AMX int8/bf16 tile GEMM — enable, fix 5 bugs, optimize (0→197 GMAC/s) + vertical-SIMD primitives#217

Merged

AdaWorldAPI merged 19 commits into

claude/wonderful-hawking-lodtql

Jun 14, 2026

Commits on Jun 13, 2026

feat(simd): native AVX2 U16x16 (__m256i) — TD-T22 lowering
claude
committed
feat(simd): FastScan PQ4-ADC helpers on U16x16/U8x32
claude
committed
feat(simd): F32x8::mul_add — 8-wide FMA companion to F32x16::mul_add
claude
committed
feat(simd): F32x8::cmp_gt_mask — 8-wide ordered-GT movemask
claude
committed
feat(simd): U8x32::from_ptr — zero-overhead unchecked hot-loop load
claude
committed
fix(backend): wire gemm_bf16 / gemm_i8 / gemv to real SIMD dispatch (audit TD-T2/3/5/7)
claude
committed
bench(simd): amx_gemm_bench probe (amx_available + int8-GEMM throughput) + fmt
claude
committed
fix(amx): enable AMX (arch_prctl syscall) + 4 tile-kernel correctness bugs
claude
committed

Commits on Jun 14, 2026