aditya/llama.cpp
1
0
Fork 0
mirror of https://git.adityakumar.xyz/llama.cpp.git synced 2024-11-09 15:29:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Don't use vdotq_s32 if it's not available (#139)

Browse source 
* Don't use vdotq_s32 if it's not available

`dotprod` extensions aren't available on some ARM CPUs (e.g. Raspberry Pi 4), so check for them and only use them if they're available.

Reintroduces the code removed in 84d9015 if `__ARM_FEATURE_DOTPROD` isn't defined.

* Update ggml.c

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Ronsor 2023-03-14 12:34:37 -07:00 committed by GitHub
parent 60f819a2b1
commit 47857e564c
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 32 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

33
ggml.c
View file

@ -1359,8 +1359,8 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
const int8x16_t v1_1hs = vsubq_s8(v1_1h, s8b);
#if defined(__ARM_FEATURE_DOTPROD)
// dot product into int16x8_t
// assume that vdotq_s32 is always available, if not, should check for __ARM_FEATURE_DOTPROD
int32x4_t p_0 = vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls);
int32x4_t p_1 = vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls);
@ -1374,6 +1374,37 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
#else
sum0 += d0_0*d1_0*(vgetq_lane_s32(p_0, 0) + vgetq_lane_s32(p_0, 1) + vgetq_lane_s32(p_0, 2) + vgetq_lane_s32(p_0, 3));
sum1 += d0_1*d1_1*(vgetq_lane_s32(p_1, 0) + vgetq_lane_s32(p_1, 1) + vgetq_lane_s32(p_1, 2) + vgetq_lane_s32(p_1, 3));
#endif
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
const int16x8_t pl_0 = vaddq_s16(pl0l, pl0h);
const int16x8_t ph_0 = vaddq_s16(ph0l, ph0h);
const int16x8_t pl_1 = vaddq_s16(pl1l, pl1h);
const int16x8_t ph_1 = vaddq_s16(ph1l, ph1h);
const int16x8_t p_0 = vaddq_s16(pl_0, ph_0);
const int16x8_t p_1 = vaddq_s16(pl_1, ph_1);
// scalar
#if defined(__ARM_FEATURE_QRDMX)
sum0 += d0_0*d1_0*vaddvq_s16(p_0);
sum1 += d0_1*d1_1*vaddvq_s16(p_1);
#else
sum0 += d0_0*d1_0*(vgetq_lane_s16(p_0, 0) + vgetq_lane_s16(p_0, 1) + vgetq_lane_s16(p_0, 2) + vgetq_lane_s16(p_0, 3) + vgetq_lane_s16(p_0, 4) + vgetq_lane_s16(p_0, 5) + vgetq_lane_s16(p_0, 6) + vgetq_lane_s16(p_0, 7));
sum1 += d0_1*d1_1*(vgetq_lane_s16(p_1, 0) + vgetq_lane_s16(p_1, 1) + vgetq_lane_s16(p_1, 2) + vgetq_lane_s16(p_1, 3) + vgetq_lane_s16(p_1, 4) + vgetq_lane_s16(p_1, 5) + vgetq_lane_s16(p_1, 6) + vgetq_lane_s16(p_1, 7));
#endif
#endif
}