Pr 973 bench FMLAL 8-accumulator unroll#977

Draft

lerman25 wants to merge 25 commits into

mainfrom

pr-973-bench-fmlal-x2

Commits on Jun 1, 2026

Add design spec for SQ8↔FP16 ARM SIMD kernels [MOD-14972]

authored and
Ubuntu
committed
Add implementation plan for SQ8↔FP16 ARM SIMD kernels [MOD-14972]

authored and
Ubuntu
committed
Add NEON_HP SQ8↔FP16 IP kernel header [MOD-14972]

dor-forer
authored and
Ubuntu
committed
Add NEON_HP SQ8↔FP16 L2 kernel header [MOD-14972]

authored and
Ubuntu
committed
Wire NEON_HP SQ8↔FP16 choosers [MOD-14972]

authored and
Ubuntu
committed
Dispatch SQ8↔FP16 to NEON_HP tier on AArch64 [MOD-14972]

authored and
Ubuntu
committed
Extend SQ8↔FP16 tier-walk tests with NEON_HP [MOD-14972]

authored and
Ubuntu
committed
Add SVE SQ8↔FP16 IP kernel header [MOD-14972]

dor-forer
authored and
Ubuntu
committed
Add SVE SQ8↔FP16 L2 kernel header [MOD-14972]

dor-forer
authored and
Ubuntu
committed
Wire SVE/SVE2 SQ8↔FP16 choosers [MOD-14972]

authored and
Ubuntu
committed
Dispatch SQ8↔FP16 to SVE/SVE2 tiers on AArch64 [MOD-14972]

authored and
Ubuntu
committed
Extend SQ8↔FP16 tier-walk tests with SVE/SVE2 [MOD-14972]

authored and
Ubuntu
committed
Register ARM SQ8↔FP16 microbenchmarks [MOD-14972]

authored and
Ubuntu
committed
Add missing alignment=0 assertions to SQ8↔FP16 ARM tier-walk tests [MOD-14972]
Ubuntu
committed
Fix SVE SQ8↔FP16 kernel: use svzip1 to correct FP16→FP32 widening [MOD-14972]
Ubuntu
committed
Optimize ARM SQ8↔FP16 kernels and align with codebase conventions [MOD-14972]
Ubuntu
committed
Apply clang-format [MOD-14972]
Ubuntu
committed
Trim PR churn: remove docs, dispatcher comments, and test verbosity [MOD-14972]
Ubuntu
committed
Apply clang-format 18.1.8 (matches CI) [MOD-14972]
Ubuntu
committed
bench: register spaces_sq8_fp16 in benchmark setups
lerman25
committed
perf(arm): optimize SQ8<->FP16 NEON_HP widening and add SVE2 FMLALB/FMLALT kernel
lerman25
committed
style: clang-format SVE2.cpp
lerman25
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Pr 973 bench FMLAL 8-accumulator unroll#977

Pr 973 bench FMLAL 8-accumulator unroll#977
lerman25 wants to merge 25 commits into
mainfrom
pr-973-bench-fmlal-x2

Commits on Jun 1, 2026

Add design spec for SQ8↔FP16 ARM SIMD kernels [MOD-14972]

Add implementation plan for SQ8↔FP16 ARM SIMD kernels [MOD-14972]

Add NEON_HP SQ8↔FP16 IP kernel header [MOD-14972]

Add NEON_HP SQ8↔FP16 L2 kernel header [MOD-14972]

Wire NEON_HP SQ8↔FP16 choosers [MOD-14972]

Dispatch SQ8↔FP16 to NEON_HP tier on AArch64 [MOD-14972]

Extend SQ8↔FP16 tier-walk tests with NEON_HP [MOD-14972]

Add SVE SQ8↔FP16 IP kernel header [MOD-14972]

Add SVE SQ8↔FP16 L2 kernel header [MOD-14972]

Wire SVE/SVE2 SQ8↔FP16 choosers [MOD-14972]

Dispatch SQ8↔FP16 to SVE/SVE2 tiers on AArch64 [MOD-14972]

Extend SQ8↔FP16 tier-walk tests with SVE/SVE2 [MOD-14972]

Register ARM SQ8↔FP16 microbenchmarks [MOD-14972]

Add missing alignment=0 assertions to SQ8↔FP16 ARM tier-walk tests [MOD-14972]

Fix SVE SQ8↔FP16 kernel: use svzip1 to correct FP16→FP32 widening [MOD-14972]

Optimize ARM SQ8↔FP16 kernels and align with codebase conventions [MOD-14972]

Apply clang-format [MOD-14972]

Trim PR churn: remove docs, dispatcher comments, and test verbosity [MOD-14972]

Apply clang-format 18.1.8 (matches CI) [MOD-14972]

bench: register spaces_sq8_fp16 in benchmark setups

perf(arm): optimize SQ8<->FP16 NEON_HP widening and add SVE2 FMLALB/FMLALT kernel

style: clang-format SVE2.cpp

Commits on Jun 2, 2026

Add FMLAL (FEAT_FHM) widening-FMA NEON kernel for SQ8<->FP16

Format

Widen NEON FMLAL main loop to 8 accumulators (32 lanes/iter)