1749788 - Make wasm simd lichess/stockfish fast with AVX2

Status: RESOLVED FIXED

(Core :: JavaScript: WebAssembly, enhancement, P3)

Description

[Yury Delendik (:yury)]

The wasm version source code of lichess engine is available at https://github.com/hi-ogawa/Stockfish . Unfortunately, it relies on threads to be present, means it needs to be run and benchmarked in browser context. After my preliminary runs, the enabling AVX does not provides meaningful win in performance (~1%). Here is the list of SIMD ops used in the code minus ops we already implemented for AVX, or irrelevant:

• f64x2.mul
• f64x2.replace_lane
• i16x8.add, i16x8.sub
• i16x8.extend_high_i8x16_s, i16x8.extend_low_i8x16_s
• i16x8.max_s, i16x8.min_s
• i32x4.abs, i32x4.sub
• i32x4.dot_i16x8_s
• i32x4.max_s, i32x4.max_u, i32x4.min_s
• i32x4.replace_lane
• i32x4.shl, i32x4.shr_s, i32x4.shr_u w/const
• i64x2.add, i64x2.sub
• i64x2.shl, i64x2.shr_s w/const
• i8x16.narrow_i16x8_u
• i8x16.replace_lane
• i8x16.shuffle
• v128.bitselect

Comment 1

[Yury Delendik (:yury)]

Attachment: Bug 1749788 - Implement simple binary SIMD operations for AVX encoding. r?jseward

Comment 2

[Yury Delendik (:yury)]

Attachment: Bug 1749788 - Fix replace_lane and i64x2.splat for AVX. r?jseward

• Implement {i8x16,i16x8,i32x4,f32x4,f64x2}.replace_lane
• Fix i64x2.splat for AVX2 and make it faster for non-AVX2

Depends on D139051

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/ae6d99fecf42
Implement simple binary SIMD operations for AVX encoding. r=jseward
https://hg.mozilla.org/integration/autoland/rev/99317db06b7c
Fix replace_lane and i64x2.splat for AVX. r=jseward

Comment 4

[Atila Butkovits]

https://hg.mozilla.org/mozilla-central/rev/ae6d99fecf42
https://hg.mozilla.org/mozilla-central/rev/99317db06b7c