Skip to content

Latest commit

 

History

History
12 lines (7 loc) · 726 Bytes

README.md

File metadata and controls

12 lines (7 loc) · 726 Bytes

Introduction

A demo of cusolver stream on solving eigen problem of a large number of small matrix.

Conclusion

After analyses with Nsight system, I found that either Jacobi or divide and conquer method is unable to employ multi-stream. As there are unavoidable pageable memory copies in these two function, there is no way to realize overlap calculation.

cusolver_timeline_from_nsight

Besides, it is different in cublas where overlap calculation can be watched in Nsight System clearly. cublas_timeline_from_nsight