Skip to content

Commit c02aac6

Browse files
compiladecompilade
committed
wip : faster 1.625 bpw AVX2 vec_dot
Not using a lookup table anymore makes it match q4_0 speed. * gguf-py : fix formatting * llama : remove spaces on empty line
1 parent 29b85e6 commit c02aac6

File tree

2 files changed

+58
-31
lines changed

2 files changed

+58
-31
lines changed

‎ggml-quants.c

Lines changed: 56 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -11371,40 +11371,68 @@ void ggml_vec_dot_q1_3_q8_0(int n, float * restrict s, size_t bs, const void * r
1137111371
__m256 accumf = _mm256_setzero_ps();
1137211372

1137311373
for (int i = 0; i < nb; ++i) {
11374-
{
11375-
__m256i x0 = _mm256_set_epi32(q1_3_grid[x[i].q[7]], q1_3_grid[x[i].q[6]],
11376-
q1_3_grid[x[i].q[5]], q1_3_grid[x[i].q[4]],
11377-
q1_3_grid[x[i].q[3]], q1_3_grid[x[i].q[2]],
11378-
q1_3_grid[x[i].q[1]], q1_3_grid[x[i].q[0]]);
11379-
__m256i y0 = _mm256_lddqu_si256((const __m256i_u *) (y[2*i].qs));
11380-
11381-
__m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(y[2*i].d));
11382-
11383-
__m256 q = mul_sum_i8_pairs_float(x0, y0);
11384-
11385-
accumf = _mm256_fmadd_ps(d, q, accumf);
11386-
}
11374+
// __m128i x12b = _mm_maskload_epi32((const int32_t *) x[i].q, _mm_set_epi32(0, -1, -1, -1));
11375+
// __m128i x12b = _mm_insert_epi8(x12a, x[i].qs[0], 12);
11376+
// WARNING: reading 3 bytes further than necessary. It's faster than the above on my CPU, though.
11377+
__m128i x12b = _mm_loadu_si128((const __m128i_u *) x[i].q);
11378+
__m256i x12 = MM256_SET_M128I(x12b, x12b);
1138711379

1138811380
{
11389-
__m256i x1 = _mm256_castsi128_si256(_mm_set_epi32(q1_3_grid[x[i].q[11]], q1_3_grid[x[i].q[10]],
11390-
q1_3_grid[x[i].q[9]], q1_3_grid[x[i].q[8]]));
11391-
__m256i x2 = _mm256_cvtepu8_epi16(_mm_maskload_epi32((const int32_t *) x[i].q, _mm_set_epi32(0, -1, -1, -1)));
11381+
__m256i x0l = _mm256_shuffle_epi8(x12, _mm256_set_epi8(5, -1, 5, -1, 5, -1, 5, -1,
11382+
4, -1, 4, -1, 4, -1, 4, -1,
11383+
1, -1, 1, -1, 1, -1, 1, -1,
11384+
0, -1, 0, -1, 0, -1, 0, -1));
11385+
__m256i x0h = _mm256_shuffle_epi8(x12, _mm256_set_epi8(7, -1, 7, -1, 7, -1, 7, -1,
11386+
6, -1, 6, -1, 6, -1, 6, -1,
11387+
3, -1, 3, -1, 3, -1, 3, -1,
11388+
2, -1, 2, -1, 2, -1, 2, -1));
11389+
__m256i x1l = _mm256_shuffle_epi8(x12, _mm256_set_epi8(7, -1, 6, -1, 5, -1, 4, -1,
11390+
3, -1, 2, -1, 1, -1, 0, -1,
11391+
9, -1, 9, -1, 9, -1, 9, -1,
11392+
8, -1, 8, -1, 8, -1, 8, -1));
11393+
__m256i x1h = _mm256_shuffle_epi8(x12, _mm256_set_epi8(12, -1, 12, -1, 12, -1, 12, -1,
11394+
11, -1, 10, -1, 9, -1, 8, -1,
11395+
11, -1, 11, -1, 11, -1, 11, -1,
11396+
10, -1, 10, -1, 10, -1, 10, -1));
11397+
const __m256i shift0 = _mm256_set_epi16(3, 9, 27, 81,
11398+
3, 9, 27, 81,
11399+
3, 9, 27, 81,
11400+
3, 9, 27, 81);
11401+
const __m256i shift1l = _mm256_set_epi16(1, 1, 1, 1,
11402+
1, 1, 1, 1,
11403+
3, 9, 27, 81,
11404+
3, 9, 27, 81);
11405+
const __m256i shift1h = _mm256_set_epi16(3, 9, 27, 81,
11406+
1, 1, 1, 1,
11407+
3, 9, 27, 81,
11408+
3, 9, 27, 81);
11409+
x0l = _mm256_mullo_epi16(x0l, shift0);
11410+
x0h = _mm256_mullo_epi16(x0h, shift0);
11411+
x1l = _mm256_mullo_epi16(x1l, shift1l);
11412+
x1h = _mm256_mullo_epi16(x1h, shift1h);
11413+
x0l = _mm256_mulhi_epu16(x0l, _mm256_set1_epi16(3));
11414+
x0h = _mm256_mulhi_epu16(x0h, _mm256_set1_epi16(3));
11415+
x1l = _mm256_mulhi_epu16(x1l, _mm256_set1_epi16(3));
11416+
x1h = _mm256_mulhi_epu16(x1h, _mm256_set1_epi16(3));
11417+
x0l = _mm256_sub_epi16(x0l, _mm256_set1_epi16(1));
11418+
x0h = _mm256_sub_epi16(x0h, _mm256_set1_epi16(1));
11419+
x1l = _mm256_sub_epi16(x1l, _mm256_set1_epi16(1));
11420+
x1h = _mm256_sub_epi16(x1h, _mm256_set1_epi16(1));
11421+
11422+
__m256i x0 = _mm256_packs_epi16(x0l, x0h);
11423+
__m256i x1 = _mm256_packs_epi16(x1l, x1h);
11424+
11425+
__m256i y0 = _mm256_lddqu_si256((const __m256i_u *) (y[2*i + 0].qs));
1139211426
__m256i y1 = _mm256_lddqu_si256((const __m256i_u *) (y[2*i + 1].qs));
1139311427

11394-
x2 = _mm256_mulhi_epu16(x2, _mm256_set1_epi16(3 << 8));
11395-
x2 = _mm256_sub_epi16(x2, _mm256_set1_epi16(1));
11396-
11397-
// TODO: reduce shuffling
11398-
x2 = _mm256_packs_epi16(x2, _mm256_setzero_si256());
11399-
x2 = _mm256_permute4x64_epi64(x2, _MM_SHUFFLE(3, 1, 2, 0));
11400-
__m128i x2_l = _mm_insert_epi32(_mm256_castsi256_si128(x2), q1_3_grid[x[i].qs[0]], 3);
11401-
x1 = _mm256_inserti128_si256(x1, x2_l, 1);
11402-
11403-
__m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(y[2*i + 1].d));
11428+
__m256 d0 = _mm256_set1_ps(GGML_FP16_TO_FP32(y[2*i].d));
11429+
__m256 d1 = _mm256_set1_ps(GGML_FP16_TO_FP32(y[2*i + 1].d));
1140411430

11405-
__m256 q = mul_sum_i8_pairs_float(x1, y1);
11431+
__m256 q0 = mul_sum_i8_pairs_float(x0, y0);
11432+
__m256 q1 = mul_sum_i8_pairs_float(x1, y1);
1140611433

11407-
accumf = _mm256_fmadd_ps(d, q, accumf);
11434+
accumf = _mm256_fmadd_ps(d0, q0, accumf);
11435+
accumf = _mm256_fmadd_ps(d1, q1, accumf);
1140811436
}
1140911437
}
1141011438

‎gguf-py/gguf/quants.py

Lines changed: 2 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -148,9 +148,9 @@ def __quantize_q1_3_rows(n: np.ndarray) -> np.ndarray:
148148
q48 = np.sum(q48 * pow3.reshape((1, 1, 4)), axis=2, keepdims=True).reshape((n_blocks, 12))
149149
q4 = np.sum(q4 * pow3.reshape((1, 4)), axis=1, keepdims=True)
150150
q48 = q48 + (q12 * 81)
151-
q = np.concatenate([q48, q4], axis=1);
151+
q = np.concatenate([q48, q4], axis=1)
152152
q = ((q.astype(np.uint16) * 256) // 243).astype(np.uint8)
153-
q = np.where(q != 0, q + 1, 0);
153+
q = np.where(q != 0, q + 1, 0)
154154

155155
return q.reshape(__quantize_q1_3_shape_change(shape))
156156

@@ -170,4 +170,3 @@ def quantize_q1_3(data: np.ndarray):
170170
return __quantize_q1_3_lazy(data)
171171
else:
172172
return __quantize_q1_3_array(data)
173-

0 commit comments

Comments
 (0)