Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 06:39:25 +01:00
Code Issues Packages Projects Releases Wiki Activity

Revert "ggml : load data into int8x16x4_t using vld4q_s8 on arm64 (#1738)"

Browse Source 
This reverts commit 8432d4d9f7.
This commit is contained in:
Georgi Gerganov 2023-06-08 20:48:14 +03:00
parent 8432d4d9f7
commit 0bf7cf1b29
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
k_quants.c
View File

@ -1259,8 +1259,8 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
for (int j = 0; j < QK_K/128; ++j) { for (int j = 0; j < QK_K/128; ++j) {
const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32; const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32;
const int8x16x4_t q8bytes_1 = vld4q_s8(q8); q8 += 64; const int8x16x4_t q8bytes_1 = vld1q_s8_x4(q8); q8 += 64;
const int8x16x4_t q8bytes_2 = vld4q_s8(q8); q8 += 64; const int8x16x4_t q8bytes_2 = vld1q_s8_x4(q8); q8 += 64;
q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2); q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2); q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
@ -1788,7 +1788,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
for (int j = 0; j < QK_K/64; ++j) { for (int j = 0; j < QK_K/64; ++j) {
const uint8x16x2_t q5bits = vld1q_u8_x2(q5); q5 += 32; const uint8x16x2_t q5bits = vld1q_u8_x2(q5); q5 += 32;
const int8x16x4_t q8bytes = vld4q_s8(q8); q8 += 64; const int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64;
q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
@ -2020,8 +2020,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
for (int j = 0; j < QK_K/128; ++j) { for (int j = 0; j < QK_K/128; ++j) {
uint8x16x2_t qhbits = vld1q_u8_x2(qh); qh += 32; uint8x16x2_t qhbits = vld1q_u8_x2(qh); qh += 32;
uint8x16x4_t q6bits = vld4q_u8(q6); q6 += 64; uint8x16x4_t q6bits = vld1q_u8_x4(q6); q6 += 64;
int8x16x4_t q8bytes = vld4q_s8(q8); q8 += 64; int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64;
q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
@ -2064,7 +2064,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
scale += 2; scale += 2;
#endif #endif
q8bytes = vld4q_s8(q8); q8 += 64; q8bytes = vld1q_s8_x4(q8); q8 += 64;
shifted = vshrq_n_u8(qhbits.val[0], 4); shifted = vshrq_n_u8(qhbits.val[0], 4);
q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);