k_quants : add AVX support to dot functions with QK_K as 64 (#2339)

* add AVX to ggml_vec_dot_q2_K_q8_K()

* add AVX to ggml_vec_dot_q3_K_q8_K()

* add AVX to ggml_vec_dot_q4_K_q8_K()

* add AVX to ggml_vec_dot_q5_K_q8_K()

* add AVX to ggml_vec_dot_q6_K_q8_K()

* refactor AVX code in ggml_vec_dot_q6_K_q8_K()

1 files changed, 325 insertions, 0 deletions

diff --git a/k_quants.c b/k_quants.c
index e790abf..e792c57 100644
--- a/k_quants.c
+++ b/k_quants.c
@@ -1666,6 +1666,62 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) + summs;
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t ud, um;
+    const uint8_t * restrict db = (const uint8_t *)&ud;
+    const uint8_t * restrict mb = (const uint8_t *)&um;
+
+    float summs = 0;
+
+    // TODO: optimize this
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+        const float dmin = -y[i].d * ggml_fp16_to_fp32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
+        ud = (sc[0] >> 0) & 0x0f0f0f0f;
+        um = (sc[0] >> 4) & 0x0f0f0f0f;
+
+        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
+        summs += dmin * smin;
+
+        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
+        const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+        const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+        const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+        const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m256i p_0 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
+        const __m256i p_1 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
+        const __m256i p_2 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
+        const __m256i p_3 = _mm256_set_m128i(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc);
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
 #else
 
     float sumf = 0;
@@ -2295,6 +2351,93 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint64_t aux64;
+
+    uint16_t aux16[2];
+    const int8_t * aux8 = (const int8_t *)aux16;
+
+    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 uint16_t a = *(const uint16_t *)x[i].scales;
+        aux16[0] = a & 0x0f0f;
+        aux16[1] = (a >> 4) & 0x0f0f;
+
+        const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8);
+        const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8);
+        const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8);
+        const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8);
+
+        memcpy(&aux64, x[i].hmask, 8);
+
+        __m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
+        __m128i q3h_1 = _mm_srli_epi16(q3h_0, 2);
+        __m128i q3h_2 = _mm_srli_epi16(q3h_0, 4);
+        __m128i q3h_3 = _mm_srli_epi16(q3h_0, 6);
+        q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2);
+        q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2);
+        q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2);
+        q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2);
+
+        // load low 2 bits
+        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
+
+        // prepare low and high bits
+        const __m128i q3l_0 = _mm_and_si128(q3bits, m3);
+        const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3);
+        const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3);
+        const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), 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 _mm_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 __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1));
+
+        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+
+        // multiply with scales
+        p16_0 = _mm_madd_epi16(scale_0, p16_0);
+        p16_1 = _mm_madd_epi16(scale_1, p16_1);
+        p16_2 = _mm_madd_epi16(scale_2, p16_2);
+        p16_3 = _mm_madd_epi16(scale_3, p16_3);
+
+        p16_0 = _mm_add_epi32(p16_0, p16_2);
+        p16_1 = _mm_add_epi32(p16_1, p16_3);
+        __m256i p16 = _mm256_set_m128i(p16_1, p16_0);
+
+        // multiply with block scale and accumulate
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];
@@ -2781,6 +2924,60 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc) - summs;
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    uint16_t aux16[2];
+    const uint8_t * scales = (const uint8_t *)aux16;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = ggml_fp16_to_fp32(x[i].d[0]) * y[i].d;
+        const float m = ggml_fp16_to_fp32(x[i].d[1]) * y[i].d;
+        const __m256 vd = _mm256_set1_ps(d);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0);
+        const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1);
+        const __m128i q4_0 = _mm_and_si128(q4bits_0, m4);
+        const __m128i q4_1 = _mm_and_si128(q4bits_1, m4);
+        const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4);
+        const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
+        const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_1, p32_0))), acc);
+
+        const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
+        const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_set_m128i(p32_3, p32_2))), acc);
+
+    }
+
+    *s = hsum_float_8(acc) - summs;
+
 #else
 
     uint8_t aux8[QK_K];
@@ -3295,6 +3492,63 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mone  = _mm_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
+
+        const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]);
+        const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]);
+        const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]);
+        const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]);
+
+        int64_t aux64;
+        memcpy(&aux64, x[i].qh, 8);
+        const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64);
+        const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2);
+
+        const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4);
+        const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4);
+        const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4);
+        const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4);
+
+        const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4);
+        const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4);
+        const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4);
+        const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1)));
+        const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0)));
+        const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1)));
+        const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0)));
+        const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1)));
+
+        const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
+        const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_set_m128i(dot_1, dot_0))), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t aux8[QK_K];
@@ -3857,6 +4111,77 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
     *s = hsum_float_8(acc);
 
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(3);
+    const __m128i m32s = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * ggml_fp16_to_fp32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
+        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
+        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
+        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
+        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
+
+        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
+        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
+
+        const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4);
+        const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4);
+        const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4);
+        const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4);
+
+        const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0);
+        const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1);
+        const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2);
+        const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3);
+
+        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
+        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
+
+        __m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0));
+        __m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1));
+        __m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0));
+        __m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1));
+
+        __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
+        __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
+        __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
+        __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
+
+        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+
+        p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+        p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
+        p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+        p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
+
+        sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+        sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(_mm256_set_m128i(sumi_1, sumi_0))), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
 #else
 
     int8_t  aux8[QK_K];