XNNPACK

XNNPACK copied to clipboard

Enable wasmusdot qd8_f32_qc8w gemm/igemm in gemm-config

3

This PR is a follow-up work for https://github.com/google/XNNPACK/pull/6771 Performance data for reference: https://github.com/google/XNNPACK/pull/6739

fanchenkong1

Copybara import of the project: -- 423f8fe98268e10270517c51a26f7dab74a0946a by kaustubh-raste : Resolved conflicts FUTURE_COPYBARA_INTEGRATE_REVIEW=https://github.com/google/XNNPACK/pull/6533 from imaginationtech:img_patch11 9352759cdd40ef60502de3bf38d1e4db0cb458e0

copybara-service[bot]

S32 pop count - microkernel and test

3

parth8mcw

parth8mcw

Use the header-driven approach to drive the packw/packb/packx/zerob tests and benchmarks; remove the yaml files, generation scripts, and "do not edit" comments as appropriate.

copybara-service[bot]

More accurate set of coefficients for `f32-vtanh`. The new kernels are 5-10% slower, but are now accurate to within 4 ulp, without exception.

copybara-service[bot]

Copybara import of the project: -- 28f3b893272b739b9283e195b4448c305bfa978f by Chenyu Yang : add qs8 qc8w gemm avxvnni prfm e2e bench FUTURE_COPYBARA_INTEGRATE_REVIEW=https://github.com/google/XNNPACK/pull/6898 from Ch3nYuY:qs8-gemm-avxvnni-prfm-e2e-bench 28f3b893272b739b9283e195b4448c305bfa978f

copybara-service[bot]

Fix dropped benchmark

copybara-service[bot]

Add QS8 QC8W GEMM AVXVNNI e2e Bench

1

Ch3nYuY

LokeshReddyOVS-MCW