gridbased search for batched gemm

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

gridbased search for batched gemm #1362

Closed aazz44ss closed 4 days ago

aazz44ss commented 1 week ago

Support both batch grid and remap batch to MN in gridbased algorithm. We'll first search [m,n,b,k] in grid, if (b != 1 && grid_b == 1), it means no batch grid in logic, then we remap batch to MN to search [m*b,n,1,k] or [m,n*b,1,k] in grid.

aazz44ss commented 1 week ago

[----------] Global test environment tear-down [==========] 50056 tests from 13 test suites ran. (998990 ms total) [ PASSED ] 50056 tests. hipBLASLt version: 1200

command line: ./hipblaslt-test