]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
hexagon: eliminate scalar VTCM loads via HVX splat helpers (#22993)
authorTrivikram Reddy <redacted>
Wed, 13 May 2026 00:28:02 +0000 (19:28 -0500)
committerGitHub <redacted>
Wed, 13 May 2026 00:28:02 +0000 (17:28 -0700)
commit856c3adac1709be15e1ea2529a0e89f742d25fe0
treee990e82ebf4138a0557d7af8c0edc613a6c1b93b
parenta9883db8ee021cf16783016a60996d41820b5195
hexagon: eliminate scalar VTCM loads via HVX splat helpers (#22993)

* hexagon: add hvx_vec_repl helpers and use those for splat-from-vtcm usecase

* hmx-mm: optimize per-group scale handling

* hmx-fa: optimize slope load from vtcm

* hmx-fa: use aligned access where possible in hmx-utils

* hexagon: add hvx_vec_repl_2x_f16 helper and consolidate repl helpers

---------

Co-authored-by: Max Krasnyansky <redacted>
ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c
ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c
ggml/src/ggml-hexagon/htp/hmx-utils.h
ggml/src/ggml-hexagon/htp/hvx-repl.h [new file with mode: 0644]
ggml/src/ggml-hexagon/htp/hvx-utils.h
scripts/snapdragon/adb/run-completion.sh