Hi,
MAGMA really seems great. However, I noticed that on devices with compute capability <1.3 (no double precision, afaik) the behavior is as misleading as supplied tests are.
A lot of tests, including the zgemm and dgemm tests pass without any warning, as they compare only MAGMA against CudaBLAS results. Comparing them against the CPU-Lapack results instantly reveals that the results are just wrong.
I'd also suggest to supply a warning if the double-prec routines are issued on such a device.
Thank you for your outstanding work!
Best,
Frank
double precision routines on compute capability <1.3 devices
-
Stan Tomov
- Posts: 283
- Joined: Fri Aug 21, 2009 10:39 pm
Re: double precision routines on compute capability <1.3 dev
Hi,
Thank you for pointing this out. We will put this on our to-do-list, or at least to think about it. I am not sure how CUBLAS handles it. We can handle the MAGMA BLAS similarly to CUBLAS for unsupported hardware. The MAGMA (high-level) algorithms are based on BLAS so for those algorithms we can leave it up to the BLAS implementation to decide how to handle unsupported hardware.
Stan
Thank you for pointing this out. We will put this on our to-do-list, or at least to think about it. I am not sure how CUBLAS handles it. We can handle the MAGMA BLAS similarly to CUBLAS for unsupported hardware. The MAGMA (high-level) algorithms are based on BLAS so for those algorithms we can leave it up to the BLAS implementation to decide how to handle unsupported hardware.
Stan