Balasubramanian, Vignesh 37f255821a Optimal rerouting of GEMV inputs to avoid packing ...

- Added conditional swapping of input matrices and their
  strides for GEMV, based on whether transpose is toggled
  specifically for the matrix, namely the B matrix when m=1
  and the A matrix when n=1.

- This swapping ensures that we reroute the inputs to use the
  alternative variant(code-path) in order to avoid packing cost
  for the matrix, through logical transposition.

- Currently, this optimization is enabled only when no post-ops
  are involved. With post-ops, there is a need to update the
  incoming data(from the user) in some scenarios, which will be
  dealt with later.

AMD-Internal: [CPUPL-7323]

Co-authored-by: Vignesh Balasubramanian <vignbala@amd.com>

2025-09-03 09:15:59 +05:30

..

aocl_gemm

Optimal rerouting of GEMV inputs to avoid packing

2025-09-03 09:15:59 +05:30

gemmd

Code cleanup: Copyright notices

2024-08-05 15:35:08 -04:00

CMakeLists.txt

Miscellaneous changes

2025-08-26 16:37:43 +01:00