]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
hexagon: add HTP kernel for GGML_OP_GATED_DELTA_NET (llama/22837)
authorYanzhao Wang <redacted>
Sat, 9 May 2026 00:12:04 +0000 (17:12 -0700)
committerGeorgi Gerganov <redacted>
Thu, 14 May 2026 18:26:48 +0000 (21:26 +0300)
commit42aea65eda1b4100bedb42ea1cc7344da9f1054a
tree628138f6bc6b185b664ead91c8c944393bf62920
parent892f786a653d19c2f474b3a9c56d9d4d7be2fb1f
hexagon: add HTP kernel for GGML_OP_GATED_DELTA_NET (llama/22837)

Implement the Gated Delta Net recurrence on HVX with:
- 4-row fused kernels for PP (prompt processing) path
- 8-row fused kernels for TG (token generation) path, reducing
  K/Q/gate vector reload overhead by 2x
- Separate PP/TG thread functions for I-cache isolation
- VTCM state scratchpad with DMA in/out for TG single-cycle access
- Vectorized gate exp via hvx_exp_f32
ggml/src/ggml-hexagon/ggml-hexagon.cpp
ggml/src/ggml-hexagon/htp/CMakeLists.txt
ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c [new file with mode: 0644]
ggml/src/ggml-hexagon/htp/htp-ctx.h
ggml/src/ggml-hexagon/htp/htp-ops.h
ggml/src/ggml-hexagon/htp/main.c