Skip to content

ipred_filter AVX2

Henrik Gramner requested to merge gramner/dav1d:ipred_filter_avx2 into master

Skylake-X benchmark results:

intra_pred_filter_w4_8bpc_c: 400.6
intra_pred_filter_w4_8bpc_avx2: 53.8
intra_pred_filter_w8_8bpc_c: 839.9
intra_pred_filter_w8_8bpc_avx2: 126.3
intra_pred_filter_w16_8bpc_c: 1541.6
intra_pred_filter_w16_8bpc_avx2: 187.6
intra_pred_filter_w32_8bpc_c: 3742.5
intra_pred_filter_w32_8bpc_avx2: 443.4

The SIMD is unavoidably very shuffle-heavy which is the main bottleneck.

Merge request reports