#endif
}
-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) {
+//
+// GEMV/GEMM templates
+//
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+
+// GEMV for 8x blocks of 32 4-bit quants with a single scale factor per block
+template<typename block_tx8>
+static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
+ static_assert(
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>,
+ "Unsupported block type");
+
const int qk = QK8_0;
const int nb = n / qk;
- const int ncols_interleaved = 8;
- const int blocklen = 8;
-
- assert (n % qk == 0);
- assert (nc % ncols_interleaved == 0);
- UNUSED(s);
UNUSED(bs);
- UNUSED(vx);
- UNUSED(vy);
- UNUSED(nr);
- UNUSED(nc);
- UNUSED(nb);
- UNUSED(ncols_interleaved);
- UNUSED(blocklen);
-#if defined(__AVX2__)
- // Lookup table to convert signed nibbles to signed bytes
- __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
- signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
__m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
__m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
// Permute mask used for easier vector processing at later stages
const __m256i m4b = _mm256_set1_epi8(0x0F);
- int64_t b_nb = n / QK4_0;
+ int64_t b_nb = n / 32;
- const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+ const block_tx8 * b_ptr_start = (const block_tx8 *)vx;
const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy;
// Process Q8_0 blocks one by one
// Pointers to LHS blocks of block_q8_0 format
const block_q8_0 * a_ptr = a_ptr_start + (y * nb);
- // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+ // Take group of eight blocks at each pass of the loop and perform dot product operation
for (int64_t x = 0; x < nc / 8; x++) {
// Pointers to RHS blocks
- const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+ const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
// Master FP accumulator
__m256 acc_row = _mm256_setzero_ps();
for (int64_t b = 0; b < nb; b++) {
- // Load 8 blocks of Q4_0 interleaved as 8 bytes (B0 - B7)
+ // Load 8 blocks of 32 interleaved as 8 bytes (B0 - B7)
const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1);
const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2);
const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31)
const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31)
- // Load the scale values for the 8 blocks interleaved in block_q4_0x8
- const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+ // Load the scale values for the 8 blocks interleaved in block_tx8
+ __m256 col_scale_f32;
+ if constexpr (
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>) {
+ col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+ }
// Load and convert to FP32 scale from block_q8_0
const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
_mm256_storeu_ps(s + (y * nr + x * 8), acc_row);
}
}
- return;
-
-#endif
- 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) {
- const int qk = QK_K;
- const int nb = n / qk;
- const int ncols_interleaved = 8;
- const int blocklen = 8;
- static const uint32_t kmask1 = 0x3f3f3f3f;
- static const uint32_t kmask2 = 0x0f0f0f0f;
- static const uint32_t kmask3 = 0x03030303;
-
- assert (n % qk == 0);
- assert (nc % ncols_interleaved == 0);
+// GEMM for 8x blocks of 32 4-bit quants with a single scale factor per block
+template<typename block_tx8>
+static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
+ static_assert(
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>,
+ "Unsupported block type");
- UNUSED(s);
- UNUSED(bs);
- UNUSED(vx);
- UNUSED(vy);
- UNUSED(nr);
- UNUSED(nc);
- UNUSED(nb);
- UNUSED(ncols_interleaved);
- UNUSED(blocklen);
+ const int qk = QK8_0;
+ const int nb = n / qk;
-#if defined(__AVX2__)
- // Lookup table to convert signed nibbles to signed bytes
- __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
- signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
- // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
- __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
- __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
- // Permute mask used for easier vector processing at later stages
- __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+ const block_tx8 * b_ptr_start = (const block_tx8 *)vx;
+ const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
- // Mask to extract nibbles from bytes
+ int64_t b_nb = n / 32;
+ int64_t y = 0;
+ // Mask to mask out nibbles from packed bytes
const __m256i m4b = _mm256_set1_epi8(0x0F);
+ const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
+ // Permute mask used for easier vector processing at later stages
+ __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
+ int64_t xstart = 0;
+ int anr = nr - nr%16; // Used to align nr with boundary of 16
+#ifdef __AVX512F__
+ int anc = nc - nc%16; // Used to align nc with boundary of 16
+ // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+ const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
+ // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
+ __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
- int64_t b_nb = n / QK_K;
-
- const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
- const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
+ // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+ for (; y < anr / 4; y += 4) {
- // Process Q8_K blocks one by one
- for (int64_t y = 0; y < nr; y++) {
+ const block_q8_0x4 * a_ptrs[4];
- // Pointers to LHS blocks of block_q8_K format
- const block_q8_K * a_ptr = a_ptr_start + (y * nb);
+ a_ptrs[0] = a_ptr_start + (y * nb);
+ for (int i = 0; i < 3; ++i) {
+ a_ptrs[i + 1] = a_ptrs[i] + nb;
+ }
- // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
- for (int64_t x = 0; x < nc / 8; x++) {
+ // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
+ for (int64_t x = 0; x < anc / 8; x += 2) {
- // Pointers to RHS blocks
- const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+ const block_tx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+ const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
// Master FP accumulators
- __m256 acc_row = _mm256_setzero_ps();
- __m256 acc_min_rows = _mm256_setzero_ps();
+ __m512 acc_rows[16];
+ for (int i = 0; i < 16; i++) {
+ acc_rows[i] = _mm512_setzero_ps();
+ }
for (int64_t b = 0; b < nb; b++) {
+ // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+ const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+ const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+ const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+ const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+ const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+ const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+ const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+ const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+ // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
+ const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+ const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+ const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+ const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+ const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+ const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+ const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+ const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+ const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+ const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
- // Load and convert to FP32 scale from block_q8_K
- const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
-
- // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
- // col_scale_f32 rearranged so as to multiply with appropriate quants
- const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
- const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+ // 4-bit -> 8-bit - Sign is maintained
+ const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+ const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
- __m256i iacc_b = _mm256_setzero_si256();
- __m256i iacc_min_b = _mm256_setzero_si256();
+ const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+ const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
- const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
- __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
- q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
+ const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+ const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
- // Processes two sub blocks from each Q4_K in each iteration
- for (int sb = 0; sb < QK_K / 64; sb++) {
+ const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+ const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
- // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
- const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
- const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
- const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
- const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
- const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
- const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
- const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
- const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
+ // Shuffle pattern one - right side input
+ const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+ const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
- // 4-bit -> 8-bit
- // Values of the first sub block of eight block_q4_K structures for the sb loop
- const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
- const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
- const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
- const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
- const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
- const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
- const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
- const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
+ const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+ const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
- // Values of the second sub block of eight block_q4_K structures when sb = 1
- const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
- const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
- const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
- const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
- const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
- const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
- const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
- const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
+ const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+ const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
- uint32_t utmp_0[4], utmp_1[4];
+ const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+ const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
- // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
- // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
- memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
- utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
- const uint32_t uaux_0 = utmp_0[1] & kmask1;
- utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
- utmp_0[2] = uaux_0;
- utmp_0[0] &= kmask1;
+ // Shuffle pattern two - right side input
- // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
- memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
- utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
- const uint32_t uaux_1 = utmp_1[1] & kmask1;
- utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
- utmp_1[2] = uaux_1;
- utmp_1[0] &= kmask1;
+ const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+ const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
- // Scales of first sub block in the sb loop
- const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
- __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
- __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
+ const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+ const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
- // Scales of second sub block in the sb loop
- __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
- __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
- __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
+ const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+ const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
- // Mins of first and second sub block of Q4_K block are arranged side by side
- __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
+ const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+ const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
- // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
- __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
- __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
- __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
- __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
+ // Scale values - Load the weight scale values of two block_tx8
+ __m512 col_scale_f32;
+ if constexpr (
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>) {
+ col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+ }
- lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
- lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
- lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
- lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
+ // Process LHS in pairs of rows
+ for (int rp = 0; rp < 4; rp++) {
- // Dot product done within 32 bit lanes and accumulated in the same vector
- // First done for first sub block and thenn for second sub block in each sb
- // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
- // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
- // ...........................................................................
- // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
+ // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+ // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+ __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+ __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+ __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+ __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+ __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+ __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+ __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+ __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+ __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+ __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+ __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+ __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+ __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+ __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+ __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+ __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+ __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+ __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+ __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+ __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
- __m256i iacc_0 = _mm256_setzero_si256();
- __m256i iacc_1 = _mm256_setzero_si256();
+ // Shuffle pattern one - left side input
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
+ const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+ const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
+ const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+ const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
+ const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+ const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
+ const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+ const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
- iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
+ // Shuffle pattern two - left side input
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
+ const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+ const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
+ const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+ const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
+ const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+ const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
+ const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+ const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
- iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
+ // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+ // Resembles MMLAs into 2x2 matrices in ARM Version
+ const __m512i zero = _mm512_setzero_epi32();
+ __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
+ __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
+ __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
+ __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
+ __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
+ __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
+ __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
+ __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
- // Accumulate the iacc value for one sb
- __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+ __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+ __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+ __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
- // Broadcast the bsums of the two sub blocks of the iteration of Q8_K across the vector
- // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
- __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
- __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
- q8s = _mm256_bsrli_epi128(q8s, 4);
- // Accumulate for the complete block
- iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
- iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+ // Straighten out to make 4 row vectors
+ __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+ __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+ // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+ const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
+ const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+ // Multiply with appropiate scales and accumulate
+ acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+ acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+ acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+ acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
}
+ }
- // Multiply-Add with scale values for the complete super block
- acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
- acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+ // Store the accumulated values
+ for (int i = 0; i < 16; i++) {
+ _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+ }
+ }
+ }
+
+ // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+ for (; y < nr / 4; y ++) {
+ const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+ // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
+ for (int64_t x = 0; x < anc / 8; x += 2) {
+
+ const block_tx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+ const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+ // Master FP accumulators
+ __m512 acc_rows[4];
+ for (int i = 0; i < 4; i++) {
+ acc_rows[i] = _mm512_setzero_ps();
}
- // Accumulated output values permuted so as to be stored in appropriate order post accumulation
- acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
- _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+ for (int64_t b = 0; b < nb; b++) {
+ // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+ const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+ const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+ const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+ const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+ const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+ const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+ const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+ const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+ // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+ const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+ const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+ const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+ const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+ const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+ const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+ const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+ const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+ const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+ const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+ // 4-bit -> 8-bit - Sign is maintained
+ const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+ const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+
+ const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+ const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+
+ const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+ const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+
+ const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+ const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+
+ // Shuffle pattern one - right side input
+ const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+ const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+
+ const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+ const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+
+ const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+ const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+
+ const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+ const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+
+ // Shuffle pattern two - right side input
+
+ const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+ const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+
+ const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+ const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+
+ const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+ const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+
+ const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+ const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+
+
+ // Scale values - Load the weight scale values of two block_tx8
+ __m512 col_scale_f32;
+ if constexpr (
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>) {
+ col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+ }
+
+ // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+ // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+ __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+ __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+ __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+ __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+ __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+ __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+ __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+ __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+ __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+ __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+ __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+ __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+
+ __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+ __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+ __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+ __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+ __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+ __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+ __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+ __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+
+ // Shuffle pattern one - left side input
+
+ const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+ const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+ const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+ const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+ const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+ const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+ const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+ const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+ // Shuffle pattern two - left side input
+
+ const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+ const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+ const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+ const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+ const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+ const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+ const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+ const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+ // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+ // Resembles MMLAs into 2x2 matrices in ARM Version
+ const __m512i zero = _mm512_setzero_epi32();
+ __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
+ __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
+ __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
+ __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
+ __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
+ __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
+ __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
+ __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+ __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+ __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+ __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+ // Straighten out to make 4 row vectors
+ __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+ __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+ // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+ const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
+ const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+ // Multiply with appropiate scales and accumulate
+ acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+ acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+ acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+ acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+ }
+
+ // Store the accumulated values
+ for (int i = 0; i < 4; i++) {
+ _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+ }
}
}
+ if (anc != nc) {
+ xstart = anc/8;
+ y = 0;
+ }
+#endif // __AVX512F__
-#else
- UNUSED(kmask1);
- UNUSED(kmask2);
- UNUSED(kmask3);
- ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+ // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+
+ for (; y < anr / 4; y += 4) {
+ const block_q8_0x4 * a_ptrs[4];
+
+ a_ptrs[0] = a_ptr_start + (y * nb);
+ for (int i = 0; i < 3; ++i) {
+ a_ptrs[i + 1] = a_ptrs[i] + nb;
+ }
+
+ // Take group of eight block_tx8 structures at each pass of the loop and perform dot product operation
+ for (int64_t x = xstart; x < nc / 8; x++) {
+
+ const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+ // Master FP accumulators
+ __m256 acc_rows[16];
+ for (int i = 0; i < 16; i++) {
+ acc_rows[i] = _mm256_setzero_ps();
+ }
+
+ for (int64_t b = 0; b < nb; b++) {
+ // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+ const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+ const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+ const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+ const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+ // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+ const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+ const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+ // 4-bit -> 8-bit - Sign is maintained
+ const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+ const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+ const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+ const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+ const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+ const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+ const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+ const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+ // Shuffle pattern one - right side input
+ const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+ const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+ const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+ const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+ const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+ const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+ const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+ const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+ // Shuffle pattern two - right side input
+
+ const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+ const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+ const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+ const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+ const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+ const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+ const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+ const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+ // Scale values - Load the wight scale values of block_tx8
+ __m256 col_scale_f32;
+ if constexpr (
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>) {
+ col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+ }
+
+ // Process LHS in groups of four
+ for (int rp = 0; rp < 4; rp++) {
+ // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+ // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+ __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+ __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+ __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+ __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+ __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+ __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+ __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+ __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+ __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+ __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+ __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+ __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+ // Shuffle pattern one - left side input
+ const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+ const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+ const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+ const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+ const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+ const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+ const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+ const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+ // Shuffle pattern two - left side input
+ const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+ const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+ const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+ const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+ const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+ const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+ const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+ const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+ // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+ // Resembles MMLAs into 2x2 matrices in ARM Version
+ const __m256i zero = _mm256_setzero_si256();
+ __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
+ __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
+ __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
+ __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
+ __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
+ __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
+ __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
+ __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+ __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+ __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+ __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+ // Straighten out to make 4 row vectors
+ __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+ __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+ __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+ __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+ // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+ const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+
+ // Multiply with appropiate scales and accumulate
+ acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+ acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+ acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+ acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+ }
+ }
+
+ // Store the accumulated values
+ for (int i = 0; i < 16; i++) {
+ _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+ }
+ }
+ }
+
+ // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+ for (; y < nr / 4; y ++) {
+ const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+ // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+ for (int64_t x = xstart; x < nc / 8; x++) {
+ const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+ // Master FP accumulators
+ __m256 acc_rows[4];
+ for (int i = 0; i < 4; i++) {
+ acc_rows[i] = _mm256_setzero_ps();
+ }
+
+ for (int64_t b = 0; b < nb; b++) {
+ // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+ const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+ const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+ const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+ const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+ // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+ const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+ const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+ const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+ // 4-bit -> 8-bit - Sign is maintained
+ const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+ const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+ const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+ const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+ const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+ const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+ const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+ const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+ // Shuffle pattern one - right side input
+ const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+ const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+ const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+ const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+ const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+ const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+ const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+ const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+ // Shuffle pattern two - right side input
+
+ const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+ const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+ const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+ const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+ const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+ const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+ const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+ const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+ // Scale values - Load the wight scale values of block_tx8
+ __m256 col_scale_f32;
+ if constexpr (
+ std::is_same_v<block_tx8, block_q4_0x8> ||
+ std::is_same_v<block_tx8, block_iq4_nlx8>) {
+ col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+ }
+
+ // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+ // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+ __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+ __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+ __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+ __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+ __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+ __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+ __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+ __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+ __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+ __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+ __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+ __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+ // Shuffle pattern one - left side input
+
+ const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+ const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+ const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+ const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+ const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+ const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+ const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+ const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+ // Shuffle pattern two - left side input
+
+ const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+ const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+ const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+ const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+ const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+ const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+ const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+ const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+ // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+ // Resembles MMLAs into 2x2 matrices in ARM Version
+ const __m256i zero = _mm256_setzero_si256();
+ __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
+ __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
+ __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
+ __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
+ __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
+ __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
+ __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
+ __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+ __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+ __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+ __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+ // Straighten out to make 4 row vectors
+ __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+ __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+ __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+ __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+ // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+ const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+
+ // Multiply with appropiate scales and accumulate
+ acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+ acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+ acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+ acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+ }
+
+ // Store the accumulated values
+ for (int i = 0; i < 4; i++) {
+ _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+ }
+ }
+ }
+}
+
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+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) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+ {
+ // Lookup table to convert signed nibbles to signed bytes
+ __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+ signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+ gemv_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+ return;
+ }
#endif
+
+ ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
-void ggml_gemv_q2_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) {
+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) {
const int qk = QK_K;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
+ static const uint32_t kmask1 = 0x3f3f3f3f;
+ static const uint32_t kmask2 = 0x0f0f0f0f;
+ static const uint32_t kmask3 = 0x03030303;
assert (n % qk == 0);
assert (nc % ncols_interleaved == 0);
// Lookup table to convert signed nibbles to signed bytes
__m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
- // Shuffle masks to rearrange delta values to multiply with appropriate scales
+ // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
__m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+ __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
// Permute mask used for easier vector processing at later stages
__m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
- const __m256i m3b = _mm256_set1_epi8(3);
- const __m128i m4b_sse = _mm_set1_epi8(0xF);
-
- //Mask to get appropriate scales
- __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
- __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
+ // Mask to extract nibbles from bytes
+ const __m256i m4b = _mm256_set1_epi8(0x0F);
int64_t b_nb = n / QK_K;
- const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+ const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
// Process Q8_K blocks one by one
// Pointers to LHS blocks of block_q8_K format
const block_q8_K * a_ptr = a_ptr_start + (y * nb);
- // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
- for(int64_t x = 0; x < nc / 8; x++) {
+ // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
+ for (int64_t x = 0; x < nc / 8; x++) {
// Pointers to RHS blocks
- const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+ const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
// Master FP accumulators
__m256 acc_row = _mm256_setzero_ps();
for (int64_t b = 0; b < nb; b++) {
- // Load and convert to FP32 delta from block_q8_K
+ // Load and convert to FP32 scale from block_q8_K
const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
- // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+ // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
// col_scale_f32 rearranged so as to multiply with appropriate quants
const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
__m256i iacc_b = _mm256_setzero_si256();
__m256i iacc_min_b = _mm256_setzero_si256();
- // Processes eight sub blocks from each Q2_K in each iteration
- for(int sb = 0; sb < QK_K / 128; sb++) {
+ const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
+ __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
+ q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
- // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
- const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+ // Processes two sub blocks from each Q4_K in each iteration
+ for (int sb = 0; sb < QK_K / 64; sb++) {
+
+ // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+ const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
- // 2-bit -> 8-bit
- // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
- const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
- const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
- const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
- const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
-
- const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
- const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
- const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
- const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
-
- const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
- const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
- const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
- const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
-
- const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
- const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
- const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
- const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
-
- // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
- const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
- const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
- const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
- const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
-
- const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
- const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
- const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
- const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
-
- const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
- const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
- const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
- const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
-
- const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
- const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
- const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
- const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
-
- //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
- //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
-
- const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
- const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
- const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
- const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
-
- // Extract scales which is lower half from mins_and_scales
- const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
- const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
- const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
- const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
-
- // Extract mins which is upper half from mins_and_scales
- const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
- const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
- const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
- const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
-
- // Scales of sub blocks in the sb loop
- // Scales of the 0th sub block from each super block
- __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
- __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
-
- // Scales of the 1st sub block from each super block
- __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
- __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
-
- // Scales of the 2nd sub block from each super block
- __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
- __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
-
- // Scales of the 3rd sub block from each super block
- __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
- __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
-
- // Scales of the 4th sub block from each super block
- __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
- __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
-
- // Scales of the 5th sub block from each super block
- __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
- __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
-
- // Scales of the 6th sub block from each super block
- __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
- __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
-
- // Scales of the 7th sub block from each super block
- __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
- __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
-
- // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
- __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
- __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
- __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
- __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
- __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
- __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
- __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
- __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
-
- lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
- lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
- lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
- lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
- lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
- lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
- lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
- lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
-
- __m256i iacc_0 = _mm256_setzero_si256();
- __m256i iacc_1 = _mm256_setzero_si256();
- __m256i iacc_2 = _mm256_setzero_si256();
- __m256i iacc_3 = _mm256_setzero_si256();
- __m256i iacc_4 = _mm256_setzero_si256();
- __m256i iacc_5 = _mm256_setzero_si256();
- __m256i iacc_6 = _mm256_setzero_si256();
- __m256i iacc_7 = _mm256_setzero_si256();
-
- // Dot product done within 32 bit lanes and accumulated in the same vector
- // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop) // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
- // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
- // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
- // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
-
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
-
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
- iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
-
- iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
-
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
-
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
- iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
-
- iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
-
- iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
- iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
-
- iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
- iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
-
- iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
-
- iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
- iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
-
- iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
- iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
-
- iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
-
- iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
- iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
-
- iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
- iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
-
- iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
-
- iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
- iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
-
- iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
- iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
-
- iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
-
- iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
- iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
-
- iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
- iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
-
- iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
-
- iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
- iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
-
- iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
- iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
-
- iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
-
- // Accumulate the iacc value for one sb
- __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
-
- __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
- __m256i q8s = _mm256_castsi128_si256(q8sums);
- q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
-
- // Broadcast the bsums of the two corresponding subblocks of q8_k
- // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
- __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
- __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
- __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
- __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
-
- __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
-
- // Accumulate for the complete block
- iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
- iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
- }
-
- //Multiply-Add with scale values for complete super block
- acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
- acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
- }
- // Accumulated output values permuted so as to be stored in appropriate order post accumulation
- acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
- _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
- }
- }
-#else
-
- ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
-
-#endif
-}
-
-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) {
- const int qk = QK8_0;
- const int nb = n / qk;
- const int ncols_interleaved = 8;
- const int blocklen = 8;
-
- assert (n % qk == 0);
- assert (nr % 4 == 0);
- assert (nc % ncols_interleaved == 0);
-
- UNUSED(s);
- UNUSED(bs);
- UNUSED(vx);
- UNUSED(vy);
- UNUSED(nr);
- UNUSED(nc);
- UNUSED(nb);
- UNUSED(ncols_interleaved);
- UNUSED(blocklen);
-
-#if defined(__AVX2__) || defined(__AVX512F__)
- {
- const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
- const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
- int64_t b_nb = n / QK4_0;
- int64_t y = 0;
- // Mask to mask out nibbles from packed bytes
- const __m256i m4b = _mm256_set1_epi8(0x0F);
- const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
- // Lookup table to convert signed nibbles to signed bytes
- __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
- signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
- // Permute mask used for easier vector processing at later stages
- __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
- int64_t xstart = 0;
- int anr = nr - nr%16; // Used to align nr with boundary of 16
- #ifdef __AVX512F__
- int anc = nc - nc%16; // Used to align nc with boundary of 16
- // Mask to mask out nibbles from packed bytes expanded to 512 bit length
- const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
- // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
- __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
-
- // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
- for (; y < anr / 4; y += 4) {
-
- const block_q8_0x4 * a_ptrs[4];
-
- a_ptrs[0] = a_ptr_start + (y * nb);
- for (int i = 0; i < 3; ++i) {
- a_ptrs[i + 1] = a_ptrs[i] + nb;
- }
-
- // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
- for (int64_t x = 0; x < anc / 8; x += 2) {
-
- const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
- const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
-
- // Master FP accumulators
- __m512 acc_rows[16];
- for (int i = 0; i < 16; i++) {
- acc_rows[i] = _mm512_setzero_ps();
- }
-
- for (int64_t b = 0; b < nb; b++) {
- // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
- const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
- const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
- const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
- const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
-
- const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
- const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
- const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
- const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
-
- // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
-
- const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
- const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
-
- const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
- const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
- const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
- const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
-
- // 4-bit -> 8-bit - Sign is maintained
- const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
- const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
-
- const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
- const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
-
- const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
- const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
-
- const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
- const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
-
- // Shuffle pattern one - right side input
- const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
- const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
-
- const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
- const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
-
- const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
- const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
-
- const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
- const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
-
- // Shuffle pattern two - right side input
-
- const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
- const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
-
- const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
- const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
-
- const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
- const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
-
- const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
- const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
-
- // Scale values - Load the weight scale values of two block_q4_0x8
- const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
-
- // Process LHS in pairs of rows
- for (int rp = 0; rp < 4; rp++) {
-
- // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
- // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
- __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
- __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
- __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
- __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
- __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
- __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
- __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
- __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
- __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
- __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
- __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
- __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
-
- __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
- __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
- __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
- __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
- __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
- __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
- __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
- __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
-
- // Shuffle pattern one - left side input
-
- const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
- const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
-
- const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
- const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
-
- const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
- const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
-
- const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
- const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
-
- // Shuffle pattern two - left side input
-
- const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
- const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
-
- const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
- const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
-
- const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
- const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
-
- const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
- const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
-
- // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
- // Resembles MMLAs into 2x2 matrices in ARM Version
- const __m512i zero = _mm512_setzero_epi32();
- __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
- __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
- __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
- __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
- __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
- __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
- __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
- __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
-
- // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
- __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
- __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
- __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
- __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
-
-
- // Straighten out to make 4 row vectors
- __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
- __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
- __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
- __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
-
- // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
- const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
- const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
-
- // Multiply with appropiate scales and accumulate
- acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
- acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
- acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
- acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
- }
- }
-
- // Store the accumulated values
- for (int i = 0; i < 16; i++) {
- _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
- }
- }
- }
- // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
- for (; y < nr / 4; y ++) {
-
- const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
-
- // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
- for (int64_t x = 0; x < anc / 8; x += 2) {
-
- const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
- const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
-
- // Master FP accumulators
- __m512 acc_rows[4];
- for (int i = 0; i < 4; i++) {
- acc_rows[i] = _mm512_setzero_ps();
- }
-
- for (int64_t b = 0; b < nb; b++) {
- // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
- const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
- const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
- const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
- const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
-
- const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
- const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
- const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
- const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
-
- // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
-
- const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
- const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+ // 4-bit -> 8-bit
+ // Values of the first sub block of eight block_q4_K structures for the sb loop
+ const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
+ const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
+ const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
+ const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
+ const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
+ const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
+ const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
+ const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
- const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
- const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
- const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
- const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+ // Values of the second sub block of eight block_q4_K structures when sb = 1
+ const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
+ const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
+ const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
+ const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
+ const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
+ const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
+ const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
+ const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
- // 4-bit -> 8-bit - Sign is maintained
- const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
- const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+ uint32_t utmp_0[4], utmp_1[4];
- const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
- const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+ // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
+ // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
+ memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
+ utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux_0 = utmp_0[1] & kmask1;
+ utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
+ utmp_0[2] = uaux_0;
+ utmp_0[0] &= kmask1;
- const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
- const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+ // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
+ memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
+ utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux_1 = utmp_1[1] & kmask1;
+ utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
+ utmp_1[2] = uaux_1;
+ utmp_1[0] &= kmask1;
- const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
- const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+ // Scales of first sub block in the sb loop
+ const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
+ __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
+ __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
- // Shuffle pattern one - right side input
- const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
- const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+ // Scales of second sub block in the sb loop
+ __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
+ __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
+ __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
- const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
- const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+ // Mins of first and second sub block of Q4_K block are arranged side by side
+ __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
- const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
- const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+ // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+ __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
+ __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
+ __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
+ __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
- const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
- const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+ lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
+ lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
+ lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
+ lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
- // Shuffle pattern two - right side input
+ // Dot product done within 32 bit lanes and accumulated in the same vector
+ // First done for first sub block and thenn for second sub block in each sb
+ // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+ // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+ // ...........................................................................
+ // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
- const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
- const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
- const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
- const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+ __m256i iacc_0 = _mm256_setzero_si256();
+ __m256i iacc_1 = _mm256_setzero_si256();
- const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
- const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
- const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
- const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
- // Scale values - Load the weight scale values of two block_q4_0x8
- const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
- // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
- // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
- __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
- __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
- __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
- __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
- __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
- __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
- __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
- __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
- __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
- __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
- __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
- __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+ iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
- __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
- __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
- __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
- __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
- __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
- __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
- __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
- __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
- // Shuffle pattern one - left side input
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
- const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
- const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
- const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
- const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
- const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
- const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+ iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
- const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
- const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+ // Accumulate the iacc value for one sb
+ __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
- // Shuffle pattern two - left side input
+ // Broadcast the bsums of the two sub blocks of the iteration of Q8_K across the vector
+ // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
+ __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
+ __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
+ q8s = _mm256_bsrli_epi128(q8s, 4);
- const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
- const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+ // Accumulate for the complete block
+ iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+ iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+ }
- const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
- const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+ // Multiply-Add with scale values for the complete super block
+ acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+ acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
- const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
- const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+ }
- const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
- const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+ // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+ acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+ _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+ }
+ }
- // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
- // Resembles MMLAs into 2x2 matrices in ARM Version
- const __m512i zero = _mm512_setzero_epi32();
- __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
- __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
- __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
- __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
- __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
- __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
- __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
- __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
+#else
+ UNUSED(kmask1);
+ UNUSED(kmask2);
+ UNUSED(kmask3);
+ ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+#endif
+}
- // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
- __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
- __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
- __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
- __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+void ggml_gemv_iq4_nl_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) {
+#if defined(__AVX2__)
+ __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
+ signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+ gemv_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
- // Straighten out to make 4 row vectors
- __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
- __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
- __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
- __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+ return;
+#endif
- // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
- const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
- const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+ ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
+}
- // Multiply with appropiate scales and accumulate
- acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
- acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
- acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
- acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
- }
+void ggml_gemv_q2_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) {
+ const int qk = QK_K;
+ const int nb = n / qk;
+ const int ncols_interleaved = 8;
+ const int blocklen = 8;
- // Store the accumulated values
- for (int i = 0; i < 4; i++) {
- _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
- }
- }
- }
- if (anc != nc) {
- xstart = anc/8;
- y = 0;
- }
- #endif // __AVX512F__
+ assert (n % qk == 0);
+ assert (nc % ncols_interleaved == 0);
- // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+ UNUSED(s);
+ UNUSED(bs);
+ UNUSED(vx);
+ UNUSED(vy);
+ UNUSED(nr);
+ UNUSED(nc);
+ UNUSED(nb);
+ UNUSED(ncols_interleaved);
+ UNUSED(blocklen);
- for (; y < anr / 4; y += 4) {
- const block_q8_0x4 * a_ptrs[4];
+#if defined(__AVX2__)
+ // Lookup table to convert signed nibbles to signed bytes
+ __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+ signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+ // Shuffle masks to rearrange delta values to multiply with appropriate scales
+ __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+ // Permute mask used for easier vector processing at later stages
+ __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
- a_ptrs[0] = a_ptr_start + (y * nb);
- for (int i = 0; i < 3; ++i) {
- a_ptrs[i + 1] = a_ptrs[i] + nb;
- }
+ const __m256i m3b = _mm256_set1_epi8(3);
+ const __m128i m4b_sse = _mm_set1_epi8(0xF);
- // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
- for (int64_t x = xstart; x < nc / 8; x++) {
+ //Mask to get appropriate scales
+ __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
+ __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
- const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+ int64_t b_nb = n / QK_K;
- // Master FP accumulators
- __m256 acc_rows[16];
- for (int i = 0; i < 16; i++) {
- acc_rows[i] = _mm256_setzero_ps();
- }
+ const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+ const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
- for (int64_t b = 0; b < nb; b++) {
- // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
- const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
- const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
- const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
- const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+ // Process Q8_K blocks one by one
+ for (int64_t y = 0; y < nr; y++) {
- // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+ // Pointers to LHS blocks of block_q8_K format
+ const block_q8_K * a_ptr = a_ptr_start + (y * nb);
- // 4-bit -> 8-bit - Sign is maintained
- const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
- const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+ // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
+ for(int64_t x = 0; x < nc / 8; x++) {
- const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
- const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+ // Pointers to RHS blocks
+ const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
- const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
- const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+ // Master FP accumulators
+ __m256 acc_row = _mm256_setzero_ps();
+ __m256 acc_min_rows = _mm256_setzero_ps();
- const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
- const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+ for (int64_t b = 0; b < nb; b++) {
- // Shuffle pattern one - right side input
- const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
- const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+ // Load and convert to FP32 delta from block_q8_K
+ const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
- const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
- const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+ // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+ // col_scale_f32 rearranged so as to multiply with appropriate quants
+ const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
+ const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
- const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
- const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+ __m256i iacc_b = _mm256_setzero_si256();
+ __m256i iacc_min_b = _mm256_setzero_si256();
- const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
- const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+ // Processes eight sub blocks from each Q2_K in each iteration
+ for(int sb = 0; sb < QK_K / 128; sb++) {
- // Shuffle pattern two - right side input
+ // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+ const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+ const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
+ const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
+ const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
+ const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
+ const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
+ const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
+ const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
- const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
- const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+ // 2-bit -> 8-bit
+ // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
+ const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
+ const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
+ const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
+ const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
- const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
- const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+ const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
+ const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
+ const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
+ const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
- const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
- const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+ const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
+ const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
+ const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
+ const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
- const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
- const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+ const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
+ const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
+ const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
+ const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
- // Scale values - Load the wight scale values of block_q4_0x8
- const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+ // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
+ const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
+ const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
+ const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
+ const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
- // Process LHS in groups of four
- for (int rp = 0; rp < 4; rp++) {
- // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
- // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
- __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
- __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
- __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
- __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
- __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
- __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
- __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
- __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
- __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
- __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
- __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
- __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+ const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
+ const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
+ const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
+ const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
- // Shuffle pattern one - left side input
- const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
- const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+ const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
+ const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
+ const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
+ const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
- const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
- const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+ const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
+ const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
+ const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
+ const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
- const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
- const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+ //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+ //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
- const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
- const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+ const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+ const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
- // Shuffle pattern two - left side input
- const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
- const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+ // Extract scales which is lower half from mins_and_scales
+ const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+ const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+ const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+ const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
- const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
- const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+ // Extract mins which is upper half from mins_and_scales
+ const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+ const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+ const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+ const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
- const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
- const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+ // Scales of sub blocks in the sb loop
+ // Scales of the 0th sub block from each super block
+ __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
+ __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
- const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
- const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+ // Scales of the 1st sub block from each super block
+ __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
+ __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
- // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
- // Resembles MMLAs into 2x2 matrices in ARM Version
- const __m256i zero = _mm256_setzero_si256();
- __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
- __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
- __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
- __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
- __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
- __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
- __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
- __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+ // Scales of the 2nd sub block from each super block
+ __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
+ __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
- // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
- __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
- __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
- __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
- __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+ // Scales of the 3rd sub block from each super block
+ __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
+ __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
- // Straighten out to make 4 row vectors
- __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
- __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
- __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
- __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+ // Scales of the 4th sub block from each super block
+ __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
+ __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
- // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
- const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+ // Scales of the 5th sub block from each super block
+ __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
+ __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
- // Multiply with appropiate scales and accumulate
- acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
- acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
- acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
- acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
- }
- }
+ // Scales of the 6th sub block from each super block
+ __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
+ __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
- // Store the accumulated values
- for (int i = 0; i < 16; i++) {
- _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
- }
- }
- }
+ // Scales of the 7th sub block from each super block
+ __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
+ __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
- // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
- for (; y < nr / 4; y ++) {
+ // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+ __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
+ __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
+ __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
+ __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
+ __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
+ __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
+ __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
+ __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
- const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+ lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
+ lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
+ lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
+ lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
+ lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
+ lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
+ lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
+ lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
- // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
- for (int64_t x = xstart; x < nc / 8; x++) {
+ __m256i iacc_0 = _mm256_setzero_si256();
+ __m256i iacc_1 = _mm256_setzero_si256();
+ __m256i iacc_2 = _mm256_setzero_si256();
+ __m256i iacc_3 = _mm256_setzero_si256();
+ __m256i iacc_4 = _mm256_setzero_si256();
+ __m256i iacc_5 = _mm256_setzero_si256();
+ __m256i iacc_6 = _mm256_setzero_si256();
+ __m256i iacc_7 = _mm256_setzero_si256();
- const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+ // Dot product done within 32 bit lanes and accumulated in the same vector
+ // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop) // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+ // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+ // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
+ // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
- // Master FP accumulators
- __m256 acc_rows[4];
- for (int i = 0; i < 4; i++) {
- acc_rows[i] = _mm256_setzero_ps();
- }
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
- for (int64_t b = 0; b < nb; b++) {
- // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
- const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
- const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
- const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
- const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
- // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
- const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
- const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
- const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+ iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
- // 4-bit -> 8-bit - Sign is maintained
- const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
- const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
- const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
- const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
- const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
- const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+ iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
- const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
- const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
- // Shuffle pattern one - right side input
- const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
- const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
- const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
- const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+ iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
- const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
- const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
- const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
- const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
- // Shuffle pattern two - right side input
+ iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
- const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
- const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
- const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
- const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
- const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
- const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+ iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
- const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
- const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
- // Scale values - Load the wight scale values of block_q4_0x8
- const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
- // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
- // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
- __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
- __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
- __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
- __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
- __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
- __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
- __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
- __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
- __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
- __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
- __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
- __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+ iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
- // Shuffle pattern one - left side input
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
- const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
- const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
- const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
- const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+ iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
- const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
- const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
- const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
- const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
- // Shuffle pattern two - left side input
+ iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
- const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
- const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+ // Accumulate the iacc value for one sb
+ __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
- const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
- const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+ __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
+ __m256i q8s = _mm256_castsi128_si256(q8sums);
+ q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
- const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
- const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+ // Broadcast the bsums of the two corresponding subblocks of q8_k
+ // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
+ __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
+ __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
+ __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
+ __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
- const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
- const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+ __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
- // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
- // Resembles MMLAs into 2x2 matrices in ARM Version
- const __m256i zero = _mm256_setzero_si256();
- __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
- __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
- __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
- __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
- __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
- __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
- __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
- __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
+ // Accumulate for the complete block
+ iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+ iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+ }
- // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
- __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
- __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
- __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
- __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+ //Multiply-Add with scale values for complete super block
+ acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+ acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+ }
+ // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+ acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+ _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+ }
+ }
+#else
+ ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
- // Straighten out to make 4 row vectors
- __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
- __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
- __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
- __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+#endif
+}
- // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
- const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+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) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+ {
+ // Lookup table to convert signed nibbles to signed bytes
+ __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+ signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
- // Multiply with appropiate scales and accumulate
- acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
- acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
- acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
- acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
- }
+ gemm_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
- // Store the accumulated values
- for (int i = 0; i < 4; i++) {
- _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
- }
- }
- }
return;
}
+#endif // defined(__AVX2__) || defined(__AVX512F__)
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
}
#endif
}
+void ggml_gemm_iq4_nl_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) {
+#if defined(__AVX2__) || defined(__AVX512F__)
+ {
+ __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
+ signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+
+ gemm_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
+
+ return;
+ }
+#endif // defined(__AVX2__) || defined(__AVX512F__)
+
+ ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
void ggml_gemm_q2_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) {
const int qk = QK_K;
const int nb = n / qk;
overview / pkg / ggml / sources / llama.cpp / commitdiff