Cusparse on multiply devices

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello everyone,
I’ll use cusparseScsrmv on multiply devices. I have a question about the cusparseHandle_t.
Do i need to initialize the library ones and use it for every devices or does it need to be initialized on every device?

Also I’ve done some research to find out what algorithm is used in cusparseScsrmv, and I’ve found this paper: http://www.nvidia.com/docs/IO/66889/nvr-2008-004.pdf.
Is this the algorithm used in Cusparse?

Thanks
Stole

2

Hello Stole,

You need to initialize the CUSPARSE library handle (cusparseHandle_t) separately for each device. The CUSPARSE and CUBLAS libraries are similar so you can also glance at the CUBLAS documentation, Section 2.2 “CUBLAS Context” (CUDA Toolkit 4.0 RC2).

A good reference for the sparse matrix-vector multiplication (in different formats, including CSR) is Efficient Sparse Matrix-Vector Multiplication on CUDA | Research

Perhaps this is exactly the link you have posted, but I am not able to click on it.

Hope it helps,
Maxim

3

Thanks Maxim, it helped a lot.