aditya/llama.cpp
1
0
Fork 0
mirror of https://git.adityakumar.xyz/llama.cpp.git synced 2024-11-13 00:39:47 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Revert "10% performance boost on ARM"

Browse source 
This reverts commit 113a9e83eb.

There are some reports for illegal instruction.
Moved this stuff to vdotq_s32 branch until resolve
This commit is contained in:
Georgi Gerganov 2023-03-13 01:28:08 +02:00
parent 54a0e66ea0
commit c80e2a8f2a
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 0 additions and 17 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

17
ggml.c
View file

@ -1360,22 +1360,6 @@ 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); const int8x16_t v1_1hs = vsubq_s8(v1_1h, s8b);
// dot product into int16x8_t // dot product into int16x8_t
#if defined(__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);
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_s32(p_0);
sum1 += d0_1*d1_1*vaddvq_s32(p_1);
#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 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 pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
@ -1404,7 +1388,6 @@ inline static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void
#else #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)); 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)); 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 #endif
} }