/* GPUrand */
void GPUrand(const GPUtype &OUT) {
curandStatus_t status;
gpuTYPE_t type = gm->gputype.getType(OUT);
gm->gmat->control.cacheClean();
const void *gpuptr = gm->gputype.getGPUptr(OUT); // pointer to GPU memory
int numel = gm->gputype.getNumel(OUT); // number of elements
int datasize = gm->gputype.getDataSize(OUT); // bytes for each element
gen = 0;
// implement recovery procedure
// try and if error try again
// init curand
if (curandCreateGenerator(&gen,CURAND_RNG_PSEUDO_DEFAULT)!=CURAND_STATUS_SUCCESS) {
mexErrMsgTxt(ERROR_CURAND_INIT);
}
//if (curandCreateGenerator(&gen,CURAND_RNG_QUASI_DEFAULT)!=CURAND_STATUS_SUCCESS) {
// mexErrMsgTxt(ERROR_CURAND_INIT);
//}
// seed
seed++;
if (curandSetPseudoRandomGeneratorSeed(gen, time(NULL)+seed)!=CURAND_STATUS_SUCCESS) {
mexErrMsgTxt(ERROR_CURAND_SEED);
}
if (type == gpuFLOAT) {
status = curandGenerateUniform(gen, (float *) gpuptr, numel);
} else if (type == gpuCFLOAT) {
status = curandGenerateUniform(gen, (float *) gpuptr, numel*2);
} else if (type == gpuDOUBLE) {
status = curandGenerateUniformDouble(gen, (double *) gpuptr, numel);
} else if (type == gpuCDOUBLE) {
status = curandGenerateUniformDouble(gen, (double *) gpuptr, numel*2);
}
if (status!=CURAND_STATUS_SUCCESS) {
curandDestroyGenerator(gen);
mexErrMsgTxt(ERROR_CURAND_GEN);
}
// destroy
if (curandDestroyGenerator(gen)!=CURAND_STATUS_SUCCESS) {
mexErrMsgTxt(ERROR_CURAND_DESTROY);
}
}
void Caffe::SetDevice(const int device_id) {
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == device_id) {
return;
}
// The call to cudaSetDevice must come before any calls to Get, which
// may perform initialization using the GPU.
CUDA_CHECK(cudaSetDevice(device_id));
if (Get().cublas_handle_) CUBLAS_CHECK(cublasDestroy(Get().cublas_handle_));
if (Get().cusparse_descr_)CUSPARSE_CHECK(cusparseDestroyMatDescr(Get().cusparse_descr_));
if (Get().cusparse_handle_)CUSPARSE_CHECK(cusparseDestroy(Get().cusparse_handle_));
if (Get().curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(Get().curand_generator_));
}
CUSPARSE_CHECK(cusparseCreate(&Get().cusparse_handle_));
CUSPARSE_CHECK(cusparseCreateMatDescr(&Get().cusparse_descr_));
// cusparseSetMatType(cusparse_descr_,CUSPARSE_MATRIX_TYPE_GENERAL);
// cusparseSetMatIndexBase(cusparse_descr_,CUSPARSE_INDEX_BASE_ZERO);
LOG(INFO)<<"set descr";
CUBLAS_CHECK(cublasCreate(&Get().cublas_handle_));
CURAND_CHECK(curandCreateGenerator(&Get().curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(Get().curand_generator_,
cluster_seedgen()));
}
curand_generator::
~curand_generator()
{
if (handle)
{
curandDestroyGenerator((curandGenerator_t)handle);
}
}
Caffe::~Caffe() {
if (cusparse_descr_) CUSPARSE_CHECK(cusparseDestroyMatDescr(cusparse_descr_));
if (cublas_handle_) CUBLAS_CHECK(cublasDestroy(cublas_handle_));
if (cusparse_handle_) CUSPARSE_CHECK(cusparseDestroy(cusparse_handle_));
if (curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(curand_generator_));
}
}
void Caffe::set_random_seed(unsigned int seed)
{
CURAND_CHECK(curandDestroyGenerator(Get().curand_generator_));
CURAND_CHECK(curandCreateGenerator(&Get().curand_generator_, CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(Get().curand_generator_, seed));
VSL_CHECK(vslDeleteStream(&Get().vsl_stream_));
VSL_CHECK(vslNewStream(&Get().vsl_stream_, VSL_BRNG_MT19937, seed));
}
Caffe::~Caffe()
{
if (cublas_handle_)
CUBLAS_CHECK(cublasDestroy(cublas_handle_));
if (curand_generator_)
CURAND_CHECK(curandDestroyGenerator(curand_generator_));
if (vsl_stream_)
VSL_CHECK(vslDeleteStream(&vsl_stream_));
}
WIE::Random::~Random()
{
assert(generator);
curandDestroyGenerator(generator);
generator = NULL;
assert(samples);
cudaFree(samples);
samples = NULL;
}
Caffe::~Caffe() {
// Make sure all device contexts and
// dependent memory blocks are freed properly
device_contexts_.clear();
#ifdef USE_CUDA
if (cublas_handle_)
CUBLAS_CHECK(cublasDestroy(cublas_handle_));
if (curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(curand_generator_));
}
#endif // USE_CUDA
}
Caffe::~Caffe() {
for (vector<cublasHandle_t>& group_cublas_handles : cublas_handles_) {
for (cublasHandle_t h : group_cublas_handles) {
if (h) {
CUBLAS_CHECK(cublasDestroy(h));
}
}
}
for_each(curand_generators_.begin(), curand_generators_.end(), [](curandGenerator_t h) {
if (h) {
CURAND_CHECK(curandDestroyGenerator(h));
}
});
}
void Dragon::set_device(const int device_id) {
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == device_id) return;
// The call to cudaSetDevice must come before any calls to Get, which
// may perform initialization using the GPU.
// reset Device must reset handle and generator???
CUDA_CHECK(cudaSetDevice(device_id));
if (Get().cublas_handle) cublasDestroy_v2(Get().cublas_handle);
if (Get().curand_generator) curandDestroyGenerator(Get().curand_generator);
cublasCreate_v2(&Get().cublas_handle);
curandCreateGenerator(&Get().curand_generator, CURAND_RNG_PSEUDO_DEFAULT);
curandSetPseudoRandomGeneratorSeed(Get().curand_generator, cluster_seedgen());
}
void Caffe::set_random_seed(const unsigned int seed) {
// Curand seed
// Yangqing's note: simply setting the generator seed does not seem to
// work on the tesla K20s, so I wrote the ugly reset thing below.
if (Get().curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(curand_generator()));
CURAND_CHECK(curandCreateGenerator(&Get().curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(curand_generator(),
seed));
} else {
LOG(ERROR) << "Curand not available. Skipping setting the curand seed.";
}
// RNG seed
Get().random_generator_.reset(new RNG(seed));
}
void Caffe::SetDevice(const int device_id) {
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == device_id) {
return;
}
if (Get().cublas_handle_) CUBLAS_CHECK(cublasDestroy(Get().cublas_handle_));
if (Get().curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(Get().curand_generator_));
}
CUDA_CHECK(cudaSetDevice(device_id));
CUBLAS_CHECK(cublasCreate(&Get().cublas_handle_));
CURAND_CHECK(curandCreateGenerator(&Get().curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(Get().curand_generator_,
cluster_seedgen()));
}
void Engine::SetDevice(const int device_id) {
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == device_id) {
return;
}
// The call to cudaSetDevice must come before any calls to Get, which
// may perform initialization using the GPU.
CUDA_CHECK(cudaSetDevice(device_id));
if (Get().cublas_handle_) CUBLAS_CHECK(cublasDestroy(Get().cublas_handle_));
if (Get().curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(Get().curand_generator_));
}
CUBLAS_CHECK(cublasCreate(&Get().cublas_handle_));
CURAND_CHECK(curandCreateGenerator(&Get().curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(Get().curand_generator_,
cluster_seedgen()));
}
void Caffe::SetSlaveDevice(const int slave_device_id) {
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == slave_device_id) {
return;
}
if (Get().slave_cublas_handle_) CUBLAS_CHECK(cublasDestroy(Get().slave_cublas_handle_));
if (Get().slave_curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(Get().slave_curand_generator_));
}
CUDA_CHECK(cudaSetDevice(slave_device_id));
CUDA_CHECK(cudaStreamCreate (&Get().slave_cu_stream_));
CUBLAS_CHECK(cublasCreate(&Get().slave_cublas_handle_));
CUBLAS_CHECK(cublasSetStream(Get().slave_cublas_handle_, Get().slave_cu_stream_));
CURAND_CHECK(curandCreateGenerator(&Get().slave_curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(curandSetPseudoRandomGeneratorSeed(Get().slave_curand_generator_,
cluster_seedgen()));
Get().slave_device_id_ = slave_device_id;
CUDA_CHECK(cudaSetDevice(current_device));
Caffe::set_gpu_mode(Caffe::MASTER_SLAVE);
}
void Caffe::SetDevice(const int device_id) {
std::vector<int> devices;
devices.push_back(device_id);
Caffe::SetDevices(devices);
Get().default_device_context_ = GetDeviceContext(device_id);
if (Get().default_device_context_->backend() == Backend::BACKEND_CUDA) {
#ifdef USE_CUDA
int current_device;
CUDA_CHECK(cudaGetDevice(¤t_device));
if (current_device == device_id) {
return;
}
// The call to cudaSetDevice must come before any calls to Get, which
// may perform initialization using the GPU.
CUDA_CHECK(cudaSetDevice(device_id));
if (Get().cublas_handle_)
CUBLAS_CHECK(cublasDestroy(Get().cublas_handle_));
if (Get().curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(Get().curand_generator_));
}
CUBLAS_CHECK(cublasCreate(&Get().cublas_handle_));
CURAND_CHECK(
curandCreateGenerator(&Get().curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
CURAND_CHECK(
curandSetPseudoRandomGeneratorSeed(Get().curand_generator_,
cluster_seedgen()));
#endif // USE_CUDA
} else {
#ifdef USE_GREENTEA
#ifdef USE_CLBLAS
clblasSetup();
#endif // USE_CLBLAS
#endif // USE_GREENTEA
}
}
///////////////////////////////////////////////////////////////////////////////
// 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);
}
Engine::~Engine() {
if (cublas_handle_) CUBLAS_CHECK(cublasDestroy(cublas_handle_));
if (curand_generator_) {
CURAND_CHECK(curandDestroyGenerator(curand_generator_));
}
}
/* GPUrandn */
void GPUrandn(const GPUtype &OUT) {
curandStatus_t status;
gpuTYPE_t type = gm->gputype.getType(OUT);
gm->gmat->control.cacheClean();
const void *gpuptr = gm->gputype.getGPUptr(OUT); // pointer to GPU memory
int numel = gm->gputype.getNumel(OUT); // number of elements
int datasize = gm->gputype.getDataSize(OUT); // bytes for each element
gen = 0;
// implement recovery procedure
// try and if error try again
// init curand
if (curandCreateGenerator(&gen,CURAND_RNG_PSEUDO_DEFAULT)!=CURAND_STATUS_SUCCESS) {
mexErrMsgTxt(ERROR_CURAND_INIT);
}
//if (curandCreateGenerator(&gen,CURAND_RNG_QUASI_DEFAULT)!=CURAND_STATUS_SUCCESS) {
// mexErrMsgTxt(ERROR_CURAND_INIT);
//}
// randn requires even numbers
// we split the execution in 2 parts (overlap if not even)
// seed
seed++;
if (curandSetPseudoRandomGeneratorSeed(gen, time(NULL)+seed)!=CURAND_STATUS_SUCCESS) {
mexErrMsgTxt(ERROR_CURAND_SEED);
}
unsigned int n = 0;
if (type == gpuFLOAT) {
n = numel;
} else if (type == gpuCFLOAT) {
n = numel*2;
} else if (type == gpuDOUBLE) {
n = numel;
} else if (type == gpuCDOUBLE) {
n = numel*2;
}
unsigned int even = (n%2) == 0;
unsigned int offset = 0;
unsigned int mysize = 0;
unsigned int iter = 1;
if (!even) {
n = n-1;
iter = 2;
}
if (type == gpuFLOAT) {
float mean = 0.0;
float std = 1.0;
status = curandGenerateNormal(gen, (float *) gpuptr, n, mean, std);
if (!even) {
float *devData;
if((cudaMalloc((void **)&devData, 4 * sizeof(float))) != cudaSuccess) {
status = CURAND_STATUS_LAUNCH_FAILURE;
} else {
status = curandGenerateNormal(gen, devData, 4, mean, std);
if (status==CURAND_STATUS_SUCCESS) {
void *dst = (void *) ((UINTPTR gpuptr)+n*datasize);
if (cudaMemcpy(dst, (void *) devData, datasize, cudaMemcpyDeviceToDevice)!=cudaSuccess) {
status = CURAND_STATUS_LAUNCH_FAILURE;
}
}
if(cudaFree(devData) != cudaSuccess) {
status = CURAND_STATUS_LAUNCH_FAILURE;
}
}
}
} else if (type == gpuCFLOAT) {
float mean = 0.0;
float std = 1.0;
status = curandGenerateNormal(gen, (float *) gpuptr, n, mean, std);
} else if (type == gpuDOUBLE) {
double mean = 0.0;
double std = 1.0;
status = curandGenerateNormalDouble(gen, (double *) gpuptr, n, mean, std);
if (!even) {
double *devData;
if((cudaMalloc((void **)&devData, 4 * sizeof(double))) != cudaSuccess) {
status = CURAND_STATUS_LAUNCH_FAILURE;
} else {
status = curandGenerateNormalDouble(gen, devData, 4, mean, std);
if (status==CURAND_STATUS_SUCCESS) {
void *dst = (void *) ((UINTPTR gpuptr)+n*datasize);
if (cudaMemcpy(dst, (void *) devData, datasize, cudaMemcpyDeviceToDevice)!=cudaSuccess) {
status = CURAND_STATUS_LAUNCH_FAILURE;
}
}
if(cudaFree(devData) != cudaSuccess) {
status = CURAND_STATUS_LAUNCH_FAILURE;
}
}
}
} else if (type == gpuCDOUBLE) {
double mean = 0.0;
double std = 1.0;
status = curandGenerateNormalDouble(gen, (double *) gpuptr, n, mean, std);
}
if (status!=CURAND_STATUS_SUCCESS) {
curandDestroyGenerator(gen);
mexErrMsgTxt(ERROR_CURAND_GEN);
}
// destroy
if (curandDestroyGenerator(gen)!=CURAND_STATUS_SUCCESS) {
mexErrMsgTxt(ERROR_CURAND_DESTROY);
}
}
Dragon::~Dragon(){
if (cublas_handle) cublasDestroy_v2(cublas_handle);
if (curand_generator) curandDestroyGenerator(curand_generator);
}
prngenerator_cuda<TFloat>::~prngenerator_cuda() {
CurandSafeCall(curandDestroyGenerator(_dev_bulk_prng_engine));
CudaSafeCall(cudaFree(_dev_prng_engines));
}
ModelWPAMGPU::~ModelWPAMGPU()
{
curandDestroyGenerator(gen);
}