Search found 279 matches

by Stan Tomov
Thu Jun 25, 2020 1:24 am
Forum: User discussion
Topic: Compiling MAGMA on Summit with PGI compiler
Replies: 4
Views: 221

Re: Compiling MAGMA on Summit with PGI compiler

The LAPACK used has underscores added to the names, like sgerqf_, so when referenced by MAGMA it also has to be with underscores added, otherwise you get that sgerqf (and the other LAPACK routines used) is not defined. To fix this, you can add options -DADD_ to a few places in the make.inc file, nam...
by Stan Tomov
Fri Jun 19, 2020 3:49 pm
Forum: User discussion
Topic: Looking for the SpMV GPU kernel implementing CSR-I
Replies: 4
Views: 313

Re: Looking for the SpMV GPU kernel implementing CSR-I

The magma_d_spmv is a wrapper that is pre-tuned to call the "fastest" implementation (based on architecture, sizes, formats, etc.). You can take a look at it in file magma_d_blaswrapper.cpp. So, if the matrix is in CSR format for example, we currently call cuSparse.
by Stan Tomov
Fri Jun 19, 2020 1:46 pm
Forum: User discussion
Topic: Looking for the SpMV GPU kernel implementing CSR-I
Replies: 4
Views: 313

Re: Looking for the SpMV GPU kernel implementing CSR-I

We haven't considered adding it. MAGMA has CSR5 storage and kernel optimized for it.
by Stan Tomov
Fri May 08, 2020 1:08 pm
Forum: User discussion
Topic: spare linear solver on multiple GPUs
Replies: 1
Views: 394

Re: spare linear solver on multiple GPUs

The main focus of MAGMA is dense linear algebra, providing LAPACK for GPUs. MAGMA has also a sparse iterative solvers component but current support there is for single GPU only. There is a hipMAGMA port of MAGMA for AMD GPUs that ports CUDA to HIP.
by Stan Tomov
Wed Mar 11, 2020 4:54 pm
Forum: User discussion
Topic: Limitations on precision
Replies: 5
Views: 750

Re: Limitations on precision

You may also want to try the 2-stage reduction algorithms, e.g. ./testing_dsyevdx_2stage -JV --niter 2 -n 7000 These are much faster especially for the large sizes that you target. Maybe also using multiple GPUs would help (adding "--ngpu 4" option). Also, you can try with ESSL. There is make.inc ex...
by Stan Tomov
Wed Mar 11, 2020 1:04 pm
Forum: User discussion
Topic: Limitations on precision
Replies: 5
Views: 750

Re: Limitations on precision

Some of these errors seem to be large and inconsistent. This is what I get on one of our systems with V100 and Intel CPU. [tomov@a04 testing]$ ./testing_dsyevd -JV --niter 5 -c -l -n 7000 % MAGMA 2.5.2 svn compiled for CUDA capability >= 7.0, 32-bit magma_int_t, 64-bit pointer. % CUDA runtime 9020, ...
by Stan Tomov
Tue Feb 25, 2020 12:51 pm
Forum: User discussion
Topic: External GPU on thunderbolt - and CUDA/MAGMA?
Replies: 2
Views: 807

Re: External GPU on thunderbolt - and CUDA/MAGMA?

I am not sure about the exact mechanism of connecting them and what GPUs can be used as eGPUs, but if one makes the setup the issues of using the GPUs will be similar to systems where you have powerful GPUs connected to a "slow" host CPU (due to slow CPU or slow CPU-GPU data transfers). MAGMA can be...
by Stan Tomov
Tue Feb 25, 2020 12:13 pm
Forum: User discussion
Topic: Very large sparse eigensolves
Replies: 2
Views: 603

Re: Very large sparse eigensolves

Looks like you need about 15 GB just for the matrix. You could use MAGMA and 32GB GPU to solve such problems. MAGMA implements the LOBPCG method, but that has to be adjusted for your problem - the matrix has to be definite and you would look for the smallest (or largest) eigenstates. One way is fold...
by Stan Tomov
Tue Jan 07, 2020 11:36 am
Forum: User discussion
Topic: Best library for O(100k) linear system
Replies: 2
Views: 1028

Re: Best library for O(100k) linear system

Dear Radek,
You can try the SLATE library:
https://bitbucket.org/icl/slate/
which provides the ScaLAPACK functionalities with support for GPU use.
Stan
by Stan Tomov
Wed Dec 18, 2019 12:56 am
Forum: User discussion
Topic: low performance running mixed precision lu factorization
Replies: 11
Views: 1474

Re: low performance running mixed precision lu factorization

The slow CPU will affect performance since MAGMA still uses CPUs for part of the computation. We can tune for this case or use other codes that are GPU only, but these are not connected yet to the mixed-precision solvers.