MAGMA Forum

Shouldn’t matter, as long as they support CUDA. See Nvidia’s site.

Mark

If you want all the eigenvalues, use range = MagmaRangeAll, and you can ignore il, iu, vl, vu by passing in dummy values. If you want a subset of eigenvalues, use range = MagmaRangeV, and pass in vl, vu as the range of eigenvalues to find, e.g., to find all eigenvalues in half-open range (0.0, 1.0],...

LU is a direct method, not an iterative method (see Matlab code below). It forms matrices L, U such that PA = LU (exactly, in exact arithmetic). L has unit diagonal, U has non-zero diagonal (if it has a zero, then A is singular). There is no tolerance level epsilon in the algorithm. % LU factorizati...

If you want to write GPU kernels to compute the mean and stddev, you can model them after magmablas/dnrm2.cu, which computes the 2-norm of each column of an m-by-n matrix dA on the GPU. (You can ignore the adjust functions for updating 2-norms.) It is easiest to do in 2 kernels, one for mean, one fo...

MAGMA doesn't have functions to do this. It has the Level 1 BLAS functions to find vector norms:

• magma_dasum = sum_i abs( x_ i )
• magma_idamax = argmax_i abs( x_i )
• magma_dnrm2 = norm( x )_2 = sqrt( sum_i abs( x_i )^2 )

But not the mean and standard deviation.
-mark

We apologize for the delay in posts from the past several months. The forum software was reconfigured, causing messages to get stuck in the moderation queue, unbeknownst to us. There were quite a lot of spam messages to clean out, and we may have inadvertently skipped over a legitimate message in th...

lwork is the length of the work array that is passed to MAGMA's routine. Per the documentation, it should be: - If JOBZ = MagmaNoVec and N > 1, LWORK >= N + N*NB. - If JOBZ = MagmaVec and N > 1, LWORK >= max( N + N*NB, 2*N + N**2 ). NB can be obtained through magma_get_zhetrd_nb(N). I guess whoever ...

MAGMA doesn't have anything built in it to inject faults. You can try doing a bit flip; the result will depend on which bit is flipped. If it is in the floating point mantissa, the accuracy will suffer. If it is in the floating point exponent, probably the result will be entirely wrong, since that i...

I think this issue has been resolved in MAGMA's bitbucket repository (https://bitbucket.org/icl/magma). We should be having a bug fix release soon to address it.

-mark

[Sorry for the delay in posting. A misconfiguration left messages in the moderation queue.]

MAGMA is a hybrid CPU-GPU library. It does some computation on the CPU, and some on the GPU, trying to overlap those two computations and leverage the strengths of each. Thus, it can run into MKL errors on the CPU. As for your error, gesdd cannot work if there are NaNs. The SVD is not defined in tha...