search

DrTimothyAldenDavis / SuiteSparse

The official SuiteSparse library: a suite of sparse matrix algorithms authored or co-authored by Tim Davis, Texas A&M University.
https://people.engr.tamu.edu/davis/suitesparse.html
Other
1.15k stars 259 forks source link

Is it possible to use GPU for solving with FP32 quantization in CHOLMOD? #798

Open pcbcos opened 5 months ago

pcbcos commented 5 months ago

Is your feature request related to a problem? Please describe.

Currently, CHOLMOD's calculations on GPU can only use FP64, which is double precision. However, in recent years, the FP64 computation speed of NVIDIA consumer-grade graphics cards is 1/64 of their FP32 speed, which would result in slow solving.

Describe the solution you'd like

Supporting FP32 in the GPU module of CHOLMOD could potentially significantly improve computational speed, albeit at the cost of sacrificing some precision.

DrTimothyAldenDavis commented 5 months ago

Yes, it's possible. I only recently added FP32 support to the CPU methods in CHOLMOD. Adding them to the CUDA kernels is something I hope to do in the future.