Skip to content

Fix for Int4 per-channel SME GEMM kernel failing with n > 64

Anitha Raj requested to merge fix_qsi4cx_sme into main

In kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa:

  • Fix the offset calculation
  • Fix pointer increments in the matmul

Add new shapes to unit tests, to test n > 64

Resolves: #KLEIDIAI-405, #COMPMID-7918

Signed-off-by: Anitha Raj anitha.raj@arm.com

Merge request reports