Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-dwconv/up-sse.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <xmmintrin.h> |
| 13 | |
| 14 | #include <xnnpack/dwconv.h> |
| 15 | |
| 16 | |
| 17 | void xnn_f32_dwconv_ukernel_up8x9__sse( |
| 18 | size_t channels, |
| 19 | size_t output_width, |
| 20 | const float** input, |
| 21 | const float* weights, |
| 22 | float* output, |
| 23 | size_t input_stride, |
| 24 | size_t output_increment, |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame^] | 25 | const union xnn_f32_minmax_params params[restrict static 1]) |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 26 | { |
| 27 | assert(channels != 0); |
| 28 | assert(output_width != 0); |
| 29 | |
| 30 | const __m128 vmax = _mm_load_ps(params->sse.max); |
| 31 | const __m128 vmin = _mm_load_ps(params->sse.min); |
| 32 | do { |
| 33 | const float* i0 = input[0]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 34 | assert(i0 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 35 | const float* i1 = input[1]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 36 | assert(i1 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 37 | const float* i2 = input[2]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 38 | assert(i2 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 39 | const float* i3 = input[3]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 40 | assert(i3 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 41 | const float* i4 = input[4]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 42 | assert(i4 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 43 | const float* i5 = input[5]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 44 | assert(i5 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 45 | const float* i6 = input[6]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 46 | assert(i6 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 47 | const float* i7 = input[7]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 48 | assert(i7 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 49 | const float* i8 = input[8]; |
Marat Dukhan | 6866099 | 2020-02-03 13:31:12 -0800 | [diff] [blame] | 50 | assert(i8 != NULL); |
Marat Dukhan | 5098c3e | 2019-11-07 12:01:19 -0800 | [diff] [blame] | 51 | input = (const float**) ((uintptr_t) input + input_stride); |
| 52 | |
| 53 | size_t c = channels; |
| 54 | const float* w = weights; |
| 55 | for (; c >= 8; c -= 8) { |
| 56 | __m128 vacc0123p0 = _mm_load_ps(w); |
| 57 | __m128 vacc4567p0 = _mm_load_ps(w + 4); |
| 58 | |
| 59 | |
| 60 | const __m128 vi0x0123 = _mm_loadu_ps(i0); |
| 61 | const __m128 vi0x4567 = _mm_loadu_ps(i0 + 4); |
| 62 | i0 += 8; |
| 63 | |
| 64 | const __m128 vk0x0123 = _mm_load_ps(w + 8); |
| 65 | const __m128 vk0x4567 = _mm_load_ps(w + 12); |
| 66 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); |
| 67 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi0x4567, vk0x4567)); |
| 68 | |
| 69 | const __m128 vi1x0123 = _mm_loadu_ps(i1); |
| 70 | const __m128 vi1x4567 = _mm_loadu_ps(i1 + 4); |
| 71 | i1 += 8; |
| 72 | |
| 73 | const __m128 vk1x0123 = _mm_load_ps(w + 16); |
| 74 | const __m128 vk1x4567 = _mm_load_ps(w + 20); |
| 75 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); |
| 76 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi1x4567, vk1x4567)); |
| 77 | |
| 78 | const __m128 vi2x0123 = _mm_loadu_ps(i2); |
| 79 | const __m128 vi2x4567 = _mm_loadu_ps(i2 + 4); |
| 80 | i2 += 8; |
| 81 | |
| 82 | const __m128 vk2x0123 = _mm_load_ps(w + 24); |
| 83 | const __m128 vk2x4567 = _mm_load_ps(w + 28); |
| 84 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); |
| 85 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi2x4567, vk2x4567)); |
| 86 | |
| 87 | const __m128 vi3x0123 = _mm_loadu_ps(i3); |
| 88 | const __m128 vi3x4567 = _mm_loadu_ps(i3 + 4); |
| 89 | i3 += 8; |
| 90 | |
| 91 | const __m128 vk3x0123 = _mm_load_ps(w + 32); |
| 92 | const __m128 vk3x4567 = _mm_load_ps(w + 36); |
| 93 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); |
| 94 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi3x4567, vk3x4567)); |
| 95 | |
| 96 | const __m128 vi4x0123 = _mm_loadu_ps(i4); |
| 97 | const __m128 vi4x4567 = _mm_loadu_ps(i4 + 4); |
| 98 | i4 += 8; |
| 99 | |
| 100 | const __m128 vk4x0123 = _mm_load_ps(w + 40); |
| 101 | const __m128 vk4x4567 = _mm_load_ps(w + 44); |
| 102 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); |
| 103 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi4x4567, vk4x4567)); |
| 104 | |
| 105 | const __m128 vi5x0123 = _mm_loadu_ps(i5); |
| 106 | const __m128 vi5x4567 = _mm_loadu_ps(i5 + 4); |
| 107 | i5 += 8; |
| 108 | |
| 109 | const __m128 vk5x0123 = _mm_load_ps(w + 48); |
| 110 | const __m128 vk5x4567 = _mm_load_ps(w + 52); |
| 111 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); |
| 112 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi5x4567, vk5x4567)); |
| 113 | |
| 114 | const __m128 vi6x0123 = _mm_loadu_ps(i6); |
| 115 | const __m128 vi6x4567 = _mm_loadu_ps(i6 + 4); |
| 116 | i6 += 8; |
| 117 | |
| 118 | const __m128 vk6x0123 = _mm_load_ps(w + 56); |
| 119 | const __m128 vk6x4567 = _mm_load_ps(w + 60); |
| 120 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); |
| 121 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi6x4567, vk6x4567)); |
| 122 | |
| 123 | const __m128 vi7x0123 = _mm_loadu_ps(i7); |
| 124 | const __m128 vi7x4567 = _mm_loadu_ps(i7 + 4); |
| 125 | i7 += 8; |
| 126 | |
| 127 | const __m128 vk7x0123 = _mm_load_ps(w + 64); |
| 128 | const __m128 vk7x4567 = _mm_load_ps(w + 68); |
| 129 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); |
| 130 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi7x4567, vk7x4567)); |
| 131 | |
| 132 | const __m128 vi8x0123 = _mm_loadu_ps(i8); |
| 133 | const __m128 vi8x4567 = _mm_loadu_ps(i8 + 4); |
| 134 | i8 += 8; |
| 135 | |
| 136 | const __m128 vk8x0123 = _mm_load_ps(w + 72); |
| 137 | const __m128 vk8x4567 = _mm_load_ps(w + 76); |
| 138 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); |
| 139 | vacc4567p0 = _mm_add_ps(vacc4567p0, _mm_mul_ps(vi8x4567, vk8x4567)); |
| 140 | |
| 141 | w += 80; |
| 142 | |
| 143 | |
| 144 | __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); |
| 145 | __m128 vacc4567 = _mm_max_ps(vacc4567p0, vmin); |
| 146 | vacc0123 = _mm_min_ps(vacc0123, vmax); |
| 147 | vacc4567 = _mm_min_ps(vacc4567, vmax); |
| 148 | |
| 149 | _mm_storeu_ps(output, vacc0123); |
| 150 | _mm_storeu_ps(output + 4, vacc4567); |
| 151 | output += 8; |
| 152 | } |
| 153 | for (; c >= 4; c -= 4) { |
| 154 | __m128 vacc0123p0 = _mm_load_ps(w); |
| 155 | |
| 156 | const __m128 vi0x0123 = _mm_loadu_ps(i0); |
| 157 | i0 += 4; |
| 158 | |
| 159 | const __m128 vk0x0123 = _mm_load_ps(w + 8); |
| 160 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); |
| 161 | |
| 162 | const __m128 vi1x0123 = _mm_loadu_ps(i1); |
| 163 | i1 += 4; |
| 164 | |
| 165 | const __m128 vk1x0123 = _mm_load_ps(w + 16); |
| 166 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); |
| 167 | |
| 168 | const __m128 vi2x0123 = _mm_loadu_ps(i2); |
| 169 | i2 += 4; |
| 170 | |
| 171 | const __m128 vk2x0123 = _mm_load_ps(w + 24); |
| 172 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); |
| 173 | |
| 174 | const __m128 vi3x0123 = _mm_loadu_ps(i3); |
| 175 | i3 += 4; |
| 176 | |
| 177 | const __m128 vk3x0123 = _mm_load_ps(w + 32); |
| 178 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); |
| 179 | |
| 180 | const __m128 vi4x0123 = _mm_loadu_ps(i4); |
| 181 | i4 += 4; |
| 182 | |
| 183 | const __m128 vk4x0123 = _mm_load_ps(w + 40); |
| 184 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); |
| 185 | |
| 186 | const __m128 vi5x0123 = _mm_loadu_ps(i5); |
| 187 | i5 += 4; |
| 188 | |
| 189 | const __m128 vk5x0123 = _mm_load_ps(w + 48); |
| 190 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); |
| 191 | |
| 192 | const __m128 vi6x0123 = _mm_loadu_ps(i6); |
| 193 | i6 += 4; |
| 194 | |
| 195 | const __m128 vk6x0123 = _mm_load_ps(w + 56); |
| 196 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); |
| 197 | |
| 198 | const __m128 vi7x0123 = _mm_loadu_ps(i7); |
| 199 | i7 += 4; |
| 200 | |
| 201 | const __m128 vk7x0123 = _mm_load_ps(w + 64); |
| 202 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); |
| 203 | |
| 204 | const __m128 vi8x0123 = _mm_loadu_ps(i8); |
| 205 | i8 += 4; |
| 206 | |
| 207 | const __m128 vk8x0123 = _mm_load_ps(w + 72); |
| 208 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); |
| 209 | |
| 210 | w += 4; |
| 211 | |
| 212 | |
| 213 | __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); |
| 214 | vacc0123 = _mm_min_ps(vacc0123, vmax); |
| 215 | |
| 216 | _mm_storeu_ps(output, vacc0123); |
| 217 | output += 4; |
| 218 | } |
| 219 | if XNN_UNLIKELY(c != 0) { |
| 220 | __m128 vacc0123p0 = _mm_load_ps(w); |
| 221 | |
| 222 | const __m128 vi0x0123 = _mm_loadu_ps(i0); |
| 223 | const __m128 vk0x0123 = _mm_load_ps(w + 8); |
| 224 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi0x0123, vk0x0123)); |
| 225 | |
| 226 | const __m128 vi1x0123 = _mm_loadu_ps(i1); |
| 227 | const __m128 vk1x0123 = _mm_load_ps(w + 16); |
| 228 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi1x0123, vk1x0123)); |
| 229 | |
| 230 | const __m128 vi2x0123 = _mm_loadu_ps(i2); |
| 231 | const __m128 vk2x0123 = _mm_load_ps(w + 24); |
| 232 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi2x0123, vk2x0123)); |
| 233 | |
| 234 | const __m128 vi3x0123 = _mm_loadu_ps(i3); |
| 235 | const __m128 vk3x0123 = _mm_load_ps(w + 32); |
| 236 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi3x0123, vk3x0123)); |
| 237 | |
| 238 | const __m128 vi4x0123 = _mm_loadu_ps(i4); |
| 239 | const __m128 vk4x0123 = _mm_load_ps(w + 40); |
| 240 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi4x0123, vk4x0123)); |
| 241 | |
| 242 | const __m128 vi5x0123 = _mm_loadu_ps(i5); |
| 243 | const __m128 vk5x0123 = _mm_load_ps(w + 48); |
| 244 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi5x0123, vk5x0123)); |
| 245 | |
| 246 | const __m128 vi6x0123 = _mm_loadu_ps(i6); |
| 247 | const __m128 vk6x0123 = _mm_load_ps(w + 56); |
| 248 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi6x0123, vk6x0123)); |
| 249 | |
| 250 | const __m128 vi7x0123 = _mm_loadu_ps(i7); |
| 251 | const __m128 vk7x0123 = _mm_load_ps(w + 64); |
| 252 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); |
| 253 | |
| 254 | const __m128 vi8x0123 = _mm_loadu_ps(i8); |
| 255 | const __m128 vk8x0123 = _mm_load_ps(w + 72); |
| 256 | vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi8x0123, vk8x0123)); |
| 257 | |
| 258 | |
| 259 | __m128 vacc0123 = _mm_max_ps(vacc0123p0, vmin); |
| 260 | vacc0123 = _mm_min_ps(vacc0123, vmax); |
| 261 | |
| 262 | if (c & 2) { |
| 263 | _mm_storel_pi((__m64*) output, vacc0123); |
| 264 | vacc0123 = _mm_movehl_ps(vacc0123, vacc0123); |
| 265 | output += 2; |
| 266 | } |
| 267 | if (c & 1) { |
| 268 | _mm_store_ss(output, vacc0123); |
| 269 | output += 1; |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | output = (float*) ((uintptr_t) output + output_increment); |
| 274 | } while (--output_width != 0); |
| 275 | } |