tree 1240af2e798cc9c5909f4cbf5213079882cd942d
parent 6834b537f81e00bc36cd9593e06bdf1ec634f8b0
author Luc Trudeau <luc@trud.ca> 1520393747 -0500
committer Luc Trudeau <luc@trud.ca> 1520645609 +0000

[CFL] Faster AVX2 Average Subtract

Based on the observation that for small blocks AVX2 does not outperform
SSE2, we call the SSE2 code for block widths 4 and 8.

For widths 16 and 32, the AVX2 version is optimized by:
  * Summing over two rows in the summing loop;
  * Operating over the full 256bit registers in the summing loop;
  * Using more accumulators to reduce coupling operations;
  * Leveraging chained hadd calls in the fill function.

AVX2/CFLSubAvgTest
4x4: C time = 384 us, SIMD time = 153 us (~2.5x)
8x8: C time = 805 us, SIMD time = 229 us (~3.5x)
16x16: C time = 2757 us, SIMD time = 775 us (~3.6x)
32x32: C time = 10035 us, SIMD time = 2524 us (~4x)

Change-Id: I683994026c1f1626828e90949cd0bd911b46ed5e