1// Copyright 2020 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$assert ROW_TILE >= 1 7$assert ACCUMULATORS >= 1 8#include <assert.h> 9 10#include <tmmintrin.h> 11 12#include <xnnpack/dwconv.h> 13#include <xnnpack/math.h> 14 15 16void xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_${ROW_TILE}x4${"_acc%d" % ACCUMULATORS if ACCUMULATORS > 1 else ""}( 17 size_t input_height, 18 size_t input_width, 19 const float* input, 20 const float* weights, 21 const float* zero, 22 float* output, 23 uint32_t padding_top, 24 const union xnn_f32_chw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 25{ 26 assert(input_height != 0); 27 assert(input_width != 0); 28 assert(input_width % sizeof(float) == 0); 29 assert(padding_top == 1); 30 31 const __m128 vmask = _mm_load_ps((const float*) params->sse.mask); 32 const __m128 vmax = _mm_load_ps(params->sse.max); 33 const __m128 vmin = _mm_load_ps(params->sse.min); 34 35 const __m128 vbias = _mm_load1_ps(weights); 36 const __m128 vk00 = _mm_load1_ps(weights + 1); 37 const __m128 vk01 = _mm_load1_ps(weights + 2); 38 const __m128 vk02 = _mm_load1_ps(weights + 3); 39 const __m128 vk10 = _mm_load1_ps(weights + 4); 40 const __m128 vk11 = _mm_load1_ps(weights + 5); 41 const __m128 vk12 = _mm_load1_ps(weights + 6); 42 const __m128 vk20 = _mm_load1_ps(weights + 7); 43 const __m128 vk21 = _mm_load1_ps(weights + 8); 44 const __m128 vk22 = _mm_load1_ps(weights + 9); 45 46 const size_t input_decrement = round_up_po2(input_width, 4 * sizeof(float)); 47 48 const float* i0 = zero; 49 const float* i1 = input; 50 $for M in range(2, 2 + ROW_TILE): 51 const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_width); 52 53 float* o0 = output; 54 $for M in range(1, ROW_TILE): 55 float* o${M} = (float*) ((uintptr_t) o${M-1} + input_width); 56 57 size_t output_height = input_height; 58 do { 59 $for M in range(2, 2 + ROW_TILE): 60 if XNN_UNPREDICTABLE(output_height < ${M}) { 61 i${M} = zero; 62 $if M <= ROW_TILE: 63 o${M-1} = o${M-2}; 64 } 65 66 $for M in range(2 + ROW_TILE): 67 __m128 vi${M}x0123 = _mm_setzero_ps(); 68 69 $for M in range(2 + ROW_TILE): 70 __m128 vi${M}x4567 = _mm_loadu_ps(i${M}); 71 i${M} += 4; 72 73 size_t w = input_width; 74 for (; w > 4 * sizeof(float); w -= 4 * sizeof(float)) { 75 $for M in range(2 + ROW_TILE): 76 const __m128 vi${M}x89AB = _mm_loadu_ps(i${M}); 77 i${M} += 4; 78 79 $for K in range(3): 80 $for M in range(ROW_TILE): 81 $if K == 0: 82 __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${M+K}x4567, vk${K}1)); 83 $elif K < ACCUMULATORS: 84 __m128 vo${M}p${K} = _mm_mul_ps(vi${M+K}x4567, vk${K}1); 85 $else: 86 vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x4567, vk${K}1)); 87 88 $for M in range(2 + ROW_TILE): 89 const __m128 vi${M}x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi${M}x4567), _mm_castps_si128(vi${M}x0123), 12)); 90 91 $for K in range(3): 92 $for M in range(ROW_TILE): 93 $if K+3 < ACCUMULATORS: 94 __m128 vo${M}p${K+3} = _mm_mul_ps(vi${M+K}x3456, vk${K}0); 95 $else: 96 vo${M}p${(K+3) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+3) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x3456, vk${K}0)); 97 98 $for M in range(2 + ROW_TILE): 99 vi${M}x0123 = vi${M}x4567; 100 101 $for M in range(2 + ROW_TILE): 102 const __m128 vi${M}x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi${M}x89AB), _mm_castps_si128(vi${M}x4567), 4)); 103 104 $for K in range(3): 105 $for M in range(ROW_TILE): 106 vo${M}p${(K+6) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+6) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x5678, vk${K}2)); 107 108 $for M in range(2 + ROW_TILE): 109 vi${M}x4567 = vi${M}x89AB; 110 111 $if ACCUMULATORS > 1: 112 $ACC_SLICE = 1 113 $while ACC_SLICE < ACCUMULATORS: 114 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 115 $if A + ACC_SLICE < ACCUMULATORS: 116 $for M in range(ROW_TILE): 117 vo${M}p${A} = _mm_add_ps(vo${M}p${A}, vo${M}p${A + ACC_SLICE}); 118 $ACC_SLICE *= 2 119 120 $for M in range(ROW_TILE): 121 __m128 vo${M} = _mm_max_ps(vo${M}p0, vmin); 122 123 $for M in range(ROW_TILE): 124 vo${M} = _mm_min_ps(vo${M}, vmax); 125 126 $for M in reversed(range(ROW_TILE)): 127 _mm_storeu_ps(o${M}, vo${M}); 128 o${M} += 4; 129 } 130 // Always process the last block of 1..4 pixels. 131 assert(w >= 1 * sizeof(float)); 132 assert(w <= 4 * sizeof(float)); 133 { 134 $for M in range(2 + ROW_TILE): 135 vi${M}x4567 = _mm_and_ps(vmask, vi${M}x4567); 136 137 $for K in range(3): 138 $for M in range(ROW_TILE): 139 $if K == 0: 140 __m128 vo${M}p0 = _mm_add_ps(vbias, _mm_mul_ps(vi${M+K}x4567, vk${K}1)); 141 $elif K < ACCUMULATORS: 142 __m128 vo${M}p${K} = _mm_mul_ps(vi${M+K}x4567, vk${K}1); 143 $else: 144 vo${M}p${K % ACCUMULATORS} = _mm_add_ps(vo${M}p${K % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x4567, vk${K}1)); 145 146 $for M in range(2 + ROW_TILE): 147 const __m128 vi${M}x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi${M}x4567), _mm_castps_si128(vi${M}x0123), 12)); 148 149 $for K in range(3): 150 $for M in range(ROW_TILE): 151 $if K+3 < ACCUMULATORS: 152 __m128 vo${M}p${K+3} = _mm_mul_ps(vi${M+K}x3456, vk${K}0); 153 $else: 154 vo${M}p${(K+3) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+3) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x3456, vk${K}0)); 155 156 const __m128i vzero = _mm_setzero_si128(); 157 $for M in range(2 + ROW_TILE): 158 const __m128 vi${M}x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi${M}x4567), 4)); 159 160 $for K in range(3): 161 $for M in range(ROW_TILE): 162 vo${M}p${(K+6) % ACCUMULATORS} = _mm_add_ps(vo${M}p${(K+6) % ACCUMULATORS}, _mm_mul_ps(vi${M+K}x5678, vk${K}2)); 163 164 $if ACCUMULATORS > 1: 165 $ACC_SLICE = 1 166 $while ACC_SLICE < ACCUMULATORS: 167 $for A in range(0, ACCUMULATORS, ACC_SLICE * 2): 168 $if A + ACC_SLICE < ACCUMULATORS: 169 $for M in range(ROW_TILE): 170 vo${M}p${A} = _mm_add_ps(vo${M}p${A}, vo${M}p${A + ACC_SLICE}); 171 $ACC_SLICE *= 2 172 173 $for M in range(ROW_TILE): 174 __m128 vo${M} = _mm_max_ps(vo${M}p0, vmin); 175 176 $for M in range(ROW_TILE): 177 vo${M} = _mm_min_ps(vo${M}, vmax); 178 179 if XNN_LIKELY(w == 4 * sizeof(float)) { 180 $for M in reversed(range(ROW_TILE)): 181 _mm_storeu_ps(o${M}, vo${M}); 182 o${M} += 4; 183 } else { 184 if (w & (2 * sizeof(float))) { 185 $for M in reversed(range(ROW_TILE)): 186 _mm_storel_pi((__m64*) o${M}, vo${M}); 187 o${M} += 2; 188 189 $for M in range(ROW_TILE): 190 vo${M} = _mm_movehl_ps(vo${M}, vo${M}); 191 } 192 if (w & (1 * sizeof(float))) { 193 $for M in reversed(range(ROW_TILE)): 194 _mm_store_ss(o${M}, vo${M}); 195 o${M} += 1; 196 } 197 } 198 } 199 200 i0 = (const float*) ((uintptr_t) i${ROW_TILE} - input_decrement); 201 i1 = (const float*) ((uintptr_t) i${ROW_TILE+1} - input_decrement); 202 $for M in range(2, 2 + ROW_TILE): 203 i${M} = (const float*) ((uintptr_t) i${M-1} + input_width); 204 205 $if ROW_TILE > 1: 206 o0 = o${ROW_TILE - 1}; 207 $for M in range(1, ROW_TILE): 208 o${M} = (float*) ((uintptr_t) o${M-1} + input_width); 209 210 $if ROW_TILE > 1: 211 output_height = doz(output_height, ${ROW_TILE}); 212 } while (${"--" if ROW_TILE == 1 else ""}output_height != 0); 213} 214