b9928

ggml-org/llama.cppb9928Jul 8, 2026by github-actions[bot]same day as b9927

What changed since b9927

low risk

b9928 adds Hexagon backend performance work: new VTCM layouts and improved MUL_MAT, MUL_MAT_ID, and FLASH_ATTN_EXT pipelines.

New features

  • Hexagon kernels gain new VTCM memory layouts and reworked pipelines for MUL_MAT, MUL_MAT_ID, and FLASH_ATTN_EXT.

Improvements

  • Kernel parameter computation refactored onto a shared, more generic layout builder.
  • DMA pipelining and prefetching improved, including earlier weight prefetch in the synchronization fallback path.
  • HMX functions hardened with explicit compiler barriers and inline assembly replacing intrinsics.
  • Inner kernel loops unrolled and simplified for better throughput.
  • VTCM allocation for matrix multiply now supports overlapping regions.

Upgrade notes

This is a Hexagon-backend-only optimization release; no CLI, API, or configuration changes are introduced relative to b9927. Standard binary swap is sufficient, and only Qualcomm Hexagon users will see behavioral differences.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Focuses on performance improvements to the Hexagon backend, specifically regarding memory management and kernel pipelines.

Key Highlights

  • New VTCM layouts and improved pipelines for MUL_MAT, MUL_MAT_ID, and FLASH_ATTN_EXT
  • Refactored kernel parameter computation using common layout builder
  • Improved DMA pipelining and prefetching with explicit compiler barriers

Release Notes (excerpt)

<details open> hexagon: new vtcm layouts and improved pipelines for MUL_MAT, MUL_MAT_ID and FLASH_ATTN_EXT (#25425) * hex-fa: refactor kernel param compute to use common layout builder * hmx: add explicit compiler barriers to make hmx funcs more robust * hex-vtcm: more generic vtcm layout builder for mm and flash-attn kernels * hex-hmx: unroll inner kernels * hex-hmx: use inline asm instead of intrinsics to avoid compiler issues * hex-hmx: define inline asm macros and simplify code…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories