[ermig1979]

Summary

Adds an AMX-INT8 implementation of SynetDynamicQuantizedInnerProduct16bGemmV0 and dispatches it ahead of AVX-512 VNNI.

• A is quantized per row and reordered into 16×64 blocks. B keeps the VNNI panel layout and is padded to a multiple of 64 along K so AMX tiles can consume the tail.
• The product uses 16×16×64 _tile_dpbusd microkernels (32×32 output tiles), matching AmxBf16::SynetDynamicQuantizedInnerProduct32fGemmV0. The last K chunk reuses that body, then dequantizes, adds bias, and applies the activation with AVX-512 (FP32 or BF16 output).
• M == 1 stays on the AVX-512 VNNI kernel. The int32 accumulator is a separate buffer whose stride is rounded up to 32, and the AMX scratch buffer is allocated whenever it is reserved.

Release notes for 8.0.167 and the VS2022 AMX project list the new sources. The 16b auto test now runs the AMX path.

Test

Release build with -DSIMD_AMXBF16=ON on AMX-INT8 hardware:

./Test -r=.. -fi=SynetDynamicQuantizedInnerProduct16bForward -tt=1 -ts=1

Result: all tests finished successfully. On 333×555×444, AMX was about 1.4×–2.1× faster than AVX-512 VNNI (common time 0.155 ms vs 0.276 ms).

  

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.