Skip to content

ggml-cpu: unpack the Q4_K scales/mins with SIMD in the AVX2 ggml_vec_dot_q4_K_q8_K #35

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
codspeed-hq wants to merge 1 commit into
base: master
Choose a base branch
from
Open

ggml-cpu: unpack the Q4_K scales/mins with SIMD in the AVX2 ggml_vec_dot_q4_K_q8_K #35

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 36 additions & 7 deletions ggml/src/ggml-cpu/arch/x86/quants.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2056,8 +2056,34 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

#if defined __AVX2__

UNUSED(kmask1);
UNUSED(kmask2);
UNUSED(kmask3);
UNUSED(utmp);

const __m256i m4 = _mm256_set1_epi8(0xF);

// Unpack the 12 packed 6-bit scales/mins of a superblock with SIMD instead
// of the scalar utmp[] shuffle. The scalar version needs ~25 GP
// instructions plus a store/reload of utmp[] to get the 16 bytes back into
// a vector register; the sequence below does it with two byte shuffles and
// a handful of masked shifts, entirely in registers.
//
// Input bytes u[0..11] (u[0..3] = utmp[0], u[4..7] = utmp[1], u[8..11] = utmp[2])
// produce the 16 bytes { scales[0..7], mins[0..7] }:
// out[ 0.. 3] = u[0..3] & 0x3f
// out[ 4.. 7] = (u[8..11] & 0x0f) | ((u[0..3] & 0xc0) >> 2)
// out[ 8..11] = u[4..7] & 0x3f
// out[12..15] = (u[8..11] >> 4) | ((u[4..7] & 0xc0) >> 2)
const __m128i sc_shuf_a = _mm_setr_epi8( 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 8, 9, 10, 11);
const __m128i sc_shuf_b = _mm_setr_epi8(-1, -1, -1, -1, 0, 1, 2, 3, -1, -1, -1, -1, 4, 5, 6, 7);
const __m128i sc_mask_a = _mm_setr_epi8(0x3f, 0x3f, 0x3f, 0x3f, 0x0f, 0x0f, 0x0f, 0x0f,
0x3f, 0x3f, 0x3f, 0x3f, 0x00, 0x00, 0x00, 0x00);
const __m128i sc_mask_b = _mm_setr_epi8(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x0f, 0x0f, 0x0f, 0x0f);
const __m128i sc_mask_c = _mm_setr_epi8(0x00, 0x00, 0x00, 0x00, 0x30, 0x30, 0x30, 0x30,
0x00, 0x00, 0x00, 0x00, 0x30, 0x30, 0x30, 0x30);

__m256 acc = _mm256_setzero_ps();
__m128 acc_m = _mm_setzero_ps();

Expand All @@ -2066,17 +2092,20 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);

memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
// reads 16 bytes starting at scales[0]; the trailing 4 bytes belong to
// qs[] of the same (144 byte) superblock, so this never reads out of bounds
const __m128i sc_raw = _mm_loadu_si128((const __m128i *)x[i].scales);
const __m128i sc_a = _mm_shuffle_epi8(sc_raw, sc_shuf_a);
const __m128i sc_b = _mm_shuffle_epi8(sc_raw, sc_shuf_b);
const __m128i sc_mins = _mm_or_si128(
_mm_or_si128(_mm_and_si128(sc_a, sc_mask_a),
_mm_and_si128(_mm_srli_epi16(sc_a, 4), sc_mask_b)),
_mm_and_si128(_mm_srli_epi16(sc_b, 2), sc_mask_c));

const uint8_t * GGML_RESTRICT q4 = x[i].qs;
const int8_t * GGML_RESTRICT q8 = y[i].qs;

const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
const __m256i mins_and_scales = _mm256_cvtepu8_epi16(sc_mins);

const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
Expand Down
Loading