diff options
Diffstat (limited to 'ggml')
| -rw-r--r-- | ggml/include/ggml.h | 4 | ||||
| -rw-r--r-- | ggml/src/ggml-common.h | 14 | ||||
| -rw-r--r-- | ggml/src/ggml-quants.c | 1 | ||||
| -rw-r--r-- | ggml/src/ggml.c | 26 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_mul_mat.cpp | 189 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.cpp | 142 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.h | 6 |
7 files changed, 366 insertions, 16 deletions
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 2358fb76..99c39b9c 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -408,8 +408,10 @@ extern "C" { GGML_TYPE_IQ4_KSS = 146, GGML_TYPE_Q4_0_R4 = 202, + GGML_TYPE_Q5_0_R4 = 206, GGML_TYPE_Q8_0_R4 = 208, GGML_TYPE_IQ4_NL_X4 = 220, + GGML_TYPE_Q6_0_R4 = 233, GGML_TYPE_COUNT, }; @@ -471,7 +473,9 @@ extern "C" { // GGML_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors GGML_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors + GGML_FTYPE_MOSTLY_Q5_0_R4 = 208, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_NL_X4 = 219, // except 1d tensors + GGML_FTYPE_MOSTLY_Q6_0_R4 = 227, // except 1d tensors }; // available tensor operations: diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 2af3323d..fb87a602 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -182,6 +182,13 @@ typedef struct { } block_q5_0; static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); +typedef struct { + ggml_half d[4]; // delta + uint8_t qh[QK5_0/2]; // 5-th bit of quants + uint8_t qs[QK5_0*2]; // nibbles / quants +} block_q5_0_r4; +static_assert(sizeof(block_q5_0_r4) == 4*sizeof(ggml_half) + QK5_0*2 + QK5_0/2, "wrong q5_0_r4 block size/padding"); + #define QK5_1 32 typedef struct { GGML_SCALE_TYPE1(m, dm); @@ -198,6 +205,13 @@ typedef struct { } block_q6_0; static_assert(sizeof(block_q6_0) == sizeof(ggml_half) + QK6_0/2 + QK6_0/4, "wrong q6_0 block size/padding"); +typedef struct { + ggml_half d[4]; // delta + uint8_t qh[QK6_0]; // 5+6-th bit of quants + uint8_t qs[QK6_0*2]; // nibbles / quants +} block_q6_0_r4; +static_assert(sizeof(block_q6_0_r4) == 4*sizeof(ggml_half) + QK6_0*2 + QK6_0, "wrong q6_0_r4 block size/padding"); + #define QK8_0 32 typedef struct { ggml_half d; // delta diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 3140fc19..1953fb7e 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -15198,6 +15198,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte case GGML_TYPE_IQ4_KSS: break; case GGML_TYPE_IQ4_NL_X4: break; case GGML_TYPE_Q4_0_R4: break; + case GGML_TYPE_Q5_0_R4: break; case GGML_TYPE_Q8_0_R4: break; case GGML_TYPE_Q4_0_4_4: case GGML_TYPE_Q4_0_4_8: diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index fd65ae67..0eb76a07 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1296,6 +1296,23 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_Q5_0_R4] = { + .type_name = "q5_0_r4", + .blck_size = QK5_0, + .type_size = sizeof(block_q5_0), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_q5_0_r4, + .from_float = quantize_row_q5_0_r4, + .from_float_ref = (ggml_from_float_t)quantize_row_q5_0_r4_ref, + .vec_dot = vec_dot_q5_0_r4_q8_0, +#if GGML_USE_IQK_MULMAT && defined __AVX2__ + .vec_dot_type = GGML_TYPE_Q8_1, +#else + .vec_dot_type = GGML_TYPE_Q8_0, +#endif + .nrows = 1, + .row_meta_size = 0, + }, }; // For internal test use @@ -3956,6 +3973,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ4_NL: wtype = GGML_TYPE_IQ4_NL; break; case GGML_FTYPE_MOSTLY_IQ4_NL_X4: wtype = GGML_TYPE_IQ4_NL_X4;break; case GGML_FTYPE_MOSTLY_Q4_0_R4: wtype = GGML_TYPE_Q4_0_R4; break; + case GGML_FTYPE_MOSTLY_Q5_0_R4: wtype = GGML_TYPE_Q5_0_R4; break; case GGML_FTYPE_MOSTLY_Q8_0_R4: wtype = GGML_TYPE_Q8_0_R4; break; case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break; case GGML_FTYPE_MOSTLY_IQ4_KS: wtype = GGML_TYPE_IQ4_KS; break; @@ -10482,6 +10500,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -10927,6 +10946,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -11069,6 +11089,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -14257,6 +14278,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -14639,6 +14661,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -14915,6 +14938,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -15518,6 +15542,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q5_0_R4: case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: @@ -22347,6 +22372,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_NL_X4: result = quantize_iq4_nl_x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_0_R4: result = quantize_q4_0_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q5_0_R4: result = quantize_q5_0_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q8_0_R4: result = quantize_q8_0_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_KS: result = quantize_iq4_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index bbf7e379..4cdc1a08 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -2283,6 +2283,123 @@ static void mul_mat_q4_0_r4_q8_1(int n, const void * vx, size_t bx, const DataIn } #endif +template <int nrc_y> +static void mul_mat_q5_0_r4_q8_1_avx2(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + GGML_ASSERT(nrc_x%8 == 0); + Q8<nrc_y, block_q8_1_x4> q8(info); + auto m4 = _mm256_set1_epi8(0xf); + auto m5 = _mm256_set1_epi8(0x10); + auto m1 = _mm256_set1_epi16(1); + int nb = n / QK5_0; + GGML_ASSERT(nb%4 == 0); + __m256 acc[nrc_y] = {}; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_q5_0_r4 * iq5 = (const block_q5_0_r4 *)((const char *)vx + ix*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq5[4*ib4+k].d)); + auto scales = _mm256_set_m128(scales128, scales128); + auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-8.f)); + auto bits1 = _mm256_loadu_si256((const __m256i *)iq5[4*ib4+k].qs+0); + auto bits2 = _mm256_loadu_si256((const __m256i *)iq5[4*ib4+k].qs+1); + auto hbits = _mm_loadu_si128((const __m128i *)iq5[4*ib4+k].qh); + auto hb = MM256_SET_M128I(_mm_srli_epi16(hbits, 1), hbits); + auto q1 = _mm256_and_si256(bits1, m4) | _mm256_and_si256(_mm256_slli_epi16(hb, 4), m5); + auto q2 = _mm256_and_si256(bits2, m4) | _mm256_and_si256(_mm256_slli_epi16(hb, 2), m5); + auto q3 = _mm256_and_si256(_mm256_srli_epi16(bits1, 4), m4) | _mm256_and_si256(hb, m5); + auto q4 = _mm256_and_si256(_mm256_srli_epi16(bits2, 4), m4) | _mm256_and_si256(_mm256_srli_epi16(hb, 2), m5);; + 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(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k]))); + acc[iy] = _mm256_fmadd_ps(d4d8, _mm256_cvtepi32_ps(sumi), acc[iy]); + acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k+4])), 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_q5_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_q5_0_r4_q8_1_avx2<1>(n, vx, bx, info, nrc_x); + } else { + GGML_ASSERT(nrc_x%8 == 0); + Q8<nrc_y, block_q8_1_x4> q8(info); + auto m4 = _mm512_set1_epi8(0xf); + auto m5 = _mm512_set1_epi8(0x10); + int nb = n / QK5_0; + GGML_ASSERT(nb%4 == 0); + __m512 acc[2*nrc_y] = {}; + __m512i qx[4]; + for (int ix = 0; ix < nrc_x; ix += 8) { + const block_q5_0_r4 * iq5l = (const block_q5_0_r4 *)((const char *)vx + (ix+0)*bx); + const block_q5_0_r4 * iq5h = (const block_q5_0_r4 *)((const char *)vx + (ix+4)*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq5l[4*ib4+k].d)); + auto scales1 = _mm256_set_m128(scales128, scales128); + scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq5h[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(-8.f)); + auto bits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq5l[4*ib4+k].qs+0)), + _mm256_loadu_si256((const __m256i *)iq5h[4*ib4+k].qs+0), 1); + auto bits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq5l[4*ib4+k].qs+1)), + _mm256_loadu_si256((const __m256i *)iq5h[4*ib4+k].qs+1), 1); + auto hbits1 = _mm_loadu_si128((const __m128i *)iq5l[4*ib4+k].qh); + auto hbits2 = _mm_loadu_si128((const __m128i *)iq5h[4*ib4+k].qh); + auto hb1 = MM256_SET_M128I(_mm_srli_epi16(hbits1, 1), hbits1); + auto hb2 = MM256_SET_M128I(_mm_srli_epi16(hbits2, 1), hbits2); + auto hb = _mm512_inserti32x8(_mm512_castsi256_si512(hb1), hb2, 1); + qx[0] = _mm512_and_si512(bits1, m4) | _mm512_and_si512(_mm512_slli_epi16(hb, 4), m5); + qx[1] = _mm512_and_si512(bits2, m4) | _mm512_and_si512(_mm512_slli_epi16(hb, 2), m5);; + //qx[2] = _mm512_and_si512(_mm512_srli_epi16(bits1, 4), m4) | _mm512_and_si512(_mm512_slli_epi16(hb, 2), m5); + qx[2] = _mm512_and_si512(_mm512_srli_epi16(bits1, 4), m4) | _mm512_and_si512(hb, m5); + qx[3] = _mm512_and_si512(_mm512_srli_epi16(bits2, 4), m4) | _mm512_and_si512(_mm512_srli_epi16(hb, 2), m5); + 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])); + 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]); + } + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + auto sum512 = _mm512_add_ps(acc[2*iy+0], acc[2*iy+1]); + 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)); + info.store(ix+0, iy, sum1); + info.store(ix+4, iy, sum2); + } + } + } +} +#else +template <int nrc_y> +static void mul_mat_q5_0_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + mul_mat_q5_0_r4_q8_1_avx2<nrc_y>(n, vx, bx, info, nrc_x); +} +#endif + #ifdef HAVE_FANCY_SIMD template <int nrc_y> static void mul_mat_q8_0_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { @@ -4398,6 +4515,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { mm.funcs[7] = mul_mat_q4_0_r4_q8_1<8>; expected_typeB = GGML_TYPE_Q8_1; break; + case GGML_TYPE_Q5_0_R4: + assert (ne00 % QK4_NL == 0); + mm.funcs[0] = mul_mat_q5_0_r4_q8_1<1>; + mm.funcs[1] = mul_mat_q5_0_r4_q8_1<2>; + mm.funcs[2] = mul_mat_q5_0_r4_q8_1<3>; + mm.funcs[3] = mul_mat_q5_0_r4_q8_1<4>; + mm.funcs[4] = mul_mat_q5_0_r4_q8_1<5>; + mm.funcs[5] = mul_mat_q5_0_r4_q8_1<6>; + mm.funcs[6] = mul_mat_q5_0_r4_q8_1<7>; + mm.funcs[7] = mul_mat_q5_0_r4_q8_1<8>; + expected_typeB = GGML_TYPE_Q8_1; + break; case GGML_TYPE_Q8_0_R4: assert (ne00 % QK4_NL == 0); mm.funcs[0] = mul_mat_q8_0_r4_q8_1<1>; @@ -6952,6 +7081,55 @@ void mul_mat_q4_0_r4_q8_0(int n, const void * vx, size_t bx, const DataInfo& inf } template <int nrc_y> +void mul_mat_q5_0_r4_q8_0(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_0_x4> q8(info); + auto m4 = vdupq_n_u8(0x0f); + auto m5 = vdupq_n_u8(0x10); + auto m16 = vdupq_n_s8(-16); + int nb = n / QK5_0; + GGML_ASSERT(nb%4 == 0); + int8x16_t qx[8]; + float32x4_t acc[nrc_y] = {}; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_q5_0_r4 * iq5 = (const block_q5_0_r4 *)((const char *)vx + ix*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales = vcvt_f32_f16(vld1_f16((const float16_t *)iq5[4*ib4+k].d)); + auto lbits = vld1q_u8_x4(iq5[4*ib4+k].qs); + auto hbits = vld1q_u8(iq5[4*ib4+k].qh); + qx[0] = vaddq_s8(vandq_u8(lbits.val[0], m4) | vandq_u8(vshlq_n_u8(hbits, 4), m5), m16); // 0...3 + qx[1] = vaddq_s8(vandq_u8(lbits.val[1], m4) | vandq_u8(vshlq_n_u8(hbits, 3), m5), m16); // 16..19 + qx[2] = vaddq_s8(vandq_u8(lbits.val[2], m4) | vandq_u8(vshlq_n_u8(hbits, 2), m5), m16); // 4...7 + qx[3] = vaddq_s8(vandq_u8(lbits.val[3], m4) | vandq_u8(vshlq_n_u8(hbits, 1), m5), m16); // 20..23 + qx[4] = vaddq_s8(vshrq_n_u8(lbits.val[0], 4)| vandq_u8(hbits, m5), m16); // 8..11 + qx[5] = vaddq_s8(vshrq_n_u8(lbits.val[1], 4)| vandq_u8(vshrq_n_u8(hbits, 1), m5), m16); // 24..27 + qx[6] = vaddq_s8(vshrq_n_u8(lbits.val[2], 4)| vandq_u8(vshrq_n_u8(hbits, 2), m5), m16); // 12..15 + qx[7] = vaddq_s8(vshrq_n_u8(lbits.val[3], 4)| vandq_u8(vshrq_n_u8(hbits, 3), m5), m16); // 28..31 + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = vld1q_s8_x2(q8.y[iy][ib4].qs+32*k); + auto sumi = vdupq_n_s32(0); + sumi = vdotq_laneq_s32(sumi, qx[0], y.val[0], 0); + sumi = vdotq_laneq_s32(sumi, qx[1], y.val[1], 0); + sumi = vdotq_laneq_s32(sumi, qx[2], y.val[0], 1); + sumi = vdotq_laneq_s32(sumi, qx[3], y.val[1], 1); + sumi = vdotq_laneq_s32(sumi, qx[4], y.val[0], 2); + sumi = vdotq_laneq_s32(sumi, qx[5], y.val[1], 2); + sumi = vdotq_laneq_s32(sumi, qx[6], y.val[0], 3); + sumi = vdotq_laneq_s32(sumi, qx[7], y.val[1], 3); + auto d4d8 = vmulq_f32(scales, vdupq_n_f32(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k]))); + acc[iy] = vfmaq_f32(acc[iy], d4d8, vcvtq_f32_s32(sumi)); + } + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + info.store(ix, iy, acc[iy]); + acc[iy] = vdupq_n_f32(0.f); + } + } +} + +template <int nrc_y> void mul_mat_q8_0_r4_q8_0(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_0_x4> q8(info); @@ -7179,6 +7357,17 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) { m.funcs[7] = mul_mat_q4_0_r4_q8_0<8>; expected_Btype = GGML_TYPE_Q8_0; break; + case GGML_TYPE_Q5_0_R4: + m.funcs[0] = mul_mat_q5_0_r4_q8_0<1>; + m.funcs[1] = mul_mat_q5_0_r4_q8_0<2>; + m.funcs[2] = mul_mat_q5_0_r4_q8_0<3>; + m.funcs[3] = mul_mat_q5_0_r4_q8_0<4>; + m.funcs[4] = mul_mat_q5_0_r4_q8_0<5>; + m.funcs[5] = mul_mat_q5_0_r4_q8_0<6>; + m.funcs[6] = mul_mat_q5_0_r4_q8_0<7>; + m.funcs[7] = mul_mat_q5_0_r4_q8_0<8>; + expected_Btype = GGML_TYPE_Q8_0; + break; case GGML_TYPE_Q8_0_R4: m.funcs[0] = mul_mat_q8_0_r4_q8_0<1>; m.funcs[1] = mul_mat_q8_0_r4_q8_0<2>; diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index 811a9fe9..eafb2887 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -3224,12 +3224,24 @@ static void repack_q4_0(int nrows, int n_per_row, const block_q4_0 * x, block_iq for (int row = 0; row < nrows; row += 4) { for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k; for (int ib = 0; ib < nblock; ++ib) { - for (int k = 0; k < 4; ++k) y[ib].d[k] = x4[k][ib].d; - for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) { - y[ib].qs[4*k+i+ 0] = (x4[k][ib].qs[i+0] & 0xf) | ((x4[k][ib].qs[i+ 8] & 0x0f) << 4); // 0....3 + 8...11 from each row - y[ib].qs[4*k+i+16] = (x4[k][ib].qs[i+0] >> 4) | ((x4[k][ib].qs[i+ 8] & 0xf0)); // 16...19 + 24...27 from each row - y[ib].qs[4*k+i+32] = (x4[k][ib].qs[i+4] & 0xf) | ((x4[k][ib].qs[i+12] & 0x0f) << 4); // 4....7 + 12...15 from each row - y[ib].qs[4*k+i+48] = (x4[k][ib].qs[i+4] >> 4) | ((x4[k][ib].qs[i+12] & 0xf0)); // 20...23 + 28...31 from each row + //for (int k = 0; k < 4; ++k) y[ib].d[k] = x4[k][ib].d; + //for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) { + // y[ib].qs[4*k+i+ 0] = (x4[k][ib].qs[i+0] & 0xf) | ((x4[k][ib].qs[i+ 8] & 0x0f) << 4); // 0....3 + 8...11 from each row + // y[ib].qs[4*k+i+16] = (x4[k][ib].qs[i+0] >> 4) | ((x4[k][ib].qs[i+ 8] & 0xf0)); // 16...19 + 24...27 from each row + // y[ib].qs[4*k+i+32] = (x4[k][ib].qs[i+4] & 0xf) | ((x4[k][ib].qs[i+12] & 0x0f) << 4); // 4....7 + 12...15 from each row + // y[ib].qs[4*k+i+48] = (x4[k][ib].qs[i+4] >> 4) | ((x4[k][ib].qs[i+12] & 0xf0)); // 20...23 + 28...31 from each row + //} + for (int k = 0; k < 4; ++k) { + y[ib].d[k] = x4[k][ib].d; + for (int l = 0; l < 4; ++l) { + // l = 0 -> 0, 8 with shift 0 -> 4*(l/2), 4*(l/2)+8 with shift 4*(l%2) + // l = 1 -> 0, 8 with shift 4 + // l = 2 -> 4, 12 with shift 0 + // l = 3 -> 4, 12 with shift 4 + for (int i = 0; i < 4; ++i) { + y[ib].qs[4*k+i+16*l] = ((x4[k][ib].qs[i+4*(l/2)] >> 4*(l%2)) & 0xf) | (((x4[k][ib].qs[i+4*(l/2)+8] >> 4*(l%2)) & 0xf) << 4); + } + } } } x += 4*nblock; @@ -3254,21 +3266,18 @@ size_t quantize_q4_0_r4(const float * src, void * dst, int64_t nrows, int64_t n_ void dequantize_row_q4_0_r4(const block_iq4_nl_x4 * x, float * y, int64_t k) { // we assume we are called with 4 rows int n_per_row = k/4; - int nb = n_per_row/QK4_NL; + int nb = n_per_row/QK4_0; float * yk[4]; for (int k = 0; k < 4; ++k) yk[k] = y + k*n_per_row; for (int ib = 0; ib < nb; ++ib) { for (int k = 0; k < 4; ++k) { float scale = GGML_FP16_TO_FP32(x[ib].d[k]); - for (int i = 0; i < 4; ++i) { - yk[k][QK4_NL*ib+i+ 0] = scale * ((x[ib].qs[4*k+i+ 0] & 0xf) - 8); - yk[k][QK4_NL*ib+i+ 8] = scale * ((x[ib].qs[4*k+i+ 0] >> 4) - 8); - yk[k][QK4_NL*ib+i+16] = scale * ((x[ib].qs[4*k+i+16] & 0xf) - 8); - yk[k][QK4_NL*ib+i+24] = scale * ((x[ib].qs[4*k+i+16] >> 4) - 8); - yk[k][QK4_NL*ib+i+ 4] = scale * ((x[ib].qs[4*k+i+32] & 0xf) - 8); - yk[k][QK4_NL*ib+i+12] = scale * ((x[ib].qs[4*k+i+32] >> 4) - 8); - yk[k][QK4_NL*ib+i+20] = scale * ((x[ib].qs[4*k+i+48] & 0xf) - 8); - yk[k][QK4_NL*ib+i+28] = scale * ((x[ib].qs[4*k+i+48] >> 4) - 8); + for (int l = 0; l < 4; ++l) { + int ll = 16*(l%2) + 4*(l/2); + for (int i = 0; i < 4; ++i) { + yk[k][QK4_0*ib+i+ll+0] = scale * ((x[ib].qs[4*k+i+16*l] & 0xf) - 8); + yk[k][QK4_0*ib+i+ll+8] = scale * ((x[ib].qs[4*k+i+16*l] >> 4) - 8); + } } } } @@ -3365,3 +3374,104 @@ void vec_dot_q8_0_r4_q8_0(int n, float * s, size_t bs, const void * vx, size_t b GGML_UNUSED(bx); GGML_UNUSED(by); } + +// +// ========================================= q5_0_r4 +// +void quantize_row_q5_0_r4_ref(const float * x, block_q5_0_r4 * y, int64_t k) { + // we assume we are called with 4 rows + quantize_q5_0_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_q5_0_r4(const float * x, void * y, int64_t k) { + // we assume we are called with 4 rows + quantize_q5_0_r4(x, y, 4, k/4, nullptr); +} + +static inline void convert_q5_0(const block_q5_0& x, uint8_t * L) { + uint32_t qh; + memcpy(&qh, x.qh, sizeof(qh)); + + for (int j = 0; j < QK5_0/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; + + L[j ] = (x.qs[j] & 0x0F) | xh_0; + L[j + QK4_0/2] = (x.qs[j] >> 4) | xh_1; + } +} + +static void repack_q5_0(int nrows, int n_per_row, const block_q5_0 * x, block_q5_0_r4 * y) { + GGML_ASSERT(nrows%4 == 0); + GGML_ASSERT(n_per_row%QK5_0 == 0); + int nblock = n_per_row/QK5_0; + const block_q5_0 * x4[4]; + uint8_t L[QK5_0]; + for (int row = 0; row < nrows; row += 4) { + for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k; + for (int ib = 0; ib < nblock; ++ib) { + std::memset(y[ib].qh, 0, QK5_0/2); + for (int k = 0; k < 4; ++k) { + y[ib].d[k] = x4[k][ib].d; + convert_q5_0(x4[k][ib], L); + for (int l = 0; l < 4; ++l) { + int l1 = 4*(l/2) + 16*(l%2), l2 = l1 + 8; + for (int i = 0; i < 4; ++i) { + y[ib].qs[4*k+i+16*l] = (L[i + l1] & 0xf) | ((L[i + l2] & 0xf) << 4); + y[ib].qh[4*k+i] |= ((L[i + l1] >> 4) | ((L[i + l2] >> 4) << 4)) << l; + } + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_q5_0_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) { + GGML_ASSERT(nrows%4 == 0); + auto row_size_0 = ggml_row_size(GGML_TYPE_Q5_0, n_per_row); + std::vector<char> qtmp(4*row_size_0); + char * qrow = (char *)dst; + for (int row = 0; row < nrows; row += 4) { + quantize_q5_0(src, qtmp.data(), 4, n_per_row, imatrix); + repack_q5_0(4, n_per_row, (const block_q5_0 *)qtmp.data(), (block_q5_0_r4 *)qrow); + src += 4*n_per_row; + qrow += 4*row_size_0; + } + return nrows*row_size_0; +} + +void dequantize_row_q5_0_r4(const block_q5_0_r4 * x, float * y, int64_t k) { + // we assume we are called with 4 rows + int n_per_row = k/4; + int nb = n_per_row/QK8_0; + float * yk[4]; + for (int k = 0; k < 4; ++k) yk[k] = y + k*n_per_row; + for (int ib = 0; ib < nb; ++ib) { + for (int k = 0; k < 4; ++k) { + float d = GGML_FP16_TO_FP32(x[ib].d[k]); + float m = -16*d; + for (int l = 0; l < 4; ++l) { + int ll = 16*(l%2) + 4*(l/2); + for (int i = 0; i < 4; ++i) { + yk[k][QK4_0*ib+i+ll+0] = d * ((x[ib].qs[4*k+i+16*l] & 0xf) | (((x[ib].qh[4*k+i] >> (l+0)) & 1) << 4)) + m; + yk[k][QK4_0*ib+i+ll+8] = d * ((x[ib].qs[4*k+i+16*l] >> 4) | (((x[ib].qh[4*k+i] >> (l+4)) & 1) << 4)) + m; + } + } + } + } +} + +void vec_dot_q5_0_r4_q8_0(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) { +#if GGML_USE_IQK_MULMAT + if (iqk_mul_mat(1, 1, n, GGML_TYPE_Q5_0_R4, vx, 0, GGML_TYPE_Q8_0, vy, 0, s, 0, 0, 1)) { + return; + } +#endif + GGML_ASSERT(n%QK4_NL == 0); + GGML_ASSERT(nrc == 1); + GGML_UNUSED(bs); + GGML_UNUSED(bx); + GGML_UNUSED(by); +} diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h index 53caed4c..24c241a2 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -81,6 +81,12 @@ size_t quantize_q8_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT ds void dequantize_row_q8_0_r4(const block_q8_0_x4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_q8_0_r4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void quantize_row_q5_0_r4_ref(const float * GGML_RESTRICT x, block_q5_0_r4 * GGML_RESTRICT y, int64_t k); +void quantize_row_q5_0_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_q5_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +void dequantize_row_q5_0_r4(const block_q5_0_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_q5_0_r4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); + #ifdef __cplusplus } #endif |