diff options
| -rw-r--r-- | ggml/src/iqk/iqk_mul_mat.cpp | 97 |
1 files changed, 53 insertions, 44 deletions
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index ca75e0fd..f95ce061 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -2481,9 +2481,61 @@ static void mul_mat_iq4_nl_r4_q8_1(int n, const void * vx, size_t bx, const Data } #endif +template <int nrc_y> +static void mul_mat_q4_0_r4_q8_1_avx2(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + GGML_ASSERT(nrc_x%4 == 0); + Q8<nrc_y, block_q8_1_x4> q8(info); + auto m4 = _mm256_set1_epi8(0xf); + auto m1 = _mm256_set1_epi16(1); + int nb = n / QK4_NL; + GGML_ASSERT(nb%4 == 0); + __m256 acc[nrc_y] = {}; + float d8[8*nrc_y]; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_iq4_nl_r4 * iq4 = (const block_iq4_nl_r4 *)((const char *)vx + ix*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int iy = 0; iy < nrc_y; ++iy) { + auto scales = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)q8.y[iy][ib4].d)); + _mm256_storeu_ps(d8 + 8*iy, scales); + } + for (int k = 0; k < 4; ++k) { + auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4[4*ib4+k].d)); + auto scales = _mm256_set_m128(scales128, scales128); + auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-4.f)); + auto bits1 = _mm256_loadu_si256((const __m256i *)iq4[4*ib4+k].qs+0); + auto bits2 = _mm256_loadu_si256((const __m256i *)iq4[4*ib4+k].qs+1); + auto q1 = _mm256_and_si256(bits1, m4); + auto q2 = _mm256_and_si256(bits2, m4); + auto q3 = _mm256_and_si256(_mm256_srli_epi16(bits1, 4), m4); + auto q4 = _mm256_and_si256(_mm256_srli_epi16(bits2, 4), m4); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k); + auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(q1, _mm256_shuffle_epi32(y, 0x00)), + _mm256_maddubs_epi16(q2, _mm256_shuffle_epi32(y, 0x55))); + auto sumi2 = _mm256_add_epi16(_mm256_maddubs_epi16(q3, _mm256_shuffle_epi32(y, 0xaa)), + _mm256_maddubs_epi16(q4, _mm256_shuffle_epi32(y, 0xff))); + auto sumi = _mm256_madd_epi16(m1, _mm256_add_epi16(sumi1, sumi2)); + auto d4d8 = _mm256_mul_ps(scales, _mm256_set1_ps(d8[8*iy+k])); + acc[iy] = _mm256_fmadd_ps(d4d8, _mm256_cvtepi32_ps(sumi), acc[iy]); + acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(d8[8*iy+4+k]), acc[iy]); + } + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1)); + info.store(ix, iy, sum); + acc[iy] = _mm256_setzero_ps(); + } + } +} + #ifdef HAVE_FANCY_SIMD template <int nrc_y> static void mul_mat_q4_0_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + if constexpr (nrc_y == 1) { + mul_mat_q4_0_r4_q8_1_avx2<1>(n, vx, bx, info, nrc_x); + return; + } GGML_ASSERT(nrc_x%8 == 0); Q8<nrc_y, block_q8_1_x4> q8(info); auto m4 = _mm512_set1_epi8(0xf); @@ -2537,50 +2589,7 @@ static void mul_mat_q4_0_r4_q8_1(int n, const void * vx, size_t bx, const DataIn #else template <int nrc_y> static void mul_mat_q4_0_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { - GGML_ASSERT(nrc_x%4 == 0); - Q8<nrc_y, block_q8_1_x4> q8(info); - auto m4 = _mm256_set1_epi8(0xf); - auto m1 = _mm256_set1_epi16(1); - int nb = n / QK4_NL; - GGML_ASSERT(nb%4 == 0); - __m256 acc[nrc_y] = {}; - float d8[8*nrc_y]; - for (int ix = 0; ix < nrc_x; ix += 4) { - const block_iq4_nl_r4 * iq4 = (const block_iq4_nl_r4 *)((const char *)vx + ix*bx); - for (int ib4 = 0; ib4 < nb/4; ++ib4) { - for (int iy = 0; iy < nrc_y; ++iy) { - auto scales = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)q8.y[iy][ib4].d)); - _mm256_storeu_ps(d8 + 8*iy, scales); - } - for (int k = 0; k < 4; ++k) { - auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq4[4*ib4+k].d)); - auto scales = _mm256_set_m128(scales128, scales128); - auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-4.f)); - auto bits1 = _mm256_loadu_si256((const __m256i *)iq4[4*ib4+k].qs+0); - auto bits2 = _mm256_loadu_si256((const __m256i *)iq4[4*ib4+k].qs+1); - auto q1 = _mm256_and_si256(bits1, m4); - auto q2 = _mm256_and_si256(bits2, m4); - auto q3 = _mm256_and_si256(_mm256_srli_epi16(bits1, 4), m4); - auto q4 = _mm256_and_si256(_mm256_srli_epi16(bits2, 4), m4); - for (int iy = 0; iy < nrc_y; ++iy) { - auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k); - auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(q1, _mm256_shuffle_epi32(y, 0x00)), - _mm256_maddubs_epi16(q2, _mm256_shuffle_epi32(y, 0x55))); - auto sumi2 = _mm256_add_epi16(_mm256_maddubs_epi16(q3, _mm256_shuffle_epi32(y, 0xaa)), - _mm256_maddubs_epi16(q4, _mm256_shuffle_epi32(y, 0xff))); - auto sumi = _mm256_madd_epi16(m1, _mm256_add_epi16(sumi1, sumi2)); - auto d4d8 = _mm256_mul_ps(scales, _mm256_set1_ps(d8[8*iy+k])); - acc[iy] = _mm256_fmadd_ps(d4d8, _mm256_cvtepi32_ps(sumi), acc[iy]); - acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(d8[8*iy+4+k]), acc[iy]); - } - } - } - for (int iy = 0; iy < nrc_y; ++iy) { - auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1)); - info.store(ix, iy, sum); - acc[iy] = _mm256_setzero_ps(); - } - } + mul_mat_q4_0_r4_q8_1_avx2<nrc_y>(n, vx, bx, info, nrc_x); } #endif |