upstream-archive/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-07-23 11:10:57 -05:00
Code Issues Packages Projects Releases Wiki Activity

ggml : use ggml_vqtbl1q_u8 for 32-bit compat (#0)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-07-10 11:06:42 +03:00
parent 289ecb05c9
commit 7695a53312
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cpu/arch/arm/quants.c
View File

@@ -263,13 +263,13 @@ void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
const uint8x16_t raw16 = vcombine_u8(raw, raw); const uint8x16_t raw16 = vcombine_u8(raw, raw);
// First 16 elements: replicate bytes 0-3, shift, mask, subtract 1 // First 16 elements: replicate bytes 0-3, shift, mask, subtract 1
uint8x16_t bytes0 = vqtbl1q_u8(raw16, idx_lo); uint8x16_t bytes0 = ggml_vqtbl1q_u8(raw16, idx_lo);
int8x16_t qv0 = vsubq_s8( int8x16_t qv0 = vsubq_s8(
vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes0, shifts), mask2)), vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes0, shifts), mask2)),
one); one);
// Second 16 elements: replicate bytes 4-7, shift, mask, subtract 1 // Second 16 elements: replicate bytes 4-7, shift, mask, subtract 1
uint8x16_t bytes1 = vqtbl1q_u8(raw16, idx_hi); uint8x16_t bytes1 = ggml_vqtbl1q_u8(raw16, idx_hi);
int8x16_t qv1 = vsubq_s8( int8x16_t qv1 = vsubq_s8(
vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes1, shifts), mask2)), vreinterpretq_s8_u8(vandq_u8(vshlq_u8(bytes1, shifts), mask2)),
one); one);