linalg/x86_64: Intel AMX and AVX-VNNI int8/bf16 GEMM kernels#2339

Open

czoli1976 wants to merge 21 commits into

sonos:mainfrom

czoli1976:feature/x86-amx-avxvnni-gemm

Commits on Jun 19, 2026

linalg/build.rs: add Intel AMX int8 assembler probe

czoli1976
authored and
kali
committed
linalg/x86_64: add Intel AMX int8 GEMM kernel (avx512amx_mmm_i32_8x8)

czoli1976
authored and
kali
committed
linalg: add amx_i32 Criterion microbench (AVX2 / VNNI / AMX i8)

czoli1976
authored and
kali
committed
linalg/core: wire PackedAmxA through OptMatMulPack and ungated re-export

czoli1976
authored and
kali
committed
linalg/x86_64: prefetch hints in AMX i8 inner K loop

czoli1976
authored and
kali
committed
linalg/x86_64: add 16x16 AMX int8 GEMM kernel (4x mul-adds per tdpbssd)

czoli1976
authored and
kali
committed
linalg/bench: include avx512amx_mmm_i32_16x16 in amx_i32 microbench

czoli1976
authored and
kali
committed
linalg/x86_64: AMX 16x16 prefetch follows oneDNN cache-aware pattern

czoli1976
authored and
kali
committed
linalg/x86_64: shape-adaptive AMX dispatch (16x16 for big, 8x8 for small)

czoli1976
authored and
kali
committed
linalg/x86_64: add CPUID-leaf-4 cache-size detection for AMX dispatch

czoli1976
authored and
kali
committed
linalg/x86_64: add AMX bf16 16x16 kernel for f32 matmul (TDPBF16PS)

czoli1976
authored and
kali
committed
linalg/bench: include avx512amx_mmm_f32_16x16 in amx_f32 microbench

czoli1976
authored and
kali
committed
linalg/x86_64: add AVX-VNNI ymm int8 GEMM kernel (avxvnni_mmm_i32_8x8)

czoli1976
authored and
kali
committed
linalg/x86_64: boost AMX 16x16 kernels + add avxvnni_i32 microbench

czoli1976
authored and
kali
committed
linalg/x86_64: fix swapped operands in AMX 16x16 sub fused-op handlers

czoli1976
authored and
kali
committed
linalg/x86_64: add AVX-512 VNNI zmm 16x16 int8 GEMM kernel

czoli1976
authored and
kali
committed
linalg/x86_64: document the int8 GEMM kernel cascade

czoli1976
authored and
kali
committed
linalg/x86_64: add AMX validation + benchmark runbook for AMX-capable hosts

czoli1976
authored and
kali
committed
linalg/x86_64: add AMX int8/bf16 GEMM validation + benchmark results

czoli1976
authored and
kali
committed
linalg/x86_64: rustfmt the AMX kernels, benches, and build probes

czoli1976
authored and
kali
committed
linalg/x86_64: build on the 1.91 MSRV and the cfg-off path

czoli1976
authored and
kali
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

linalg/x86_64: Intel AMX and AVX-VNNI int8/bf16 GEMM kernels#2339

linalg/x86_64: Intel AMX and AVX-VNNI int8/bf16 GEMM kernels#2339
czoli1976 wants to merge 21 commits into
sonos:mainfrom
czoli1976:feature/x86-amx-avxvnni-gemm

Commits on Jun 19, 2026

linalg/build.rs: add Intel AMX int8 assembler probe

linalg/x86_64: add Intel AMX int8 GEMM kernel (avx512amx_mmm_i32_8x8)

linalg: add amx_i32 Criterion microbench (AVX2 / VNNI / AMX i8)

linalg/core: wire PackedAmxA through OptMatMulPack and ungated re-export

linalg/x86_64: prefetch hints in AMX i8 inner K loop

linalg/x86_64: add 16x16 AMX int8 GEMM kernel (4x mul-adds per tdpbssd)

linalg/bench: include avx512amx_mmm_i32_16x16 in amx_i32 microbench

linalg/x86_64: AMX 16x16 prefetch follows oneDNN cache-aware pattern

linalg/x86_64: shape-adaptive AMX dispatch (16x16 for big, 8x8 for small)

linalg/x86_64: add CPUID-leaf-4 cache-size detection for AMX dispatch

linalg/x86_64: add AMX bf16 16x16 kernel for f32 matmul (TDPBF16PS)

linalg/bench: include avx512amx_mmm_f32_16x16 in amx_f32 microbench

linalg/x86_64: add AVX-VNNI ymm int8 GEMM kernel (avxvnni_mmm_i32_8x8)

linalg/x86_64: boost AMX 16x16 kernels + add avxvnni_i32 microbench

linalg/x86_64: fix swapped operands in AMX 16x16 sub fused-op handlers

linalg/x86_64: add AVX-512 VNNI zmm 16x16 int8 GEMM kernel

linalg/x86_64: document the int8 GEMM kernel cascade

linalg/x86_64: add AMX validation + benchmark runbook for AMX-capable hosts

linalg/x86_64: add AMX int8/bf16 GEMM validation + benchmark results

linalg/x86_64: rustfmt the AMX kernels, benches, and build probes

linalg/x86_64: build on the 1.91 MSRV and the cfg-off path