extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-01-01 03:28:57 +01:00
Code Issues Packages Projects Releases Wiki Activity

ggml : use vaddvq_f32 for slightly more efficient reduce

Browse Source
This commit is contained in:
Georgi Gerganov 2022-12-23 13:48:19 +02:00
parent 466ceebb78
commit 7282e2109e
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 3 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
ggml.c
View File

@ -373,8 +373,7 @@ inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float
sum2 = vaddq_f32(sum2, sum3);
sum0 = vaddq_f32(sum0, sum2);
float32x2_t sumf32 = vadd_f32(vget_low_f32(sum0), vget_high_f32(sum0));
sumf = vget_lane_f32(sumf32, 0) + vget_lane_f32(sumf32, 1);
sumf = vaddvq_f32(sum0);
// leftovers
for (int i = n16; i < n; ++i) {
@ -557,9 +556,7 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t
// reduce sum0f32 and sum1f32 to sumf
sum0f32 = vaddq_f32(sum0f32, sum1f32);
float32x2_t sumf32 = vadd_f32(vget_low_f32(sum0f32), vget_high_f32(sum0f32));
sumf = vget_lane_f32(sumf32, 0) + vget_lane_f32(sumf32, 1);
sumf = vaddvq_f32(sum0f32);
#else
float32x4_t sum0 = vdupq_n_f32(0);
float32x4_t sum1 = vdupq_n_f32(0);
@ -611,9 +608,7 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t
sum4 = vaddq_f32(sum4, sum6);
sum0 = vaddq_f32(sum0, sum4);
// reduce sum0 to sumf
float32x2_t sumf32 = vadd_f32(vget_low_f32(sum0), vget_high_f32(sum0));
sumf = vget_lane_f32(sumf32, 0) + vget_lane_f32(sumf32, 1);
sumf = vaddvq_f32(sum0);
#endif
// leftovers