Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml : fix ARM NEON nvfp4 dot product on non-dotprod targets (#21559)

Browse Source
This commit is contained in:
Richard Davison
2026-04-14 14:23:45 +03:00
committed by GitHub
parent acc37a42ea
commit 2e05f06ffb
2 changed files with 43 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/ggml-cpu-impl.h
+10
View File
@@ -306,6 +306,7 @@ inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
#if !defined(__ARM_FEATURE_DOTPROD)
// NOTE: this fallback produces the same total sum as native vdotq_s32 but with different per-lane grouping — do not use when individual lane values matter.
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
@@ -319,6 +320,15 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
#endif // !defined(__ARM_FEATURE_DOTPROD)
static inline int32x4_t ggml_nvfp4_dot8(const int8x8_t q4_lo, const int8x8_t q8_lo,
const int8x8_t q4_hi, const int8x8_t q8_hi) {
const int16x8_t p_lo = vmull_s8(q4_lo, q8_lo);
const int16x8_t p_hi = vmull_s8(q4_hi, q8_hi);
const int32x4_t sum_lo = vpaddlq_s16(p_lo);
const int32x4_t sum_hi = vpaddlq_s16(p_hi);
return vaddq_s32(sum_lo, sum_hi);
}
#endif // defined(__ARM_NEON)
#ifdef __wasm_simd128__