ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 23:24:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

Make iq4_nl_r4 work with row size that are not a multiple of 128

Browse Source 
... on Zen4
This commit is contained in:
Iwan Kawrakow
2025-01-30 12:57:42 +02:00
parent cc39e3a90c
commit 4ecba36ebf
1 changed files with 44 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -2474,44 +2474,63 @@ static void mul_mat_iq4_nl_r4_q8_1(int n, const void * vx, size_t bx, const Data
auto m4 = _mm512_set1_epi8(0xf);
auto values = load_iq4nl_values_512();
int nb = n / QK4_NL;
GGML_ASSERT(nb%4 == 0);
__m512 acc[2*nrc_y] = {};
__m512i qx[4];
float d8[8*nrc_y];
auto prepare = [&qx, &m4, &values] (const block_iq4_nl_r4& iq4l, const block_iq4_nl_r4& iq4h) {
auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4l.d));
auto scales1 = _mm256_set_m128(scales128, scales128);
scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4h.d));
auto scales2 = _mm256_set_m128(scales128, scales128);
auto scales = _mm512_insertf32x8(_mm512_castps256_ps512(scales1), scales2, 1);
auto bits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l.qs+0)),
_mm256_loadu_si256((const __m256i *)iq4h.qs+0), 1);
auto bits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l.qs+1)),
_mm256_loadu_si256((const __m256i *)iq4h.qs+1), 1);
qx[0] = _mm512_shuffle_epi8(values, _mm512_and_si512(bits1, m4));
qx[1] = _mm512_shuffle_epi8(values, _mm512_and_si512(bits2, m4));
qx[2] = _mm512_shuffle_epi8(values, _mm512_and_si512(_mm512_srli_epi16(bits1, 4), m4));
qx[3] = _mm512_shuffle_epi8(values, _mm512_and_si512(_mm512_srli_epi16(bits2, 4), m4));
return scales;
};
auto dot = [&qx] (__m256i y8) {
auto y = _mm512_inserti32x8(_mm512_castsi256_si512(y8), y8, 1);
auto sumi = _mm512_setzero_si512();
sumi = _mm512_dpbusd_epi32(sumi, qx[0], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x00)));
sumi = _mm512_dpbusd_epi32(sumi, qx[1], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x55)));
sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa)));
sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff)));
return sumi;
};
for (int ix = 0; ix < nrc_x; ix += 8) {
const block_iq4_nl_r4 * iq4l = (const block_iq4_nl_r4 *)((const char *)vx + (ix+0)*bx);
const block_iq4_nl_r4 * iq4h = (const block_iq4_nl_r4 *)((const char *)vx + (ix+4)*bx);
for (int ib4 = 0; ib4 < nb/4; ++ib4) {
for (int iy = 0; iy < nrc_y; ++iy) {
_mm256_storeu_ps(d8+8*iy, _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)q8.y[iy][ib4].d)));
}
for (int k = 0; k < 4; ++k) {
auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4l[4*ib4+k].d));
auto scales1 = _mm256_set_m128(scales128, scales128);
scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4h[4*ib4+k].d));
auto scales2 = _mm256_set_m128(scales128, scales128);
auto scales = _mm512_insertf32x8(_mm512_castps256_ps512(scales1), scales2, 1);
auto scales_m = _mm512_mul_ps(scales, _mm512_set1_ps(-64.f));
auto bits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[4*ib4+k].qs+0)),
_mm256_loadu_si256((const __m256i *)iq4h[4*ib4+k].qs+0), 1);
auto bits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[4*ib4+k].qs+1)),
_mm256_loadu_si256((const __m256i *)iq4h[4*ib4+k].qs+1), 1);
qx[0] = _mm512_shuffle_epi8(values, _mm512_and_si512(bits1, m4));
qx[1] = _mm512_shuffle_epi8(values, _mm512_and_si512(bits2, m4));
qx[2] = _mm512_shuffle_epi8(values, _mm512_and_si512(_mm512_srli_epi16(bits1, 4), m4));
qx[3] = _mm512_shuffle_epi8(values, _mm512_and_si512(_mm512_srli_epi16(bits2, 4), m4));
auto scales = prepare(iq4l[4*ib4+k], iq4h[4*ib4+k]);
for (int iy = 0; iy < nrc_y; ++iy) {
auto y8 = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k);
auto y = _mm512_inserti32x8(_mm512_castsi256_si512(y8), y8, 1);
auto sumi = _mm512_setzero_si512();
sumi = _mm512_dpbusd_epi32(sumi, qx[0], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x00)));
sumi = _mm512_dpbusd_epi32(sumi, qx[1], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x55)));
sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa)));
sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff)));
auto dy = _mm512_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k]));
auto sumi = dot(_mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k));
auto dy = _mm512_set1_ps(d8[8*iy+k]);
acc[2*iy+0] = _mm512_fmadd_ps(_mm512_mul_ps(scales, dy), _mm512_cvtepi32_ps(sumi), acc[2*iy+0]);
acc[2*iy+1] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k+4])), acc[2*iy+1]);
acc[2*iy+1] = _mm512_fmadd_ps(scales, _mm512_set1_ps(d8[8*iy+k+4]), acc[2*iy+1]);
}
}
}
for (int ib = 4*(nb/4); ib < nb; ++ib) {
auto scales = prepare(iq4l[ib], iq4h[ib]);
for (int iy = 0; iy < nrc_y; ++iy) {
auto qy = (const block_q8_1 *)q8.y[iy];
auto sumi = dot(_mm256_loadu_si256((const __m256i*)qy[ib].qs));
auto dy = _mm512_set1_ps(GGML_FP16_TO_FP32(qy[ib].d));
acc[2*iy+0] = _mm512_fmadd_ps(_mm512_mul_ps(scales, dy), _mm512_cvtepi32_ps(sumi), acc[2*iy+0]);
acc[2*iy+1] = _mm512_fmadd_ps(scales, _mm512_set1_ps(GGML_FP16_TO_FP32(qy[ib].s)), acc[2*iy+1]);
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
auto sum512 = _mm512_add_ps(acc[2*iy+0], acc[2*iy+1]);
auto sum512 = _mm512_fmadd_ps(_mm512_set1_ps(-64.f), acc[2*iy+1], acc[2*iy+0]);
acc[2*iy+0] = acc[2*iy+1] = _mm512_setzero_ps();
auto sum1 = _mm_add_ps(_mm512_extractf32x4_ps(sum512, 0), _mm512_extractf32x4_ps(sum512, 1));
auto sum2 = _mm_add_ps(_mm512_extractf32x4_ps(sum512, 2), _mm512_extractf32x4_ps(sum512, 3));