]> git.djapps.eu Git - pkg/ggml/sources/whisper.cpp/commit
ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (llama/24699)
authorTarek Dakhran <redacted>
Fri, 26 Jun 2026 07:41:56 +0000 (09:41 +0200)
committerGeorgi Gerganov <redacted>
Fri, 26 Jun 2026 13:03:57 +0000 (16:03 +0300)
commit5e7ef7e4b04ce57a23110ab4674aafa6fb7c1a12
tree32bb8cfdb25276fd09c5549957e9506c6441103d
parent02be871218519e473fa780d1de3c5340502b3302
ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (llama/24699)

* ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32

2D convolutions with kernel size 9 produced different results on SVE
enabled ARM devices. After debugging it turned out that ggml_vec_dot_f32
was using data from inactive lanes.

Use svmla_f32_m(pg, sum1, ax1, ay1) so inactive lanes retain sum1.

* cont : clean-up

---------

Co-authored-by: Georgi Gerganov <redacted>
ggml/src/ggml-cpu/vec.cpp