// Auto-generated file. Do not edit! // Template: src/x32-packw/sse2.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_x8__sse2_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 == 8); assert(kr == 1); assert(sr == 1); 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 8 const float* w0 = (const float*) weights; size_t n = nc; for (; n >= 8; n -= 8) { if XNN_LIKELY(b != NULL) { const __m128 vb0123 = _mm_loadu_ps(b); const __m128 vb4567 = _mm_loadu_ps(b + 4); b += 8; _mm_store_ps(packed_w, vb0123); _mm_store_ps(packed_w + 4, vb4567); } else { const __m128 vzero = _mm_setzero_ps(); _mm_store_ps(packed_w, vzero); _mm_store_ps(packed_w + 4, vzero); } packed_w += 8; 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; 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); size_t k = kc; // KC multiple of 4 for (; k >= 4; k -= 4) { const __m128 v0x0123 = _mm_loadu_ps(w0); w0 += 4; const __m128 v1x0123 = _mm_loadu_ps(w1); w1 += 4; const __m128 v2x0123 = _mm_loadu_ps(w2); w2 += 4; const __m128 v3x0123 = _mm_loadu_ps(w3); w3 += 4; const __m128 v4x0123 = _mm_loadu_ps(w4); w4 += 4; const __m128 v5x0123 = _mm_loadu_ps(w5); w5 += 4; const __m128 v6x0123 = _mm_loadu_ps(w6); w6 += 4; const __m128 v7x0123 = _mm_loadu_ps(w7); w7 += 4; const __m128 v01x0_01x1 = _mm_unpacklo_ps(v0x0123, v1x0123); const __m128 v23x0_23x1 = _mm_unpacklo_ps(v2x0123, v3x0123); const __m128 v01x2_01x3 = _mm_unpackhi_ps(v0x0123, v1x0123); const __m128 v23x2_23x3 = _mm_unpackhi_ps(v2x0123, v3x0123); const __m128 v45x0_45x1 = _mm_unpacklo_ps(v4x0123, v5x0123); const __m128 v67x0_67x1 = _mm_unpacklo_ps(v6x0123, v7x0123); const __m128 v45x2_45x3 = _mm_unpackhi_ps(v4x0123, v5x0123); const __m128 v67x2_67x3 = _mm_unpackhi_ps(v6x0123, v7x0123); 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); const __m128 v0123x0 = _mm_movelh_ps(v01x0_01x1, v23x0_23x1); const __m128 v0123x1 = _mm_movehl_ps(v23x0_23x1, v01x0_01x1); const __m128 v0123x2 = _mm_movelh_ps(v01x2_01x3, v23x2_23x3); const __m128 v0123x3 = _mm_movehl_ps(v23x2_23x3, v01x2_01x3); const __m128 v4567x0 = _mm_movelh_ps(v45x0_45x1, v67x0_67x1); const __m128 v4567x1 = _mm_movehl_ps(v67x0_67x1, v45x0_45x1); const __m128 v4567x2 = _mm_movelh_ps(v45x2_45x3, v67x2_67x3); const __m128 v4567x3 = _mm_movehl_ps(v67x2_67x3, v45x2_45x3); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); _mm_store_ps(packed_w + 8, v0123x1); _mm_store_ps(packed_w + 12, v4567x1); _mm_store_ps(packed_w + 16, v0123x2); _mm_store_ps(packed_w + 20, v4567x2); _mm_store_ps(packed_w + 24, v0123x3); _mm_store_ps(packed_w + 28, v4567x3); packed_w += 32; } // KC remainder (1..3) if XNN_UNLIKELY(k != 0) { assert(k >= 1); assert(k <= 3); switch (k) { case 1: { const __m128 v0x0 = _mm_load_ss(w0); w0 += 1; const __m128 v1x0 = _mm_load_ss(w1); w1 += 1; const __m128 v2x0 = _mm_load_ss(w2); w2 += 1; const __m128 v3x0 = _mm_load_ss(w3); w3 += 1; const __m128 v4x0 = _mm_load_ss(w4); w4 += 1; const __m128 v5x0 = _mm_load_ss(w5); w5 += 1; const __m128 v6x0 = _mm_load_ss(w6); w6 += 1; const __m128 v7x0 = _mm_load_ss(w7); w7 += 1; const __m128 v01x0 = _mm_unpacklo_ps(v0x0, v1x0); const __m128 v23x0 = _mm_unpacklo_ps(v2x0, v3x0); const __m128 v45x0 = _mm_unpacklo_ps(v4x0, v5x0); const __m128 v67x0 = _mm_unpacklo_ps(v6x0, v7x0); const __m128 v0123x0 = _mm_movelh_ps(v01x0, v23x0); const __m128 v4567x0 = _mm_movelh_ps(v45x0, v67x0); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); packed_w += 8; break; } case 2: { const __m128 v0x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w0)); w0 += 2; const __m128 v1x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w1)); w1 += 2; const __m128 v2x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w2)); w2 += 2; const __m128 v3x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w3)); w3 += 2; const __m128 v4x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w4)); w4 += 2; const __m128 v5x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w5)); w5 += 2; const __m128 v6x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w6)); w6 += 2; const __m128 v7x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w7)); w7 += 2; const __m128 v01x0_01x1 = _mm_unpacklo_ps(v0x01, v1x01); const __m128 v23x0_23x1 = _mm_unpacklo_ps(v2x01, v3x01); const __m128 v45x0_45x1 = _mm_unpacklo_ps(v4x01, v5x01); const __m128 v67x0_67x1 = _mm_unpacklo_ps(v6x01, v7x01); const __m128 v0123x0 = _mm_movelh_ps(v01x0_01x1, v23x0_23x1); const __m128 v0123x1 = _mm_movehl_ps(v23x0_23x1, v01x0_01x1); const __m128 v4567x0 = _mm_movelh_ps(v45x0_45x1, v67x0_67x1); const __m128 v4567x1 = _mm_movehl_ps(v67x0_67x1, v45x0_45x1); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); _mm_store_ps(packed_w + 8, v0123x1); _mm_store_ps(packed_w + 12, v4567x1); packed_w += 16; break; } case 3: { __m128 v0x012 = _mm_load_ss(w0 + 2); __m128 v1x012 = _mm_load_ss(w1 + 2); __m128 v2x012 = _mm_load_ss(w2 + 2); __m128 v3x012 = _mm_load_ss(w3 + 2); __m128 v4x012 = _mm_load_ss(w4 + 2); __m128 v5x012 = _mm_load_ss(w5 + 2); __m128 v6x012 = _mm_load_ss(w6 + 2); __m128 v7x012 = _mm_load_ss(w7 + 2); v0x012 = _mm_movelh_ps(v0x012, v0x012); v1x012 = _mm_movelh_ps(v1x012, v1x012); v2x012 = _mm_movelh_ps(v2x012, v2x012); v3x012 = _mm_movelh_ps(v3x012, v3x012); v4x012 = _mm_movelh_ps(v4x012, v4x012); v5x012 = _mm_movelh_ps(v5x012, v5x012); v6x012 = _mm_movelh_ps(v6x012, v6x012); v7x012 = _mm_movelh_ps(v7x012, v7x012); v0x012 = _mm_loadl_pi(v0x012, (const __m64*) w0); w0 += 3; v1x012 = _mm_loadl_pi(v1x012, (const __m64*) w1); w1 += 3; v2x012 = _mm_loadl_pi(v2x012, (const __m64*) w2); w2 += 3; v3x012 = _mm_loadl_pi(v3x012, (const __m64*) w3); w3 += 3; v4x012 = _mm_loadl_pi(v4x012, (const __m64*) w4); w4 += 3; v5x012 = _mm_loadl_pi(v5x012, (const __m64*) w5); w5 += 3; v6x012 = _mm_loadl_pi(v6x012, (const __m64*) w6); w6 += 3; v7x012 = _mm_loadl_pi(v7x012, (const __m64*) w7); w7 += 3; const __m128 v01x0_01x1 = _mm_unpacklo_ps(v0x012, v1x012); const __m128 v23x0_23x1 = _mm_unpacklo_ps(v2x012, v3x012); const __m128 v01x2 = _mm_unpackhi_ps(v0x012, v1x012); const __m128 v23x2 = _mm_unpackhi_ps(v2x012, v3x012); const __m128 v45x0_45x1 = _mm_unpacklo_ps(v4x012, v5x012); const __m128 v67x0_67x1 = _mm_unpacklo_ps(v6x012, v7x012); const __m128 v45x2 = _mm_unpackhi_ps(v4x012, v5x012); const __m128 v67x2 = _mm_unpackhi_ps(v6x012, v7x012); const __m128 v0123x0 = _mm_movelh_ps(v01x0_01x1, v23x0_23x1); const __m128 v0123x1 = _mm_movehl_ps(v23x0_23x1, v01x0_01x1); const __m128 v0123x2 = _mm_movelh_ps(v01x2, v23x2); const __m128 v4567x0 = _mm_movelh_ps(v45x0_45x1, v67x0_67x1); const __m128 v4567x1 = _mm_movehl_ps(v67x0_67x1, v45x0_45x1); const __m128 v4567x2 = _mm_movelh_ps(v45x2, v67x2); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); _mm_store_ps(packed_w + 8, v0123x1); _mm_store_ps(packed_w + 12, v4567x1); _mm_store_ps(packed_w + 16, v0123x2); _mm_store_ps(packed_w + 20, v4567x2); packed_w += 24; break; } default: XNN_UNREACHABLE; } } packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); w0 = w7; } // NC remainder (1..7) if XNN_UNLIKELY(n != 0) { assert(n >= 1); assert(n <= 7); if XNN_LIKELY(b != NULL) { size_t nb = n; do { *packed_w++ = *b++; } while (--nb != 0); packed_w += (8 - n); } else { const __m128 vzero = _mm_setzero_ps(); _mm_store_ps(packed_w, vzero); _mm_store_ps(packed_w + 4, vzero); packed_w += 8; } 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; } size_t k = kc; // KC multiple of 4 for (; k >= 4; k -= 4) { const __m128 v0x0123 = _mm_loadu_ps(w0); w0 += 4; const __m128 v1x0123 = _mm_loadu_ps(w1); w1 += 4; const __m128 v2x0123 = _mm_loadu_ps(w2); w2 += 4; const __m128 v3x0123 = _mm_loadu_ps(w3); w3 += 4; const __m128 v4x0123 = _mm_loadu_ps(w4); w4 += 4; const __m128 v5x0123 = _mm_loadu_ps(w5); w5 += 4; const __m128 v6x0123 = _mm_loadu_ps(w6); w6 += 4; const __m128 v01x0_01x1 = _mm_unpacklo_ps(v0x0123, v1x0123); const __m128 v23x0_23x1 = _mm_unpacklo_ps(v2x0123, v3x0123); const __m128 v01x2_01x3 = _mm_unpackhi_ps(v0x0123, v1x0123); const __m128 v23x2_23x3 = _mm_unpackhi_ps(v2x0123, v3x0123); const __m128 v45x0_45x1 = _mm_unpacklo_ps(v4x0123, v5x0123); const __m128 v67x0_67x1 = _mm_unpacklo_ps(v6x0123, v6x0123); const __m128 v45x2_45x3 = _mm_unpackhi_ps(v4x0123, v5x0123); const __m128 v67x2_67x3 = _mm_unpackhi_ps(v6x0123, v6x0123); 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); const __m128 v0123x0 = _mm_movelh_ps(v01x0_01x1, v23x0_23x1); const __m128 v0123x1 = _mm_movehl_ps(v23x0_23x1, v01x0_01x1); const __m128 v0123x2 = _mm_movelh_ps(v01x2_01x3, v23x2_23x3); const __m128 v0123x3 = _mm_movehl_ps(v23x2_23x3, v01x2_01x3); const __m128 v4567x0 = _mm_movelh_ps(v45x0_45x1, v67x0_67x1); const __m128 v4567x1 = _mm_movehl_ps(v67x0_67x1, v45x0_45x1); const __m128 v4567x2 = _mm_movelh_ps(v45x2_45x3, v67x2_67x3); const __m128 v4567x3 = _mm_movehl_ps(v67x2_67x3, v45x2_45x3); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); _mm_store_ps(packed_w + 8, v0123x1); _mm_store_ps(packed_w + 12, v4567x1); _mm_store_ps(packed_w + 16, v0123x2); _mm_store_ps(packed_w + 20, v4567x2); _mm_store_ps(packed_w + 24, v0123x3); _mm_store_ps(packed_w + 28, v4567x3); packed_w += 32; } // KC remainder (1..3) if XNN_UNLIKELY(k != 0) { assert(k >= 1); assert(k <= 3); switch (k) { case 1: { const __m128 v0x0 = _mm_load_ss(w0); const __m128 v1x0 = _mm_load_ss(w1); const __m128 v2x0 = _mm_load_ss(w2); const __m128 v3x0 = _mm_load_ss(w3); const __m128 v4x0 = _mm_load_ss(w4); const __m128 v5x0 = _mm_load_ss(w5); const __m128 v6x0 = _mm_load_ss(w6); const __m128 v01x0 = _mm_unpacklo_ps(v0x0, v1x0); const __m128 v23x0 = _mm_unpacklo_ps(v2x0, v3x0); const __m128 v45x0 = _mm_unpacklo_ps(v4x0, v5x0); const __m128 v67x0 = _mm_unpacklo_ps(v6x0, v6x0); const __m128 v0123x0 = _mm_movelh_ps(v01x0, v23x0); const __m128 v4567x0 = _mm_movelh_ps(v45x0, v67x0); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); packed_w += 8; break; } case 2: { const __m128 v0x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w0)); const __m128 v1x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w1)); const __m128 v2x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w2)); const __m128 v3x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w3)); const __m128 v4x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w4)); const __m128 v5x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w5)); const __m128 v6x01 = _mm_castsi128_ps(_mm_loadl_epi64((const __m128i*) w6)); const __m128 v01x0_01x1 = _mm_unpacklo_ps(v0x01, v1x01); const __m128 v23x0_23x1 = _mm_unpacklo_ps(v2x01, v3x01); const __m128 v45x0_45x1 = _mm_unpacklo_ps(v4x01, v5x01); const __m128 v67x0_67x1 = _mm_unpacklo_ps(v6x01, v6x01); const __m128 v0123x0 = _mm_movelh_ps(v01x0_01x1, v23x0_23x1); const __m128 v0123x1 = _mm_movehl_ps(v23x0_23x1, v01x0_01x1); const __m128 v4567x0 = _mm_movelh_ps(v45x0_45x1, v67x0_67x1); const __m128 v4567x1 = _mm_movehl_ps(v67x0_67x1, v45x0_45x1); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); _mm_store_ps(packed_w + 8, v0123x1); _mm_store_ps(packed_w + 12, v4567x1); packed_w += 16; break; } case 3: { __m128 v0x012 = _mm_load_ss(w0 + 2); __m128 v1x012 = _mm_load_ss(w1 + 2); __m128 v2x012 = _mm_load_ss(w2 + 2); __m128 v3x012 = _mm_load_ss(w3 + 2); __m128 v4x012 = _mm_load_ss(w4 + 2); __m128 v5x012 = _mm_load_ss(w5 + 2); __m128 v6x012 = _mm_load_ss(w6 + 2); v0x012 = _mm_movelh_ps(v0x012, v0x012); v1x012 = _mm_movelh_ps(v1x012, v1x012); v2x012 = _mm_movelh_ps(v2x012, v2x012); v3x012 = _mm_movelh_ps(v3x012, v3x012); v4x012 = _mm_movelh_ps(v4x012, v4x012); v5x012 = _mm_movelh_ps(v5x012, v5x012); v6x012 = _mm_movelh_ps(v6x012, v6x012); v0x012 = _mm_loadl_pi(v0x012, (const __m64*) w0); v1x012 = _mm_loadl_pi(v1x012, (const __m64*) w1); v2x012 = _mm_loadl_pi(v2x012, (const __m64*) w2); v3x012 = _mm_loadl_pi(v3x012, (const __m64*) w3); v4x012 = _mm_loadl_pi(v4x012, (const __m64*) w4); v5x012 = _mm_loadl_pi(v5x012, (const __m64*) w5); v6x012 = _mm_loadl_pi(v6x012, (const __m64*) w6); const __m128 v01x0_01x1 = _mm_unpacklo_ps(v0x012, v1x012); const __m128 v23x0_23x1 = _mm_unpacklo_ps(v2x012, v3x012); const __m128 v01x2 = _mm_unpackhi_ps(v0x012, v1x012); const __m128 v23x2 = _mm_unpackhi_ps(v2x012, v3x012); const __m128 v45x0_45x1 = _mm_unpacklo_ps(v4x012, v5x012); const __m128 v67x0_67x1 = _mm_unpacklo_ps(v6x012, v6x012); const __m128 v45x2 = _mm_unpackhi_ps(v4x012, v5x012); const __m128 v67x2 = _mm_unpackhi_ps(v6x012, v6x012); const __m128 v0123x0 = _mm_movelh_ps(v01x0_01x1, v23x0_23x1); const __m128 v0123x1 = _mm_movehl_ps(v23x0_23x1, v01x0_01x1); const __m128 v0123x2 = _mm_movelh_ps(v01x2, v23x2); const __m128 v4567x0 = _mm_movelh_ps(v45x0_45x1, v67x0_67x1); const __m128 v4567x1 = _mm_movehl_ps(v67x0_67x1, v45x0_45x1); const __m128 v4567x2 = _mm_movelh_ps(v45x2, v67x2); _mm_store_ps(packed_w, v0123x0); _mm_store_ps(packed_w + 4, v4567x0); _mm_store_ps(packed_w + 8, v0123x1); _mm_store_ps(packed_w + 12, v4567x1); _mm_store_ps(packed_w + 16, v0123x2); _mm_store_ps(packed_w + 20, v4567x2); packed_w += 24; break; } default: XNN_UNREACHABLE; } } packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); } weights += nc * kc; } while (--g != 0); }