From db16e2831c0f344f041af3d067db81c42b16eb22 Mon Sep 17 00:00:00 2001 From: xctan Date: Mon, 28 Jul 2025 23:40:24 +0800 Subject: [PATCH] ggml-cpu : deduplicate scalar implementations (#14897) * remove redundant code in riscv * remove redundant code in arm * remove redundant code in loongarch * remove redundant code in ppc * remove redundant code in s390 * remove redundant code in wasm * remove redundant code in x86 * remove fallback headers * fix x86 ggml_vec_dot_q8_0_q8_0 --- ggml/src/ggml-cpu/arch/arm/quants.c | 667 ++----------------- ggml/src/ggml-cpu/arch/arm/repack.cpp | 300 +-------- ggml/src/ggml-cpu/arch/loongarch/quants.c | 659 +++---------------- ggml/src/ggml-cpu/arch/powerpc/quants.c | 699 +++----------------- ggml/src/ggml-cpu/arch/riscv/quants.c | 398 ++--------- ggml/src/ggml-cpu/arch/riscv/repack.cpp | 61 +- ggml/src/ggml-cpu/arch/s390/quants.c | 367 ++--------- ggml/src/ggml-cpu/arch/wasm/quants.c | 368 ++--------- ggml/src/ggml-cpu/arch/x86/quants.c | 767 +++------------------- ggml/src/ggml-cpu/arch/x86/repack.cpp | 265 +------- 10 files changed, 563 insertions(+), 3988 deletions(-) diff --git a/ggml/src/ggml-cpu/arch/arm/quants.c b/ggml/src/ggml-cpu/arch/arm/quants.c index 3e2d3d03d..c6d1d852e 100644 --- a/ggml/src/ggml-cpu/arch/arm/quants.c +++ b/ggml/src/ggml-cpu/arch/arm/quants.c @@ -1236,44 +1236,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = sumf; #else - const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243}; - - float sumf = 0.0f; - - for (int i = 0; i < nb; ++i) { - int sum = 0; - - for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) { - for (size_t l = 0; l < 5; ++l) { - for (size_t m = 0; m < 32; ++m) { - uint8_t q = x[i].qs[j + m] * pow3[l]; - uint16_t xi = ((uint16_t) q * 3) >> 8; - sum += (xi - 1) * y[i].qs[j*5 + l*32 + m]; - } - } - } - for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) { - for (size_t l = 0; l < 5; ++l) { - for (size_t m = 0; m < 16; ++m) { - uint8_t q = x[i].qs[j + m] * pow3[l]; - uint16_t xi = ((uint16_t) q * 3) >> 8; - sum += (xi - 1) * y[i].qs[j*5 + l*16 + m]; - } - } - } - - for (size_t l = 0; l < 4; ++l) { - for (size_t j = 0; j < sizeof(x->qh); ++j) { - uint8_t q = x[i].qh[j] * pow3[l]; - uint16_t xi = ((uint16_t) q * 3) >> 8; - sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j]; - } - } - - sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d); - } - - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1381,25 +1347,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = sumf; #else - float sumf = 0.0f; - - for (int i = 0; i < nb; ++i) { - int32_t sumi = 0; - - for (size_t j = 0; j < sizeof(x->qs); j += 32) { - for (size_t l = 0; l < 4; ++l) { - for (size_t k = 0; k < 32; ++k) { - sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1); - } - } - } - - const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); - - sumf += (float) sumi * d; - } - - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1729,45 +1680,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sum; #else - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < 16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = 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); - - int isum = 0; - int is = 0; - int d; - for (int k = 0; k < QK_K/128; ++k) { - int shift = 0; - for (int j = 0; j < 4; ++j) { - d = sc[is++] & 0xF; - int isuml = 0; - for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - d = sc[is++] & 0xF; - isuml = 0; - for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - shift += 2; - q8 += 32; - } - q2 += 32; - } - sumf += dall * isum - dmin * summs; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2057,68 +1973,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sum; #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2431,61 +2291,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2578,66 +2391,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3093,47 +2854,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi } *s = sum; #else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3229,34 +2953,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.25f * sumf; #else - - uint32_t aux32[2]; - const uint8_t * aux8 = (const uint8_t *)aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(aux32, q2, 2*sizeof(uint32_t)); - q2 += 4; - const uint32_t ls = 2*(aux32[1] >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3327,42 +3027,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = 0.125f * sumf; #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const uint8_t * GGML_RESTRICT sc = x[i].scales; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; - const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls1; - sumi = 0; - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls2; - q2 += 4; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3455,45 +3123,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = 0.125f * sumf; #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint8_t * qh = x[i].qh; - const uint8_t * signs = qs + QK_K/8; - - int bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf); - int ls2 = 1 + 2*(x[i].scales[ib32] >> 4); - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += ls1 * sumi1 + ls2 * sumi2; - qs += 4; - signs += 4; - } - - sumf += d * bsum; - } - - *s = 0.125f * sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3553,36 +3186,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.5f * sumf; #else - - uint32_t aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t); - const uint32_t ls = 2*(aux32 >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]); - const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]); - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - q3 += 8; - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.25f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3689,48 +3296,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = sumf; #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT qs = x[i].qs; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const uint8_t * GGML_RESTRICT signs = x[i].signs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { - const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; - const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls1; - sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls2; - } - sumf += d * bsum; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3793,36 +3362,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = sumf; #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint16_t * qh = x[i].qh; - - int sumi = 0, sumi1 = 0; - for (int ib = 0; ib < QK_K/32; ++ib) { - const int ls = 2*((qh[ib] >> 12) & 7) + 1; - const int delta = qh[ib] & 0x8000 ? -1 : 1; - int lsum = 0; - for (int l = 0; l < 4; ++l) { - const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); - for (int j = 0; j < 8; ++j) { - lsum += q8[j] * grid[j]; - } - q8 += 8; - } - sumi += ls * lsum; - sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]); - qs += 4; - } - - sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); - } - - *s = sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3912,52 +3455,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = sumf; #else - - int sum1[2], sum2[2], delta[4]; - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint8_t * qh = x[i].qh; - const uint16_t * sc = (const uint16_t *)x[i].scales; - - scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - - int sumi1 = 0, sumi2 = 0; - for (int ib = 0; ib < QK_K/32; ++ib) { - delta[0] = qh[0] & 0x08 ? -1 : 1; - delta[1] = qh[0] & 0x80 ? -1 : 1; - delta[2] = qh[1] & 0x08 ? -1 : 1; - delta[3] = qh[1] & 0x80 ? -1 : 1; - sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0; - for (int l = 0; l < 4; ++l) { - const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700))); - int lsum1 = 0, lsum2 = 0; - for (int j = 0; j < 8; ++j) { - lsum1 += q8[j] * grid[j]; - lsum2 += q8[j]; - } - q8 += 8; - sum1[l/2] += lsum1; - sum2[l/2] += lsum2*delta[l]; - } - - const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1; - const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1; - - sumi1 += sum1[0] * ls1 + sum1[1] * ls2; - sumi2 += sum2[0] * ls1 + sum2[1] * ls2; - qs += 4; - qh += 2; - } - - sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); - } - - *s = sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(scale); + ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -4078,37 +3580,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = sumf; #else - float sumf = 0; - for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; - uint16_t h = x[ibl].scales_h; - const uint8_t * qs = x[ibl].qs; - const int8_t * q8 = y[ibl].qs; - for (int ib = 0; ib < QK_K/32; ib += 2) { - const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); - const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); - h >>= 4; - const float d1 = d4d8*(ls1 - 32); - const float d2 = d4d8*(ls2 - 32); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d1 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - sumi1 = sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d2 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - } - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 2f8bc9e25..fdd0a513b 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -86,35 +86,9 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR } } #else - // scalar - const int blck_size_interleave = 4; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } + UNUSED(nb); + UNUSED(y); + ggml_quantize_mat_q8_0_4x4_generic(x, vy, k); #endif } @@ -205,35 +179,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR } #else - // scalar - const int blck_size_interleave = 8; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } + UNUSED(nb); + UNUSED(y); + ggml_quantize_mat_q8_0_4x8_generic(x, vy, k); #endif } @@ -295,29 +243,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo } return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - float sumf[4]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } + ggml_gemv_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -383,29 +309,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo } return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - float sumf[4]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } + ggml_gemv_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -497,31 +401,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #endif // #if defined(__ARM_FEATURE_SVE) #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) - { - float sumf[8]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } - } + ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -591,31 +471,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const } return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - { - float sumf[4]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; - const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } - } + ggml_gemv_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -1096,40 +952,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo ); return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - { - float sumf[4][4]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; - } - } - } - } + ggml_gemm_q4_0_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -1550,38 +1373,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo ); return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - float sumf[4][4]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; - } - } - } + ggml_gemm_q4_0_4x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -2019,38 +1811,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) - float sumf[4][8]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; - } - } - } + ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -2126,38 +1887,5 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const } return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - { - float sumf[4][4]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; - const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; - } - } - } - } + ggml_gemm_iq4_nl_4x4_q8_0_generic(n, s, bs, vx, vy, nr, nc); } diff --git a/ggml/src/ggml-cpu/arch/loongarch/quants.c b/ggml/src/ggml-cpu/arch/loongarch/quants.c index 7908da4d1..0f9af7bf5 100644 --- a/ggml/src/ggml-cpu/arch/loongarch/quants.c +++ b/ggml/src/ggml-cpu/arch/loongarch/quants.c @@ -821,24 +821,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = hsum_float_8(acc) + summs; -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F); - const int v1 = (x[ib].qs[j] >> 4); - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_0_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) { @@ -883,30 +874,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = hsum_float_8(acc); -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - - const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16); - const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16); - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_1_q8_1(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) { @@ -954,30 +930,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = hsum_float_8(acc) + summs; -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0; - const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1; - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q8_0_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) { @@ -1016,18 +977,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = hsum_float_8(acc); -#endif - for (; ib < nb; ++ib) { - int sumi = 0; - - for (int j = 0; j < qk; j++) { - sumi += x[ib].qs[j]*y[ib].qs[j]; - } - - sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q2_K_q8_K(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) { @@ -1103,45 +1061,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc); #else - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < 16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = 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); - - int isum = 0; - int is = 0; - int d; - for (int k = 0; k < QK_K/128; ++k) { - int shift = 0; - for (int j = 0; j < 4; ++j) { - d = sc[is++] & 0xF; - int isuml = 0; - for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - d = sc[is++] & 0xF; - isuml = 0; - for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - shift += 2; - q8 += 32; - } - q2 += 32; - } - sumf += dall * isum - dmin * summs; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1239,70 +1162,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc); #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_q4_K_q8_K(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) { @@ -1391,61 +1257,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc) + ((v4f32)acc_m)[0]; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1541,66 +1360,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc) + ((v4f32)acc_m)[0]; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1678,47 +1445,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc); #else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1815,34 +1545,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.125f * hsum_float_8(accumf); #else - - uint32_t aux32[2]; - const uint8_t * aux8 = (const uint8_t *)aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(aux32, q2, 2*sizeof(uint32_t)); - q2 += 4; - const uint32_t ls = 2*(aux32[1] >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1978,42 +1684,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = 0.125f * hsum_float_8(accumf); #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const uint8_t * GGML_RESTRICT sc = x[i].scales; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; - const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls1; - sumi = 0; - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls2; - q2 += 4; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2105,47 +1779,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = 0.125f * hsum_float_8(accumf); #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint8_t * qh = x[i].qh; - const uint8_t * signs = qs + QK_K/8; - - int bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf); - int ls2 = 1 + 2*(x[i].scales[ib32] >> 4); - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += ls1 * sumi1 + ls2 * sumi2; - qs += 4; - signs += 4; - } - - sumf += d * bsum; - } - - *s = 0.125f * sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_iq3_xxs_q8_K(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) { @@ -2209,36 +1847,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.25f * hsum_float_8(accumf); #else - - uint32_t aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t); - const uint32_t ls = 2*(aux32 >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]); - const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]); - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - q3 += 8; - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.25f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2338,48 +1950,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(accumf); #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT qs = x[i].qs; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const uint8_t * GGML_RESTRICT signs = x[i].signs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { - const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; - const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls1; - sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls2; - } - sumf += d * bsum; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2460,36 +2034,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(accum) + IQ1S_DELTA * accum1; #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint16_t * qh = x[i].qh; - - int sumi = 0, sumi1 = 0; - for (int ib = 0; ib < QK_K/32; ++ib) { - const int ls = 2*((qh[ib] >> 12) & 7) + 1; - const int delta = qh[ib] & 0x8000 ? -1 : 1; - int lsum = 0; - for (int l = 0; l < 4; ++l) { - const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); - for (int j = 0; j < 8; ++j) { - lsum += q8[j] * grid[j]; - } - q8 += 8; - } - sumi += ls * lsum; - sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]); - qs += 4; - } - - sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); - } - - *s = sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2603,37 +2151,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = hsum_float_8(accum); #else - float sumf = 0; - for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; - uint16_t h = x[ibl].scales_h; - const uint8_t * qs = x[ibl].qs; - const int8_t * q8 = y[ibl].qs; - for (int ib = 0; ib < QK_K/32; ib += 2) { - const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); - const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); - h >>= 4; - const float d1 = d4d8*(ls1 - 32); - const float d2 = d4d8*(ls2 - 32); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d1 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - sumi1 = sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d2 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - } - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/powerpc/quants.c b/ggml/src/ggml-cpu/arch/powerpc/quants.c index 053d5cbdc..49aae7a23 100644 --- a/ggml/src/ggml-cpu/arch/powerpc/quants.c +++ b/ggml/src/ggml-cpu/arch/powerpc/quants.c @@ -201,24 +201,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = vec_extract(vsumf0, 0); -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F) - 8; - const int v1 = (x[ib].qs[j] >> 4) - 8; - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); - } - *s = sumf; +#else + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q4_1_q8_1(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) { @@ -278,24 +268,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = vec_extract(vsumf0, 0); -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F); - const int v1 = (x[ib].qs[j] >> 4); - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_0_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) { @@ -360,30 +340,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = vec_extract(vsumf0, 0); -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - - const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16); - const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16); - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; - } - *s = sumf; +#else + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_1_q8_1(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) { @@ -451,30 +415,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = vec_extract(vsumf0, 0); -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0; - const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1; - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q8_0_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) { @@ -535,18 +484,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = vec_extract(vsumf0, 0); -#endif - for (; ib < nb; ++ib) { - int sumi = 0; - - for (int j = 0; j < qk; j++) { - sumi += x[ib].qs[j]*y[ib].qs[j]; - } - - sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q2_K_q8_K(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) { @@ -695,45 +641,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = vec_extract(vsumf0, 0); #else - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < 16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = 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); - - int isum = 0; - int is = 0; - int d; - for (int k = 0; k < QK_K/128; ++k) { - int shift = 0; - for (int j = 0; j < 4; ++j) { - d = sc[is++] & 0xF; - int isuml = 0; - for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - d = sc[is++] & 0xF; - isuml = 0; - for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - shift += 2; - q8 += 32; - } - q2 += 32; - } - sumf += dall * isum - dmin * summs; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -907,70 +818,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = vec_extract(vsumf0, 0); #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_q4_K_q8_K(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) { @@ -1130,61 +984,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = vec_extract(vsumf0, 0); #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1342,66 +1149,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = vec_extract(vsumf0, 0); #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1556,47 +1311,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = vec_extract(vsumf0, 0); #else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1737,34 +1455,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.125f * vec_extract(vsumf0, 0); #else - - uint32_t aux32[2]; - const uint8_t * aux8 = (const uint8_t *)aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(aux32, q2, 2*sizeof(uint32_t)); - q2 += 4; - const uint32_t ls = 2*(aux32[1] >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1869,42 +1563,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = 0.125f * vec_extract(vsumf0, 0); #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const uint8_t * GGML_RESTRICT sc = x[i].scales; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; - const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls1; - sumi = 0; - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls2; - q2 += 4; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2030,47 +1692,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = 0.125f * vec_extract(vsumf0, 0); #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint8_t * qh = x[i].qh; - const uint8_t * signs = qs + QK_K/8; - - int bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf); - int ls2 = 1 + 2*(x[i].scales[ib32] >> 4); - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += ls1 * sumi1 + ls2 * sumi2; - qs += 4; - signs += 4; - } - - sumf += d * bsum; - } - - *s = 0.125f * sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_iq3_xxs_q8_K(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) { @@ -2172,36 +1798,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.25f * vec_extract(vsumf0, 0); #else - - uint32_t aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t); - const uint32_t ls = 2*(aux32 >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]); - const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]); - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - q3 += 8; - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.25f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2327,48 +1927,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = vec_extract(vsumf0, 0); #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT qs = x[i].qs; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const uint8_t * GGML_RESTRICT signs = x[i].signs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { - const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; - const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls1; - sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls2; - } - sumf += d * bsum; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2481,36 +2043,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = vec_extract(vsumf0, 0); #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint16_t * qh = x[i].qh; - - int sumi = 0, sumi1 = 0; - for (int ib = 0; ib < QK_K/32; ++ib) { - const int ls = 2*((qh[ib] >> 12) & 7) + 1; - const int delta = qh[ib] & 0x8000 ? -1 : 1; - int lsum = 0; - for (int l = 0; l < 4; ++l) { - const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); - for (int j = 0; j < 8; ++j) { - lsum += q8[j] * grid[j]; - } - q8 += 8; - } - sumi += ls * lsum; - sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]); - qs += 4; - } - - sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); - } - - *s = sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2581,17 +2117,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v sumf = vec_extract(vsumf0, 0); -#endif - for (; ib < nb; ++ib) { - const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < QK4_NL/2; ++j) { - sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; - sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4]; - } - sumf += d * (sumi1 + sumi2); - } *s = sumf; +#else + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_iq4_xs_q8_K(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) { @@ -2696,37 +2230,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = vec_extract(vsumf0, 0); #else - float sumf = 0; - for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; - uint16_t h = x[ibl].scales_h; - const uint8_t * qs = x[ibl].qs; - const int8_t * q8 = y[ibl].qs; - for (int ib = 0; ib < QK_K/32; ib += 2) { - const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); - const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); - h >>= 4; - const float d1 = d4d8*(ls1 - 32); - const float d2 = d4d8*(ls2 - 32); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d1 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - sumi1 = sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d2 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - } - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/riscv/quants.c b/ggml/src/ggml-cpu/arch/riscv/quants.c index 8b64d8adc..6c74417c9 100644 --- a/ggml/src/ggml-cpu/arch/riscv/quants.c +++ b/ggml/src/ggml-cpu/arch/riscv/quants.c @@ -116,6 +116,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i //===================================== Dot products ================================= void ggml_vec_dot_q4_0_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) { +#if defined(__riscv_v) const int qk = QK8_0; const int nb = n / qk; @@ -132,7 +133,6 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi int ib = 0; float sumf = 0; -#if defined(__riscv_v) size_t vl = qk / 2; for (; ib < nb; ++ib) { @@ -164,27 +164,14 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F) - 8; - const int v1 = (x[ib].qs[j] >> 4) - 8; - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); - } - *s = sumf; +#else + ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q4_1_q8_1(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) { +#if defined(__riscv_v) const int qk = QK8_1; const int nb = n / qk; @@ -201,7 +188,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi int ib = 0; float sumf = 0; -#if defined(__riscv_v) size_t vl = qk / 2; for (; ib < nb; ++ib) { @@ -229,27 +215,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F); - const int v1 = (x[ib].qs[j] >> 4); - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_0_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) { +#if defined(__riscv_v) const int qk = QK8_0; const int nb = n / qk; @@ -267,7 +240,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const block_q5_0 * GGML_RESTRICT x = vx; const block_q8_0 * GGML_RESTRICT y = vy; -#if defined(__riscv_v) size_t vl; size_t vlenb = __riscv_vlenb(); @@ -297,33 +269,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - - const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16); - const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16); - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; - } - *s = sumf; +#else + ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_1_q8_1(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) { +#if defined(__riscv_v) const int qk = QK8_1; const int nb = n / qk; @@ -341,7 +294,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const block_q5_1 * GGML_RESTRICT x = vx; const block_q8_1 * GGML_RESTRICT y = vy; -#if defined(__riscv_v) size_t vl; size_t vlenb = __riscv_vlenb(); @@ -370,30 +322,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0; - const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1; - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q8_0_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) { @@ -431,18 +363,17 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } -#endif - for (; ib < nb; ++ib) { - int sumi = 0; - - for (int j = 0; j < qk; j++) { - sumi += x[ib].qs[j]*y[ib].qs[j]; - } - - sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); - } - *s = sumf; +#else + + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + + ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q2_K_q8_K(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) { @@ -738,44 +669,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi #else - float sumf = 0; + UNUSED(x); + UNUSED(y); + UNUSED(nb); - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < 16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = 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); - - int isum = 0; - int is = 0; - int d; - for (int k = 0; k < QK_K/128; ++k) { - int shift = 0; - for (int j = 0; j < 4; ++j) { - d = sc[is++] & 0xF; - int isuml = 0; - for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - d = sc[is++] & 0xF; - isuml = 0; - for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - shift += 2; - q8 += 32; - } - q2 += 32; - } - sumf += dall * isum - dmin * summs; - } - *s = sumf; + ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1147,68 +1045,14 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1534,60 +1378,15 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi #else - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; + UNUSED(x); + UNUSED(y); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(nb); + UNUSED(utmp); - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1698,65 +1497,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi #else - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; + UNUSED(x); + UNUSED(y); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(nb); + UNUSED(utmp); - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2024,46 +1773,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi #else - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); + UNUSED(x); + UNUSED(y); + UNUSED(nb); - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/riscv/repack.cpp b/ggml/src/ggml-cpu/arch/riscv/repack.cpp index 45c91a694..2a35ff9ad 100644 --- a/ggml/src/ggml-cpu/arch/riscv/repack.cpp +++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp @@ -112,31 +112,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo } #endif - { - float sumf[8]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } - } + ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -361,37 +337,6 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo return; } -#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) - float sumf[4][8]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; - } - } - } +#endif + ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } diff --git a/ggml/src/ggml-cpu/arch/s390/quants.c b/ggml/src/ggml-cpu/arch/s390/quants.c index a840219a4..7e4229d0e 100644 --- a/ggml/src/ggml-cpu/arch/s390/quants.c +++ b/ggml/src/ggml-cpu/arch/s390/quants.c @@ -172,24 +172,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = acc[0] + acc[1] + acc[2] + acc[3]; -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F) - 8; - const int v1 = (x[ib].qs[j] >> 4) - 8; - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q4_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q4_1_q8_1(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) { @@ -239,24 +230,15 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs; -#endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F); - const int v1 = (x[ib].qs[j] >> 4); - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q8_0_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) { @@ -298,18 +280,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = acc[0] + acc[1] + acc[2] + acc[3]; -#endif - for (; ib < nb; ++ib) { - int sumi = 0; - - for (int j = 0; j < qk; j++) { - sumi += x[ib].qs[j]*y[ib].qs[j]; - } - - sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q3_K_q8_K(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) { @@ -442,70 +421,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sum; #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_q4_K_q8_K(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) { @@ -600,61 +522,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -767,66 +642,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -969,47 +792,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sum; #else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1186,17 +972,15 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]); } -#endif - for (; ib < nb; ++ib) { - const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < QK4_NL/2; ++j) { - sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; - sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4]; - } - sumf += d * (sumi1 + sumi2); - } *s = sumf; +#else + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_iq4_nl_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_iq4_xs_q8_K(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) { @@ -1264,37 +1048,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = sumf; #else - float sumf = 0; - for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; - uint16_t h = x[ibl].scales_h; - const uint8_t * qs = x[ibl].qs; - const int8_t * q8 = y[ibl].qs; - for (int ib = 0; ib < QK_K/32; ib += 2) { - const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); - const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); - h >>= 4; - const float d1 = d4d8*(ls1 - 32); - const float d2 = d4d8*(ls2 - 32); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d1 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - sumi1 = sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d2 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - } - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/wasm/quants.c b/ggml/src/ggml-cpu/arch/wasm/quants.c index b0904d8a3..74a359e6d 100644 --- a/ggml/src/ggml-cpu/arch/wasm/quants.c +++ b/ggml/src/ggml-cpu/arch/wasm/quants.c @@ -435,30 +435,15 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - - const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16); - const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16); - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q5_1_q8_1(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) { @@ -545,30 +530,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs; -#endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0; - const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1; - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(sumf); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q8_0_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) { @@ -628,18 +598,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); -#endif - for (; ib < nb; ++ib) { - int sumi = 0; - - for (int j = 0; j < qk; j++) { - sumi += x[ib].qs[j]*y[ib].qs[j]; - } - - sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); - } - *s = sumf; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + UNUSED(sumf); + ggml_vec_dot_q8_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); +#endif } void ggml_vec_dot_q2_K_q8_K(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) { @@ -755,45 +722,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < 16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = 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); - - int isum = 0; - int is = 0; - int d; - for (int k = 0; k < QK_K/128; ++k) { - int shift = 0; - for (int j = 0; j < 4; ++j) { - d = sc[is++] & 0xF; - int isuml = 0; - for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - d = sc[is++] & 0xF; - isuml = 0; - for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - shift += 2; - q8 += 32; - } - q2 += 32; - } - sumf += dall * isum - dmin * summs; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -902,68 +834,12 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1089,61 +965,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1279,66 +1108,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1435,47 +1212,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = sumf; #else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/x86/quants.c b/ggml/src/ggml-cpu/arch/x86/quants.c index e7527c00a..30fd59f70 100644 --- a/ggml/src/ggml-cpu/arch/x86/quants.c +++ b/ggml/src/ggml-cpu/arch/x86/quants.c @@ -702,7 +702,6 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const block_q8_1 * GGML_RESTRICT y = vy; int ib = 0; - float sumf = 0; #if defined(__AVX2__) || defined(__AVX__) // Initialize accumulator with zeros @@ -737,26 +736,14 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi #endif } - sumf = hsum_float_8(acc) + summs; - + *s = hsum_float_8(acc) + summs; +#else + UNUSED(nb); + UNUSED(x); + UNUSED(y); + UNUSED(ib); + ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - for (; ib < nb; ++ib) { - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const int v0 = (x[ib].qs[j] & 0x0F); - const int v1 = (x[ib].qs[j] >> 4); - - sumi0 += (v0 * y[ib].qs[j]); - sumi1 += (v1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - - *s = sumf; } void ggml_vec_dot_q5_0_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) { @@ -764,7 +751,6 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const int nb = n / qk; int ib = 0; - float sumf = 0; assert(n % qk == 0); assert(qk == QK5_0); @@ -799,7 +785,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi acc = _mm256_fmadd_ps(d, q, acc); } - sumf = hsum_float_8(acc); + *s = hsum_float_8(acc); #elif defined(__AVX__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); @@ -830,32 +816,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc); } - sumf = hsum_float_8(acc); - + *s = hsum_float_8(acc); +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - - const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16); - const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16); - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; - } - - *s = sumf; } void ggml_vec_dot_q5_1_q8_1(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) { @@ -863,7 +831,6 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const int nb = n / qk; int ib = 0; - float sumf = 0; assert(n % qk == 0); assert(qk == QK5_1); @@ -901,7 +868,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); } - sumf = hsum_float_8(acc) + summs; + *s = hsum_float_8(acc) + summs; #elif defined(__AVX__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); @@ -935,32 +902,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc); } - sumf = hsum_float_8(acc) + summs; - + *s = hsum_float_8(acc) + summs; +#else + UNUSED(nb); + UNUSED(ib); + UNUSED(x); + UNUSED(y); + ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - for (; ib < nb; ++ib) { - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); - - int sumi0 = 0; - int sumi1 = 0; - - for (int j = 0; j < qk/2; ++j) { - const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; - const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - - const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0; - const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1; - - sumi0 += (x0 * y[ib].qs[j]); - sumi1 += (x1 * y[ib].qs[j + qk/2]); - } - - int sumi = sumi0 + sumi1; - sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); - } - - *s = sumf; } void ggml_vec_dot_q8_0_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) { @@ -1017,7 +966,6 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } sumf = hsum_float_8(accum); - #endif for (; ib < nb; ++ib) { int sumi = 0; @@ -1157,44 +1105,10 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(sumf); #else - const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243}; - - float sumf = 0.0f; - - for (int i = 0; i < nb; ++i) { - int sum = 0; - - for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) { - for (size_t l = 0; l < 5; ++l) { - for (size_t m = 0; m < 32; ++m) { - uint8_t q = x[i].qs[j + m] * pow3[l]; - uint16_t xi = ((uint16_t) q * 3) >> 8; - sum += (xi - 1) * y[i].qs[j*5 + l*32 + m]; - } - } - } - for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) { - for (size_t l = 0; l < 5; ++l) { - for (size_t m = 0; m < 16; ++m) { - uint8_t q = x[i].qs[j + m] * pow3[l]; - uint16_t xi = ((uint16_t) q * 3) >> 8; - sum += (xi - 1) * y[i].qs[j*5 + l*16 + m]; - } - } - } - - for (size_t l = 0; l < 4; ++l) { - for (size_t j = 0; j < sizeof(x->qh); ++j) { - uint8_t q = x[i].qh[j] * pow3[l]; - uint16_t xi = ((uint16_t) q * 3) >> 8; - sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j]; - } - } - - sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d); - } - - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1257,25 +1171,10 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(sumf); #else - float sumf = 0.0f; - - for (int i = 0; i < nb; ++i) { - int32_t sumi = 0; - - for (size_t j = 0; j < sizeof(x->qs); j += 32) { - for (size_t l = 0; l < 4; ++l) { - for (size_t k = 0; k < 32; ++k) { - sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1); - } - } - } - - const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); - - sumf += (float) sumi * d; - } - - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1464,45 +1363,10 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc); #else - - float sumf = 0; - - for (int i = 0; i < nb; ++i) { - - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; - const uint8_t * sc = x[i].scales; - - int summs = 0; - for (int j = 0; j < 16; ++j) { - summs += y[i].bsums[j] * (sc[j] >> 4); - } - - const float dall = 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); - - int isum = 0; - int is = 0; - int d; - for (int k = 0; k < QK_K/128; ++k) { - int shift = 0; - for (int j = 0; j < 4; ++j) { - d = sc[is++] & 0xF; - int isuml = 0; - for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - d = sc[is++] & 0xF; - isuml = 0; - for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); - isum += d * isuml; - shift += 2; - q8 += 32; - } - q2 += 32; - } - sumf += dall * isum - dmin * summs; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -1769,70 +1633,13 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc); #else - // scalar version - // This function is written like this so the compiler can manage to vectorize most of it - // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the - // manually vectorized version above. Every other version I tried would run at least 4 times slower. - // The ideal situation would be if we could just write the code once, and the compiler would - // automatically produce the best possible set of machine instructions, instead of us having to manually - // write vectorized versions for AVX, ARM_NEON, etc. - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - uint32_t auxs[4]; - const int8_t * scales = (const int8_t*)auxs; - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].hmask; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; - for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); - a += 32; m <<= 1; - q3 += 32; - } - a = aux8; - - memcpy(auxs, x[i].scales, 12); - uint32_t tmp = auxs[2]; - auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); - auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); - auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); - auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); - for (int j = 0; j < QK_K/16; ++j) { - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; - + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_q4_K_q8_K(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) { @@ -2002,61 +1809,14 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m); #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - a += 32; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - a += 32; q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2259,66 +2019,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc) + summs; #else - - const uint8_t * scales = (const uint8_t*)&utmp[0]; - const uint8_t * mins = (const uint8_t*)&utmp[2]; - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].qs; - const uint8_t * GGML_RESTRICT hm = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - uint8_t m = 1; - for (int j = 0; j < QK_K/64; ++j) { - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); - for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); - a += 32; m <<= 1; - q4 += 32; - } - 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; - - int sumi = 0; - for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/32; ++j) { - int32_t scale = scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; - sumf -= dmin * sumi; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + UNUSED(utmp); + ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2520,47 +2228,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi *s = hsum_float_8(acc); #else - - int8_t aux8[QK_K]; - int16_t aux16[8]; - float sums [8]; - int32_t aux32[8]; - memset(sums, 0, 8*sizeof(float)); - - float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * GGML_RESTRICT q4 = x[i].ql; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * GGML_RESTRICT a = aux8; - for (int j = 0; j < QK_K; j += 128) { - for (int l = 0; l < 32; ++l) { - a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; - a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; - a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; - a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; - } - a += 128; - q4 += 64; - qh += 32; - } - a = aux8; - int is = 0; - for (int j = 0; j < QK_K/16; ++j) { - int scale = x[i].scales[is++]; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; - for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; - q8 += 8; a += 8; - } - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - } - for (int l = 0; l < 8; ++l) sumf += sums[l]; - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -2712,34 +2383,10 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.125f * hsum_float_8(accumf); #else - - uint32_t aux32[2]; - const uint8_t * aux8 = (const uint8_t *)aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(aux32, q2, 2*sizeof(uint32_t)); - q2 += 4; - const uint32_t ls = 2*(aux32[1] >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3033,42 +2680,10 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = 0.125f * hsum_float_8(accumf); #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * GGML_RESTRICT q2 = x[i].qs; - const uint8_t * GGML_RESTRICT sc = x[i].scales; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; - const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls1; - sumi = 0; - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); - const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; - for (int j = 0; j < 8; ++j) { - sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += sumi * ls2; - q2 += 4; - } - sumf += d * bsum; - } - *s = 0.125f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3250,47 +2865,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = 0.125f * hsum_float_8(accumf); #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint8_t * qh = x[i].qh; - const uint8_t * signs = qs + QK_K/8; - - int bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf); - int ls2 = 1 + 2*(x[i].scales[ib32] >> 4); - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < 2; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - for (int l = 2; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); - for (int j = 0; j < 8; ++j) { - sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - } - bsum += ls1 * sumi1 + ls2 * sumi2; - qs += 4; - signs += 4; - } - - sumf += d * bsum; - } - - *s = 0.125f * sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif - } void ggml_vec_dot_iq3_xxs_q8_K(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) { @@ -3410,36 +2989,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const *s = 0.25f * hsum_float_8(accumf); #else - - uint32_t aux32; - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT q3 = x[i].qs; - const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { - memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t); - const uint32_t ls = 2*(aux32 >> 28) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]); - const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]); - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - q3 += 8; - bsum += sumi * ls; - } - sumf += d * bsum; - } - *s = 0.25f * sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3646,48 +3199,10 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(accumf); #else - - float sumf = 0.f; - for (int i = 0; i < nb; ++i) { - const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * GGML_RESTRICT qs = x[i].qs; - const uint8_t * GGML_RESTRICT qh = x[i].qh; - const uint8_t * GGML_RESTRICT signs = x[i].signs; - const int8_t * GGML_RESTRICT q8 = y[i].qs; - int32_t bsum = 0; - for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { - const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; - const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; - int32_t sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls1; - sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); - for (int j = 0; j < 4; ++j) { - sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); - sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); - } - q8 += 8; - } - qs += 8; - signs += 4; - bsum += sumi * ls2; - } - sumf += d * bsum; - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -3811,36 +3326,10 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(accum) + IQ1S_DELTA * accum1; #else - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint16_t * qh = x[i].qh; - - int sumi = 0, sumi1 = 0; - for (int ib = 0; ib < QK_K/32; ++ib) { - const int ls = 2*((qh[ib] >> 12) & 7) + 1; - const int delta = qh[ib] & 0x8000 ? -1 : 1; - int lsum = 0; - for (int l = 0; l < 4; ++l) { - const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); - for (int j = 0; j < 8; ++j) { - lsum += q8[j] * grid[j]; - } - q8 += 8; - } - sumi += ls * lsum; - sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]); - qs += 4; - } - - sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); - } - - *s = sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -4043,52 +3532,11 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2); #else - - int sum1[2], sum2[2], delta[4]; - - float sumf = 0; - for (int i = 0; i < nb; i++) { - - const int8_t * q8 = y[i].qs; - const uint8_t * qs = x[i].qs; - const uint8_t * qh = x[i].qh; - const uint16_t * sc = (const uint16_t *)x[i].scales; - - scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - - int sumi1 = 0, sumi2 = 0; - for (int ib = 0; ib < QK_K/32; ++ib) { - delta[0] = qh[0] & 0x08 ? -1 : 1; - delta[1] = qh[0] & 0x80 ? -1 : 1; - delta[2] = qh[1] & 0x08 ? -1 : 1; - delta[3] = qh[1] & 0x80 ? -1 : 1; - sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0; - for (int l = 0; l < 4; ++l) { - const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700))); - int lsum1 = 0, lsum2 = 0; - for (int j = 0; j < 8; ++j) { - lsum1 += q8[j] * grid[j]; - lsum2 += q8[j]; - } - q8 += 8; - sum1[l/2] += lsum1; - sum2[l/2] += lsum2*delta[l]; - } - - const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1; - const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1; - - sumi1 += sum1[0] * ls1 + sum1[1] * ls2; - sumi2 += sum2[0] * ls1 + sum2[1] * ls2; - qs += 4; - qh += 2; - } - - sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); - } - - *s = sumf; - + UNUSED(x); + UNUSED(y); + UNUSED(nb); + UNUSED(scale); + ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } @@ -4275,37 +3723,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v *s = hsum_float_8(accum); #else - float sumf = 0; - for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; - uint16_t h = x[ibl].scales_h; - const uint8_t * qs = x[ibl].qs; - const int8_t * q8 = y[ibl].qs; - for (int ib = 0; ib < QK_K/32; ib += 2) { - const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); - const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); - h >>= 4; - const float d1 = d4d8*(ls1 - 32); - const float d2 = d4d8*(ls2 - 32); - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d1 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - sumi1 = sumi2 = 0; - for (int j = 0; j < 16; ++j) { - sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; - sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; - } - sumf += d2 * (sumi1 + sumi2); - qs += 16; - q8 += 32; - } - } - *s = sumf; + UNUSED(x); + UNUSED(y); + UNUSED(nb); + ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc); #endif } diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp index c00c1e541..1982cfef9 100644 --- a/ggml/src/ggml-cpu/arch/x86/repack.cpp +++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp @@ -281,35 +281,9 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR } #else - // scalar - const int blck_size_interleave = 8; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } + UNUSED(nb); + UNUSED(y); + ggml_quantize_mat_q8_0_4x8_generic(x, vy, k); #endif } @@ -531,49 +505,9 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR } #else - - // scalar - const int blck_size_interleave = 8; - float srcv[4][QK_K]; - float iscale[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - float max = 0; - - for (int j = 0; j < QK_K; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK_K + j]; - // Update the maximum value of the corresponding super block - if(amax < fabsf(srcv[row_iter][j])) { - amax = fabsf(srcv[row_iter][j]); - max = srcv[row_iter][j]; - } - } - - iscale[row_iter] = amax ? -127.f/max : 0; - - y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; - } - - for (int j = 0; j < QK_K / 4; j++) { - y[i].bsums[j] = 0; - } - - // Quants values are interleaved in sequence of eight bytes from corresponding super blocks - // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving - // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on - for (int j = 0; j < QK_K * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); - - float x0 = srcv[src_id][src_offset] * iscale[src_id]; - y[i].qs[j] = nearest_int(x0); - y[i].bsums[index] += y[i].qs[j]; - } - } + UNUSED(nb); + UNUSED(y); + ggml_quantize_mat_q8_K_4x8_generic(x, vy, k); #endif } @@ -689,31 +623,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo return; #endif - { - float sumf[8]; - int sumi; - - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); - } - } - } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; - } - } + ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -932,61 +842,10 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } #else - - float sumf[8]; - float sum_minf[8]; - uint32_t utmp[32]; - int sumi1; - int sumi2; - int sumi; - - const block_q8_K * a_ptr = (const block_q8_K *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); - - for (int j = 0; j < ncols_interleaved; j++) { - sumf[j] = 0.0; - sum_minf[j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int sb = 0; sb < 8; sb++) { - memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); - utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); - const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; - utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); - utmp[sb * 4 + 2] = uaux_0; - utmp[sb * 4 + 0] &= kmask1; - } - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; - uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; - for (int j = 0; j < ncols_interleaved; j++) { - sumi1 = 0; - sumi2 = 0; - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); - sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]); - sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]); - sumi1 = sumi1 * scales_0[j]; - sumi2 = sumi2 * scales_1[j]; - sumi += sumi1 + sumi2; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; - } - } - for (int sb = 0; sb < 8; sb++) { - uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; - for (int j = 0; j < ncols_interleaved; j++) { - sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; - } - } - } - for (int j = 0; j < ncols_interleaved; j++) { - s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; - } - } + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); #endif } @@ -1735,38 +1594,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo } #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) - float sumf[4][8]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; - } - } - } + ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -3216,70 +3044,9 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } #else - - float sumf[4][8]; - float sum_minf[4][8]; - uint32_t utmp[32]; - int sumi1; - int sumi2; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumf[m][j] = 0.0; - sum_minf[m][j] = 0.0; - } - } - for (int l = 0; l < nb; l++) { - for (int sb = 0; sb < 8; sb++) { - memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); - utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); - const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; - utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); - utmp[sb * 4 + 2] = uaux_0; - utmp[sb * 4 + 0] &= kmask1; - } - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; - uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi1 = 0; - sumi2 = 0; - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); - sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]); - sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]); - sumi1 = sumi1 * scales_0[j]; - sumi2 = sumi2 * scales_1[j]; - sumi += sumi1 + sumi2; - } - sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; - } - } - } - for (int sb = 0; sb < 8; sb++) { - uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; - for(int m = 0; m < 4; m++) { - const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); - for(int j = 0; j < ncols_interleaved; j++) { - sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; - } - } - } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; - } - } - } - } + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); #endif }