Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-10-30 06:30:15 +01:00
Code Issues Packages Projects Releases Wiki Activity

Use vdotq_s32 to improve performance (#67)

Browse Source 
* 10% performance boost on ARM

* Back to original change
This commit is contained in:
Georgi Gerganov 2023-03-13 18:36:44 +02:00 committed by GitHub
parent 63fd76fbb0
commit 84d9015c4a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 9 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
ggml.c
View File

@ -1360,34 +1360,20 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
const int8x16_t v1_1hs = vsubq_s8(v1_1h, s8b);
// dot product into int16x8_t
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));
// 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);
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);
p_0 = vdotq_s32(p_0, v0_0hs, v1_0hs);
p_1 = vdotq_s32(p_1, v0_1hs, v1_1hs);
// scalar
#if defined(__ARM_FEATURE_QRDMX)
sum0 += d0_0*d1_0*vaddvq_s16(p_0);
sum1 += d0_1*d1_1*vaddvq_s16(p_1);
sum0 += d0_0*d1_0*vaddvq_s32(p_0);
sum1 += d0_1*d1_1*vaddvq_s32(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));
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
}