v);
// compare results (L1 norm)
bool
status = CompareWithGold(width, height, stride, h_uGold, h_vGold, h_u, h_v);
WriteFloFile("FlowGPU.flo", width, height, stride, h_u, h_v);
WriteFloFile("FlowCPU.flo", width, height, stride, h_uGold, h_vGold);
// free resources
delete [] h_uGold;
delete [] h_vGold;
delete [] h_u;
delete [] h_v;
delete [] h_source;
delete [] h_target;
// 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();
// report self-test status
exit(status ? EXIT_SUCCESS : EXIT_FAILURE);
}