1778751 - Implement BFloat16 product instruction

Status: RESOLVED FIXED

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

Description

[Yury Delendik (:yury)]

See https://github.com/WebAssembly/relaxed-simd/issues/77

Instruction name: f32x4.relaxed_dot_bf16x8_add_f32x4, and its code is reserved for 0x114.

Since we have no concrete plans for AVX512 support and had not consider ARM64 extensions yet: the first implementation will be just optimized version of reference lowering of:

    y = f32x4.relaxed_dot_bf16x8_add_f32x4(a, b, c):
        const v128_t a_lo = wasm_i32x4_shl(a, 16)
        const v128_t b_lo = wasm_i32x4_shl(b, 16)
        const v128_t a_hi = wasm_v128_and(a, wasm_i32x4_const_splat(0xFFFF0000))
        const v128_t b_hi = wasm_v128_and(b, wasm_i32x4_const_splat(0xFFFF0000))
        y = f32x4.relaxed_fma(a_lo, b_lo, c)
        y = f32x4.relaxed_fma(a_hi, b_hi, y)

Comment 1

[Yury Delendik (:yury)]

Attachment: Bug 1778751 - Implement BFloat16 product instruction. r?jseward

Comment 2

[Yury Delendik (:yury)]

https://github.com/WebAssembly/relaxed-simd/pull/88

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/7d238f018de3
Implement BFloat16 product instruction. r=jseward

Comment 4

[Norisz Fay [:noriszfay]]

https://hg.mozilla.org/mozilla-central/rev/7d238f018de3