diff --git a/Makefile b/Makefile index a163bde40..bda11791d 100644 --- a/Makefile +++ b/Makefile @@ -43,8 +43,11 @@ endif # keep standard at C11 and C++11 # -Ofast tends to produce faster code, but may not be available for some compilers. -#OPT = -Ofast +ifdef LLAMA_FAST +OPT = -Ofast +else OPT = -O3 +endif CFLAGS = -I. $(OPT) -std=c11 -fPIC CXXFLAGS = -I. -I./examples $(OPT) -std=c++11 -fPIC LDFLAGS = diff --git a/k_quants.c b/k_quants.c index 42fa9a402..32eea3660 100644 --- a/k_quants.c +++ b/k_quants.c @@ -1495,7 +1495,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri *s = sum; -#elif defined z__AVX2__ +#elif defined __AVX2__ const __m256i m3 = _mm256_set1_epi8(3); @@ -1584,6 +1584,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri } #endif +#if QK_K == 256 void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { assert(n % QK_K == 0); @@ -1884,6 +1885,242 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri } +#else + +void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + assert(n % QK_K == 0); + + const block_q3_K * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + +#ifdef __ARM_NEON + + uint32_t aux[3]; + uint32_t utmp[4]; + + const uint8x16_t m3b = vdupq_n_u8(0x3); +#ifdef __ARM_FEATURE_DOTPROD + const int32x4_t vzero = vdupq_n_s32(0); +#endif + + const uint8x16_t m0 = vdupq_n_u8(1); + const uint8x16_t m1 = vshlq_n_u8(m0, 1); + const uint8x16_t m2 = vshlq_n_u8(m0, 2); + const uint8x16_t m3 = vshlq_n_u8(m0, 3); + const int8_t m32 = 32; + + int8x16x4_t q3bytes; + + float sum = 0; + + for (int i = 0; i < nb; ++i) { + + const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + + const uint8_t * restrict q3 = x[i].qs; + const uint8_t * restrict qh = x[i].hmask; + const int8_t * restrict q8 = y[i].qs; + + uint8x16x2_t qhbits = vld1q_u8_x2(qh); + + uint8x16x4_t q3h; + + int32_t isum = 0; + + // Set up scales + memcpy(aux, x[i].scales, 12); + utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); + utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); + utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); + utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + + int8_t * scale = (int8_t *)utmp; + for (int j = 0; j < 16; ++j) scale[j] -= m32; + + for (int j = 0; j < QK_K/128; ++j) { + + const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32; + const int8x16x4_t q8bytes_1 = vld1q_s8_x4(q8); q8 += 64; + const int8x16x4_t q8bytes_2 = vld1q_s8_x4(q8); q8 += 64; + + q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2); + q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2); + q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1); + q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1); + + q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0])); + q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1])); + q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2])); + q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3])); + +#if defined(__ARM_FEATURE_DOTPROD) + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0]; + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1]; + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2]; + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3]; +#else + int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_1.val[0])), + vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_1.val[0]))); + int16x8_t p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_1.val[1])), + vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_1.val[1]))); + int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_1.val[2])), + vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_1.val[2]))); + int16x8_t p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_1.val[3])), + vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_1.val[3]))); + isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3]; +#endif + scale += 4; + + q3h.val[0] = vbicq_u8(m2, qhbits.val[0]); + q3h.val[1] = vbicq_u8(m2, qhbits.val[1]); + q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1); + q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1); + + q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0])); + q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1])); + q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2])); + q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3])); + +#if defined(__ARM_FEATURE_DOTPROD) + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0]; + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1]; + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2]; + isum += vaddvq_s32(vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3]; +#else + p0 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[0]), vget_low_s8 (q8bytes_2.val[0])), + vmull_s8(vget_high_s8(q3bytes.val[0]), vget_high_s8(q8bytes_2.val[0]))); + p1 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[1]), vget_low_s8 (q8bytes_2.val[1])), + vmull_s8(vget_high_s8(q3bytes.val[1]), vget_high_s8(q8bytes_2.val[1]))); + p2 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[2]), vget_low_s8 (q8bytes_2.val[2])), + vmull_s8(vget_high_s8(q3bytes.val[2]), vget_high_s8(q8bytes_2.val[2]))); + p3 = vaddq_s16(vmull_s8(vget_low_s8 (q3bytes.val[3]), vget_low_s8 (q8bytes_2.val[3])), + vmull_s8(vget_high_s8(q3bytes.val[3]), vget_high_s8(q8bytes_2.val[3]))); + isum += vaddvq_s16(p0) * scale[0] + vaddvq_s16(p1) * scale[1] + vaddvq_s16(p2) * scale[2] + vaddvq_s16(p3) * scale[3]; +#endif + scale += 4; + + if (j == 0) { + qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4); + qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4); + } + + } + sum += d * isum; + + } + + *s = sum; + +#elif defined __AVX2__ + + const __m256i m3 = _mm256_set1_epi8(3); + const __m256i m1 = _mm256_set1_epi8(1); + + __m256 acc = _mm256_setzero_ps(); + + uint64_t aux64; + + for (int i = 0; i < nb; ++i) { + + const float d = y[i].d * ggml_fp16_to_fp32(x[i].d); + + const uint8_t * restrict q3 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + const __m256i scale_0 = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0])); + const __m256i scale_1 = _mm256_set_m128i(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2])); + + // Set up scales + memcpy(&aux64, x[i].hmask, 8); + + __m256i q3h_0 = _mm256_set_m128i(_mm_set_epi64x(aux64 >> 3, aux64 >> 2), _mm_set_epi64x(aux64 >> 1, aux64 >> 0)); + __m256i q3h_1 = _mm256_set_m128i(_mm_set_epi64x(aux64 >> 7, aux64 >> 6), _mm_set_epi64x(aux64 >> 5, aux64 >> 4)); + q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2); + q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2); + + // load low 2 bits + const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3); + + // prepare low and high bits + const __m256i q3l_0 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q3bits, 2), q3bits), m3); + const __m256i q3l_1 = _mm256_and_si256(_mm256_set_m128i(_mm_srli_epi16(q3bits, 6), _mm_srli_epi16(q3bits, 4)), m3); + + // load Q8 quants + const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0)); + const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32)); + + // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16, + // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set, + // and 2 if the high bit was set) + const __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0); + const __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1); + + __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0); + __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1); + + p16_0 = _mm256_sub_epi16(p16_0, q8s_0); + p16_1 = _mm256_sub_epi16(p16_1, q8s_1); + + // multiply with scales + p16_0 = _mm256_madd_epi16(scale_0, p16_0); + p16_1 = _mm256_madd_epi16(scale_1, p16_1); + + p16_0 = _mm256_add_epi32(p16_0, p16_1); + + // multiply with block scale and accumulate + acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16_0), acc); + + } + + *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 * restrict q3 = x[i].qs; + const uint8_t * restrict hm = x[i].hmask; + const int8_t * restrict q8 = y[i].qs; + int8_t * restrict a = aux8; + for (int l = 0; l < 8; ++l) { + a[l+ 0] = (int8_t)((q3[l+0] >> 0) & 3) - (hm[l] & 0x01 ? 0 : 4); + a[l+ 8] = (int8_t)((q3[l+8] >> 0) & 3) - (hm[l] & 0x02 ? 0 : 4); + a[l+16] = (int8_t)((q3[l+0] >> 2) & 3) - (hm[l] & 0x04 ? 0 : 4); + a[l+24] = (int8_t)((q3[l+8] >> 2) & 3) - (hm[l] & 0x08 ? 0 : 4); + a[l+32] = (int8_t)((q3[l+0] >> 4) & 3) - (hm[l] & 0x10 ? 0 : 4); + a[l+40] = (int8_t)((q3[l+8] >> 4) & 3) - (hm[l] & 0x20 ? 0 : 4); + a[l+48] = (int8_t)((q3[l+0] >> 6) & 3) - (hm[l] & 0x40 ? 0 : 4); + a[l+56] = (int8_t)((q3[l+8] >> 6) & 3) - (hm[l] & 0x80 ? 0 : 4); + } + + memset(aux32, 0, 8*sizeof(int32_t)); + for (int j = 0; j < QK_K/16; ++j) { + for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; + q8 += 8; a += 8; + for (int l = 0; l < 8; ++l) aux16[l] += q8[l] * a[l]; + q8 += 8; a += 8; + int sc = x[i].scales[j]; + for (int l = 0; l < 8; ++l) aux32[l] += sc * aux16[l]; + } + const float d = ggml_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; + +#endif + +} +#endif + #if QK_K == 256 void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { assert(n % QK_K == 0);