Search found 5 matches

by dsambit
Tue Oct 16, 2018 10:51 pm
Forum: User discussion
Topic: Create distributed matrix on gpus with no cpu to gpu copy
Replies: 9
Views: 1829

Re: Create distributed matrix on gpus with no cpu to gpu cop

Thank you for your suggestions. I have enough information now to get started with something similar to Stan's suggestion: magma_dgemm( MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, dA, ldda, dA, ldda, c_zero, dwork, n, queue ); magma_dpotrf_gpu(MagmaUpper, n, dwork, n, &info); magma_dtrsm( MagmaRigh...
by dsambit
Tue Oct 16, 2018 3:36 pm
Forum: User discussion
Topic: Create distributed matrix on gpus with no cpu to gpu copy
Replies: 9
Views: 1829

Re: Create distributed matrix on gpus with no cpu to gpu cop

Hi Mark, I don't see a reason to create an explicit inverse here. Inverses are usually slower and less accurate than doing a solve. Just do a triangular solve (trsm) with L on the right. I.e., solve We do not use triangular solve currently as our X matrix is a parallel distributed memory (M_global x...
by dsambit
Tue Oct 16, 2018 1:50 pm
Forum: User discussion
Topic: Create distributed matrix on gpus with no cpu to gpu copy
Replies: 9
Views: 1829

Re: Create distributed matrix on gpus with no cpu to gpu cop

Hi Mark, See the code in magma/testing/testing_dpotrf_mgpu.cpp for an example of setting up the matrices. The CPU calls magma_dpotrf_mgpu, not the GPU. There is no need for unified memory. Thanks for the clarification and directing me to the appropriate example code. We're also working on a multi-GP...
by dsambit
Mon Oct 15, 2018 12:41 am
Forum: User discussion
Topic: Create distributed matrix on gpus with no cpu to gpu copy
Replies: 9
Views: 1829

Re: Create distributed matrix on gpus with no cpu to gpu cop

Hi Stan, Thanks for your reply. I think I can work with the nb provided by magma_get_dpotrf_nb(n). how do you have the matrix distributed and how do you describe the distribution? This is what I am really confused about. Let's say I have the nb blocks of memory across n gpus, but it seems somehow I ...
by dsambit
Sun Oct 14, 2018 6:45 pm
Forum: User discussion
Topic: Create distributed matrix on gpus with no cpu to gpu copy
Replies: 9
Views: 1829

Create distributed matrix on gpus with no cpu to gpu copy

Hi, I want to use _mgpu functions in magma as we need to work with matrix sizes of N=40,000 and more whose memory can't fit in a single gpu. Lets say I am working with k gpus, and I am able to independently compute N/k different column blocks of the matrix directly on k gpus. Next, lets say I want t...