nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (#24699)

Browse Source 
* 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 <ggerganov@gmail.com>
This commit is contained in:
Tarek Dakhran
2026-06-26 09:41:56 +02:00
committed by GitHub
parent 1a87dcdc45
commit c16c35b814
2 changed files with 61 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cpu/vec.cpp
View File

@@ -75,12 +75,12 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
ay1 = GGML_F32_VEC_LOAD(y + i);
sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
}
// maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
// maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmla on available elements only
if (np2 < n) {
svbool_t pg = svwhilelt_b32(np2, n);
ax1 = svld1_f32(pg, x + np2);
ay1 = svld1_f32(pg, y + np2);
sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
sum1 = svmla_f32_m(pg, sum1, ax1, ay1);
}
// reduce sum1,sum2 to sum1
GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);