search

ROCm / hipBLASLt

hipBLASLt is a library that provides general matrix-matrix operations with a flexible API and extends functionalities beyond a traditional BLAS library
https://rocm.docs.amd.com/projects/hipBLASLt/en/latest/index.html
MIT License
64 stars 89 forks source link

Fix UnrollLoopSwapGlobalReadOrder=1 bug for small GRVW. #1392

Open hcman2 opened 18 hours ago

hcman2 commented 18 hours ago

The GRVWA/B index may jump in different way. We cannot re-use the G2L VGPRs if index mismatched.

hcman2 commented 18 hours ago
image

This is an example that the G2LA/B vgpr index jump in different way.

hcman2 commented 15 hours ago

gfx94x tox passed