happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Clang-format

This commit is contained in:
Alberto Cabrera 2026-02-04 19:46:55 +00:00
parent 02afbc3715
commit 67eae511ae
2 changed files with 35 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-cpu/arch/arm/repack.cpp
View File

@ -3724,7 +3724,7 @@ void ggml_gemm_q6_K_8x4_q8_K(int n,
int32x4_t acc_s32[acc_size];
for (int i = 0; i < acc_size; i++) {
acc_s32[i] = vdupq_n_s32(0);
acc_s32[i] = vdupq_n_s32(0);
}
int16_t q6_scales[8 * 16];
@ -3751,8 +3751,8 @@ void ggml_gemm_q6_K_8x4_q8_K(int n,
// 4 rows * 16 elements per scale
// 4 reads of 16 bytes each
constexpr int reads_per_sb = 4;
int8x16_t q8_l[reads_per_sb];
int8x16_t q8_h[reads_per_sb];
int8x16_t q8_l[reads_per_sb];
int8x16_t q8_h[reads_per_sb];
for (int k = 0; k < reads_per_sb; k++) {
q8_l[k] = vld1q_s8(q8_base_l + 16 * k);
q8_h[k] = vld1q_s8(q8_base_h + 16 * k);
@ -3788,11 +3788,9 @@ void ggml_gemm_q6_K_8x4_q8_K(int n,
const uint8x16_t hbit_hi_4567 = vandq_u8(q6_qh_4567[k], mask_hi);
const int8x16_t q6_0123_lo = vsubq_s8(
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_0123[k], m4b), hbit_lo_0123, 4)),
m32s);
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_0123[k], m4b), hbit_lo_0123, 4)), m32s);
const int8x16_t q6_0123_hi = vsubq_s8(
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_0123[k], 4), hbit_hi_0123)),
m32s);
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_0123[k], 4), hbit_hi_0123)), m32s);
acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q6_0123_lo, q8_l[k], 0); // 0..3 r0 c0123
acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q6_0123_lo, q8_l[k], 1); // 0..3 r1 c0123
@ -3805,11 +3803,9 @@ void ggml_gemm_q6_K_8x4_q8_K(int n,
acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q6_0123_hi, q8_h[k], 3); // 64..67 r3 c0123
const int8x16_t q6_4567_lo = vsubq_s8(
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_4567[k], m4b), hbit_lo_4567, 4)),
m32s);
vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_4567[k], m4b), hbit_lo_4567, 4)), m32s);
const int8x16_t q6_4567_hi = vsubq_s8(
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_4567[k], 4), hbit_hi_4567)),
m32s);
vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_4567[k], 4), hbit_hi_4567)), m32s);
acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q6_4567_lo, q8_l[k], 0); // 0..3 r0 c4567
acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q6_4567_lo, q8_l[k], 1); // 0..3 r1 c4567

52
ggml/src/ggml-cpu/repack.cpp
View File

@ -258,12 +258,12 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR
template <int N, int M>
static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int n,
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
constexpr int ncols_interleaved = N;
constexpr int blocklen = M;
const int qk = QK_K;
@ -334,7 +334,8 @@ static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int n,
sumi_h += q_h * a_h;
}
sumf[j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
sumf[j] +=
(sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
}
}
}
@ -347,12 +348,12 @@ static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int n,
template <int N, int M>
static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int n,
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
constexpr int ncols_interleaved = N;
constexpr int blocklen = M;
const int qk = QK_K;
@ -409,17 +410,19 @@ static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int n,
const int l_4 = b_ptr[l].ql[ql_pos] & 0xF;
const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
const int qh_idx_l = qh_half_l + ((base_l + i) % 32);
const int qh_chunk_l = qh_idx_l / blocklen;
const int qh_pos_l = qh_idx_l % blocklen;
const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
const int qh_idx_l = qh_half_l + ((base_l + i) % 32);
const int qh_chunk_l = qh_idx_l / blocklen;
const int qh_pos_l = qh_idx_l % blocklen;
const int qh_offset_l =
qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
const int qh_idx_h = qh_half_h + ((base_h + i) % 32);
const int qh_chunk_h = qh_idx_h / blocklen;
const int qh_pos_h = qh_idx_h % blocklen;
const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
const int qh_idx_h = qh_half_h + ((base_h + i) % 32);
const int qh_chunk_h = qh_idx_h / blocklen;
const int qh_pos_h = qh_idx_h % blocklen;
const int qh_offset_h =
qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
const int q_l = ((hi_2_l << 4) | l_4) - 32;
const int q_h = ((hi_2_h << 4) | hi_4) - 32;
@ -431,7 +434,8 @@ static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int n,
sumi_h += q_h * q8_h;
}
sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
a_ptr[l].d[m];
}
}
}