/* Computes the Frobenius norm of a matrix. */
double nrm2(Mat mA) {
const int n2 = MatN2(mA);
const void* a = MatElems(mA);
const bool dev = MatDev(mA);
double norm;
switch (MatElemSize(mA)) {
case 4:
if (dev) {
float norm32;
cublasSnrm2(g_cublasHandle, n2, a, 1, (float*)&norm32);
norm = norm32;
} else {
norm = cblas_snrm2(n2, a, 1);
}
break;
case 8:
if (dev) {
cublasDnrm2(g_cublasHandle, n2, a, 1, (double*)&norm);
} else {
norm = cblas_dnrm2(n2, a, 1);
}
break;
}
return norm;
}
float CUDABLAS1::nrm2( IndexType n, const float* x_d, IndexType incx, SyncToken* syncToken )
{
LAMA_CHECK_CUDA_ACCESS
cudaStream_t stream = NULL;
if ( syncToken )
{
CUDAStreamSyncToken* cudaStreamSyncToken = dynamic_cast<CUDAStreamSyncToken*>( syncToken );
LAMA_ASSERT_DEBUG( cudaStreamSyncToken, "no cuda stream sync token provided" )
stream = cudaStreamSyncToken->getCUDAStream();
}
cublasSetKernelStream( stream );
LAMA_CHECK_CUBLAS_ERROR
float res = cublasSnrm2( n, x_d, incx );
// No error check here possible as kernel is started asynchronously
if ( !syncToken )
{
cudaStreamSynchronize( 0 );
LAMA_CHECK_CUDA_ERROR
}
static double cuda_norm(long size, const float* src1)
{
#if 1
// cublasSnrm2 produces NaN in some situations
// e.g. nlinv -g -i8 utests/data/und2x2 o
// git rev: ab28a9a953a80d243511640b23501f964a585349
// printf("cublas: %f\n", cublasSnrm2(size, src1, 1));
// printf("GPU norm (sdot: %f)\n", sqrt(cuda_sdot(size, src1, src1)));
return sqrt(cuda_sdot(size, src1, src1));
#else
return cublasSnrm2(size, src1, 1);
#endif
}
float cunorm2 (const Darray<float>& ary)
{
ary.deviceSet();
float ret;
CUBLAS_SAFE_CALL(
cublasSnrm2 (DeviceManager::handle,
ary.size(),
ary.dev_data,
1,
&ret)
);
return ret;
}
CAMLprim value spoc_cublasSnrm2 (value n, value x, value incx, value dev){
CAMLparam4(n,x,incx, dev);
CAMLlocal4(dev_vec_array, dev_vec, res, gi);
CUdeviceptr d_A;
int id;
float result;
GET_VEC(x, d_A);
CUBLAS_GET_CONTEXT;
result = cublasSnrm2(Int_val(n), (float*)d_A, Int_val(incx));
CUBLAS_CHECK_CALL(cublasGetError());
res = caml_copy_double((double)result);
CUDA_RESTORE_CONTEXT;
CAMLreturn((res));
}
float magma_snrm2(
magma_int_t n,
const float *dx, magma_int_t incx )
{
return cublasSnrm2( n, dx, incx );
}
//
// Overloaded function for dispatching to
// * CUBLAS backend, and
// * float value-type.
//
inline float nrm2( const int n, const float* x, const int incx ) {
return cublasSnrm2( n, x, incx );
}
