// Auto-generated file. Do not edit! // Template: src/x32-packw/s4-avx.c.in // Generator: tools/xngen // // Copyright 2023 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include #include #include "xnnpack/packw.h" #include "xnnpack/prefetch.h" void xnn_x32_packw_gemm_goi_ukernel_x16s4__avx_u4_prfm( size_t g, size_t nc, size_t kc, size_t nr, size_t kr, size_t sr, const uint32_t* weights, const uint32_t* bias, const void* scale, uint32_t* packed_weights, size_t extra_bytes, const void* params) { assert(g != 0); assert(nc != 0); assert(kc != 0); assert(nr == 16); assert(kr == 1); assert(sr == 4); assert(weights != NULL); assert(packed_weights != NULL); const float* b = (const float*) bias; float* packed_w = (float*) packed_weights; do { // NC main loop multiple of 16 const float* w0 = (const float*) weights; size_t n = nc; for (; n >= 16; n -= 16) { if XNN_LIKELY(b != NULL) { const __m256 vb0 = _mm256_loadu_ps(b); const __m256 vb8 = _mm256_loadu_ps(b + 8); _mm256_store_ps(packed_w, vb0); _mm256_store_ps(packed_w + 8, vb8); b += 16; } else { const __m256 vzero = _mm256_setzero_ps(); _mm256_store_ps(packed_w, vzero); _mm256_store_ps(packed_w + 8, vzero); } packed_w += 16; const float* w1 = w0 + kc; const float* w2 = w1 + kc; const float* w3 = w2 + kc; const float* w4 = w3 + kc; const float* w5 = w4 + kc; const float* w6 = w5 + kc; const float* w7 = w6 + kc; const float* w8 = w7 + kc; const float* w9 = w8 + kc; const float* w10 = w9 + kc; const float* w11 = w10 + kc; const float* w12 = w11 + kc; const float* w13 = w12 + kc; const float* w14 = w13 + kc; const float* w15 = w14 + kc; xnn_prefetch_to_l1((const int8_t*) w0); xnn_prefetch_to_l1((const int8_t*) w0 + 64); xnn_prefetch_to_l1((const int8_t*) w1); xnn_prefetch_to_l1((const int8_t*) w1 + 64); xnn_prefetch_to_l1((const int8_t*) w2); xnn_prefetch_to_l1((const int8_t*) w2 + 64); xnn_prefetch_to_l1((const int8_t*) w3); xnn_prefetch_to_l1((const int8_t*) w3 + 64); xnn_prefetch_to_l1((const int8_t*) w4); xnn_prefetch_to_l1((const int8_t*) w4 + 64); xnn_prefetch_to_l1((const int8_t*) w5); xnn_prefetch_to_l1((const int8_t*) w5 + 64); xnn_prefetch_to_l1((const int8_t*) w6); xnn_prefetch_to_l1((const int8_t*) w6 + 64); xnn_prefetch_to_l1((const int8_t*) w7); xnn_prefetch_to_l1((const int8_t*) w7 + 64); xnn_prefetch_to_l1((const int8_t*) w8); xnn_prefetch_to_l1((const int8_t*) w8 + 64); xnn_prefetch_to_l1((const int8_t*) w9); xnn_prefetch_to_l1((const int8_t*) w9 + 64); xnn_prefetch_to_l1((const int8_t*) w10); xnn_prefetch_to_l1((const int8_t*) w10 + 64); xnn_prefetch_to_l1((const int8_t*) w11); xnn_prefetch_to_l1((const int8_t*) w11 + 64); xnn_prefetch_to_l1((const int8_t*) w12); xnn_prefetch_to_l1((const int8_t*) w12 + 64); xnn_prefetch_to_l1((const int8_t*) w13); xnn_prefetch_to_l1((const int8_t*) w13 + 64); xnn_prefetch_to_l1((const int8_t*) w14); xnn_prefetch_to_l1((const int8_t*) w14 + 64); xnn_prefetch_to_l1((const int8_t*) w15); xnn_prefetch_to_l1((const int8_t*) w15 + 64); // KC main loop multiple of 4 size_t k = kc; for (; k >= 4; k -= 4) { // Read blocks of 4x4 // a b c d // e f g h // i j k l // m n o p // Load first 4 rows of N into low part of each register __m256 v0x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w0)); w0 += 4; __m256 v1x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w1)); w1 += 4; __m256 v2x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w2)); w2 += 4; __m256 v3x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w3)); w3 += 4; __m256 v8x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w8)); w8 += 4; __m256 v9x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w9)); w9 += 4; __m256 v10x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w10)); w10 += 4; __m256 v11x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w11)); w11 += 4; // Load next 4 rows of N into the high part of each register v0x0123 = _mm256_insertf128_ps(v0x0123, _mm_loadu_ps(w4), 1); w4 += 4; v1x0123 = _mm256_insertf128_ps(v1x0123, _mm_loadu_ps(w5), 1); w5 += 4; v2x0123 = _mm256_insertf128_ps(v2x0123, _mm_loadu_ps(w6), 1); w6 += 4; v3x0123 = _mm256_insertf128_ps(v3x0123, _mm_loadu_ps(w7), 1); w7 += 4; v8x0123 = _mm256_insertf128_ps(v8x0123, _mm_loadu_ps(w12), 1); w12 += 4; v9x0123 = _mm256_insertf128_ps(v9x0123, _mm_loadu_ps(w13), 1); w13 += 4; v10x0123 = _mm256_insertf128_ps(v10x0123, _mm_loadu_ps(w14), 1); w14 += 4; v11x0123 = _mm256_insertf128_ps(v11x0123, _mm_loadu_ps(w15), 1); w15 += 4; // Apply SR4 shuffle v1x0123 = _mm256_permute_ps(v1x0123, _MM_SHUFFLE(0, 3, 2, 1)); v2x0123 = _mm256_permute_ps(v2x0123, _MM_SHUFFLE(1, 0, 3, 2)); v3x0123 = _mm256_permute_ps(v3x0123, _MM_SHUFFLE(2, 1, 0, 3)); v9x0123 = _mm256_permute_ps(v9x0123, _MM_SHUFFLE(0, 3, 2, 1)); v10x0123 = _mm256_permute_ps(v10x0123, _MM_SHUFFLE(1, 0, 3, 2)); v11x0123 = _mm256_permute_ps(v11x0123, _MM_SHUFFLE(2, 1, 0, 3)); // Transpose 2x2 const __m256 vtmp0x0123 = _mm256_unpacklo_ps(v0x0123, v1x0123); // a e b f from row 0, 1 const __m256 vtmp1x0123 = _mm256_unpacklo_ps(v2x0123, v3x0123); // i m j n from row 2, 3 const __m256 vtmp2x0123 = _mm256_unpackhi_ps(v0x0123, v1x0123); // c g d h from row 0, 1 const __m256 vtmp3x0123 = _mm256_unpackhi_ps(v2x0123, v3x0123); // k o l p from row 2, 3 const __m256 vtmp8x0123 = _mm256_unpacklo_ps(v8x0123, v9x0123); // a e b f from row 0, 1 const __m256 vtmp9x0123 = _mm256_unpacklo_ps(v10x0123, v11x0123); // i m j n from row 2, 3 const __m256 vtmp10x0123 = _mm256_unpackhi_ps(v8x0123, v9x0123); // c g d h from row 0, 1 const __m256 vtmp11x0123 = _mm256_unpackhi_ps(v10x0123, v11x0123); // k o l p from row 2, 3 xnn_prefetch_to_l1((const int8_t*) w0 + 128); xnn_prefetch_to_l1((const int8_t*) w1 + 128); xnn_prefetch_to_l1((const int8_t*) w2 + 128); xnn_prefetch_to_l1((const int8_t*) w3 + 128); xnn_prefetch_to_l1((const int8_t*) w4 + 128); xnn_prefetch_to_l1((const int8_t*) w5 + 128); xnn_prefetch_to_l1((const int8_t*) w6 + 128); xnn_prefetch_to_l1((const int8_t*) w7 + 128); xnn_prefetch_to_l1((const int8_t*) w8 + 128); xnn_prefetch_to_l1((const int8_t*) w9 + 128); xnn_prefetch_to_l1((const int8_t*) w10 + 128); xnn_prefetch_to_l1((const int8_t*) w11 + 128); xnn_prefetch_to_l1((const int8_t*) w12 + 128); xnn_prefetch_to_l1((const int8_t*) w13 + 128); xnn_prefetch_to_l1((const int8_t*) w14 + 128); xnn_prefetch_to_l1((const int8_t*) w15 + 128); // Transpose 4x4 v0x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp0x0123), _mm256_castps_pd(vtmp1x0123))); // a e i m from row 0, 1 v1x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp0x0123), _mm256_castps_pd(vtmp1x0123))); // b f j n from row 0, 1 v2x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp2x0123), _mm256_castps_pd(vtmp3x0123))); // c g k o from row 2, 3 v3x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp2x0123), _mm256_castps_pd(vtmp3x0123))); // d h l p from row 2, 3 v8x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp8x0123), _mm256_castps_pd(vtmp9x0123))); // a e i m from row 0, 1 v9x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp8x0123), _mm256_castps_pd(vtmp9x0123))); // b f j n from row 0, 1 v10x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp10x0123), _mm256_castps_pd(vtmp11x0123))); // c g k o from row 2, 3 v11x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp10x0123), _mm256_castps_pd(vtmp11x0123))); // d h l p from row 2, 3 _mm256_store_ps(packed_w, v0x0123); _mm256_store_ps(packed_w + 8, v8x0123); _mm256_store_ps(packed_w + 16, v1x0123); _mm256_store_ps(packed_w + 24, v9x0123); _mm256_store_ps(packed_w + 32, v2x0123); _mm256_store_ps(packed_w + 40, v10x0123); _mm256_store_ps(packed_w + 48, v3x0123); _mm256_store_ps(packed_w + 56, v11x0123); packed_w += 64; } // KC remainder (1..3) if XNN_UNLIKELY(k != 0) { assert(k >= 1); assert(k <= 3); __m128 v0 = _mm_undefined_ps(); __m128 v1 = _mm_undefined_ps(); __m128 v2 = _mm_undefined_ps(); __m128 v3 = _mm_undefined_ps(); __m128 v4 = _mm_undefined_ps(); __m128 v5 = _mm_undefined_ps(); __m128 v6 = _mm_undefined_ps(); __m128 v7 = _mm_undefined_ps(); __m128 v8 = _mm_undefined_ps(); __m128 v9 = _mm_undefined_ps(); __m128 v10 = _mm_undefined_ps(); __m128 v11 = _mm_undefined_ps(); __m128 v12 = _mm_undefined_ps(); __m128 v13 = _mm_undefined_ps(); __m128 v14 = _mm_undefined_ps(); __m128 v15 = _mm_undefined_ps(); switch (k) { case 1: // Read blocks of 4x1 // a // e // i // m v0 = _mm_load_ss(w0); w0 += 1; v1 = _mm_load_ss(w1); w1 += 1; v2 = _mm_load_ss(w2); w2 += 1; v3 = _mm_load_ss(w3); w3 += 1; v4 = _mm_load_ss(w4); w4 += 1; v5 = _mm_load_ss(w5); w5 += 1; v6 = _mm_load_ss(w6); w6 += 1; v7 = _mm_load_ss(w7); w7 += 1; v8 = _mm_load_ss(w8); w8 += 1; v9 = _mm_load_ss(w9); w9 += 1; v10 = _mm_load_ss(w10); w10 += 1; v11 = _mm_load_ss(w11); w11 += 1; v12 = _mm_load_ss(w12); w12 += 1; v13 = _mm_load_ss(w13); w13 += 1; v14 = _mm_load_ss(w14); w14 += 1; v15 = _mm_load_ss(w15); w15 += 1; break; case 2: // Read blocks of 4x2 // a b // e f // i j // m n v0 = _mm_castpd_ps(_mm_load_sd((const double*) w0)); w0 += 2; v1 = _mm_castpd_ps(_mm_load_sd((const double*) w1)); w1 += 2; v2 = _mm_castpd_ps(_mm_load_sd((const double*) w2)); w2 += 2; v3 = _mm_castpd_ps(_mm_load_sd((const double*) w3)); w3 += 2; v4 = _mm_castpd_ps(_mm_load_sd((const double*) w4)); w4 += 2; v5 = _mm_castpd_ps(_mm_load_sd((const double*) w5)); w5 += 2; v6 = _mm_castpd_ps(_mm_load_sd((const double*) w6)); w6 += 2; v7 = _mm_castpd_ps(_mm_load_sd((const double*) w7)); w7 += 2; v8 = _mm_castpd_ps(_mm_load_sd((const double*) w8)); w8 += 2; v9 = _mm_castpd_ps(_mm_load_sd((const double*) w9)); w9 += 2; v10 = _mm_castpd_ps(_mm_load_sd((const double*) w10)); w10 += 2; v11 = _mm_castpd_ps(_mm_load_sd((const double*) w11)); w11 += 2; v12 = _mm_castpd_ps(_mm_load_sd((const double*) w12)); w12 += 2; v13 = _mm_castpd_ps(_mm_load_sd((const double*) w13)); w13 += 2; v14 = _mm_castpd_ps(_mm_load_sd((const double*) w14)); w14 += 2; v15 = _mm_castpd_ps(_mm_load_sd((const double*) w15)); w15 += 2; break; case 3: { // Read blocks of 4x3 // a b c // e f g // i j k // m n o const __m128 v0lo = _mm_castpd_ps(_mm_load_sd((const double*) w0)); const __m128 v0hi = _mm_load_ss(w0 + 2); v0 = _mm_movelh_ps(v0lo, v0hi); w0 += 3; const __m128 v1lo = _mm_castpd_ps(_mm_load_sd((const double*) w1)); const __m128 v1hi = _mm_load_ss(w1 + 2); v1 = _mm_movelh_ps(v1lo, v1hi); w1 += 3; const __m128 v2lo = _mm_castpd_ps(_mm_load_sd((const double*) w2)); const __m128 v2hi = _mm_load_ss(w2 + 2); v2 = _mm_movelh_ps(v2lo, v2hi); w2 += 3; const __m128 v3lo = _mm_castpd_ps(_mm_load_sd((const double*) w3)); const __m128 v3hi = _mm_load_ss(w3 + 2); v3 = _mm_movelh_ps(v3lo, v3hi); w3 += 3; const __m128 v4lo = _mm_castpd_ps(_mm_load_sd((const double*) w4)); const __m128 v4hi = _mm_load_ss(w4 + 2); v4 = _mm_movelh_ps(v4lo, v4hi); w4 += 3; const __m128 v5lo = _mm_castpd_ps(_mm_load_sd((const double*) w5)); const __m128 v5hi = _mm_load_ss(w5 + 2); v5 = _mm_movelh_ps(v5lo, v5hi); w5 += 3; const __m128 v6lo = _mm_castpd_ps(_mm_load_sd((const double*) w6)); const __m128 v6hi = _mm_load_ss(w6 + 2); v6 = _mm_movelh_ps(v6lo, v6hi); w6 += 3; const __m128 v7lo = _mm_castpd_ps(_mm_load_sd((const double*) w7)); const __m128 v7hi = _mm_load_ss(w7 + 2); v7 = _mm_movelh_ps(v7lo, v7hi); w7 += 3; const __m128 v8lo = _mm_castpd_ps(_mm_load_sd((const double*) w8)); const __m128 v8hi = _mm_load_ss(w8 + 2); v8 = _mm_movelh_ps(v8lo, v8hi); w8 += 3; const __m128 v9lo = _mm_castpd_ps(_mm_load_sd((const double*) w9)); const __m128 v9hi = _mm_load_ss(w9 + 2); v9 = _mm_movelh_ps(v9lo, v9hi); w9 += 3; const __m128 v10lo = _mm_castpd_ps(_mm_load_sd((const double*) w10)); const __m128 v10hi = _mm_load_ss(w10 + 2); v10 = _mm_movelh_ps(v10lo, v10hi); w10 += 3; const __m128 v11lo = _mm_castpd_ps(_mm_load_sd((const double*) w11)); const __m128 v11hi = _mm_load_ss(w11 + 2); v11 = _mm_movelh_ps(v11lo, v11hi); w11 += 3; const __m128 v12lo = _mm_castpd_ps(_mm_load_sd((const double*) w12)); const __m128 v12hi = _mm_load_ss(w12 + 2); v12 = _mm_movelh_ps(v12lo, v12hi); w12 += 3; const __m128 v13lo = _mm_castpd_ps(_mm_load_sd((const double*) w13)); const __m128 v13hi = _mm_load_ss(w13 + 2); v13 = _mm_movelh_ps(v13lo, v13hi); w13 += 3; const __m128 v14lo = _mm_castpd_ps(_mm_load_sd((const double*) w14)); const __m128 v14hi = _mm_load_ss(w14 + 2); v14 = _mm_movelh_ps(v14lo, v14hi); w14 += 3; const __m128 v15lo = _mm_castpd_ps(_mm_load_sd((const double*) w15)); const __m128 v15hi = _mm_load_ss(w15 + 2); v15 = _mm_movelh_ps(v15lo, v15hi); w15 += 3; break; } default: XNN_UNREACHABLE; } // Apply SR4 shuffle v1 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v1), _MM_SHUFFLE(0, 3, 2, 1))); v2 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v2), _MM_SHUFFLE(1, 0, 3, 2))); v3 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v3), _MM_SHUFFLE(2, 1, 0, 3))); v5 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v5), _MM_SHUFFLE(0, 3, 2, 1))); v6 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v6), _MM_SHUFFLE(1, 0, 3, 2))); v7 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v7), _MM_SHUFFLE(2, 1, 0, 3))); v9 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v9), _MM_SHUFFLE(0, 3, 2, 1))); v10 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v10), _MM_SHUFFLE(1, 0, 3, 2))); v11 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v11), _MM_SHUFFLE(2, 1, 0, 3))); v13 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v13), _MM_SHUFFLE(0, 3, 2, 1))); v14 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v14), _MM_SHUFFLE(1, 0, 3, 2))); v15 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v15), _MM_SHUFFLE(2, 1, 0, 3))); // Transpose 2x2 const __m128 vtmp0 = _mm_unpacklo_ps(v0, v1); // a e b f from row 0, 1 const __m128 vtmp1 = _mm_unpacklo_ps(v2, v3); // i m j n from row 2, 3 const __m128 vtmp2 = _mm_unpackhi_ps(v0, v1); // c g d h from row 0, 1 const __m128 vtmp3 = _mm_unpackhi_ps(v2, v3); // k o l p from row 2, 3 const __m128 vtmp4 = _mm_unpacklo_ps(v4, v5); // a e b f from row 0, 1 const __m128 vtmp5 = _mm_unpacklo_ps(v6, v7); // i m j n from row 2, 3 const __m128 vtmp6 = _mm_unpackhi_ps(v4, v5); // c g d h from row 0, 1 const __m128 vtmp7 = _mm_unpackhi_ps(v6, v7); // k o l p from row 2, 3 const __m128 vtmp8 = _mm_unpacklo_ps(v8, v9); // a e b f from row 0, 1 const __m128 vtmp9 = _mm_unpacklo_ps(v10, v11); // i m j n from row 2, 3 const __m128 vtmp10 = _mm_unpackhi_ps(v8, v9); // c g d h from row 0, 1 const __m128 vtmp11 = _mm_unpackhi_ps(v10, v11); // k o l p from row 2, 3 const __m128 vtmp12 = _mm_unpacklo_ps(v12, v13); // a e b f from row 0, 1 const __m128 vtmp13 = _mm_unpacklo_ps(v14, v15); // i m j n from row 2, 3 const __m128 vtmp14 = _mm_unpackhi_ps(v12, v13); // c g d h from row 0, 1 const __m128 vtmp15 = _mm_unpackhi_ps(v14, v15); // k o l p from row 2, 3 // Transpose 4x4 v0 = _mm_movelh_ps(vtmp0, vtmp1); // a e i m from row 0, 1 v1 = _mm_movehl_ps(vtmp1, vtmp0); // b f j n from row 0, 1 v2 = _mm_movelh_ps(vtmp2, vtmp3); // c g k o from row 2, 3 v3 = _mm_movehl_ps(vtmp3, vtmp2); // d h l p from row 2, 3 v4 = _mm_movelh_ps(vtmp4, vtmp5); // a e i m from row 0, 1 v5 = _mm_movehl_ps(vtmp5, vtmp4); // b f j n from row 0, 1 v6 = _mm_movelh_ps(vtmp6, vtmp7); // c g k o from row 2, 3 v7 = _mm_movehl_ps(vtmp7, vtmp6); // d h l p from row 2, 3 v8 = _mm_movelh_ps(vtmp8, vtmp9); // a e i m from row 0, 1 v9 = _mm_movehl_ps(vtmp9, vtmp8); // b f j n from row 0, 1 v10 = _mm_movelh_ps(vtmp10, vtmp11); // c g k o from row 2, 3 v11 = _mm_movehl_ps(vtmp11, vtmp10); // d h l p from row 2, 3 v12 = _mm_movelh_ps(vtmp12, vtmp13); // a e i m from row 0, 1 v13 = _mm_movehl_ps(vtmp13, vtmp12); // b f j n from row 0, 1 v14 = _mm_movelh_ps(vtmp14, vtmp15); // c g k o from row 2, 3 v15 = _mm_movehl_ps(vtmp15, vtmp14); // d h l p from row 2, 3 _mm_store_ps(packed_w, v0); _mm_store_ps(packed_w + 4, v4); _mm_store_ps(packed_w + 8, v8); _mm_store_ps(packed_w + 12, v12); _mm_store_ps(packed_w + 16, v1); _mm_store_ps(packed_w + 20, v5); _mm_store_ps(packed_w + 24, v9); _mm_store_ps(packed_w + 28, v13); _mm_store_ps(packed_w + 32, v2); _mm_store_ps(packed_w + 36, v6); _mm_store_ps(packed_w + 40, v10); _mm_store_ps(packed_w + 44, v14); _mm_store_ps(packed_w + 48, v3); _mm_store_ps(packed_w + 52, v7); _mm_store_ps(packed_w + 56, v11); _mm_store_ps(packed_w + 60, v15); packed_w += 64; } packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); w0 = w15; } // NC remainder (1..15) if XNN_UNLIKELY(n != 0) { assert(n >= 1); assert(n <= 15); if XNN_LIKELY(b != NULL) { size_t nb = n; do { *packed_w++ = *b++; } while (--nb != 0); packed_w += (16 - n); } else { const __m128 vzero = _mm_setzero_ps(); _mm_store_ps(packed_w, vzero); _mm_store_ps(packed_w + 4, vzero); _mm_store_ps(packed_w + 8, vzero); _mm_store_ps(packed_w + 12, vzero); packed_w += 16; } // NR remainder has less than 16 rows so last row is not loaded const float* w1 = w0 + kc; if XNN_UNPREDICTABLE(n < 2) { w1 = w0; } const float* w2 = w1 + kc; if XNN_UNPREDICTABLE(n <= 2) { w2 = w1; } const float* w3 = w2 + kc; if XNN_UNPREDICTABLE(n < 4) { w3 = w2; } const float* w4 = w3 + kc; if XNN_UNPREDICTABLE(n <= 4) { w4 = w3; } const float* w5 = w4 + kc; if XNN_UNPREDICTABLE(n < 6) { w5 = w4; } const float* w6 = w5 + kc; if XNN_UNPREDICTABLE(n <= 6) { w6 = w5; } const float* w7 = w6 + kc; if XNN_UNPREDICTABLE(n < 8) { w7 = w6; } const float* w8 = w7 + kc; if XNN_UNPREDICTABLE(n <= 8) { w8 = w7; } const float* w9 = w8 + kc; if XNN_UNPREDICTABLE(n < 10) { w9 = w8; } const float* w10 = w9 + kc; if XNN_UNPREDICTABLE(n <= 10) { w10 = w9; } const float* w11 = w10 + kc; if XNN_UNPREDICTABLE(n < 12) { w11 = w10; } const float* w12 = w11 + kc; if XNN_UNPREDICTABLE(n <= 12) { w12 = w11; } const float* w13 = w12 + kc; if XNN_UNPREDICTABLE(n < 14) { w13 = w12; } const float* w14 = w13 + kc; if XNN_UNPREDICTABLE(n <= 14) { w14 = w13; } // KC main loop multiple of 4 size_t k = kc; for (; k >= 4; k -= 4) { // Read blocks of 4x4 // a b c d // e f g h // i j k l // m n o p // Load first 4 rows of N into low part of each register __m256 v0x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w0)); w0 += 4; __m256 v1x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w1)); w1 += 4; __m256 v2x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w2)); w2 += 4; __m256 v3x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w3)); w3 += 4; __m256 v8x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w8)); w8 += 4; __m256 v9x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w9)); w9 += 4; __m256 v10x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w10)); w10 += 4; __m256 v11x0123 = _mm256_castps128_ps256(_mm_loadu_ps(w11)); w11 += 4; // Load next 4 rows of N into the high part of each register v0x0123 = _mm256_insertf128_ps(v0x0123, _mm_loadu_ps(w4), 1); w4 += 4; v1x0123 = _mm256_insertf128_ps(v1x0123, _mm_loadu_ps(w5), 1); w5 += 4; v2x0123 = _mm256_insertf128_ps(v2x0123, _mm_loadu_ps(w6), 1); w6 += 4; v3x0123 = _mm256_insertf128_ps(v3x0123, _mm_loadu_ps(w7), 1); w7 += 4; v8x0123 = _mm256_insertf128_ps(v8x0123, _mm_loadu_ps(w12), 1); w12 += 4; v9x0123 = _mm256_insertf128_ps(v9x0123, _mm_loadu_ps(w13), 1); w13 += 4; v10x0123 = _mm256_insertf128_ps(v10x0123, _mm_loadu_ps(w14), 1); w14 += 4; // Apply SR4 shuffle v1x0123 = _mm256_permute_ps(v1x0123, _MM_SHUFFLE(0, 3, 2, 1)); v2x0123 = _mm256_permute_ps(v2x0123, _MM_SHUFFLE(1, 0, 3, 2)); v3x0123 = _mm256_permute_ps(v3x0123, _MM_SHUFFLE(2, 1, 0, 3)); v9x0123 = _mm256_permute_ps(v9x0123, _MM_SHUFFLE(0, 3, 2, 1)); v10x0123 = _mm256_permute_ps(v10x0123, _MM_SHUFFLE(1, 0, 3, 2)); v11x0123 = _mm256_permute_ps(v11x0123, _MM_SHUFFLE(2, 1, 0, 3)); // Transpose 2x2 const __m256 vtmp0x0123 = _mm256_unpacklo_ps(v0x0123, v1x0123); // a e b f from row 0, 1 const __m256 vtmp1x0123 = _mm256_unpacklo_ps(v2x0123, v3x0123); // i m j n from row 2, 3 const __m256 vtmp2x0123 = _mm256_unpackhi_ps(v0x0123, v1x0123); // c g d h from row 0, 1 const __m256 vtmp3x0123 = _mm256_unpackhi_ps(v2x0123, v3x0123); // k o l p from row 2, 3 const __m256 vtmp8x0123 = _mm256_unpacklo_ps(v8x0123, v9x0123); // a e b f from row 0, 1 const __m256 vtmp9x0123 = _mm256_unpacklo_ps(v10x0123, v11x0123); // i m j n from row 2, 3 const __m256 vtmp10x0123 = _mm256_unpackhi_ps(v8x0123, v9x0123); // c g d h from row 0, 1 const __m256 vtmp11x0123 = _mm256_unpackhi_ps(v10x0123, v11x0123); // k o l p from row 2, 3 xnn_prefetch_to_l1((const int8_t*) w0 + 128); xnn_prefetch_to_l1((const int8_t*) w1 + 128); xnn_prefetch_to_l1((const int8_t*) w2 + 128); xnn_prefetch_to_l1((const int8_t*) w3 + 128); xnn_prefetch_to_l1((const int8_t*) w4 + 128); xnn_prefetch_to_l1((const int8_t*) w5 + 128); xnn_prefetch_to_l1((const int8_t*) w6 + 128); xnn_prefetch_to_l1((const int8_t*) w7 + 128); xnn_prefetch_to_l1((const int8_t*) w8 + 128); xnn_prefetch_to_l1((const int8_t*) w9 + 128); xnn_prefetch_to_l1((const int8_t*) w10 + 128); xnn_prefetch_to_l1((const int8_t*) w11 + 128); xnn_prefetch_to_l1((const int8_t*) w12 + 128); xnn_prefetch_to_l1((const int8_t*) w13 + 128); xnn_prefetch_to_l1((const int8_t*) w14 + 128); // Transpose 4x4 v0x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp0x0123), _mm256_castps_pd(vtmp1x0123))); // a e i m from row 0, 1 v1x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp0x0123), _mm256_castps_pd(vtmp1x0123))); // b f j n from row 0, 1 v2x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp2x0123), _mm256_castps_pd(vtmp3x0123))); // c g k o from row 2, 3 v3x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp2x0123), _mm256_castps_pd(vtmp3x0123))); // d h l p from row 2, 3 v8x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp8x0123), _mm256_castps_pd(vtmp9x0123))); // a e i m from row 0, 1 v9x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp8x0123), _mm256_castps_pd(vtmp9x0123))); // b f j n from row 0, 1 v10x0123 = _mm256_castpd_ps(_mm256_unpacklo_pd(_mm256_castps_pd(vtmp10x0123), _mm256_castps_pd(vtmp11x0123))); // c g k o from row 2, 3 v11x0123 = _mm256_castpd_ps(_mm256_unpackhi_pd(_mm256_castps_pd(vtmp10x0123), _mm256_castps_pd(vtmp11x0123))); // d h l p from row 2, 3 _mm256_store_ps(packed_w, v0x0123); _mm256_store_ps(packed_w + 8, v8x0123); _mm256_store_ps(packed_w + 16, v1x0123); _mm256_store_ps(packed_w + 24, v9x0123); _mm256_store_ps(packed_w + 32, v2x0123); _mm256_store_ps(packed_w + 40, v10x0123); _mm256_store_ps(packed_w + 48, v3x0123); _mm256_store_ps(packed_w + 56, v11x0123); packed_w += 64; } // KC remainder (1..3) if XNN_UNLIKELY(k != 0) { assert(k >= 1); assert(k <= 3); __m128 v0 = _mm_undefined_ps(); __m128 v1 = _mm_undefined_ps(); __m128 v2 = _mm_undefined_ps(); __m128 v3 = _mm_undefined_ps(); __m128 v4 = _mm_undefined_ps(); __m128 v5 = _mm_undefined_ps(); __m128 v6 = _mm_undefined_ps(); __m128 v7 = _mm_undefined_ps(); __m128 v8 = _mm_undefined_ps(); __m128 v9 = _mm_undefined_ps(); __m128 v10 = _mm_undefined_ps(); __m128 v11 = _mm_undefined_ps(); __m128 v12 = _mm_undefined_ps(); __m128 v13 = _mm_undefined_ps(); __m128 v14 = _mm_undefined_ps(); __m128 v15 = _mm_undefined_ps(); switch (k) { case 1: // Read blocks of 4x1 // a // e // i // m v0 = _mm_load_ss(w0); w0 += 1; v1 = _mm_load_ss(w1); w1 += 1; v2 = _mm_load_ss(w2); w2 += 1; v3 = _mm_load_ss(w3); w3 += 1; v4 = _mm_load_ss(w4); w4 += 1; v5 = _mm_load_ss(w5); w5 += 1; v6 = _mm_load_ss(w6); w6 += 1; v7 = _mm_load_ss(w7); w7 += 1; v8 = _mm_load_ss(w8); w8 += 1; v9 = _mm_load_ss(w9); w9 += 1; v10 = _mm_load_ss(w10); w10 += 1; v11 = _mm_load_ss(w11); w11 += 1; v12 = _mm_load_ss(w12); w12 += 1; v13 = _mm_load_ss(w13); w13 += 1; v14 = _mm_load_ss(w14); w14 += 1; break; case 2: // Read blocks of 4x2 // a b // e f // i j // m n v0 = _mm_castpd_ps(_mm_load_sd((const double*) w0)); w0 += 2; v1 = _mm_castpd_ps(_mm_load_sd((const double*) w1)); w1 += 2; v2 = _mm_castpd_ps(_mm_load_sd((const double*) w2)); w2 += 2; v3 = _mm_castpd_ps(_mm_load_sd((const double*) w3)); w3 += 2; v4 = _mm_castpd_ps(_mm_load_sd((const double*) w4)); w4 += 2; v5 = _mm_castpd_ps(_mm_load_sd((const double*) w5)); w5 += 2; v6 = _mm_castpd_ps(_mm_load_sd((const double*) w6)); w6 += 2; v7 = _mm_castpd_ps(_mm_load_sd((const double*) w7)); w7 += 2; v8 = _mm_castpd_ps(_mm_load_sd((const double*) w8)); w8 += 2; v9 = _mm_castpd_ps(_mm_load_sd((const double*) w9)); w9 += 2; v10 = _mm_castpd_ps(_mm_load_sd((const double*) w10)); w10 += 2; v11 = _mm_castpd_ps(_mm_load_sd((const double*) w11)); w11 += 2; v12 = _mm_castpd_ps(_mm_load_sd((const double*) w12)); w12 += 2; v13 = _mm_castpd_ps(_mm_load_sd((const double*) w13)); w13 += 2; v14 = _mm_castpd_ps(_mm_load_sd((const double*) w14)); w14 += 2; break; case 3: { // Read blocks of 4x3 // a b c // e f g // i j k // m n o const __m128 v0lo = _mm_castpd_ps(_mm_load_sd((const double*) w0)); const __m128 v0hi = _mm_load_ss(w0 + 2); v0 = _mm_movelh_ps(v0lo, v0hi); w0 += 3; const __m128 v1lo = _mm_castpd_ps(_mm_load_sd((const double*) w1)); const __m128 v1hi = _mm_load_ss(w1 + 2); v1 = _mm_movelh_ps(v1lo, v1hi); w1 += 3; const __m128 v2lo = _mm_castpd_ps(_mm_load_sd((const double*) w2)); const __m128 v2hi = _mm_load_ss(w2 + 2); v2 = _mm_movelh_ps(v2lo, v2hi); w2 += 3; const __m128 v3lo = _mm_castpd_ps(_mm_load_sd((const double*) w3)); const __m128 v3hi = _mm_load_ss(w3 + 2); v3 = _mm_movelh_ps(v3lo, v3hi); w3 += 3; const __m128 v4lo = _mm_castpd_ps(_mm_load_sd((const double*) w4)); const __m128 v4hi = _mm_load_ss(w4 + 2); v4 = _mm_movelh_ps(v4lo, v4hi); w4 += 3; const __m128 v5lo = _mm_castpd_ps(_mm_load_sd((const double*) w5)); const __m128 v5hi = _mm_load_ss(w5 + 2); v5 = _mm_movelh_ps(v5lo, v5hi); w5 += 3; const __m128 v6lo = _mm_castpd_ps(_mm_load_sd((const double*) w6)); const __m128 v6hi = _mm_load_ss(w6 + 2); v6 = _mm_movelh_ps(v6lo, v6hi); w6 += 3; const __m128 v7lo = _mm_castpd_ps(_mm_load_sd((const double*) w7)); const __m128 v7hi = _mm_load_ss(w7 + 2); v7 = _mm_movelh_ps(v7lo, v7hi); w7 += 3; const __m128 v8lo = _mm_castpd_ps(_mm_load_sd((const double*) w8)); const __m128 v8hi = _mm_load_ss(w8 + 2); v8 = _mm_movelh_ps(v8lo, v8hi); w8 += 3; const __m128 v9lo = _mm_castpd_ps(_mm_load_sd((const double*) w9)); const __m128 v9hi = _mm_load_ss(w9 + 2); v9 = _mm_movelh_ps(v9lo, v9hi); w9 += 3; const __m128 v10lo = _mm_castpd_ps(_mm_load_sd((const double*) w10)); const __m128 v10hi = _mm_load_ss(w10 + 2); v10 = _mm_movelh_ps(v10lo, v10hi); w10 += 3; const __m128 v11lo = _mm_castpd_ps(_mm_load_sd((const double*) w11)); const __m128 v11hi = _mm_load_ss(w11 + 2); v11 = _mm_movelh_ps(v11lo, v11hi); w11 += 3; const __m128 v12lo = _mm_castpd_ps(_mm_load_sd((const double*) w12)); const __m128 v12hi = _mm_load_ss(w12 + 2); v12 = _mm_movelh_ps(v12lo, v12hi); w12 += 3; const __m128 v13lo = _mm_castpd_ps(_mm_load_sd((const double*) w13)); const __m128 v13hi = _mm_load_ss(w13 + 2); v13 = _mm_movelh_ps(v13lo, v13hi); w13 += 3; const __m128 v14lo = _mm_castpd_ps(_mm_load_sd((const double*) w14)); const __m128 v14hi = _mm_load_ss(w14 + 2); v14 = _mm_movelh_ps(v14lo, v14hi); w14 += 3; break; } default: XNN_UNREACHABLE; } // Apply SR4 shuffle v1 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v1), _MM_SHUFFLE(0, 3, 2, 1))); v2 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v2), _MM_SHUFFLE(1, 0, 3, 2))); v3 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v3), _MM_SHUFFLE(2, 1, 0, 3))); v5 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v5), _MM_SHUFFLE(0, 3, 2, 1))); v6 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v6), _MM_SHUFFLE(1, 0, 3, 2))); v7 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v7), _MM_SHUFFLE(2, 1, 0, 3))); v9 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v9), _MM_SHUFFLE(0, 3, 2, 1))); v10 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v10), _MM_SHUFFLE(1, 0, 3, 2))); v11 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v11), _MM_SHUFFLE(2, 1, 0, 3))); v13 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v13), _MM_SHUFFLE(0, 3, 2, 1))); v14 = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(v14), _MM_SHUFFLE(1, 0, 3, 2))); // Transpose 2x2 const __m128 vtmp0 = _mm_unpacklo_ps(v0, v1); // a e b f from row 0, 1 const __m128 vtmp1 = _mm_unpacklo_ps(v2, v3); // i m j n from row 2, 3 const __m128 vtmp2 = _mm_unpackhi_ps(v0, v1); // c g d h from row 0, 1 const __m128 vtmp3 = _mm_unpackhi_ps(v2, v3); // k o l p from row 2, 3 const __m128 vtmp4 = _mm_unpacklo_ps(v4, v5); // a e b f from row 0, 1 const __m128 vtmp5 = _mm_unpacklo_ps(v6, v7); // i m j n from row 2, 3 const __m128 vtmp6 = _mm_unpackhi_ps(v4, v5); // c g d h from row 0, 1 const __m128 vtmp7 = _mm_unpackhi_ps(v6, v7); // k o l p from row 2, 3 const __m128 vtmp8 = _mm_unpacklo_ps(v8, v9); // a e b f from row 0, 1 const __m128 vtmp9 = _mm_unpacklo_ps(v10, v11); // i m j n from row 2, 3 const __m128 vtmp10 = _mm_unpackhi_ps(v8, v9); // c g d h from row 0, 1 const __m128 vtmp11 = _mm_unpackhi_ps(v10, v11); // k o l p from row 2, 3 const __m128 vtmp12 = _mm_unpacklo_ps(v12, v13); // a e b f from row 0, 1 const __m128 vtmp13 = _mm_unpacklo_ps(v14, v14); // i m j n from row 2, 3 const __m128 vtmp14 = _mm_unpackhi_ps(v12, v13); // c g d h from row 0, 1 const __m128 vtmp15 = _mm_unpackhi_ps(v14, v14); // k o l p from row 2, 3 // Transpose 4x4 v0 = _mm_movelh_ps(vtmp0, vtmp1); // a e i m from row 0, 1 v1 = _mm_movehl_ps(vtmp1, vtmp0); // b f j n from row 0, 1 v2 = _mm_movelh_ps(vtmp2, vtmp3); // c g k o from row 2, 3 v3 = _mm_movehl_ps(vtmp3, vtmp2); // d h l p from row 2, 3 v4 = _mm_movelh_ps(vtmp4, vtmp5); // a e i m from row 0, 1 v5 = _mm_movehl_ps(vtmp5, vtmp4); // b f j n from row 0, 1 v6 = _mm_movelh_ps(vtmp6, vtmp7); // c g k o from row 2, 3 v7 = _mm_movehl_ps(vtmp7, vtmp6); // d h l p from row 2, 3 v8 = _mm_movelh_ps(vtmp8, vtmp9); // a e i m from row 0, 1 v9 = _mm_movehl_ps(vtmp9, vtmp8); // b f j n from row 0, 1 v10 = _mm_movelh_ps(vtmp10, vtmp11); // c g k o from row 2, 3 v11 = _mm_movehl_ps(vtmp11, vtmp10); // d h l p from row 2, 3 v12 = _mm_movelh_ps(vtmp12, vtmp13); // a e i m from row 0, 1 v13 = _mm_movehl_ps(vtmp13, vtmp12); // b f j n from row 0, 1 v14 = _mm_movelh_ps(vtmp14, vtmp15); // c g k o from row 2, 3 v15 = _mm_movehl_ps(vtmp15, vtmp14); // d h l p from row 2, 3 _mm_store_ps(packed_w, v0); _mm_store_ps(packed_w + 4, v4); _mm_store_ps(packed_w + 8, v8); _mm_store_ps(packed_w + 12, v12); _mm_store_ps(packed_w + 16, v1); _mm_store_ps(packed_w + 20, v5); _mm_store_ps(packed_w + 24, v9); _mm_store_ps(packed_w + 28, v13); _mm_store_ps(packed_w + 32, v2); _mm_store_ps(packed_w + 36, v6); _mm_store_ps(packed_w + 40, v10); _mm_store_ps(packed_w + 44, v14); _mm_store_ps(packed_w + 48, v3); _mm_store_ps(packed_w + 52, v7); _mm_store_ps(packed_w + 56, v11); _mm_store_ps(packed_w + 60, v15); packed_w += 64; } packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); } weights += nc * kc; } while (--g != 0); }