[MLAS][AArch64] SQNBitGemm CompInt8 - Use 4x2 tiles (#21380) · microsoft/onnxruntime@05fc0c6 · GitHub

Commit

[MLAS][AArch64] SQNBitGemm CompInt8 - Use 4x2 tiles (#21380)

Browse files

Update SQNBitGemm ARM NEON kernel to compute 4x2 tile of output.

Note: Also tried 2x4 and 4x4 tiles but observed the best microbenchmark results with 4x2 tiles.

Loading branch information

edgchen1 authored Jul 18, 2024

1 parent 92f66de commit 05fc0c6