/* 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);
}
}
inline void prngenerator_cuda<double>::_generate(const uint32_t n, double * output) {
// #ifdef _DEBUG
// __cudaCheckMemory();
// std::cout << "Generating " << n << " numbers." << std::endl;
// #endif
CurandSafeCall(curandGenerateUniformDouble (_dev_bulk_prng_engine,
output,
n));
}
void cuda_rand(void *ptr, int numel, bool dbl)
{
static bool is_init = false;
static curandGenerator_t stream;
if (!is_init) {
curandCreateGenerator(&stream, CURAND_RNG_PSEUDO_DEFAULT);
is_init = true;
}
if (!dbl) {
curandGenerateUniform(stream, (float *)ptr, numel);
} else {
curandGenerateUniformDouble(stream, (double *)ptr, numel);
}
return;
}
