tree b4d98f56b0baf3921cd47ba447022c486c94adb0
parent 452bb4c07a4f98630396e74cd886ef8d381f9466
author Jerome Jiang <jianj@google.com> 1779900756 -0400
committer Jerome Jiang <jianj@google.com> 1780423629 -0700

Conv horiz: expand hwy avx512

Further optimize for small blocks.
But fall back to avx2 for blocks with height 32.

- Use 8-bit pairwise multiply-accumulate (SatWidenMulPairwiseAdd)
  instead of 16-bit math for w <= 32 with even coefficients.
- Halve filter coefficients to fit in int8_t and avoid overflow,
  adjusting final scaling shift to FILTER_BITS - 1.
- Eliminate expensive 8-bit to 16-bit pixel promotion.
- Add specialized unrolled loops for w = 4, 8, 16, and 32.

All blocks now show significant speed up except for small
slow downs for block 16x32, 32x32 and 64x32

I'll further investigate these block sizes.

Size       | avx2      | avx512 (diff)
------------------------------------------
4x4        |    5.62µs |    4.03µs (-28.3%)
4x8        |    6.78µs |    5.17µs (-23.7%)
8x4        |    5.94µs |    4.03µs (-32.2%)
8x8        |    6.75µs |    5.17µs (-23.4%)
8x16       |   10.01µs |    7.66µs (-23.4%)
16x8       |    7.28µs |    6.49µs (-10.8%)
16x16      |   10.92µs |    10.47µs (-4.1%)
16x32      |   17.94µs |   19.83µs (+10.5%)
32x16      |   19.34µs |    19.59µs (+1.3%)
32x32      |   33.67µs |   38.31µs (+13.8%)
32x64      |  170.90µs |  153.10µs (-10.4%)
64x32      |   68.21µs |   76.28µs (+11.8%)
64x64      |  307.20µs |  151.80µs (-50.6%)
64x128     |  677.800s |  305.30µs (-55.0%)
128x64     |  527.90µs |  298.60µs (-43.4%)
128x128    |    1.35ms |  593.90µs (-56.1%)

Change-Id: I4134a9ca0e233855761f6b03c5f35e8fcf8e25fa