Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 1 | // Copyright 2020 Google LLC |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 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 | #include <xnnpack.h> |
| 7 | |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 8 | #include <array> |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 9 | #include <algorithm> |
| 10 | #include <functional> |
| 11 | #include <iostream> |
| 12 | #include <limits> |
| 13 | #include <random> |
| 14 | |
Marat Dukhan | 66f3ccd | 2020-09-14 16:09:38 -0700 | [diff] [blame] | 15 | #include <fp16/fp16.h> |
| 16 | |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 17 | #include "models/models.h" |
| 18 | |
| 19 | namespace models { |
| 20 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 21 | ExecutionPlan FP16MobileNetV3Small(pthreadpool_t threadpool) { |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 22 | alignas(16) static std::array<uint16_t, 150528> v0; |
| 23 | alignas(16) static std::array<uint16_t, 200704> v1; |
| 24 | alignas(16) static std::array<uint16_t, 200704> v2; |
| 25 | alignas(16) static std::array<uint16_t, 50176> v3; |
| 26 | alignas(16) static std::array<uint16_t, 16> v4; |
| 27 | alignas(16) static std::array<uint16_t, 8> v5; |
| 28 | alignas(16) static std::array<uint16_t, 16> v6; |
| 29 | alignas(16) static std::array<uint16_t, 50176> v7; |
| 30 | alignas(16) static std::array<uint16_t, 50176> v8; |
| 31 | alignas(16) static std::array<uint16_t, 225792> v9; |
| 32 | alignas(16) static std::array<uint16_t, 56448> v10; |
| 33 | alignas(16) static std::array<uint16_t, 18816> v11; |
| 34 | alignas(16) static std::array<uint16_t, 68992> v12; |
| 35 | alignas(16) static std::array<uint16_t, 68992> v13; |
| 36 | alignas(16) static std::array<uint16_t, 18816> v14; |
| 37 | alignas(16) static std::array<uint16_t, 18816> v15; |
| 38 | alignas(16) static std::array<uint16_t, 75264> v16; |
| 39 | alignas(16) static std::array<uint16_t, 75264> v17; |
| 40 | alignas(16) static std::array<uint16_t, 18816> v18; |
| 41 | alignas(16) static std::array<uint16_t, 18816> v19; |
| 42 | alignas(16) static std::array<uint16_t, 96> v20; |
| 43 | alignas(16) static std::array<uint16_t, 24> v21; |
| 44 | alignas(16) static std::array<uint16_t, 96> v22; |
| 45 | alignas(16) static std::array<uint16_t, 18816> v23; |
| 46 | alignas(16) static std::array<uint16_t, 7840> v24; |
| 47 | alignas(16) static std::array<uint16_t, 47040> v25; |
| 48 | alignas(16) static std::array<uint16_t, 47040> v26; |
| 49 | alignas(16) static std::array<uint16_t, 47040> v27; |
| 50 | alignas(16) static std::array<uint16_t, 47040> v28; |
| 51 | alignas(16) static std::array<uint16_t, 240> v29; |
| 52 | alignas(16) static std::array<uint16_t, 64> v30; |
| 53 | alignas(16) static std::array<uint16_t, 240> v31; |
| 54 | alignas(16) static std::array<uint16_t, 47040> v32; |
| 55 | alignas(16) static std::array<uint16_t, 7840> v33; |
| 56 | alignas(16) static std::array<uint16_t, 7840> v34; |
| 57 | alignas(16) static std::array<uint16_t, 47040> v35; |
| 58 | alignas(16) static std::array<uint16_t, 47040> v36; |
| 59 | alignas(16) static std::array<uint16_t, 47040> v37; |
| 60 | alignas(16) static std::array<uint16_t, 47040> v38; |
| 61 | alignas(16) static std::array<uint16_t, 240> v39; |
| 62 | alignas(16) static std::array<uint16_t, 64> v40; |
| 63 | alignas(16) static std::array<uint16_t, 240> v41; |
| 64 | alignas(16) static std::array<uint16_t, 47040> v42; |
| 65 | alignas(16) static std::array<uint16_t, 7840> v43; |
| 66 | alignas(16) static std::array<uint16_t, 7840> v44; |
| 67 | alignas(16) static std::array<uint16_t, 23520> v45; |
| 68 | alignas(16) static std::array<uint16_t, 23520> v46; |
| 69 | alignas(16) static std::array<uint16_t, 23520> v47; |
| 70 | alignas(16) static std::array<uint16_t, 23520> v48; |
| 71 | alignas(16) static std::array<uint16_t, 120> v49; |
| 72 | alignas(16) static std::array<uint16_t, 32> v50; |
| 73 | alignas(16) static std::array<uint16_t, 120> v51; |
| 74 | alignas(16) static std::array<uint16_t, 23520> v52; |
| 75 | alignas(16) static std::array<uint16_t, 9408> v53; |
| 76 | alignas(16) static std::array<uint16_t, 28224> v54; |
| 77 | alignas(16) static std::array<uint16_t, 28224> v55; |
| 78 | alignas(16) static std::array<uint16_t, 28224> v56; |
| 79 | alignas(16) static std::array<uint16_t, 28224> v57; |
| 80 | alignas(16) static std::array<uint16_t, 144> v58; |
| 81 | alignas(16) static std::array<uint16_t, 40> v59; |
| 82 | alignas(16) static std::array<uint16_t, 144> v60; |
| 83 | alignas(16) static std::array<uint16_t, 28224> v61; |
| 84 | alignas(16) static std::array<uint16_t, 9408> v62; |
| 85 | alignas(16) static std::array<uint16_t, 9408> v63; |
| 86 | alignas(16) static std::array<uint16_t, 56448> v64; |
| 87 | alignas(16) static std::array<uint16_t, 56448> v65; |
| 88 | alignas(16) static std::array<uint16_t, 14112> v66; |
| 89 | alignas(16) static std::array<uint16_t, 14112> v67; |
| 90 | alignas(16) static std::array<uint16_t, 288> v68; |
| 91 | alignas(16) static std::array<uint16_t, 72> v69; |
| 92 | alignas(16) static std::array<uint16_t, 288> v70; |
| 93 | alignas(16) static std::array<uint16_t, 14112> v71; |
| 94 | alignas(16) static std::array<uint16_t, 4704> v72; |
| 95 | alignas(16) static std::array<uint16_t, 28224> v73; |
| 96 | alignas(16) static std::array<uint16_t, 28224> v74; |
| 97 | alignas(16) static std::array<uint16_t, 28224> v75; |
| 98 | alignas(16) static std::array<uint16_t, 28224> v76; |
| 99 | alignas(16) static std::array<uint16_t, 576> v77; |
| 100 | alignas(16) static std::array<uint16_t, 144> v78; |
| 101 | alignas(16) static std::array<uint16_t, 576> v79; |
| 102 | alignas(16) static std::array<uint16_t, 28224> v80; |
| 103 | alignas(16) static std::array<uint16_t, 4704> v81; |
| 104 | alignas(16) static std::array<uint16_t, 4704> v82; |
| 105 | alignas(16) static std::array<uint16_t, 28224> v83; |
| 106 | alignas(16) static std::array<uint16_t, 28224> v84; |
| 107 | alignas(16) static std::array<uint16_t, 28224> v85; |
| 108 | alignas(16) static std::array<uint16_t, 28224> v86; |
| 109 | alignas(16) static std::array<uint16_t, 576> v87; |
| 110 | alignas(16) static std::array<uint16_t, 144> v88; |
| 111 | alignas(16) static std::array<uint16_t, 576> v89; |
| 112 | alignas(16) static std::array<uint16_t, 28224> v90; |
| 113 | alignas(16) static std::array<uint16_t, 4704> v91; |
| 114 | alignas(16) static std::array<uint16_t, 4704> v92; |
| 115 | alignas(16) static std::array<uint16_t, 28224> v93; |
| 116 | alignas(16) static std::array<uint16_t, 28224> v94; |
| 117 | alignas(16) static std::array<uint16_t, 576> v95; |
| 118 | alignas(16) static std::array<uint16_t, 1024> v96; |
| 119 | alignas(16) static std::array<uint16_t, 1024> v97; |
| 120 | alignas(16) static std::array<uint16_t, 1024> v98; |
| 121 | alignas(16) static std::array<uint16_t, 1001> v99; |
| 122 | alignas(16) static std::array<uint16_t, 432> w100; |
| 123 | alignas(16) static std::array<uint16_t, 16> w101; |
| 124 | alignas(16) static std::array<uint16_t, 144> w102; |
| 125 | alignas(16) static std::array<uint16_t, 16> w103; |
| 126 | alignas(16) static std::array<uint16_t, 128> w104; |
| 127 | alignas(16) static std::array<uint16_t, 8> w105; |
| 128 | alignas(16) static std::array<uint16_t, 128> w106; |
| 129 | alignas(16) static std::array<uint16_t, 16> w107; |
| 130 | alignas(16) static std::array<uint16_t, 256> w108; |
| 131 | alignas(16) static std::array<uint16_t, 16> w109; |
| 132 | alignas(16) static std::array<uint16_t, 1152> w110; |
| 133 | alignas(16) static std::array<uint16_t, 72> w111; |
| 134 | alignas(16) static std::array<uint16_t, 648> w112; |
| 135 | alignas(16) static std::array<uint16_t, 72> w113; |
| 136 | alignas(16) static std::array<uint16_t, 1728> w114; |
| 137 | alignas(16) static std::array<uint16_t, 24> w115; |
| 138 | alignas(16) static std::array<uint16_t, 2112> w116; |
| 139 | alignas(16) static std::array<uint16_t, 88> w117; |
| 140 | alignas(16) static std::array<uint16_t, 792> w118; |
| 141 | alignas(16) static std::array<uint16_t, 88> w119; |
| 142 | alignas(16) static std::array<uint16_t, 2112> w120; |
| 143 | alignas(16) static std::array<uint16_t, 24> w121; |
| 144 | alignas(16) static std::array<uint16_t, 2304> w122; |
| 145 | alignas(16) static std::array<uint16_t, 96> w123; |
| 146 | alignas(16) static std::array<uint16_t, 2400> w124; |
| 147 | alignas(16) static std::array<uint16_t, 96> w125; |
| 148 | alignas(16) static std::array<uint16_t, 2304> w126; |
| 149 | alignas(16) static std::array<uint16_t, 24> w127; |
| 150 | alignas(16) static std::array<uint16_t, 2304> w128; |
| 151 | alignas(16) static std::array<uint16_t, 96> w129; |
| 152 | alignas(16) static std::array<uint16_t, 3840> w130; |
| 153 | alignas(16) static std::array<uint16_t, 40> w131; |
| 154 | alignas(16) static std::array<uint16_t, 9600> w132; |
| 155 | alignas(16) static std::array<uint16_t, 240> w133; |
| 156 | alignas(16) static std::array<uint16_t, 6000> w134; |
| 157 | alignas(16) static std::array<uint16_t, 240> w135; |
| 158 | alignas(16) static std::array<uint16_t, 15360> w136; |
| 159 | alignas(16) static std::array<uint16_t, 64> w137; |
| 160 | alignas(16) static std::array<uint16_t, 15360> w138; |
| 161 | alignas(16) static std::array<uint16_t, 240> w139; |
| 162 | alignas(16) static std::array<uint16_t, 9600> w140; |
| 163 | alignas(16) static std::array<uint16_t, 40> w141; |
| 164 | alignas(16) static std::array<uint16_t, 9600> w142; |
| 165 | alignas(16) static std::array<uint16_t, 240> w143; |
| 166 | alignas(16) static std::array<uint16_t, 6000> w144; |
| 167 | alignas(16) static std::array<uint16_t, 240> w145; |
| 168 | alignas(16) static std::array<uint16_t, 15360> w146; |
| 169 | alignas(16) static std::array<uint16_t, 64> w147; |
| 170 | alignas(16) static std::array<uint16_t, 15360> w148; |
| 171 | alignas(16) static std::array<uint16_t, 240> w149; |
| 172 | alignas(16) static std::array<uint16_t, 9600> w150; |
| 173 | alignas(16) static std::array<uint16_t, 40> w151; |
| 174 | alignas(16) static std::array<uint16_t, 4800> w152; |
| 175 | alignas(16) static std::array<uint16_t, 120> w153; |
| 176 | alignas(16) static std::array<uint16_t, 3000> w154; |
| 177 | alignas(16) static std::array<uint16_t, 120> w155; |
| 178 | alignas(16) static std::array<uint16_t, 3840> w156; |
| 179 | alignas(16) static std::array<uint16_t, 32> w157; |
| 180 | alignas(16) static std::array<uint16_t, 3840> w158; |
| 181 | alignas(16) static std::array<uint16_t, 120> w159; |
| 182 | alignas(16) static std::array<uint16_t, 5760> w160; |
| 183 | alignas(16) static std::array<uint16_t, 48> w161; |
| 184 | alignas(16) static std::array<uint16_t, 6912> w162; |
| 185 | alignas(16) static std::array<uint16_t, 144> w163; |
| 186 | alignas(16) static std::array<uint16_t, 3600> w164; |
| 187 | alignas(16) static std::array<uint16_t, 144> w165; |
| 188 | alignas(16) static std::array<uint16_t, 5760> w166; |
| 189 | alignas(16) static std::array<uint16_t, 40> w167; |
| 190 | alignas(16) static std::array<uint16_t, 5760> w168; |
| 191 | alignas(16) static std::array<uint16_t, 144> w169; |
| 192 | alignas(16) static std::array<uint16_t, 6912> w170; |
| 193 | alignas(16) static std::array<uint16_t, 48> w171; |
| 194 | alignas(16) static std::array<uint16_t, 13824> w172; |
| 195 | alignas(16) static std::array<uint16_t, 288> w173; |
| 196 | alignas(16) static std::array<uint16_t, 7200> w174; |
| 197 | alignas(16) static std::array<uint16_t, 288> w175; |
| 198 | alignas(16) static std::array<uint16_t, 20736> w176; |
| 199 | alignas(16) static std::array<uint16_t, 72> w177; |
| 200 | alignas(16) static std::array<uint16_t, 20736> w178; |
| 201 | alignas(16) static std::array<uint16_t, 288> w179; |
| 202 | alignas(16) static std::array<uint16_t, 27648> w180; |
| 203 | alignas(16) static std::array<uint16_t, 96> w181; |
| 204 | alignas(16) static std::array<uint16_t, 55296> w182; |
| 205 | alignas(16) static std::array<uint16_t, 576> w183; |
| 206 | alignas(16) static std::array<uint16_t, 14400> w184; |
| 207 | alignas(16) static std::array<uint16_t, 576> w185; |
| 208 | alignas(16) static std::array<uint16_t, 82944> w186; |
| 209 | alignas(16) static std::array<uint16_t, 144> w187; |
| 210 | alignas(16) static std::array<uint16_t, 82944> w188; |
| 211 | alignas(16) static std::array<uint16_t, 576> w189; |
| 212 | alignas(16) static std::array<uint16_t, 55296> w190; |
| 213 | alignas(16) static std::array<uint16_t, 96> w191; |
| 214 | alignas(16) static std::array<uint16_t, 55296> w192; |
| 215 | alignas(16) static std::array<uint16_t, 576> w193; |
| 216 | alignas(16) static std::array<uint16_t, 14400> w194; |
| 217 | alignas(16) static std::array<uint16_t, 576> w195; |
| 218 | alignas(16) static std::array<uint16_t, 82944> w196; |
| 219 | alignas(16) static std::array<uint16_t, 144> w197; |
| 220 | alignas(16) static std::array<uint16_t, 82944> w198; |
| 221 | alignas(16) static std::array<uint16_t, 576> w199; |
| 222 | alignas(16) static std::array<uint16_t, 55296> w200; |
| 223 | alignas(16) static std::array<uint16_t, 96> w201; |
| 224 | alignas(16) static std::array<uint16_t, 55296> w202; |
| 225 | alignas(16) static std::array<uint16_t, 576> w203; |
| 226 | alignas(16) static std::array<uint16_t, 589824> w204; |
| 227 | alignas(16) static std::array<uint16_t, 1024> w205; |
| 228 | alignas(16) static std::array<uint16_t, 1025024> w206; |
| 229 | alignas(16) static std::array<uint16_t, 1001> w207; |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 230 | |
| 231 | std::random_device random_device; |
| 232 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 43cb20e | 2020-06-25 20:18:54 -0700 | [diff] [blame] | 233 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-1.0f, +1.0f), std::ref(rng)); |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 234 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 235 | std::generate(v0.begin(), v0.end(), std::ref(f16rng)); |
| 236 | std::generate(v1.begin(), v1.end(), std::ref(f16rng)); |
| 237 | std::generate(v2.begin(), v2.end(), std::ref(f16rng)); |
| 238 | std::generate(v3.begin(), v3.end(), std::ref(f16rng)); |
| 239 | std::generate(v4.begin(), v4.end(), std::ref(f16rng)); |
| 240 | std::generate(v5.begin(), v5.end(), std::ref(f16rng)); |
| 241 | std::generate(v6.begin(), v6.end(), std::ref(f16rng)); |
| 242 | std::generate(v7.begin(), v7.end(), std::ref(f16rng)); |
| 243 | std::generate(v8.begin(), v8.end(), std::ref(f16rng)); |
| 244 | std::generate(v9.begin(), v9.end(), std::ref(f16rng)); |
| 245 | std::generate(v10.begin(), v10.end(), std::ref(f16rng)); |
| 246 | std::generate(v11.begin(), v11.end(), std::ref(f16rng)); |
| 247 | std::generate(v12.begin(), v12.end(), std::ref(f16rng)); |
| 248 | std::generate(v13.begin(), v13.end(), std::ref(f16rng)); |
| 249 | std::generate(v14.begin(), v14.end(), std::ref(f16rng)); |
| 250 | std::generate(v15.begin(), v15.end(), std::ref(f16rng)); |
| 251 | std::generate(v16.begin(), v16.end(), std::ref(f16rng)); |
| 252 | std::generate(v17.begin(), v17.end(), std::ref(f16rng)); |
| 253 | std::generate(v18.begin(), v18.end(), std::ref(f16rng)); |
| 254 | std::generate(v19.begin(), v19.end(), std::ref(f16rng)); |
| 255 | std::generate(v20.begin(), v20.end(), std::ref(f16rng)); |
| 256 | std::generate(v21.begin(), v21.end(), std::ref(f16rng)); |
| 257 | std::generate(v22.begin(), v22.end(), std::ref(f16rng)); |
| 258 | std::generate(v23.begin(), v23.end(), std::ref(f16rng)); |
| 259 | std::generate(v24.begin(), v24.end(), std::ref(f16rng)); |
| 260 | std::generate(v25.begin(), v25.end(), std::ref(f16rng)); |
| 261 | std::generate(v26.begin(), v26.end(), std::ref(f16rng)); |
| 262 | std::generate(v27.begin(), v27.end(), std::ref(f16rng)); |
| 263 | std::generate(v28.begin(), v28.end(), std::ref(f16rng)); |
| 264 | std::generate(v29.begin(), v29.end(), std::ref(f16rng)); |
| 265 | std::generate(v30.begin(), v30.end(), std::ref(f16rng)); |
| 266 | std::generate(v31.begin(), v31.end(), std::ref(f16rng)); |
| 267 | std::generate(v32.begin(), v32.end(), std::ref(f16rng)); |
| 268 | std::generate(v33.begin(), v33.end(), std::ref(f16rng)); |
| 269 | std::generate(v34.begin(), v34.end(), std::ref(f16rng)); |
| 270 | std::generate(v35.begin(), v35.end(), std::ref(f16rng)); |
| 271 | std::generate(v36.begin(), v36.end(), std::ref(f16rng)); |
| 272 | std::generate(v37.begin(), v37.end(), std::ref(f16rng)); |
| 273 | std::generate(v38.begin(), v38.end(), std::ref(f16rng)); |
| 274 | std::generate(v39.begin(), v39.end(), std::ref(f16rng)); |
| 275 | std::generate(v40.begin(), v40.end(), std::ref(f16rng)); |
| 276 | std::generate(v41.begin(), v41.end(), std::ref(f16rng)); |
| 277 | std::generate(v42.begin(), v42.end(), std::ref(f16rng)); |
| 278 | std::generate(v43.begin(), v43.end(), std::ref(f16rng)); |
| 279 | std::generate(v44.begin(), v44.end(), std::ref(f16rng)); |
| 280 | std::generate(v45.begin(), v45.end(), std::ref(f16rng)); |
| 281 | std::generate(v46.begin(), v46.end(), std::ref(f16rng)); |
| 282 | std::generate(v47.begin(), v47.end(), std::ref(f16rng)); |
| 283 | std::generate(v48.begin(), v48.end(), std::ref(f16rng)); |
| 284 | std::generate(v49.begin(), v49.end(), std::ref(f16rng)); |
| 285 | std::generate(v50.begin(), v50.end(), std::ref(f16rng)); |
| 286 | std::generate(v51.begin(), v51.end(), std::ref(f16rng)); |
| 287 | std::generate(v52.begin(), v52.end(), std::ref(f16rng)); |
| 288 | std::generate(v53.begin(), v53.end(), std::ref(f16rng)); |
| 289 | std::generate(v54.begin(), v54.end(), std::ref(f16rng)); |
| 290 | std::generate(v55.begin(), v55.end(), std::ref(f16rng)); |
| 291 | std::generate(v56.begin(), v56.end(), std::ref(f16rng)); |
| 292 | std::generate(v57.begin(), v57.end(), std::ref(f16rng)); |
| 293 | std::generate(v58.begin(), v58.end(), std::ref(f16rng)); |
| 294 | std::generate(v59.begin(), v59.end(), std::ref(f16rng)); |
| 295 | std::generate(v60.begin(), v60.end(), std::ref(f16rng)); |
| 296 | std::generate(v61.begin(), v61.end(), std::ref(f16rng)); |
| 297 | std::generate(v62.begin(), v62.end(), std::ref(f16rng)); |
| 298 | std::generate(v63.begin(), v63.end(), std::ref(f16rng)); |
| 299 | std::generate(v64.begin(), v64.end(), std::ref(f16rng)); |
| 300 | std::generate(v65.begin(), v65.end(), std::ref(f16rng)); |
| 301 | std::generate(v66.begin(), v66.end(), std::ref(f16rng)); |
| 302 | std::generate(v67.begin(), v67.end(), std::ref(f16rng)); |
| 303 | std::generate(v68.begin(), v68.end(), std::ref(f16rng)); |
| 304 | std::generate(v69.begin(), v69.end(), std::ref(f16rng)); |
| 305 | std::generate(v70.begin(), v70.end(), std::ref(f16rng)); |
| 306 | std::generate(v71.begin(), v71.end(), std::ref(f16rng)); |
| 307 | std::generate(v72.begin(), v72.end(), std::ref(f16rng)); |
| 308 | std::generate(v73.begin(), v73.end(), std::ref(f16rng)); |
| 309 | std::generate(v74.begin(), v74.end(), std::ref(f16rng)); |
| 310 | std::generate(v75.begin(), v75.end(), std::ref(f16rng)); |
| 311 | std::generate(v76.begin(), v76.end(), std::ref(f16rng)); |
| 312 | std::generate(v77.begin(), v77.end(), std::ref(f16rng)); |
| 313 | std::generate(v78.begin(), v78.end(), std::ref(f16rng)); |
| 314 | std::generate(v79.begin(), v79.end(), std::ref(f16rng)); |
| 315 | std::generate(v80.begin(), v80.end(), std::ref(f16rng)); |
| 316 | std::generate(v81.begin(), v81.end(), std::ref(f16rng)); |
| 317 | std::generate(v82.begin(), v82.end(), std::ref(f16rng)); |
| 318 | std::generate(v83.begin(), v83.end(), std::ref(f16rng)); |
| 319 | std::generate(v84.begin(), v84.end(), std::ref(f16rng)); |
| 320 | std::generate(v85.begin(), v85.end(), std::ref(f16rng)); |
| 321 | std::generate(v86.begin(), v86.end(), std::ref(f16rng)); |
| 322 | std::generate(v87.begin(), v87.end(), std::ref(f16rng)); |
| 323 | std::generate(v88.begin(), v88.end(), std::ref(f16rng)); |
| 324 | std::generate(v89.begin(), v89.end(), std::ref(f16rng)); |
| 325 | std::generate(v90.begin(), v90.end(), std::ref(f16rng)); |
| 326 | std::generate(v91.begin(), v91.end(), std::ref(f16rng)); |
| 327 | std::generate(v92.begin(), v92.end(), std::ref(f16rng)); |
| 328 | std::generate(v93.begin(), v93.end(), std::ref(f16rng)); |
| 329 | std::generate(v94.begin(), v94.end(), std::ref(f16rng)); |
| 330 | std::generate(v95.begin(), v95.end(), std::ref(f16rng)); |
| 331 | std::generate(v96.begin(), v96.end(), std::ref(f16rng)); |
| 332 | std::generate(v97.begin(), v97.end(), std::ref(f16rng)); |
| 333 | std::generate(v98.begin(), v98.end(), std::ref(f16rng)); |
| 334 | std::generate(v99.begin(), v99.end(), std::ref(f16rng)); |
| 335 | std::generate(w100.begin(), w100.end(), std::ref(f16rng)); |
| 336 | std::generate(w101.begin(), w101.end(), std::ref(f16rng)); |
| 337 | std::generate(w102.begin(), w102.end(), std::ref(f16rng)); |
| 338 | std::generate(w103.begin(), w103.end(), std::ref(f16rng)); |
| 339 | std::generate(w104.begin(), w104.end(), std::ref(f16rng)); |
| 340 | std::generate(w105.begin(), w105.end(), std::ref(f16rng)); |
| 341 | std::generate(w106.begin(), w106.end(), std::ref(f16rng)); |
| 342 | std::generate(w107.begin(), w107.end(), std::ref(f16rng)); |
| 343 | std::generate(w108.begin(), w108.end(), std::ref(f16rng)); |
| 344 | std::generate(w109.begin(), w109.end(), std::ref(f16rng)); |
| 345 | std::generate(w110.begin(), w110.end(), std::ref(f16rng)); |
| 346 | std::generate(w111.begin(), w111.end(), std::ref(f16rng)); |
| 347 | std::generate(w112.begin(), w112.end(), std::ref(f16rng)); |
| 348 | std::generate(w113.begin(), w113.end(), std::ref(f16rng)); |
| 349 | std::generate(w114.begin(), w114.end(), std::ref(f16rng)); |
| 350 | std::generate(w115.begin(), w115.end(), std::ref(f16rng)); |
| 351 | std::generate(w116.begin(), w116.end(), std::ref(f16rng)); |
| 352 | std::generate(w117.begin(), w117.end(), std::ref(f16rng)); |
| 353 | std::generate(w118.begin(), w118.end(), std::ref(f16rng)); |
| 354 | std::generate(w119.begin(), w119.end(), std::ref(f16rng)); |
| 355 | std::generate(w120.begin(), w120.end(), std::ref(f16rng)); |
| 356 | std::generate(w121.begin(), w121.end(), std::ref(f16rng)); |
| 357 | std::generate(w122.begin(), w122.end(), std::ref(f16rng)); |
| 358 | std::generate(w123.begin(), w123.end(), std::ref(f16rng)); |
| 359 | std::generate(w124.begin(), w124.end(), std::ref(f16rng)); |
| 360 | std::generate(w125.begin(), w125.end(), std::ref(f16rng)); |
| 361 | std::generate(w126.begin(), w126.end(), std::ref(f16rng)); |
| 362 | std::generate(w127.begin(), w127.end(), std::ref(f16rng)); |
| 363 | std::generate(w128.begin(), w128.end(), std::ref(f16rng)); |
| 364 | std::generate(w129.begin(), w129.end(), std::ref(f16rng)); |
| 365 | std::generate(w130.begin(), w130.end(), std::ref(f16rng)); |
| 366 | std::generate(w131.begin(), w131.end(), std::ref(f16rng)); |
| 367 | std::generate(w132.begin(), w132.end(), std::ref(f16rng)); |
| 368 | std::generate(w133.begin(), w133.end(), std::ref(f16rng)); |
| 369 | std::generate(w134.begin(), w134.end(), std::ref(f16rng)); |
| 370 | std::generate(w135.begin(), w135.end(), std::ref(f16rng)); |
| 371 | std::generate(w136.begin(), w136.end(), std::ref(f16rng)); |
| 372 | std::generate(w137.begin(), w137.end(), std::ref(f16rng)); |
| 373 | std::generate(w138.begin(), w138.end(), std::ref(f16rng)); |
| 374 | std::generate(w139.begin(), w139.end(), std::ref(f16rng)); |
| 375 | std::generate(w140.begin(), w140.end(), std::ref(f16rng)); |
| 376 | std::generate(w141.begin(), w141.end(), std::ref(f16rng)); |
| 377 | std::generate(w142.begin(), w142.end(), std::ref(f16rng)); |
| 378 | std::generate(w143.begin(), w143.end(), std::ref(f16rng)); |
| 379 | std::generate(w144.begin(), w144.end(), std::ref(f16rng)); |
| 380 | std::generate(w145.begin(), w145.end(), std::ref(f16rng)); |
| 381 | std::generate(w146.begin(), w146.end(), std::ref(f16rng)); |
| 382 | std::generate(w147.begin(), w147.end(), std::ref(f16rng)); |
| 383 | std::generate(w148.begin(), w148.end(), std::ref(f16rng)); |
| 384 | std::generate(w149.begin(), w149.end(), std::ref(f16rng)); |
| 385 | std::generate(w150.begin(), w150.end(), std::ref(f16rng)); |
| 386 | std::generate(w151.begin(), w151.end(), std::ref(f16rng)); |
| 387 | std::generate(w152.begin(), w152.end(), std::ref(f16rng)); |
| 388 | std::generate(w153.begin(), w153.end(), std::ref(f16rng)); |
| 389 | std::generate(w154.begin(), w154.end(), std::ref(f16rng)); |
| 390 | std::generate(w155.begin(), w155.end(), std::ref(f16rng)); |
| 391 | std::generate(w156.begin(), w156.end(), std::ref(f16rng)); |
| 392 | std::generate(w157.begin(), w157.end(), std::ref(f16rng)); |
| 393 | std::generate(w158.begin(), w158.end(), std::ref(f16rng)); |
| 394 | std::generate(w159.begin(), w159.end(), std::ref(f16rng)); |
| 395 | std::generate(w160.begin(), w160.end(), std::ref(f16rng)); |
| 396 | std::generate(w161.begin(), w161.end(), std::ref(f16rng)); |
| 397 | std::generate(w162.begin(), w162.end(), std::ref(f16rng)); |
| 398 | std::generate(w163.begin(), w163.end(), std::ref(f16rng)); |
| 399 | std::generate(w164.begin(), w164.end(), std::ref(f16rng)); |
| 400 | std::generate(w165.begin(), w165.end(), std::ref(f16rng)); |
| 401 | std::generate(w166.begin(), w166.end(), std::ref(f16rng)); |
| 402 | std::generate(w167.begin(), w167.end(), std::ref(f16rng)); |
| 403 | std::generate(w168.begin(), w168.end(), std::ref(f16rng)); |
| 404 | std::generate(w169.begin(), w169.end(), std::ref(f16rng)); |
| 405 | std::generate(w170.begin(), w170.end(), std::ref(f16rng)); |
| 406 | std::generate(w171.begin(), w171.end(), std::ref(f16rng)); |
| 407 | std::generate(w172.begin(), w172.end(), std::ref(f16rng)); |
| 408 | std::generate(w173.begin(), w173.end(), std::ref(f16rng)); |
| 409 | std::generate(w174.begin(), w174.end(), std::ref(f16rng)); |
| 410 | std::generate(w175.begin(), w175.end(), std::ref(f16rng)); |
| 411 | std::generate(w176.begin(), w176.end(), std::ref(f16rng)); |
| 412 | std::generate(w177.begin(), w177.end(), std::ref(f16rng)); |
| 413 | std::generate(w178.begin(), w178.end(), std::ref(f16rng)); |
| 414 | std::generate(w179.begin(), w179.end(), std::ref(f16rng)); |
| 415 | std::generate(w180.begin(), w180.end(), std::ref(f16rng)); |
| 416 | std::generate(w181.begin(), w181.end(), std::ref(f16rng)); |
| 417 | std::generate(w182.begin(), w182.end(), std::ref(f16rng)); |
| 418 | std::generate(w183.begin(), w183.end(), std::ref(f16rng)); |
| 419 | std::generate(w184.begin(), w184.end(), std::ref(f16rng)); |
| 420 | std::generate(w185.begin(), w185.end(), std::ref(f16rng)); |
| 421 | std::generate(w186.begin(), w186.end(), std::ref(f16rng)); |
| 422 | std::generate(w187.begin(), w187.end(), std::ref(f16rng)); |
| 423 | std::generate(w188.begin(), w188.end(), std::ref(f16rng)); |
| 424 | std::generate(w189.begin(), w189.end(), std::ref(f16rng)); |
| 425 | std::generate(w190.begin(), w190.end(), std::ref(f16rng)); |
| 426 | std::generate(w191.begin(), w191.end(), std::ref(f16rng)); |
| 427 | std::generate(w192.begin(), w192.end(), std::ref(f16rng)); |
| 428 | std::generate(w193.begin(), w193.end(), std::ref(f16rng)); |
| 429 | std::generate(w194.begin(), w194.end(), std::ref(f16rng)); |
| 430 | std::generate(w195.begin(), w195.end(), std::ref(f16rng)); |
| 431 | std::generate(w196.begin(), w196.end(), std::ref(f16rng)); |
| 432 | std::generate(w197.begin(), w197.end(), std::ref(f16rng)); |
| 433 | std::generate(w198.begin(), w198.end(), std::ref(f16rng)); |
| 434 | std::generate(w199.begin(), w199.end(), std::ref(f16rng)); |
| 435 | std::generate(w200.begin(), w200.end(), std::ref(f16rng)); |
| 436 | std::generate(w201.begin(), w201.end(), std::ref(f16rng)); |
| 437 | std::generate(w202.begin(), w202.end(), std::ref(f16rng)); |
| 438 | std::generate(w203.begin(), w203.end(), std::ref(f16rng)); |
| 439 | std::generate(w204.begin(), w204.end(), std::ref(f16rng)); |
| 440 | std::generate(w205.begin(), w205.end(), std::ref(f16rng)); |
| 441 | std::generate(w206.begin(), w206.end(), std::ref(f16rng)); |
| 442 | std::generate(w207.begin(), w207.end(), std::ref(f16rng)); |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 443 | |
| 444 | ExecutionPlan operators; |
| 445 | xnn_status status; |
| 446 | |
| 447 | xnn_operator_t op0 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 448 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 449 | 0 /* top padding */, 1 /* right padding */, |
| 450 | 1 /* bottom padding */, 0 /* left padding */, |
| 451 | 3 /* kernel height */, 3 /* kernel width */, |
| 452 | 2 /* subsampling height */, 2 /* subsampling width */, |
| 453 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 454 | 1 /* groups */, |
| 455 | 3 /* input channels per group */, |
| 456 | 16 /* output_channels_per_group */, |
| 457 | 3 /* input pixel stride */, |
| 458 | 16 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 459 | w100.data(), w101.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 460 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 461 | 0 /* flags */, |
| 462 | &op0); |
| 463 | if (status != xnn_status_success) { |
| 464 | std::cerr << "failed to create operation #0" << std::endl; |
| 465 | return ExecutionPlan(); |
| 466 | } |
| 467 | operators.emplace_back(op0, xnn_delete_operator); |
| 468 | |
| 469 | xnn_operator_t op1 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 470 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 471 | 16 /* channels */, |
| 472 | 16 /* input stride */, |
| 473 | 16 /* output stride */, |
| 474 | 0 /* flags */, |
| 475 | &op1); |
| 476 | if (status != xnn_status_success) { |
| 477 | std::cerr << "failed to create operation #1" << std::endl; |
| 478 | return ExecutionPlan(); |
| 479 | } |
| 480 | operators.emplace_back(op1, xnn_delete_operator); |
| 481 | |
| 482 | xnn_operator_t op2 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 483 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 484 | 0 /* top padding */, 1 /* right padding */, |
| 485 | 1 /* bottom padding */, 0 /* left padding */, |
| 486 | 3 /* kernel height */, 3 /* kernel width */, |
| 487 | 2 /* subsampling height */, 2 /* subsampling width */, |
| 488 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 489 | 16 /* groups */, |
| 490 | 1 /* input channels per group */, |
| 491 | 1 /* output_channels_per_group */, |
| 492 | 16 /* input pixel stride */, |
| 493 | 16 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 494 | w102.data(), w103.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 495 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 496 | 0 /* flags */, |
| 497 | &op2); |
| 498 | if (status != xnn_status_success) { |
| 499 | std::cerr << "failed to create operation #2" << std::endl; |
| 500 | return ExecutionPlan(); |
| 501 | } |
| 502 | operators.emplace_back(op2, xnn_delete_operator); |
| 503 | |
| 504 | xnn_operator_t op3 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 505 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 506 | 16 /* channels */, 16 /* input stride */, 16 /* output stride */, |
| 507 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 508 | 0 /* flags */, |
| 509 | &op3); |
| 510 | if (status != xnn_status_success) { |
| 511 | std::cerr << "failed to create operation #3" << std::endl; |
| 512 | return ExecutionPlan(); |
| 513 | } |
| 514 | operators.emplace_back(op3, xnn_delete_operator); |
| 515 | |
| 516 | xnn_operator_t op4 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 517 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 518 | 0 /* top padding */, 0 /* right padding */, |
| 519 | 0 /* bottom padding */, 0 /* left padding */, |
| 520 | 1 /* kernel height */, 1 /* kernel width */, |
| 521 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 522 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 523 | 1 /* groups */, |
| 524 | 16 /* input channels per group */, |
| 525 | 8 /* output_channels_per_group */, |
| 526 | 16 /* input pixel stride */, |
| 527 | 8 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 528 | w104.data(), w105.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 529 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 530 | 0 /* flags */, |
| 531 | &op4); |
| 532 | if (status != xnn_status_success) { |
| 533 | std::cerr << "failed to create operation #4" << std::endl; |
| 534 | return ExecutionPlan(); |
| 535 | } |
| 536 | operators.emplace_back(op4, xnn_delete_operator); |
| 537 | |
| 538 | xnn_operator_t op5 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 539 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 540 | 0 /* top padding */, 0 /* right padding */, |
| 541 | 0 /* bottom padding */, 0 /* left padding */, |
| 542 | 1 /* kernel height */, 1 /* kernel width */, |
| 543 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 544 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 545 | 1 /* groups */, |
| 546 | 8 /* input channels per group */, |
| 547 | 16 /* output_channels_per_group */, |
| 548 | 8 /* input pixel stride */, |
| 549 | 16 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 550 | w106.data(), w107.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 551 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 552 | 0 /* flags */, |
| 553 | &op5); |
| 554 | if (status != xnn_status_success) { |
| 555 | std::cerr << "failed to create operation #5" << std::endl; |
| 556 | return ExecutionPlan(); |
| 557 | } |
| 558 | operators.emplace_back(op5, xnn_delete_operator); |
| 559 | |
| 560 | xnn_operator_t op6 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 561 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 562 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 563 | 0 /* flags */, |
| 564 | &op6); |
| 565 | if (status != xnn_status_success) { |
| 566 | std::cerr << "failed to create operation #6" << std::endl; |
| 567 | return ExecutionPlan(); |
| 568 | } |
| 569 | operators.emplace_back(op6, xnn_delete_operator); |
| 570 | |
| 571 | xnn_operator_t op7 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 572 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 573 | 0 /* top padding */, 0 /* right padding */, |
| 574 | 0 /* bottom padding */, 0 /* left padding */, |
| 575 | 1 /* kernel height */, 1 /* kernel width */, |
| 576 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 577 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 578 | 1 /* groups */, |
| 579 | 16 /* input channels per group */, |
| 580 | 16 /* output_channels_per_group */, |
| 581 | 16 /* input pixel stride */, |
| 582 | 16 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 583 | w108.data(), w109.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 584 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 585 | 0 /* flags */, |
| 586 | &op7); |
| 587 | if (status != xnn_status_success) { |
| 588 | std::cerr << "failed to create operation #7" << std::endl; |
| 589 | return ExecutionPlan(); |
| 590 | } |
| 591 | operators.emplace_back(op7, xnn_delete_operator); |
| 592 | |
| 593 | xnn_operator_t op8 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 594 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 595 | 0 /* top padding */, 0 /* right padding */, |
| 596 | 0 /* bottom padding */, 0 /* left padding */, |
| 597 | 1 /* kernel height */, 1 /* kernel width */, |
| 598 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 599 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 600 | 1 /* groups */, |
| 601 | 16 /* input channels per group */, |
| 602 | 72 /* output_channels_per_group */, |
| 603 | 16 /* input pixel stride */, |
| 604 | 72 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 605 | w110.data(), w111.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 606 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 607 | 0 /* flags */, |
| 608 | &op8); |
| 609 | if (status != xnn_status_success) { |
| 610 | std::cerr << "failed to create operation #8" << std::endl; |
| 611 | return ExecutionPlan(); |
| 612 | } |
| 613 | operators.emplace_back(op8, xnn_delete_operator); |
| 614 | |
| 615 | xnn_operator_t op9 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 616 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 617 | 0 /* top padding */, 1 /* right padding */, |
| 618 | 1 /* bottom padding */, 0 /* left padding */, |
| 619 | 3 /* kernel height */, 3 /* kernel width */, |
| 620 | 2 /* subsampling height */, 2 /* subsampling width */, |
| 621 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 622 | 72 /* groups */, |
| 623 | 1 /* input channels per group */, |
| 624 | 1 /* output_channels_per_group */, |
| 625 | 72 /* input pixel stride */, |
| 626 | 72 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 627 | w112.data(), w113.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 628 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 629 | 0 /* flags */, |
| 630 | &op9); |
| 631 | if (status != xnn_status_success) { |
| 632 | std::cerr << "failed to create operation #9" << std::endl; |
| 633 | return ExecutionPlan(); |
| 634 | } |
| 635 | operators.emplace_back(op9, xnn_delete_operator); |
| 636 | |
| 637 | xnn_operator_t op10 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 638 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 639 | 0 /* top padding */, 0 /* right padding */, |
| 640 | 0 /* bottom padding */, 0 /* left padding */, |
| 641 | 1 /* kernel height */, 1 /* kernel width */, |
| 642 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 643 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 644 | 1 /* groups */, |
| 645 | 72 /* input channels per group */, |
| 646 | 24 /* output_channels_per_group */, |
| 647 | 72 /* input pixel stride */, |
| 648 | 24 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 649 | w114.data(), w115.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 650 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 651 | 0 /* flags */, |
| 652 | &op10); |
| 653 | if (status != xnn_status_success) { |
| 654 | std::cerr << "failed to create operation #10" << std::endl; |
| 655 | return ExecutionPlan(); |
| 656 | } |
| 657 | operators.emplace_back(op10, xnn_delete_operator); |
| 658 | |
| 659 | xnn_operator_t op11 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 660 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 661 | 0 /* top padding */, 0 /* right padding */, |
| 662 | 0 /* bottom padding */, 0 /* left padding */, |
| 663 | 1 /* kernel height */, 1 /* kernel width */, |
| 664 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 665 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 666 | 1 /* groups */, |
| 667 | 24 /* input channels per group */, |
| 668 | 88 /* output_channels_per_group */, |
| 669 | 24 /* input pixel stride */, |
| 670 | 88 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 671 | w116.data(), w117.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 672 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 673 | 0 /* flags */, |
| 674 | &op11); |
| 675 | if (status != xnn_status_success) { |
| 676 | std::cerr << "failed to create operation #11" << std::endl; |
| 677 | return ExecutionPlan(); |
| 678 | } |
| 679 | operators.emplace_back(op11, xnn_delete_operator); |
| 680 | |
| 681 | xnn_operator_t op12 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 682 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 683 | 1 /* top padding */, 1 /* right padding */, |
| 684 | 1 /* bottom padding */, 1 /* left padding */, |
| 685 | 3 /* kernel height */, 3 /* kernel width */, |
| 686 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 687 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 688 | 88 /* groups */, |
| 689 | 1 /* input channels per group */, |
| 690 | 1 /* output_channels_per_group */, |
| 691 | 88 /* input pixel stride */, |
| 692 | 88 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 693 | w118.data(), w119.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 694 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 695 | 0 /* flags */, |
| 696 | &op12); |
| 697 | if (status != xnn_status_success) { |
| 698 | std::cerr << "failed to create operation #12" << std::endl; |
| 699 | return ExecutionPlan(); |
| 700 | } |
| 701 | operators.emplace_back(op12, xnn_delete_operator); |
| 702 | |
| 703 | xnn_operator_t op13 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 704 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 705 | 0 /* top padding */, 0 /* right padding */, |
| 706 | 0 /* bottom padding */, 0 /* left padding */, |
| 707 | 1 /* kernel height */, 1 /* kernel width */, |
| 708 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 709 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 710 | 1 /* groups */, |
| 711 | 88 /* input channels per group */, |
| 712 | 24 /* output_channels_per_group */, |
| 713 | 88 /* input pixel stride */, |
| 714 | 24 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 715 | w120.data(), w121.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 716 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 717 | 0 /* flags */, |
| 718 | &op13); |
| 719 | if (status != xnn_status_success) { |
| 720 | std::cerr << "failed to create operation #13" << std::endl; |
| 721 | return ExecutionPlan(); |
| 722 | } |
| 723 | operators.emplace_back(op13, xnn_delete_operator); |
| 724 | |
| 725 | xnn_operator_t op14 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 726 | status = xnn_create_add_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 727 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 728 | 0 /* flags */, |
| 729 | &op14); |
| 730 | if (status != xnn_status_success) { |
| 731 | std::cerr << "failed to create operation #14" << std::endl; |
| 732 | return ExecutionPlan(); |
| 733 | } |
| 734 | operators.emplace_back(op14, xnn_delete_operator); |
| 735 | |
| 736 | xnn_operator_t op15 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 737 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 738 | 0 /* top padding */, 0 /* right padding */, |
| 739 | 0 /* bottom padding */, 0 /* left padding */, |
| 740 | 1 /* kernel height */, 1 /* kernel width */, |
| 741 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 742 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 743 | 1 /* groups */, |
| 744 | 24 /* input channels per group */, |
| 745 | 96 /* output_channels_per_group */, |
| 746 | 24 /* input pixel stride */, |
| 747 | 96 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 748 | w122.data(), w123.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 749 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 750 | 0 /* flags */, |
| 751 | &op15); |
| 752 | if (status != xnn_status_success) { |
| 753 | std::cerr << "failed to create operation #15" << std::endl; |
| 754 | return ExecutionPlan(); |
| 755 | } |
| 756 | operators.emplace_back(op15, xnn_delete_operator); |
| 757 | |
| 758 | xnn_operator_t op16 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 759 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 760 | 96 /* channels */, |
| 761 | 96 /* input stride */, |
| 762 | 96 /* output stride */, |
| 763 | 0 /* flags */, |
| 764 | &op16); |
| 765 | if (status != xnn_status_success) { |
| 766 | std::cerr << "failed to create operation #16" << std::endl; |
| 767 | return ExecutionPlan(); |
| 768 | } |
| 769 | operators.emplace_back(op16, xnn_delete_operator); |
| 770 | |
| 771 | xnn_operator_t op17 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 772 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 773 | 1 /* top padding */, 2 /* right padding */, |
| 774 | 2 /* bottom padding */, 1 /* left padding */, |
| 775 | 5 /* kernel height */, 5 /* kernel width */, |
| 776 | 2 /* subsampling height */, 2 /* subsampling width */, |
| 777 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 778 | 96 /* groups */, |
| 779 | 1 /* input channels per group */, |
| 780 | 1 /* output_channels_per_group */, |
| 781 | 96 /* input pixel stride */, |
| 782 | 96 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 783 | w124.data(), w125.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 784 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 785 | 0 /* flags */, |
| 786 | &op17); |
| 787 | if (status != xnn_status_success) { |
| 788 | std::cerr << "failed to create operation #17" << std::endl; |
| 789 | return ExecutionPlan(); |
| 790 | } |
| 791 | operators.emplace_back(op17, xnn_delete_operator); |
| 792 | |
| 793 | xnn_operator_t op18 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 794 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 795 | 96 /* channels */, |
| 796 | 96 /* input stride */, |
| 797 | 96 /* output stride */, |
| 798 | 0 /* flags */, |
| 799 | &op18); |
| 800 | if (status != xnn_status_success) { |
| 801 | std::cerr << "failed to create operation #18" << std::endl; |
| 802 | return ExecutionPlan(); |
| 803 | } |
| 804 | operators.emplace_back(op18, xnn_delete_operator); |
| 805 | |
| 806 | xnn_operator_t op19 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 807 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 808 | 96 /* channels */, 96 /* input stride */, 96 /* output stride */, |
| 809 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 810 | 0 /* flags */, |
| 811 | &op19); |
| 812 | if (status != xnn_status_success) { |
| 813 | std::cerr << "failed to create operation #19" << std::endl; |
| 814 | return ExecutionPlan(); |
| 815 | } |
| 816 | operators.emplace_back(op19, xnn_delete_operator); |
| 817 | |
| 818 | xnn_operator_t op20 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 819 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 820 | 0 /* top padding */, 0 /* right padding */, |
| 821 | 0 /* bottom padding */, 0 /* left padding */, |
| 822 | 1 /* kernel height */, 1 /* kernel width */, |
| 823 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 824 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 825 | 1 /* groups */, |
| 826 | 96 /* input channels per group */, |
| 827 | 24 /* output_channels_per_group */, |
| 828 | 96 /* input pixel stride */, |
| 829 | 24 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 830 | w126.data(), w127.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 831 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 832 | 0 /* flags */, |
| 833 | &op20); |
| 834 | if (status != xnn_status_success) { |
| 835 | std::cerr << "failed to create operation #20" << std::endl; |
| 836 | return ExecutionPlan(); |
| 837 | } |
| 838 | operators.emplace_back(op20, xnn_delete_operator); |
| 839 | |
| 840 | xnn_operator_t op21 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 841 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 842 | 0 /* top padding */, 0 /* right padding */, |
| 843 | 0 /* bottom padding */, 0 /* left padding */, |
| 844 | 1 /* kernel height */, 1 /* kernel width */, |
| 845 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 846 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 847 | 1 /* groups */, |
| 848 | 24 /* input channels per group */, |
| 849 | 96 /* output_channels_per_group */, |
| 850 | 24 /* input pixel stride */, |
| 851 | 96 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 852 | w128.data(), w129.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 853 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 854 | 0 /* flags */, |
| 855 | &op21); |
| 856 | if (status != xnn_status_success) { |
| 857 | std::cerr << "failed to create operation #21" << std::endl; |
| 858 | return ExecutionPlan(); |
| 859 | } |
| 860 | operators.emplace_back(op21, xnn_delete_operator); |
| 861 | |
| 862 | xnn_operator_t op22 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 863 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 864 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 865 | 0 /* flags */, |
| 866 | &op22); |
| 867 | if (status != xnn_status_success) { |
| 868 | std::cerr << "failed to create operation #22" << std::endl; |
| 869 | return ExecutionPlan(); |
| 870 | } |
| 871 | operators.emplace_back(op22, xnn_delete_operator); |
| 872 | |
| 873 | xnn_operator_t op23 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 874 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 875 | 0 /* top padding */, 0 /* right padding */, |
| 876 | 0 /* bottom padding */, 0 /* left padding */, |
| 877 | 1 /* kernel height */, 1 /* kernel width */, |
| 878 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 879 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 880 | 1 /* groups */, |
| 881 | 96 /* input channels per group */, |
| 882 | 40 /* output_channels_per_group */, |
| 883 | 96 /* input pixel stride */, |
| 884 | 40 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 885 | w130.data(), w131.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 886 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 887 | 0 /* flags */, |
| 888 | &op23); |
| 889 | if (status != xnn_status_success) { |
| 890 | std::cerr << "failed to create operation #23" << std::endl; |
| 891 | return ExecutionPlan(); |
| 892 | } |
| 893 | operators.emplace_back(op23, xnn_delete_operator); |
| 894 | |
| 895 | xnn_operator_t op24 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 896 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 897 | 0 /* top padding */, 0 /* right padding */, |
| 898 | 0 /* bottom padding */, 0 /* left padding */, |
| 899 | 1 /* kernel height */, 1 /* kernel width */, |
| 900 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 901 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 902 | 1 /* groups */, |
| 903 | 40 /* input channels per group */, |
| 904 | 240 /* output_channels_per_group */, |
| 905 | 40 /* input pixel stride */, |
| 906 | 240 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 907 | w132.data(), w133.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 908 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 909 | 0 /* flags */, |
| 910 | &op24); |
| 911 | if (status != xnn_status_success) { |
| 912 | std::cerr << "failed to create operation #24" << std::endl; |
| 913 | return ExecutionPlan(); |
| 914 | } |
| 915 | operators.emplace_back(op24, xnn_delete_operator); |
| 916 | |
| 917 | xnn_operator_t op25 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 918 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 919 | 240 /* channels */, |
| 920 | 240 /* input stride */, |
| 921 | 240 /* output stride */, |
| 922 | 0 /* flags */, |
| 923 | &op25); |
| 924 | if (status != xnn_status_success) { |
| 925 | std::cerr << "failed to create operation #25" << std::endl; |
| 926 | return ExecutionPlan(); |
| 927 | } |
| 928 | operators.emplace_back(op25, xnn_delete_operator); |
| 929 | |
| 930 | xnn_operator_t op26 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 931 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 932 | 2 /* top padding */, 2 /* right padding */, |
| 933 | 2 /* bottom padding */, 2 /* left padding */, |
| 934 | 5 /* kernel height */, 5 /* kernel width */, |
| 935 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 936 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 937 | 240 /* groups */, |
| 938 | 1 /* input channels per group */, |
| 939 | 1 /* output_channels_per_group */, |
| 940 | 240 /* input pixel stride */, |
| 941 | 240 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 942 | w134.data(), w135.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 943 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 944 | 0 /* flags */, |
| 945 | &op26); |
| 946 | if (status != xnn_status_success) { |
| 947 | std::cerr << "failed to create operation #26" << std::endl; |
| 948 | return ExecutionPlan(); |
| 949 | } |
| 950 | operators.emplace_back(op26, xnn_delete_operator); |
| 951 | |
| 952 | xnn_operator_t op27 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 953 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 954 | 240 /* channels */, |
| 955 | 240 /* input stride */, |
| 956 | 240 /* output stride */, |
| 957 | 0 /* flags */, |
| 958 | &op27); |
| 959 | if (status != xnn_status_success) { |
| 960 | std::cerr << "failed to create operation #27" << std::endl; |
| 961 | return ExecutionPlan(); |
| 962 | } |
| 963 | operators.emplace_back(op27, xnn_delete_operator); |
| 964 | |
| 965 | xnn_operator_t op28 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 966 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 967 | 240 /* channels */, 240 /* input stride */, 240 /* output stride */, |
| 968 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 969 | 0 /* flags */, |
| 970 | &op28); |
| 971 | if (status != xnn_status_success) { |
| 972 | std::cerr << "failed to create operation #28" << std::endl; |
| 973 | return ExecutionPlan(); |
| 974 | } |
| 975 | operators.emplace_back(op28, xnn_delete_operator); |
| 976 | |
| 977 | xnn_operator_t op29 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 978 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 979 | 0 /* top padding */, 0 /* right padding */, |
| 980 | 0 /* bottom padding */, 0 /* left padding */, |
| 981 | 1 /* kernel height */, 1 /* kernel width */, |
| 982 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 983 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 984 | 1 /* groups */, |
| 985 | 240 /* input channels per group */, |
| 986 | 64 /* output_channels_per_group */, |
| 987 | 240 /* input pixel stride */, |
| 988 | 64 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 989 | w136.data(), w137.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 990 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 991 | 0 /* flags */, |
| 992 | &op29); |
| 993 | if (status != xnn_status_success) { |
| 994 | std::cerr << "failed to create operation #29" << std::endl; |
| 995 | return ExecutionPlan(); |
| 996 | } |
| 997 | operators.emplace_back(op29, xnn_delete_operator); |
| 998 | |
| 999 | xnn_operator_t op30 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1000 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1001 | 0 /* top padding */, 0 /* right padding */, |
| 1002 | 0 /* bottom padding */, 0 /* left padding */, |
| 1003 | 1 /* kernel height */, 1 /* kernel width */, |
| 1004 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1005 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1006 | 1 /* groups */, |
| 1007 | 64 /* input channels per group */, |
| 1008 | 240 /* output_channels_per_group */, |
| 1009 | 64 /* input pixel stride */, |
| 1010 | 240 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1011 | w138.data(), w139.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1012 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 1013 | 0 /* flags */, |
| 1014 | &op30); |
| 1015 | if (status != xnn_status_success) { |
| 1016 | std::cerr << "failed to create operation #30" << std::endl; |
| 1017 | return ExecutionPlan(); |
| 1018 | } |
| 1019 | operators.emplace_back(op30, xnn_delete_operator); |
| 1020 | |
| 1021 | xnn_operator_t op31 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1022 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1023 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1024 | 0 /* flags */, |
| 1025 | &op31); |
| 1026 | if (status != xnn_status_success) { |
| 1027 | std::cerr << "failed to create operation #31" << std::endl; |
| 1028 | return ExecutionPlan(); |
| 1029 | } |
| 1030 | operators.emplace_back(op31, xnn_delete_operator); |
| 1031 | |
| 1032 | xnn_operator_t op32 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1033 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1034 | 0 /* top padding */, 0 /* right padding */, |
| 1035 | 0 /* bottom padding */, 0 /* left padding */, |
| 1036 | 1 /* kernel height */, 1 /* kernel width */, |
| 1037 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1038 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1039 | 1 /* groups */, |
| 1040 | 240 /* input channels per group */, |
| 1041 | 40 /* output_channels_per_group */, |
| 1042 | 240 /* input pixel stride */, |
| 1043 | 40 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1044 | w140.data(), w141.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1045 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1046 | 0 /* flags */, |
| 1047 | &op32); |
| 1048 | if (status != xnn_status_success) { |
| 1049 | std::cerr << "failed to create operation #32" << std::endl; |
| 1050 | return ExecutionPlan(); |
| 1051 | } |
| 1052 | operators.emplace_back(op32, xnn_delete_operator); |
| 1053 | |
| 1054 | xnn_operator_t op33 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1055 | status = xnn_create_add_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1056 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1057 | 0 /* flags */, |
| 1058 | &op33); |
| 1059 | if (status != xnn_status_success) { |
| 1060 | std::cerr << "failed to create operation #33" << std::endl; |
| 1061 | return ExecutionPlan(); |
| 1062 | } |
| 1063 | operators.emplace_back(op33, xnn_delete_operator); |
| 1064 | |
| 1065 | xnn_operator_t op34 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1066 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1067 | 0 /* top padding */, 0 /* right padding */, |
| 1068 | 0 /* bottom padding */, 0 /* left padding */, |
| 1069 | 1 /* kernel height */, 1 /* kernel width */, |
| 1070 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1071 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1072 | 1 /* groups */, |
| 1073 | 40 /* input channels per group */, |
| 1074 | 240 /* output_channels_per_group */, |
| 1075 | 40 /* input pixel stride */, |
| 1076 | 240 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1077 | w142.data(), w143.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1078 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1079 | 0 /* flags */, |
| 1080 | &op34); |
| 1081 | if (status != xnn_status_success) { |
| 1082 | std::cerr << "failed to create operation #34" << std::endl; |
| 1083 | return ExecutionPlan(); |
| 1084 | } |
| 1085 | operators.emplace_back(op34, xnn_delete_operator); |
| 1086 | |
| 1087 | xnn_operator_t op35 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1088 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1089 | 240 /* channels */, |
| 1090 | 240 /* input stride */, |
| 1091 | 240 /* output stride */, |
| 1092 | 0 /* flags */, |
| 1093 | &op35); |
| 1094 | if (status != xnn_status_success) { |
| 1095 | std::cerr << "failed to create operation #35" << std::endl; |
| 1096 | return ExecutionPlan(); |
| 1097 | } |
| 1098 | operators.emplace_back(op35, xnn_delete_operator); |
| 1099 | |
| 1100 | xnn_operator_t op36 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1101 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1102 | 2 /* top padding */, 2 /* right padding */, |
| 1103 | 2 /* bottom padding */, 2 /* left padding */, |
| 1104 | 5 /* kernel height */, 5 /* kernel width */, |
| 1105 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1106 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1107 | 240 /* groups */, |
| 1108 | 1 /* input channels per group */, |
| 1109 | 1 /* output_channels_per_group */, |
| 1110 | 240 /* input pixel stride */, |
| 1111 | 240 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1112 | w144.data(), w145.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1113 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1114 | 0 /* flags */, |
| 1115 | &op36); |
| 1116 | if (status != xnn_status_success) { |
| 1117 | std::cerr << "failed to create operation #36" << std::endl; |
| 1118 | return ExecutionPlan(); |
| 1119 | } |
| 1120 | operators.emplace_back(op36, xnn_delete_operator); |
| 1121 | |
| 1122 | xnn_operator_t op37 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1123 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1124 | 240 /* channels */, |
| 1125 | 240 /* input stride */, |
| 1126 | 240 /* output stride */, |
| 1127 | 0 /* flags */, |
| 1128 | &op37); |
| 1129 | if (status != xnn_status_success) { |
| 1130 | std::cerr << "failed to create operation #37" << std::endl; |
| 1131 | return ExecutionPlan(); |
| 1132 | } |
| 1133 | operators.emplace_back(op37, xnn_delete_operator); |
| 1134 | |
| 1135 | xnn_operator_t op38 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1136 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1137 | 240 /* channels */, 240 /* input stride */, 240 /* output stride */, |
| 1138 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 1139 | 0 /* flags */, |
| 1140 | &op38); |
| 1141 | if (status != xnn_status_success) { |
| 1142 | std::cerr << "failed to create operation #38" << std::endl; |
| 1143 | return ExecutionPlan(); |
| 1144 | } |
| 1145 | operators.emplace_back(op38, xnn_delete_operator); |
| 1146 | |
| 1147 | xnn_operator_t op39 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1148 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1149 | 0 /* top padding */, 0 /* right padding */, |
| 1150 | 0 /* bottom padding */, 0 /* left padding */, |
| 1151 | 1 /* kernel height */, 1 /* kernel width */, |
| 1152 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1153 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1154 | 1 /* groups */, |
| 1155 | 240 /* input channels per group */, |
| 1156 | 64 /* output_channels_per_group */, |
| 1157 | 240 /* input pixel stride */, |
| 1158 | 64 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1159 | w146.data(), w147.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1160 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1161 | 0 /* flags */, |
| 1162 | &op39); |
| 1163 | if (status != xnn_status_success) { |
| 1164 | std::cerr << "failed to create operation #39" << std::endl; |
| 1165 | return ExecutionPlan(); |
| 1166 | } |
| 1167 | operators.emplace_back(op39, xnn_delete_operator); |
| 1168 | |
| 1169 | xnn_operator_t op40 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1170 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1171 | 0 /* top padding */, 0 /* right padding */, |
| 1172 | 0 /* bottom padding */, 0 /* left padding */, |
| 1173 | 1 /* kernel height */, 1 /* kernel width */, |
| 1174 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1175 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1176 | 1 /* groups */, |
| 1177 | 64 /* input channels per group */, |
| 1178 | 240 /* output_channels_per_group */, |
| 1179 | 64 /* input pixel stride */, |
| 1180 | 240 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1181 | w148.data(), w149.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1182 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 1183 | 0 /* flags */, |
| 1184 | &op40); |
| 1185 | if (status != xnn_status_success) { |
| 1186 | std::cerr << "failed to create operation #40" << std::endl; |
| 1187 | return ExecutionPlan(); |
| 1188 | } |
| 1189 | operators.emplace_back(op40, xnn_delete_operator); |
| 1190 | |
| 1191 | xnn_operator_t op41 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1192 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1193 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1194 | 0 /* flags */, |
| 1195 | &op41); |
| 1196 | if (status != xnn_status_success) { |
| 1197 | std::cerr << "failed to create operation #41" << std::endl; |
| 1198 | return ExecutionPlan(); |
| 1199 | } |
| 1200 | operators.emplace_back(op41, xnn_delete_operator); |
| 1201 | |
| 1202 | xnn_operator_t op42 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1203 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1204 | 0 /* top padding */, 0 /* right padding */, |
| 1205 | 0 /* bottom padding */, 0 /* left padding */, |
| 1206 | 1 /* kernel height */, 1 /* kernel width */, |
| 1207 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1208 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1209 | 1 /* groups */, |
| 1210 | 240 /* input channels per group */, |
| 1211 | 40 /* output_channels_per_group */, |
| 1212 | 240 /* input pixel stride */, |
| 1213 | 40 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1214 | w150.data(), w151.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1215 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1216 | 0 /* flags */, |
| 1217 | &op42); |
| 1218 | if (status != xnn_status_success) { |
| 1219 | std::cerr << "failed to create operation #42" << std::endl; |
| 1220 | return ExecutionPlan(); |
| 1221 | } |
| 1222 | operators.emplace_back(op42, xnn_delete_operator); |
| 1223 | |
| 1224 | xnn_operator_t op43 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1225 | status = xnn_create_add_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1226 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1227 | 0 /* flags */, |
| 1228 | &op43); |
| 1229 | if (status != xnn_status_success) { |
| 1230 | std::cerr << "failed to create operation #43" << std::endl; |
| 1231 | return ExecutionPlan(); |
| 1232 | } |
| 1233 | operators.emplace_back(op43, xnn_delete_operator); |
| 1234 | |
| 1235 | xnn_operator_t op44 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1236 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1237 | 0 /* top padding */, 0 /* right padding */, |
| 1238 | 0 /* bottom padding */, 0 /* left padding */, |
| 1239 | 1 /* kernel height */, 1 /* kernel width */, |
| 1240 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1241 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1242 | 1 /* groups */, |
| 1243 | 40 /* input channels per group */, |
| 1244 | 120 /* output_channels_per_group */, |
| 1245 | 40 /* input pixel stride */, |
| 1246 | 120 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1247 | w152.data(), w153.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1248 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1249 | 0 /* flags */, |
| 1250 | &op44); |
| 1251 | if (status != xnn_status_success) { |
| 1252 | std::cerr << "failed to create operation #44" << std::endl; |
| 1253 | return ExecutionPlan(); |
| 1254 | } |
| 1255 | operators.emplace_back(op44, xnn_delete_operator); |
| 1256 | |
| 1257 | xnn_operator_t op45 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1258 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1259 | 120 /* channels */, |
| 1260 | 120 /* input stride */, |
| 1261 | 120 /* output stride */, |
| 1262 | 0 /* flags */, |
| 1263 | &op45); |
| 1264 | if (status != xnn_status_success) { |
| 1265 | std::cerr << "failed to create operation #45" << std::endl; |
| 1266 | return ExecutionPlan(); |
| 1267 | } |
| 1268 | operators.emplace_back(op45, xnn_delete_operator); |
| 1269 | |
| 1270 | xnn_operator_t op46 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1271 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1272 | 2 /* top padding */, 2 /* right padding */, |
| 1273 | 2 /* bottom padding */, 2 /* left padding */, |
| 1274 | 5 /* kernel height */, 5 /* kernel width */, |
| 1275 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1276 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1277 | 120 /* groups */, |
| 1278 | 1 /* input channels per group */, |
| 1279 | 1 /* output_channels_per_group */, |
| 1280 | 120 /* input pixel stride */, |
| 1281 | 120 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1282 | w154.data(), w155.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1283 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1284 | 0 /* flags */, |
| 1285 | &op46); |
| 1286 | if (status != xnn_status_success) { |
| 1287 | std::cerr << "failed to create operation #46" << std::endl; |
| 1288 | return ExecutionPlan(); |
| 1289 | } |
| 1290 | operators.emplace_back(op46, xnn_delete_operator); |
| 1291 | |
| 1292 | xnn_operator_t op47 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1293 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1294 | 120 /* channels */, |
| 1295 | 120 /* input stride */, |
| 1296 | 120 /* output stride */, |
| 1297 | 0 /* flags */, |
| 1298 | &op47); |
| 1299 | if (status != xnn_status_success) { |
| 1300 | std::cerr << "failed to create operation #47" << std::endl; |
| 1301 | return ExecutionPlan(); |
| 1302 | } |
| 1303 | operators.emplace_back(op47, xnn_delete_operator); |
| 1304 | |
| 1305 | xnn_operator_t op48 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1306 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1307 | 120 /* channels */, 120 /* input stride */, 120 /* output stride */, |
| 1308 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 1309 | 0 /* flags */, |
| 1310 | &op48); |
| 1311 | if (status != xnn_status_success) { |
| 1312 | std::cerr << "failed to create operation #48" << std::endl; |
| 1313 | return ExecutionPlan(); |
| 1314 | } |
| 1315 | operators.emplace_back(op48, xnn_delete_operator); |
| 1316 | |
| 1317 | xnn_operator_t op49 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1318 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1319 | 0 /* top padding */, 0 /* right padding */, |
| 1320 | 0 /* bottom padding */, 0 /* left padding */, |
| 1321 | 1 /* kernel height */, 1 /* kernel width */, |
| 1322 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1323 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1324 | 1 /* groups */, |
| 1325 | 120 /* input channels per group */, |
| 1326 | 32 /* output_channels_per_group */, |
| 1327 | 120 /* input pixel stride */, |
| 1328 | 32 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1329 | w156.data(), w157.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1330 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1331 | 0 /* flags */, |
| 1332 | &op49); |
| 1333 | if (status != xnn_status_success) { |
| 1334 | std::cerr << "failed to create operation #49" << std::endl; |
| 1335 | return ExecutionPlan(); |
| 1336 | } |
| 1337 | operators.emplace_back(op49, xnn_delete_operator); |
| 1338 | |
| 1339 | xnn_operator_t op50 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1340 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1341 | 0 /* top padding */, 0 /* right padding */, |
| 1342 | 0 /* bottom padding */, 0 /* left padding */, |
| 1343 | 1 /* kernel height */, 1 /* kernel width */, |
| 1344 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1345 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1346 | 1 /* groups */, |
| 1347 | 32 /* input channels per group */, |
| 1348 | 120 /* output_channels_per_group */, |
| 1349 | 32 /* input pixel stride */, |
| 1350 | 120 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1351 | w158.data(), w159.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1352 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 1353 | 0 /* flags */, |
| 1354 | &op50); |
| 1355 | if (status != xnn_status_success) { |
| 1356 | std::cerr << "failed to create operation #50" << std::endl; |
| 1357 | return ExecutionPlan(); |
| 1358 | } |
| 1359 | operators.emplace_back(op50, xnn_delete_operator); |
| 1360 | |
| 1361 | xnn_operator_t op51 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1362 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1363 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1364 | 0 /* flags */, |
| 1365 | &op51); |
| 1366 | if (status != xnn_status_success) { |
| 1367 | std::cerr << "failed to create operation #51" << std::endl; |
| 1368 | return ExecutionPlan(); |
| 1369 | } |
| 1370 | operators.emplace_back(op51, xnn_delete_operator); |
| 1371 | |
| 1372 | xnn_operator_t op52 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1373 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1374 | 0 /* top padding */, 0 /* right padding */, |
| 1375 | 0 /* bottom padding */, 0 /* left padding */, |
| 1376 | 1 /* kernel height */, 1 /* kernel width */, |
| 1377 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1378 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1379 | 1 /* groups */, |
| 1380 | 120 /* input channels per group */, |
| 1381 | 48 /* output_channels_per_group */, |
| 1382 | 120 /* input pixel stride */, |
| 1383 | 48 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1384 | w160.data(), w161.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1385 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1386 | 0 /* flags */, |
| 1387 | &op52); |
| 1388 | if (status != xnn_status_success) { |
| 1389 | std::cerr << "failed to create operation #52" << std::endl; |
| 1390 | return ExecutionPlan(); |
| 1391 | } |
| 1392 | operators.emplace_back(op52, xnn_delete_operator); |
| 1393 | |
| 1394 | xnn_operator_t op53 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1395 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1396 | 0 /* top padding */, 0 /* right padding */, |
| 1397 | 0 /* bottom padding */, 0 /* left padding */, |
| 1398 | 1 /* kernel height */, 1 /* kernel width */, |
| 1399 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1400 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1401 | 1 /* groups */, |
| 1402 | 48 /* input channels per group */, |
| 1403 | 144 /* output_channels_per_group */, |
| 1404 | 48 /* input pixel stride */, |
| 1405 | 144 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1406 | w162.data(), w163.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1407 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1408 | 0 /* flags */, |
| 1409 | &op53); |
| 1410 | if (status != xnn_status_success) { |
| 1411 | std::cerr << "failed to create operation #53" << std::endl; |
| 1412 | return ExecutionPlan(); |
| 1413 | } |
| 1414 | operators.emplace_back(op53, xnn_delete_operator); |
| 1415 | |
| 1416 | xnn_operator_t op54 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1417 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1418 | 144 /* channels */, |
| 1419 | 144 /* input stride */, |
| 1420 | 144 /* output stride */, |
| 1421 | 0 /* flags */, |
| 1422 | &op54); |
| 1423 | if (status != xnn_status_success) { |
| 1424 | std::cerr << "failed to create operation #54" << std::endl; |
| 1425 | return ExecutionPlan(); |
| 1426 | } |
| 1427 | operators.emplace_back(op54, xnn_delete_operator); |
| 1428 | |
| 1429 | xnn_operator_t op55 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1430 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1431 | 2 /* top padding */, 2 /* right padding */, |
| 1432 | 2 /* bottom padding */, 2 /* left padding */, |
| 1433 | 5 /* kernel height */, 5 /* kernel width */, |
| 1434 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1435 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1436 | 144 /* groups */, |
| 1437 | 1 /* input channels per group */, |
| 1438 | 1 /* output_channels_per_group */, |
| 1439 | 144 /* input pixel stride */, |
| 1440 | 144 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1441 | w164.data(), w165.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1442 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1443 | 0 /* flags */, |
| 1444 | &op55); |
| 1445 | if (status != xnn_status_success) { |
| 1446 | std::cerr << "failed to create operation #55" << std::endl; |
| 1447 | return ExecutionPlan(); |
| 1448 | } |
| 1449 | operators.emplace_back(op55, xnn_delete_operator); |
| 1450 | |
| 1451 | xnn_operator_t op56 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1452 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1453 | 144 /* channels */, |
| 1454 | 144 /* input stride */, |
| 1455 | 144 /* output stride */, |
| 1456 | 0 /* flags */, |
| 1457 | &op56); |
| 1458 | if (status != xnn_status_success) { |
| 1459 | std::cerr << "failed to create operation #56" << std::endl; |
| 1460 | return ExecutionPlan(); |
| 1461 | } |
| 1462 | operators.emplace_back(op56, xnn_delete_operator); |
| 1463 | |
| 1464 | xnn_operator_t op57 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1465 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1466 | 144 /* channels */, 144 /* input stride */, 144 /* output stride */, |
| 1467 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 1468 | 0 /* flags */, |
| 1469 | &op57); |
| 1470 | if (status != xnn_status_success) { |
| 1471 | std::cerr << "failed to create operation #57" << std::endl; |
| 1472 | return ExecutionPlan(); |
| 1473 | } |
| 1474 | operators.emplace_back(op57, xnn_delete_operator); |
| 1475 | |
| 1476 | xnn_operator_t op58 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1477 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1478 | 0 /* top padding */, 0 /* right padding */, |
| 1479 | 0 /* bottom padding */, 0 /* left padding */, |
| 1480 | 1 /* kernel height */, 1 /* kernel width */, |
| 1481 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1482 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1483 | 1 /* groups */, |
| 1484 | 144 /* input channels per group */, |
| 1485 | 40 /* output_channels_per_group */, |
| 1486 | 144 /* input pixel stride */, |
| 1487 | 40 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1488 | w166.data(), w167.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1489 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1490 | 0 /* flags */, |
| 1491 | &op58); |
| 1492 | if (status != xnn_status_success) { |
| 1493 | std::cerr << "failed to create operation #58" << std::endl; |
| 1494 | return ExecutionPlan(); |
| 1495 | } |
| 1496 | operators.emplace_back(op58, xnn_delete_operator); |
| 1497 | |
| 1498 | xnn_operator_t op59 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1499 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1500 | 0 /* top padding */, 0 /* right padding */, |
| 1501 | 0 /* bottom padding */, 0 /* left padding */, |
| 1502 | 1 /* kernel height */, 1 /* kernel width */, |
| 1503 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1504 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1505 | 1 /* groups */, |
| 1506 | 40 /* input channels per group */, |
| 1507 | 144 /* output_channels_per_group */, |
| 1508 | 40 /* input pixel stride */, |
| 1509 | 144 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1510 | w168.data(), w169.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1511 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 1512 | 0 /* flags */, |
| 1513 | &op59); |
| 1514 | if (status != xnn_status_success) { |
| 1515 | std::cerr << "failed to create operation #59" << std::endl; |
| 1516 | return ExecutionPlan(); |
| 1517 | } |
| 1518 | operators.emplace_back(op59, xnn_delete_operator); |
| 1519 | |
| 1520 | xnn_operator_t op60 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1521 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1522 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1523 | 0 /* flags */, |
| 1524 | &op60); |
| 1525 | if (status != xnn_status_success) { |
| 1526 | std::cerr << "failed to create operation #60" << std::endl; |
| 1527 | return ExecutionPlan(); |
| 1528 | } |
| 1529 | operators.emplace_back(op60, xnn_delete_operator); |
| 1530 | |
| 1531 | xnn_operator_t op61 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1532 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1533 | 0 /* top padding */, 0 /* right padding */, |
| 1534 | 0 /* bottom padding */, 0 /* left padding */, |
| 1535 | 1 /* kernel height */, 1 /* kernel width */, |
| 1536 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1537 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1538 | 1 /* groups */, |
| 1539 | 144 /* input channels per group */, |
| 1540 | 48 /* output_channels_per_group */, |
| 1541 | 144 /* input pixel stride */, |
| 1542 | 48 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1543 | w170.data(), w171.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1544 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1545 | 0 /* flags */, |
| 1546 | &op61); |
| 1547 | if (status != xnn_status_success) { |
| 1548 | std::cerr << "failed to create operation #61" << std::endl; |
| 1549 | return ExecutionPlan(); |
| 1550 | } |
| 1551 | operators.emplace_back(op61, xnn_delete_operator); |
| 1552 | |
| 1553 | xnn_operator_t op62 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1554 | status = xnn_create_add_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1555 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1556 | 0 /* flags */, |
| 1557 | &op62); |
| 1558 | if (status != xnn_status_success) { |
| 1559 | std::cerr << "failed to create operation #62" << std::endl; |
| 1560 | return ExecutionPlan(); |
| 1561 | } |
| 1562 | operators.emplace_back(op62, xnn_delete_operator); |
| 1563 | |
| 1564 | xnn_operator_t op63 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1565 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1566 | 0 /* top padding */, 0 /* right padding */, |
| 1567 | 0 /* bottom padding */, 0 /* left padding */, |
| 1568 | 1 /* kernel height */, 1 /* kernel width */, |
| 1569 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1570 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1571 | 1 /* groups */, |
| 1572 | 48 /* input channels per group */, |
| 1573 | 288 /* output_channels_per_group */, |
| 1574 | 48 /* input pixel stride */, |
| 1575 | 288 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1576 | w172.data(), w173.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1577 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1578 | 0 /* flags */, |
| 1579 | &op63); |
| 1580 | if (status != xnn_status_success) { |
| 1581 | std::cerr << "failed to create operation #63" << std::endl; |
| 1582 | return ExecutionPlan(); |
| 1583 | } |
| 1584 | operators.emplace_back(op63, xnn_delete_operator); |
| 1585 | |
| 1586 | xnn_operator_t op64 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1587 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1588 | 288 /* channels */, |
| 1589 | 288 /* input stride */, |
| 1590 | 288 /* output stride */, |
| 1591 | 0 /* flags */, |
| 1592 | &op64); |
| 1593 | if (status != xnn_status_success) { |
| 1594 | std::cerr << "failed to create operation #64" << std::endl; |
| 1595 | return ExecutionPlan(); |
| 1596 | } |
| 1597 | operators.emplace_back(op64, xnn_delete_operator); |
| 1598 | |
| 1599 | xnn_operator_t op65 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1600 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1601 | 1 /* top padding */, 2 /* right padding */, |
| 1602 | 2 /* bottom padding */, 1 /* left padding */, |
| 1603 | 5 /* kernel height */, 5 /* kernel width */, |
| 1604 | 2 /* subsampling height */, 2 /* subsampling width */, |
| 1605 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1606 | 288 /* groups */, |
| 1607 | 1 /* input channels per group */, |
| 1608 | 1 /* output_channels_per_group */, |
| 1609 | 288 /* input pixel stride */, |
| 1610 | 288 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1611 | w174.data(), w175.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1612 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1613 | 0 /* flags */, |
| 1614 | &op65); |
| 1615 | if (status != xnn_status_success) { |
| 1616 | std::cerr << "failed to create operation #65" << std::endl; |
| 1617 | return ExecutionPlan(); |
| 1618 | } |
| 1619 | operators.emplace_back(op65, xnn_delete_operator); |
| 1620 | |
| 1621 | xnn_operator_t op66 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1622 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1623 | 288 /* channels */, |
| 1624 | 288 /* input stride */, |
| 1625 | 288 /* output stride */, |
| 1626 | 0 /* flags */, |
| 1627 | &op66); |
| 1628 | if (status != xnn_status_success) { |
| 1629 | std::cerr << "failed to create operation #66" << std::endl; |
| 1630 | return ExecutionPlan(); |
| 1631 | } |
| 1632 | operators.emplace_back(op66, xnn_delete_operator); |
| 1633 | |
| 1634 | xnn_operator_t op67 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1635 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1636 | 288 /* channels */, 288 /* input stride */, 288 /* output stride */, |
| 1637 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 1638 | 0 /* flags */, |
| 1639 | &op67); |
| 1640 | if (status != xnn_status_success) { |
| 1641 | std::cerr << "failed to create operation #67" << std::endl; |
| 1642 | return ExecutionPlan(); |
| 1643 | } |
| 1644 | operators.emplace_back(op67, xnn_delete_operator); |
| 1645 | |
| 1646 | xnn_operator_t op68 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1647 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1648 | 0 /* top padding */, 0 /* right padding */, |
| 1649 | 0 /* bottom padding */, 0 /* left padding */, |
| 1650 | 1 /* kernel height */, 1 /* kernel width */, |
| 1651 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1652 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1653 | 1 /* groups */, |
| 1654 | 288 /* input channels per group */, |
| 1655 | 72 /* output_channels_per_group */, |
| 1656 | 288 /* input pixel stride */, |
| 1657 | 72 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1658 | w176.data(), w177.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1659 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1660 | 0 /* flags */, |
| 1661 | &op68); |
| 1662 | if (status != xnn_status_success) { |
| 1663 | std::cerr << "failed to create operation #68" << std::endl; |
| 1664 | return ExecutionPlan(); |
| 1665 | } |
| 1666 | operators.emplace_back(op68, xnn_delete_operator); |
| 1667 | |
| 1668 | xnn_operator_t op69 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1669 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1670 | 0 /* top padding */, 0 /* right padding */, |
| 1671 | 0 /* bottom padding */, 0 /* left padding */, |
| 1672 | 1 /* kernel height */, 1 /* kernel width */, |
| 1673 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1674 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1675 | 1 /* groups */, |
| 1676 | 72 /* input channels per group */, |
| 1677 | 288 /* output_channels_per_group */, |
| 1678 | 72 /* input pixel stride */, |
| 1679 | 288 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1680 | w178.data(), w179.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1681 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 1682 | 0 /* flags */, |
| 1683 | &op69); |
| 1684 | if (status != xnn_status_success) { |
| 1685 | std::cerr << "failed to create operation #69" << std::endl; |
| 1686 | return ExecutionPlan(); |
| 1687 | } |
| 1688 | operators.emplace_back(op69, xnn_delete_operator); |
| 1689 | |
| 1690 | xnn_operator_t op70 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1691 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1692 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1693 | 0 /* flags */, |
| 1694 | &op70); |
| 1695 | if (status != xnn_status_success) { |
| 1696 | std::cerr << "failed to create operation #70" << std::endl; |
| 1697 | return ExecutionPlan(); |
| 1698 | } |
| 1699 | operators.emplace_back(op70, xnn_delete_operator); |
| 1700 | |
| 1701 | xnn_operator_t op71 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1702 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1703 | 0 /* top padding */, 0 /* right padding */, |
| 1704 | 0 /* bottom padding */, 0 /* left padding */, |
| 1705 | 1 /* kernel height */, 1 /* kernel width */, |
| 1706 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1707 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1708 | 1 /* groups */, |
| 1709 | 288 /* input channels per group */, |
| 1710 | 96 /* output_channels_per_group */, |
| 1711 | 288 /* input pixel stride */, |
| 1712 | 96 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1713 | w180.data(), w181.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1714 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1715 | 0 /* flags */, |
| 1716 | &op71); |
| 1717 | if (status != xnn_status_success) { |
| 1718 | std::cerr << "failed to create operation #71" << std::endl; |
| 1719 | return ExecutionPlan(); |
| 1720 | } |
| 1721 | operators.emplace_back(op71, xnn_delete_operator); |
| 1722 | |
| 1723 | xnn_operator_t op72 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1724 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1725 | 0 /* top padding */, 0 /* right padding */, |
| 1726 | 0 /* bottom padding */, 0 /* left padding */, |
| 1727 | 1 /* kernel height */, 1 /* kernel width */, |
| 1728 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1729 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1730 | 1 /* groups */, |
| 1731 | 96 /* input channels per group */, |
| 1732 | 576 /* output_channels_per_group */, |
| 1733 | 96 /* input pixel stride */, |
| 1734 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1735 | w182.data(), w183.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1736 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1737 | 0 /* flags */, |
| 1738 | &op72); |
| 1739 | if (status != xnn_status_success) { |
| 1740 | std::cerr << "failed to create operation #72" << std::endl; |
| 1741 | return ExecutionPlan(); |
| 1742 | } |
| 1743 | operators.emplace_back(op72, xnn_delete_operator); |
| 1744 | |
| 1745 | xnn_operator_t op73 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1746 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1747 | 576 /* channels */, |
| 1748 | 576 /* input stride */, |
| 1749 | 576 /* output stride */, |
| 1750 | 0 /* flags */, |
| 1751 | &op73); |
| 1752 | if (status != xnn_status_success) { |
| 1753 | std::cerr << "failed to create operation #73" << std::endl; |
| 1754 | return ExecutionPlan(); |
| 1755 | } |
| 1756 | operators.emplace_back(op73, xnn_delete_operator); |
| 1757 | |
| 1758 | xnn_operator_t op74 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1759 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1760 | 2 /* top padding */, 2 /* right padding */, |
| 1761 | 2 /* bottom padding */, 2 /* left padding */, |
| 1762 | 5 /* kernel height */, 5 /* kernel width */, |
| 1763 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1764 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1765 | 576 /* groups */, |
| 1766 | 1 /* input channels per group */, |
| 1767 | 1 /* output_channels_per_group */, |
| 1768 | 576 /* input pixel stride */, |
| 1769 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1770 | w184.data(), w185.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1771 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1772 | 0 /* flags */, |
| 1773 | &op74); |
| 1774 | if (status != xnn_status_success) { |
| 1775 | std::cerr << "failed to create operation #74" << std::endl; |
| 1776 | return ExecutionPlan(); |
| 1777 | } |
| 1778 | operators.emplace_back(op74, xnn_delete_operator); |
| 1779 | |
| 1780 | xnn_operator_t op75 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1781 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1782 | 576 /* channels */, |
| 1783 | 576 /* input stride */, |
| 1784 | 576 /* output stride */, |
| 1785 | 0 /* flags */, |
| 1786 | &op75); |
| 1787 | if (status != xnn_status_success) { |
| 1788 | std::cerr << "failed to create operation #75" << std::endl; |
| 1789 | return ExecutionPlan(); |
| 1790 | } |
| 1791 | operators.emplace_back(op75, xnn_delete_operator); |
| 1792 | |
| 1793 | xnn_operator_t op76 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1794 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1795 | 576 /* channels */, 576 /* input stride */, 576 /* output stride */, |
| 1796 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 1797 | 0 /* flags */, |
| 1798 | &op76); |
| 1799 | if (status != xnn_status_success) { |
| 1800 | std::cerr << "failed to create operation #76" << std::endl; |
| 1801 | return ExecutionPlan(); |
| 1802 | } |
| 1803 | operators.emplace_back(op76, xnn_delete_operator); |
| 1804 | |
| 1805 | xnn_operator_t op77 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1806 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1807 | 0 /* top padding */, 0 /* right padding */, |
| 1808 | 0 /* bottom padding */, 0 /* left padding */, |
| 1809 | 1 /* kernel height */, 1 /* kernel width */, |
| 1810 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1811 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1812 | 1 /* groups */, |
| 1813 | 576 /* input channels per group */, |
| 1814 | 144 /* output_channels_per_group */, |
| 1815 | 576 /* input pixel stride */, |
| 1816 | 144 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1817 | w186.data(), w187.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1818 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1819 | 0 /* flags */, |
| 1820 | &op77); |
| 1821 | if (status != xnn_status_success) { |
| 1822 | std::cerr << "failed to create operation #77" << std::endl; |
| 1823 | return ExecutionPlan(); |
| 1824 | } |
| 1825 | operators.emplace_back(op77, xnn_delete_operator); |
| 1826 | |
| 1827 | xnn_operator_t op78 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1828 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1829 | 0 /* top padding */, 0 /* right padding */, |
| 1830 | 0 /* bottom padding */, 0 /* left padding */, |
| 1831 | 1 /* kernel height */, 1 /* kernel width */, |
| 1832 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1833 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1834 | 1 /* groups */, |
| 1835 | 144 /* input channels per group */, |
| 1836 | 576 /* output_channels_per_group */, |
| 1837 | 144 /* input pixel stride */, |
| 1838 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1839 | w188.data(), w189.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1840 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 1841 | 0 /* flags */, |
| 1842 | &op78); |
| 1843 | if (status != xnn_status_success) { |
| 1844 | std::cerr << "failed to create operation #78" << std::endl; |
| 1845 | return ExecutionPlan(); |
| 1846 | } |
| 1847 | operators.emplace_back(op78, xnn_delete_operator); |
| 1848 | |
| 1849 | xnn_operator_t op79 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1850 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1851 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1852 | 0 /* flags */, |
| 1853 | &op79); |
| 1854 | if (status != xnn_status_success) { |
| 1855 | std::cerr << "failed to create operation #79" << std::endl; |
| 1856 | return ExecutionPlan(); |
| 1857 | } |
| 1858 | operators.emplace_back(op79, xnn_delete_operator); |
| 1859 | |
| 1860 | xnn_operator_t op80 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1861 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1862 | 0 /* top padding */, 0 /* right padding */, |
| 1863 | 0 /* bottom padding */, 0 /* left padding */, |
| 1864 | 1 /* kernel height */, 1 /* kernel width */, |
| 1865 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1866 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1867 | 1 /* groups */, |
| 1868 | 576 /* input channels per group */, |
| 1869 | 96 /* output_channels_per_group */, |
| 1870 | 576 /* input pixel stride */, |
| 1871 | 96 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1872 | w190.data(), w191.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1873 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1874 | 0 /* flags */, |
| 1875 | &op80); |
| 1876 | if (status != xnn_status_success) { |
| 1877 | std::cerr << "failed to create operation #80" << std::endl; |
| 1878 | return ExecutionPlan(); |
| 1879 | } |
| 1880 | operators.emplace_back(op80, xnn_delete_operator); |
| 1881 | |
| 1882 | xnn_operator_t op81 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1883 | status = xnn_create_add_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1884 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1885 | 0 /* flags */, |
| 1886 | &op81); |
| 1887 | if (status != xnn_status_success) { |
| 1888 | std::cerr << "failed to create operation #81" << std::endl; |
| 1889 | return ExecutionPlan(); |
| 1890 | } |
| 1891 | operators.emplace_back(op81, xnn_delete_operator); |
| 1892 | |
| 1893 | xnn_operator_t op82 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1894 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1895 | 0 /* top padding */, 0 /* right padding */, |
| 1896 | 0 /* bottom padding */, 0 /* left padding */, |
| 1897 | 1 /* kernel height */, 1 /* kernel width */, |
| 1898 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1899 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1900 | 1 /* groups */, |
| 1901 | 96 /* input channels per group */, |
| 1902 | 576 /* output_channels_per_group */, |
| 1903 | 96 /* input pixel stride */, |
| 1904 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1905 | w192.data(), w193.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1906 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1907 | 0 /* flags */, |
| 1908 | &op82); |
| 1909 | if (status != xnn_status_success) { |
| 1910 | std::cerr << "failed to create operation #82" << std::endl; |
| 1911 | return ExecutionPlan(); |
| 1912 | } |
| 1913 | operators.emplace_back(op82, xnn_delete_operator); |
| 1914 | |
| 1915 | xnn_operator_t op83 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1916 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1917 | 576 /* channels */, |
| 1918 | 576 /* input stride */, |
| 1919 | 576 /* output stride */, |
| 1920 | 0 /* flags */, |
| 1921 | &op83); |
| 1922 | if (status != xnn_status_success) { |
| 1923 | std::cerr << "failed to create operation #83" << std::endl; |
| 1924 | return ExecutionPlan(); |
| 1925 | } |
| 1926 | operators.emplace_back(op83, xnn_delete_operator); |
| 1927 | |
| 1928 | xnn_operator_t op84 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1929 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1930 | 2 /* top padding */, 2 /* right padding */, |
| 1931 | 2 /* bottom padding */, 2 /* left padding */, |
| 1932 | 5 /* kernel height */, 5 /* kernel width */, |
| 1933 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1934 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1935 | 576 /* groups */, |
| 1936 | 1 /* input channels per group */, |
| 1937 | 1 /* output_channels_per_group */, |
| 1938 | 576 /* input pixel stride */, |
| 1939 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1940 | w194.data(), w195.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1941 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1942 | 0 /* flags */, |
| 1943 | &op84); |
| 1944 | if (status != xnn_status_success) { |
| 1945 | std::cerr << "failed to create operation #84" << std::endl; |
| 1946 | return ExecutionPlan(); |
| 1947 | } |
| 1948 | operators.emplace_back(op84, xnn_delete_operator); |
| 1949 | |
| 1950 | xnn_operator_t op85 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1951 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1952 | 576 /* channels */, |
| 1953 | 576 /* input stride */, |
| 1954 | 576 /* output stride */, |
| 1955 | 0 /* flags */, |
| 1956 | &op85); |
| 1957 | if (status != xnn_status_success) { |
| 1958 | std::cerr << "failed to create operation #85" << std::endl; |
| 1959 | return ExecutionPlan(); |
| 1960 | } |
| 1961 | operators.emplace_back(op85, xnn_delete_operator); |
| 1962 | |
| 1963 | xnn_operator_t op86 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1964 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1965 | 576 /* channels */, 576 /* input stride */, 576 /* output stride */, |
| 1966 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 1967 | 0 /* flags */, |
| 1968 | &op86); |
| 1969 | if (status != xnn_status_success) { |
| 1970 | std::cerr << "failed to create operation #86" << std::endl; |
| 1971 | return ExecutionPlan(); |
| 1972 | } |
| 1973 | operators.emplace_back(op86, xnn_delete_operator); |
| 1974 | |
| 1975 | xnn_operator_t op87 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1976 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1977 | 0 /* top padding */, 0 /* right padding */, |
| 1978 | 0 /* bottom padding */, 0 /* left padding */, |
| 1979 | 1 /* kernel height */, 1 /* kernel width */, |
| 1980 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 1981 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 1982 | 1 /* groups */, |
| 1983 | 576 /* input channels per group */, |
| 1984 | 144 /* output_channels_per_group */, |
| 1985 | 576 /* input pixel stride */, |
| 1986 | 144 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 1987 | w196.data(), w197.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1988 | 0.0f /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 1989 | 0 /* flags */, |
| 1990 | &op87); |
| 1991 | if (status != xnn_status_success) { |
| 1992 | std::cerr << "failed to create operation #87" << std::endl; |
| 1993 | return ExecutionPlan(); |
| 1994 | } |
| 1995 | operators.emplace_back(op87, xnn_delete_operator); |
| 1996 | |
| 1997 | xnn_operator_t op88 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 1998 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 1999 | 0 /* top padding */, 0 /* right padding */, |
| 2000 | 0 /* bottom padding */, 0 /* left padding */, |
| 2001 | 1 /* kernel height */, 1 /* kernel width */, |
| 2002 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 2003 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 2004 | 1 /* groups */, |
| 2005 | 144 /* input channels per group */, |
| 2006 | 576 /* output_channels_per_group */, |
| 2007 | 144 /* input pixel stride */, |
| 2008 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2009 | w198.data(), w199.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2010 | 0.0f /* output min */, +0x1.00014Fp+0 /* output max */, |
| 2011 | 0 /* flags */, |
| 2012 | &op88); |
| 2013 | if (status != xnn_status_success) { |
| 2014 | std::cerr << "failed to create operation #88" << std::endl; |
| 2015 | return ExecutionPlan(); |
| 2016 | } |
| 2017 | operators.emplace_back(op88, xnn_delete_operator); |
| 2018 | |
| 2019 | xnn_operator_t op89 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2020 | status = xnn_create_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2021 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 2022 | 0 /* flags */, |
| 2023 | &op89); |
| 2024 | if (status != xnn_status_success) { |
| 2025 | std::cerr << "failed to create operation #89" << std::endl; |
| 2026 | return ExecutionPlan(); |
| 2027 | } |
| 2028 | operators.emplace_back(op89, xnn_delete_operator); |
| 2029 | |
| 2030 | xnn_operator_t op90 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2031 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2032 | 0 /* top padding */, 0 /* right padding */, |
| 2033 | 0 /* bottom padding */, 0 /* left padding */, |
| 2034 | 1 /* kernel height */, 1 /* kernel width */, |
| 2035 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 2036 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 2037 | 1 /* groups */, |
| 2038 | 576 /* input channels per group */, |
| 2039 | 96 /* output_channels_per_group */, |
| 2040 | 576 /* input pixel stride */, |
| 2041 | 96 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2042 | w200.data(), w201.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2043 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 2044 | 0 /* flags */, |
| 2045 | &op90); |
| 2046 | if (status != xnn_status_success) { |
| 2047 | std::cerr << "failed to create operation #90" << std::endl; |
| 2048 | return ExecutionPlan(); |
| 2049 | } |
| 2050 | operators.emplace_back(op90, xnn_delete_operator); |
| 2051 | |
| 2052 | xnn_operator_t op91 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2053 | status = xnn_create_add_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2054 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 2055 | 0 /* flags */, |
| 2056 | &op91); |
| 2057 | if (status != xnn_status_success) { |
| 2058 | std::cerr << "failed to create operation #91" << std::endl; |
| 2059 | return ExecutionPlan(); |
| 2060 | } |
| 2061 | operators.emplace_back(op91, xnn_delete_operator); |
| 2062 | |
| 2063 | xnn_operator_t op92 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2064 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2065 | 0 /* top padding */, 0 /* right padding */, |
| 2066 | 0 /* bottom padding */, 0 /* left padding */, |
| 2067 | 1 /* kernel height */, 1 /* kernel width */, |
| 2068 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 2069 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 2070 | 1 /* groups */, |
| 2071 | 96 /* input channels per group */, |
| 2072 | 576 /* output_channels_per_group */, |
| 2073 | 96 /* input pixel stride */, |
| 2074 | 576 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2075 | w202.data(), w203.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2076 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 2077 | 0 /* flags */, |
| 2078 | &op92); |
| 2079 | if (status != xnn_status_success) { |
| 2080 | std::cerr << "failed to create operation #92" << std::endl; |
| 2081 | return ExecutionPlan(); |
| 2082 | } |
| 2083 | operators.emplace_back(op92, xnn_delete_operator); |
| 2084 | |
| 2085 | xnn_operator_t op93 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2086 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2087 | 576 /* channels */, |
| 2088 | 576 /* input stride */, |
| 2089 | 576 /* output stride */, |
| 2090 | 0 /* flags */, |
| 2091 | &op93); |
| 2092 | if (status != xnn_status_success) { |
| 2093 | std::cerr << "failed to create operation #93" << std::endl; |
| 2094 | return ExecutionPlan(); |
| 2095 | } |
| 2096 | operators.emplace_back(op93, xnn_delete_operator); |
| 2097 | |
| 2098 | xnn_operator_t op94 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2099 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2100 | 576 /* channels */, 576 /* input stride */, 576 /* output stride */, |
| 2101 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 2102 | 0 /* flags */, |
| 2103 | &op94); |
| 2104 | if (status != xnn_status_success) { |
| 2105 | std::cerr << "failed to create operation #94" << std::endl; |
| 2106 | return ExecutionPlan(); |
| 2107 | } |
| 2108 | operators.emplace_back(op94, xnn_delete_operator); |
| 2109 | |
| 2110 | xnn_operator_t op95 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2111 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2112 | 0 /* top padding */, 0 /* right padding */, |
| 2113 | 0 /* bottom padding */, 0 /* left padding */, |
| 2114 | 1 /* kernel height */, 1 /* kernel width */, |
| 2115 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 2116 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 2117 | 1 /* groups */, |
| 2118 | 576 /* input channels per group */, |
| 2119 | 1024 /* output_channels_per_group */, |
| 2120 | 576 /* input pixel stride */, |
| 2121 | 1024 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2122 | w204.data(), w205.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2123 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 2124 | 0 /* flags */, |
| 2125 | &op95); |
| 2126 | if (status != xnn_status_success) { |
| 2127 | std::cerr << "failed to create operation #95" << std::endl; |
| 2128 | return ExecutionPlan(); |
| 2129 | } |
| 2130 | operators.emplace_back(op95, xnn_delete_operator); |
| 2131 | |
| 2132 | xnn_operator_t op96 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2133 | status = xnn_create_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2134 | 1024 /* channels */, |
| 2135 | 1024 /* input stride */, |
| 2136 | 1024 /* output stride */, |
| 2137 | 0 /* flags */, |
| 2138 | &op96); |
| 2139 | if (status != xnn_status_success) { |
| 2140 | std::cerr << "failed to create operation #96" << std::endl; |
| 2141 | return ExecutionPlan(); |
| 2142 | } |
| 2143 | operators.emplace_back(op96, xnn_delete_operator); |
| 2144 | |
| 2145 | xnn_operator_t op97 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2146 | status = xnn_create_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2147 | 1024 /* channels */, 1024 /* input stride */, 1024 /* output stride */, |
| 2148 | -std::numeric_limits<float>::infinity(), std::numeric_limits<float>::infinity(), |
| 2149 | 0 /* flags */, |
| 2150 | &op97); |
| 2151 | if (status != xnn_status_success) { |
| 2152 | std::cerr << "failed to create operation #97" << std::endl; |
| 2153 | return ExecutionPlan(); |
| 2154 | } |
| 2155 | operators.emplace_back(op97, xnn_delete_operator); |
| 2156 | |
| 2157 | xnn_operator_t op98 = nullptr; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2158 | status = xnn_create_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2159 | 0 /* top padding */, 0 /* right padding */, |
| 2160 | 0 /* bottom padding */, 0 /* left padding */, |
| 2161 | 1 /* kernel height */, 1 /* kernel width */, |
| 2162 | 1 /* subsampling height */, 1 /* subsampling width */, |
| 2163 | 1 /* dilation_height */, 1 /* dilation_width */, |
| 2164 | 1 /* groups */, |
| 2165 | 1024 /* input channels per group */, |
| 2166 | 1001 /* output_channels_per_group */, |
| 2167 | 1024 /* input pixel stride */, |
| 2168 | 1001 /* output pixel stride */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2169 | w206.data(), w207.data(), |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2170 | -std::numeric_limits<float>::infinity() /* output min */, std::numeric_limits<float>::infinity() /* output max */, |
| 2171 | 0 /* flags */, |
| 2172 | &op98); |
| 2173 | if (status != xnn_status_success) { |
| 2174 | std::cerr << "failed to create operation #98" << std::endl; |
| 2175 | return ExecutionPlan(); |
| 2176 | } |
| 2177 | operators.emplace_back(op98, xnn_delete_operator); |
| 2178 | |
| 2179 | |
| 2180 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2181 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2182 | op0, |
| 2183 | 1 /* batch size */, 224 /* input height */, 224 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2184 | v0.data() /* input */, v1.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2185 | threadpool /* threadpool */); |
| 2186 | if (status != xnn_status_success) { |
| 2187 | std::cerr << "failed to setup operation #0" << std::endl; |
| 2188 | return ExecutionPlan(); |
| 2189 | } |
| 2190 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2191 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2192 | op1, |
| 2193 | 12544 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2194 | v1.data() /* input */, v2.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2195 | threadpool /* threadpool */); |
| 2196 | if (status != xnn_status_success) { |
| 2197 | std::cerr << "failed to setup operation #1" << std::endl; |
| 2198 | return ExecutionPlan(); |
| 2199 | } |
| 2200 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2201 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2202 | op2, |
| 2203 | 1 /* batch size */, 112 /* input height */, 112 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2204 | v2.data() /* input */, v3.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2205 | threadpool /* threadpool */); |
| 2206 | if (status != xnn_status_success) { |
| 2207 | std::cerr << "failed to setup operation #2" << std::endl; |
| 2208 | return ExecutionPlan(); |
| 2209 | } |
| 2210 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2211 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2212 | op3, |
| 2213 | 1 /* batch size */, 3136 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2214 | v3.data() /* input */, v4.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2215 | threadpool /* threadpool */); |
| 2216 | if (status != xnn_status_success) { |
| 2217 | std::cerr << "failed to setup operation #3" << std::endl; |
| 2218 | return ExecutionPlan(); |
| 2219 | } |
| 2220 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2221 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2222 | op4, |
| 2223 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2224 | v4.data() /* input */, v5.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2225 | threadpool /* threadpool */); |
| 2226 | if (status != xnn_status_success) { |
| 2227 | std::cerr << "failed to setup operation #4" << std::endl; |
| 2228 | return ExecutionPlan(); |
| 2229 | } |
| 2230 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2231 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2232 | op5, |
| 2233 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2234 | v5.data() /* input */, v6.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2235 | threadpool /* threadpool */); |
| 2236 | if (status != xnn_status_success) { |
| 2237 | std::cerr << "failed to setup operation #5" << std::endl; |
| 2238 | return ExecutionPlan(); |
| 2239 | } |
| 2240 | |
| 2241 | { |
| 2242 | const size_t a_shape[] = { 1, 56, 56, 16 }; |
| 2243 | const size_t b_shape[] = { 1, 1, 1, 16 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2244 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2245 | op6, |
| 2246 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2247 | v3.data() /* a */, v6.data() /* b */, v7.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2248 | threadpool /* threadpool */); |
| 2249 | } |
| 2250 | if (status != xnn_status_success) { |
| 2251 | std::cerr << "failed to setup operation #6" << std::endl; |
| 2252 | return ExecutionPlan(); |
| 2253 | } |
| 2254 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2255 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2256 | op7, |
| 2257 | 1 /* batch size */, 56 /* input height */, 56 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2258 | v7.data() /* input */, v8.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2259 | threadpool /* threadpool */); |
| 2260 | if (status != xnn_status_success) { |
| 2261 | std::cerr << "failed to setup operation #7" << std::endl; |
| 2262 | return ExecutionPlan(); |
| 2263 | } |
| 2264 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2265 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2266 | op8, |
| 2267 | 1 /* batch size */, 56 /* input height */, 56 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2268 | v8.data() /* input */, v9.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2269 | threadpool /* threadpool */); |
| 2270 | if (status != xnn_status_success) { |
| 2271 | std::cerr << "failed to setup operation #8" << std::endl; |
| 2272 | return ExecutionPlan(); |
| 2273 | } |
| 2274 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2275 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2276 | op9, |
| 2277 | 1 /* batch size */, 56 /* input height */, 56 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2278 | v9.data() /* input */, v10.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2279 | threadpool /* threadpool */); |
| 2280 | if (status != xnn_status_success) { |
| 2281 | std::cerr << "failed to setup operation #9" << std::endl; |
| 2282 | return ExecutionPlan(); |
| 2283 | } |
| 2284 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2285 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2286 | op10, |
| 2287 | 1 /* batch size */, 28 /* input height */, 28 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2288 | v10.data() /* input */, v11.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2289 | threadpool /* threadpool */); |
| 2290 | if (status != xnn_status_success) { |
| 2291 | std::cerr << "failed to setup operation #10" << std::endl; |
| 2292 | return ExecutionPlan(); |
| 2293 | } |
| 2294 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2295 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2296 | op11, |
| 2297 | 1 /* batch size */, 28 /* input height */, 28 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2298 | v11.data() /* input */, v12.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2299 | threadpool /* threadpool */); |
| 2300 | if (status != xnn_status_success) { |
| 2301 | std::cerr << "failed to setup operation #11" << std::endl; |
| 2302 | return ExecutionPlan(); |
| 2303 | } |
| 2304 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2305 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2306 | op12, |
| 2307 | 1 /* batch size */, 28 /* input height */, 28 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2308 | v12.data() /* input */, v13.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2309 | threadpool /* threadpool */); |
| 2310 | if (status != xnn_status_success) { |
| 2311 | std::cerr << "failed to setup operation #12" << std::endl; |
| 2312 | return ExecutionPlan(); |
| 2313 | } |
| 2314 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2315 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2316 | op13, |
| 2317 | 1 /* batch size */, 28 /* input height */, 28 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2318 | v13.data() /* input */, v14.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2319 | threadpool /* threadpool */); |
| 2320 | if (status != xnn_status_success) { |
| 2321 | std::cerr << "failed to setup operation #13" << std::endl; |
| 2322 | return ExecutionPlan(); |
| 2323 | } |
| 2324 | |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2325 | { |
| 2326 | const size_t a_shape[] = { 1, 28, 28, 24 }; |
| 2327 | const size_t b_shape[] = { 1, 28, 28, 24 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2328 | status = xnn_setup_add_nd_f16( |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2329 | op14, |
| 2330 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2331 | v14.data() /* a */, v11.data() /* b */, v15.data() /* output */, |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2332 | threadpool /* threadpool */); |
| 2333 | } |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2334 | if (status != xnn_status_success) { |
| 2335 | std::cerr << "failed to setup operation #14" << std::endl; |
| 2336 | return ExecutionPlan(); |
| 2337 | } |
| 2338 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2339 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2340 | op15, |
| 2341 | 1 /* batch size */, 28 /* input height */, 28 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2342 | v15.data() /* input */, v16.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2343 | threadpool /* threadpool */); |
| 2344 | if (status != xnn_status_success) { |
| 2345 | std::cerr << "failed to setup operation #15" << std::endl; |
| 2346 | return ExecutionPlan(); |
| 2347 | } |
| 2348 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2349 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2350 | op16, |
| 2351 | 784 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2352 | v16.data() /* input */, v17.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2353 | threadpool /* threadpool */); |
| 2354 | if (status != xnn_status_success) { |
| 2355 | std::cerr << "failed to setup operation #16" << std::endl; |
| 2356 | return ExecutionPlan(); |
| 2357 | } |
| 2358 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2359 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2360 | op17, |
| 2361 | 1 /* batch size */, 28 /* input height */, 28 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2362 | v17.data() /* input */, v18.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2363 | threadpool /* threadpool */); |
| 2364 | if (status != xnn_status_success) { |
| 2365 | std::cerr << "failed to setup operation #17" << std::endl; |
| 2366 | return ExecutionPlan(); |
| 2367 | } |
| 2368 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2369 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2370 | op18, |
| 2371 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2372 | v18.data() /* input */, v19.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2373 | threadpool /* threadpool */); |
| 2374 | if (status != xnn_status_success) { |
| 2375 | std::cerr << "failed to setup operation #18" << std::endl; |
| 2376 | return ExecutionPlan(); |
| 2377 | } |
| 2378 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2379 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2380 | op19, |
| 2381 | 1 /* batch size */, 196 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2382 | v19.data() /* input */, v20.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2383 | threadpool /* threadpool */); |
| 2384 | if (status != xnn_status_success) { |
| 2385 | std::cerr << "failed to setup operation #19" << std::endl; |
| 2386 | return ExecutionPlan(); |
| 2387 | } |
| 2388 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2389 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2390 | op20, |
| 2391 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2392 | v20.data() /* input */, v21.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2393 | threadpool /* threadpool */); |
| 2394 | if (status != xnn_status_success) { |
| 2395 | std::cerr << "failed to setup operation #20" << std::endl; |
| 2396 | return ExecutionPlan(); |
| 2397 | } |
| 2398 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2399 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2400 | op21, |
| 2401 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2402 | v21.data() /* input */, v22.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2403 | threadpool /* threadpool */); |
| 2404 | if (status != xnn_status_success) { |
| 2405 | std::cerr << "failed to setup operation #21" << std::endl; |
| 2406 | return ExecutionPlan(); |
| 2407 | } |
| 2408 | |
| 2409 | { |
| 2410 | const size_t a_shape[] = { 1, 14, 14, 96 }; |
| 2411 | const size_t b_shape[] = { 1, 1, 1, 96 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2412 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2413 | op22, |
| 2414 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2415 | v19.data() /* a */, v22.data() /* b */, v23.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2416 | threadpool /* threadpool */); |
| 2417 | } |
| 2418 | if (status != xnn_status_success) { |
| 2419 | std::cerr << "failed to setup operation #22" << std::endl; |
| 2420 | return ExecutionPlan(); |
| 2421 | } |
| 2422 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2423 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2424 | op23, |
| 2425 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2426 | v23.data() /* input */, v24.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2427 | threadpool /* threadpool */); |
| 2428 | if (status != xnn_status_success) { |
| 2429 | std::cerr << "failed to setup operation #23" << std::endl; |
| 2430 | return ExecutionPlan(); |
| 2431 | } |
| 2432 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2433 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2434 | op24, |
| 2435 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2436 | v24.data() /* input */, v25.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2437 | threadpool /* threadpool */); |
| 2438 | if (status != xnn_status_success) { |
| 2439 | std::cerr << "failed to setup operation #24" << std::endl; |
| 2440 | return ExecutionPlan(); |
| 2441 | } |
| 2442 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2443 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2444 | op25, |
| 2445 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2446 | v25.data() /* input */, v26.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2447 | threadpool /* threadpool */); |
| 2448 | if (status != xnn_status_success) { |
| 2449 | std::cerr << "failed to setup operation #25" << std::endl; |
| 2450 | return ExecutionPlan(); |
| 2451 | } |
| 2452 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2453 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2454 | op26, |
| 2455 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2456 | v26.data() /* input */, v27.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2457 | threadpool /* threadpool */); |
| 2458 | if (status != xnn_status_success) { |
| 2459 | std::cerr << "failed to setup operation #26" << std::endl; |
| 2460 | return ExecutionPlan(); |
| 2461 | } |
| 2462 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2463 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2464 | op27, |
| 2465 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2466 | v27.data() /* input */, v28.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2467 | threadpool /* threadpool */); |
| 2468 | if (status != xnn_status_success) { |
| 2469 | std::cerr << "failed to setup operation #27" << std::endl; |
| 2470 | return ExecutionPlan(); |
| 2471 | } |
| 2472 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2473 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2474 | op28, |
| 2475 | 1 /* batch size */, 196 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2476 | v28.data() /* input */, v29.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2477 | threadpool /* threadpool */); |
| 2478 | if (status != xnn_status_success) { |
| 2479 | std::cerr << "failed to setup operation #28" << std::endl; |
| 2480 | return ExecutionPlan(); |
| 2481 | } |
| 2482 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2483 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2484 | op29, |
| 2485 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2486 | v29.data() /* input */, v30.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2487 | threadpool /* threadpool */); |
| 2488 | if (status != xnn_status_success) { |
| 2489 | std::cerr << "failed to setup operation #29" << std::endl; |
| 2490 | return ExecutionPlan(); |
| 2491 | } |
| 2492 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2493 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2494 | op30, |
| 2495 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2496 | v30.data() /* input */, v31.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2497 | threadpool /* threadpool */); |
| 2498 | if (status != xnn_status_success) { |
| 2499 | std::cerr << "failed to setup operation #30" << std::endl; |
| 2500 | return ExecutionPlan(); |
| 2501 | } |
| 2502 | |
| 2503 | { |
| 2504 | const size_t a_shape[] = { 1, 14, 14, 240 }; |
| 2505 | const size_t b_shape[] = { 1, 1, 1, 240 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2506 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2507 | op31, |
| 2508 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2509 | v28.data() /* a */, v31.data() /* b */, v32.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2510 | threadpool /* threadpool */); |
| 2511 | } |
| 2512 | if (status != xnn_status_success) { |
| 2513 | std::cerr << "failed to setup operation #31" << std::endl; |
| 2514 | return ExecutionPlan(); |
| 2515 | } |
| 2516 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2517 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2518 | op32, |
| 2519 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2520 | v32.data() /* input */, v33.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2521 | threadpool /* threadpool */); |
| 2522 | if (status != xnn_status_success) { |
| 2523 | std::cerr << "failed to setup operation #32" << std::endl; |
| 2524 | return ExecutionPlan(); |
| 2525 | } |
| 2526 | |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2527 | { |
| 2528 | const size_t a_shape[] = { 1, 14, 14, 40 }; |
| 2529 | const size_t b_shape[] = { 1, 14, 14, 40 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2530 | status = xnn_setup_add_nd_f16( |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2531 | op33, |
| 2532 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2533 | v33.data() /* a */, v24.data() /* b */, v34.data() /* output */, |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2534 | threadpool /* threadpool */); |
| 2535 | } |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2536 | if (status != xnn_status_success) { |
| 2537 | std::cerr << "failed to setup operation #33" << std::endl; |
| 2538 | return ExecutionPlan(); |
| 2539 | } |
| 2540 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2541 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2542 | op34, |
| 2543 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2544 | v34.data() /* input */, v35.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2545 | threadpool /* threadpool */); |
| 2546 | if (status != xnn_status_success) { |
| 2547 | std::cerr << "failed to setup operation #34" << std::endl; |
| 2548 | return ExecutionPlan(); |
| 2549 | } |
| 2550 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2551 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2552 | op35, |
| 2553 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2554 | v35.data() /* input */, v36.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2555 | threadpool /* threadpool */); |
| 2556 | if (status != xnn_status_success) { |
| 2557 | std::cerr << "failed to setup operation #35" << std::endl; |
| 2558 | return ExecutionPlan(); |
| 2559 | } |
| 2560 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2561 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2562 | op36, |
| 2563 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2564 | v36.data() /* input */, v37.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2565 | threadpool /* threadpool */); |
| 2566 | if (status != xnn_status_success) { |
| 2567 | std::cerr << "failed to setup operation #36" << std::endl; |
| 2568 | return ExecutionPlan(); |
| 2569 | } |
| 2570 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2571 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2572 | op37, |
| 2573 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2574 | v37.data() /* input */, v38.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2575 | threadpool /* threadpool */); |
| 2576 | if (status != xnn_status_success) { |
| 2577 | std::cerr << "failed to setup operation #37" << std::endl; |
| 2578 | return ExecutionPlan(); |
| 2579 | } |
| 2580 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2581 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2582 | op38, |
| 2583 | 1 /* batch size */, 196 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2584 | v38.data() /* input */, v39.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2585 | threadpool /* threadpool */); |
| 2586 | if (status != xnn_status_success) { |
| 2587 | std::cerr << "failed to setup operation #38" << std::endl; |
| 2588 | return ExecutionPlan(); |
| 2589 | } |
| 2590 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2591 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2592 | op39, |
| 2593 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2594 | v39.data() /* input */, v40.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2595 | threadpool /* threadpool */); |
| 2596 | if (status != xnn_status_success) { |
| 2597 | std::cerr << "failed to setup operation #39" << std::endl; |
| 2598 | return ExecutionPlan(); |
| 2599 | } |
| 2600 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2601 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2602 | op40, |
| 2603 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2604 | v40.data() /* input */, v41.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2605 | threadpool /* threadpool */); |
| 2606 | if (status != xnn_status_success) { |
| 2607 | std::cerr << "failed to setup operation #40" << std::endl; |
| 2608 | return ExecutionPlan(); |
| 2609 | } |
| 2610 | |
| 2611 | { |
| 2612 | const size_t a_shape[] = { 1, 14, 14, 240 }; |
| 2613 | const size_t b_shape[] = { 1, 1, 1, 240 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2614 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2615 | op41, |
| 2616 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2617 | v38.data() /* a */, v41.data() /* b */, v42.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2618 | threadpool /* threadpool */); |
| 2619 | } |
| 2620 | if (status != xnn_status_success) { |
| 2621 | std::cerr << "failed to setup operation #41" << std::endl; |
| 2622 | return ExecutionPlan(); |
| 2623 | } |
| 2624 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2625 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2626 | op42, |
| 2627 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2628 | v42.data() /* input */, v43.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2629 | threadpool /* threadpool */); |
| 2630 | if (status != xnn_status_success) { |
| 2631 | std::cerr << "failed to setup operation #42" << std::endl; |
| 2632 | return ExecutionPlan(); |
| 2633 | } |
| 2634 | |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2635 | { |
| 2636 | const size_t a_shape[] = { 1, 14, 14, 40 }; |
| 2637 | const size_t b_shape[] = { 1, 14, 14, 40 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2638 | status = xnn_setup_add_nd_f16( |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2639 | op43, |
| 2640 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2641 | v43.data() /* a */, v34.data() /* b */, v44.data() /* output */, |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2642 | threadpool /* threadpool */); |
| 2643 | } |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2644 | if (status != xnn_status_success) { |
| 2645 | std::cerr << "failed to setup operation #43" << std::endl; |
| 2646 | return ExecutionPlan(); |
| 2647 | } |
| 2648 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2649 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2650 | op44, |
| 2651 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2652 | v44.data() /* input */, v45.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2653 | threadpool /* threadpool */); |
| 2654 | if (status != xnn_status_success) { |
| 2655 | std::cerr << "failed to setup operation #44" << std::endl; |
| 2656 | return ExecutionPlan(); |
| 2657 | } |
| 2658 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2659 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2660 | op45, |
| 2661 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2662 | v45.data() /* input */, v46.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2663 | threadpool /* threadpool */); |
| 2664 | if (status != xnn_status_success) { |
| 2665 | std::cerr << "failed to setup operation #45" << std::endl; |
| 2666 | return ExecutionPlan(); |
| 2667 | } |
| 2668 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2669 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2670 | op46, |
| 2671 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2672 | v46.data() /* input */, v47.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2673 | threadpool /* threadpool */); |
| 2674 | if (status != xnn_status_success) { |
| 2675 | std::cerr << "failed to setup operation #46" << std::endl; |
| 2676 | return ExecutionPlan(); |
| 2677 | } |
| 2678 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2679 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2680 | op47, |
| 2681 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2682 | v47.data() /* input */, v48.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2683 | threadpool /* threadpool */); |
| 2684 | if (status != xnn_status_success) { |
| 2685 | std::cerr << "failed to setup operation #47" << std::endl; |
| 2686 | return ExecutionPlan(); |
| 2687 | } |
| 2688 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2689 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2690 | op48, |
| 2691 | 1 /* batch size */, 196 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2692 | v48.data() /* input */, v49.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2693 | threadpool /* threadpool */); |
| 2694 | if (status != xnn_status_success) { |
| 2695 | std::cerr << "failed to setup operation #48" << std::endl; |
| 2696 | return ExecutionPlan(); |
| 2697 | } |
| 2698 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2699 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2700 | op49, |
| 2701 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2702 | v49.data() /* input */, v50.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2703 | threadpool /* threadpool */); |
| 2704 | if (status != xnn_status_success) { |
| 2705 | std::cerr << "failed to setup operation #49" << std::endl; |
| 2706 | return ExecutionPlan(); |
| 2707 | } |
| 2708 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2709 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2710 | op50, |
| 2711 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2712 | v50.data() /* input */, v51.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2713 | threadpool /* threadpool */); |
| 2714 | if (status != xnn_status_success) { |
| 2715 | std::cerr << "failed to setup operation #50" << std::endl; |
| 2716 | return ExecutionPlan(); |
| 2717 | } |
| 2718 | |
| 2719 | { |
| 2720 | const size_t a_shape[] = { 1, 14, 14, 120 }; |
| 2721 | const size_t b_shape[] = { 1, 1, 1, 120 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2722 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2723 | op51, |
| 2724 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2725 | v48.data() /* a */, v51.data() /* b */, v52.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2726 | threadpool /* threadpool */); |
| 2727 | } |
| 2728 | if (status != xnn_status_success) { |
| 2729 | std::cerr << "failed to setup operation #51" << std::endl; |
| 2730 | return ExecutionPlan(); |
| 2731 | } |
| 2732 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2733 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2734 | op52, |
| 2735 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2736 | v52.data() /* input */, v53.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2737 | threadpool /* threadpool */); |
| 2738 | if (status != xnn_status_success) { |
| 2739 | std::cerr << "failed to setup operation #52" << std::endl; |
| 2740 | return ExecutionPlan(); |
| 2741 | } |
| 2742 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2743 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2744 | op53, |
| 2745 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2746 | v53.data() /* input */, v54.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2747 | threadpool /* threadpool */); |
| 2748 | if (status != xnn_status_success) { |
| 2749 | std::cerr << "failed to setup operation #53" << std::endl; |
| 2750 | return ExecutionPlan(); |
| 2751 | } |
| 2752 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2753 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2754 | op54, |
| 2755 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2756 | v54.data() /* input */, v55.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2757 | threadpool /* threadpool */); |
| 2758 | if (status != xnn_status_success) { |
| 2759 | std::cerr << "failed to setup operation #54" << std::endl; |
| 2760 | return ExecutionPlan(); |
| 2761 | } |
| 2762 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2763 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2764 | op55, |
| 2765 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2766 | v55.data() /* input */, v56.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2767 | threadpool /* threadpool */); |
| 2768 | if (status != xnn_status_success) { |
| 2769 | std::cerr << "failed to setup operation #55" << std::endl; |
| 2770 | return ExecutionPlan(); |
| 2771 | } |
| 2772 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2773 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2774 | op56, |
| 2775 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2776 | v56.data() /* input */, v57.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2777 | threadpool /* threadpool */); |
| 2778 | if (status != xnn_status_success) { |
| 2779 | std::cerr << "failed to setup operation #56" << std::endl; |
| 2780 | return ExecutionPlan(); |
| 2781 | } |
| 2782 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2783 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2784 | op57, |
| 2785 | 1 /* batch size */, 196 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2786 | v57.data() /* input */, v58.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2787 | threadpool /* threadpool */); |
| 2788 | if (status != xnn_status_success) { |
| 2789 | std::cerr << "failed to setup operation #57" << std::endl; |
| 2790 | return ExecutionPlan(); |
| 2791 | } |
| 2792 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2793 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2794 | op58, |
| 2795 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2796 | v58.data() /* input */, v59.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2797 | threadpool /* threadpool */); |
| 2798 | if (status != xnn_status_success) { |
| 2799 | std::cerr << "failed to setup operation #58" << std::endl; |
| 2800 | return ExecutionPlan(); |
| 2801 | } |
| 2802 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2803 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2804 | op59, |
| 2805 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2806 | v59.data() /* input */, v60.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2807 | threadpool /* threadpool */); |
| 2808 | if (status != xnn_status_success) { |
| 2809 | std::cerr << "failed to setup operation #59" << std::endl; |
| 2810 | return ExecutionPlan(); |
| 2811 | } |
| 2812 | |
| 2813 | { |
| 2814 | const size_t a_shape[] = { 1, 14, 14, 144 }; |
| 2815 | const size_t b_shape[] = { 1, 1, 1, 144 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2816 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2817 | op60, |
| 2818 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2819 | v57.data() /* a */, v60.data() /* b */, v61.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2820 | threadpool /* threadpool */); |
| 2821 | } |
| 2822 | if (status != xnn_status_success) { |
| 2823 | std::cerr << "failed to setup operation #60" << std::endl; |
| 2824 | return ExecutionPlan(); |
| 2825 | } |
| 2826 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2827 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2828 | op61, |
| 2829 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2830 | v61.data() /* input */, v62.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2831 | threadpool /* threadpool */); |
| 2832 | if (status != xnn_status_success) { |
| 2833 | std::cerr << "failed to setup operation #61" << std::endl; |
| 2834 | return ExecutionPlan(); |
| 2835 | } |
| 2836 | |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2837 | { |
| 2838 | const size_t a_shape[] = { 1, 14, 14, 48 }; |
| 2839 | const size_t b_shape[] = { 1, 14, 14, 48 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2840 | status = xnn_setup_add_nd_f16( |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2841 | op62, |
| 2842 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2843 | v62.data() /* a */, v53.data() /* b */, v63.data() /* output */, |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 2844 | threadpool /* threadpool */); |
| 2845 | } |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2846 | if (status != xnn_status_success) { |
| 2847 | std::cerr << "failed to setup operation #62" << std::endl; |
| 2848 | return ExecutionPlan(); |
| 2849 | } |
| 2850 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2851 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2852 | op63, |
| 2853 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2854 | v63.data() /* input */, v64.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2855 | threadpool /* threadpool */); |
| 2856 | if (status != xnn_status_success) { |
| 2857 | std::cerr << "failed to setup operation #63" << std::endl; |
| 2858 | return ExecutionPlan(); |
| 2859 | } |
| 2860 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2861 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2862 | op64, |
| 2863 | 196 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2864 | v64.data() /* input */, v65.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2865 | threadpool /* threadpool */); |
| 2866 | if (status != xnn_status_success) { |
| 2867 | std::cerr << "failed to setup operation #64" << std::endl; |
| 2868 | return ExecutionPlan(); |
| 2869 | } |
| 2870 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2871 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2872 | op65, |
| 2873 | 1 /* batch size */, 14 /* input height */, 14 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2874 | v65.data() /* input */, v66.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2875 | threadpool /* threadpool */); |
| 2876 | if (status != xnn_status_success) { |
| 2877 | std::cerr << "failed to setup operation #65" << std::endl; |
| 2878 | return ExecutionPlan(); |
| 2879 | } |
| 2880 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2881 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2882 | op66, |
| 2883 | 49 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2884 | v66.data() /* input */, v67.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2885 | threadpool /* threadpool */); |
| 2886 | if (status != xnn_status_success) { |
| 2887 | std::cerr << "failed to setup operation #66" << std::endl; |
| 2888 | return ExecutionPlan(); |
| 2889 | } |
| 2890 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2891 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2892 | op67, |
| 2893 | 1 /* batch size */, 49 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2894 | v67.data() /* input */, v68.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2895 | threadpool /* threadpool */); |
| 2896 | if (status != xnn_status_success) { |
| 2897 | std::cerr << "failed to setup operation #67" << std::endl; |
| 2898 | return ExecutionPlan(); |
| 2899 | } |
| 2900 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2901 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2902 | op68, |
| 2903 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2904 | v68.data() /* input */, v69.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2905 | threadpool /* threadpool */); |
| 2906 | if (status != xnn_status_success) { |
| 2907 | std::cerr << "failed to setup operation #68" << std::endl; |
| 2908 | return ExecutionPlan(); |
| 2909 | } |
| 2910 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2911 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2912 | op69, |
| 2913 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2914 | v69.data() /* input */, v70.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2915 | threadpool /* threadpool */); |
| 2916 | if (status != xnn_status_success) { |
| 2917 | std::cerr << "failed to setup operation #69" << std::endl; |
| 2918 | return ExecutionPlan(); |
| 2919 | } |
| 2920 | |
| 2921 | { |
| 2922 | const size_t a_shape[] = { 1, 7, 7, 288 }; |
| 2923 | const size_t b_shape[] = { 1, 1, 1, 288 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2924 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2925 | op70, |
| 2926 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2927 | v67.data() /* a */, v70.data() /* b */, v71.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2928 | threadpool /* threadpool */); |
| 2929 | } |
| 2930 | if (status != xnn_status_success) { |
| 2931 | std::cerr << "failed to setup operation #70" << std::endl; |
| 2932 | return ExecutionPlan(); |
| 2933 | } |
| 2934 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2935 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2936 | op71, |
| 2937 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2938 | v71.data() /* input */, v72.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2939 | threadpool /* threadpool */); |
| 2940 | if (status != xnn_status_success) { |
| 2941 | std::cerr << "failed to setup operation #71" << std::endl; |
| 2942 | return ExecutionPlan(); |
| 2943 | } |
| 2944 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2945 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2946 | op72, |
| 2947 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2948 | v72.data() /* input */, v73.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2949 | threadpool /* threadpool */); |
| 2950 | if (status != xnn_status_success) { |
| 2951 | std::cerr << "failed to setup operation #72" << std::endl; |
| 2952 | return ExecutionPlan(); |
| 2953 | } |
| 2954 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2955 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2956 | op73, |
| 2957 | 49 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2958 | v73.data() /* input */, v74.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2959 | threadpool /* threadpool */); |
| 2960 | if (status != xnn_status_success) { |
| 2961 | std::cerr << "failed to setup operation #73" << std::endl; |
| 2962 | return ExecutionPlan(); |
| 2963 | } |
| 2964 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2965 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2966 | op74, |
| 2967 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2968 | v74.data() /* input */, v75.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2969 | threadpool /* threadpool */); |
| 2970 | if (status != xnn_status_success) { |
| 2971 | std::cerr << "failed to setup operation #74" << std::endl; |
| 2972 | return ExecutionPlan(); |
| 2973 | } |
| 2974 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2975 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2976 | op75, |
| 2977 | 49 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2978 | v75.data() /* input */, v76.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2979 | threadpool /* threadpool */); |
| 2980 | if (status != xnn_status_success) { |
| 2981 | std::cerr << "failed to setup operation #75" << std::endl; |
| 2982 | return ExecutionPlan(); |
| 2983 | } |
| 2984 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2985 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2986 | op76, |
| 2987 | 1 /* batch size */, 49 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2988 | v76.data() /* input */, v77.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2989 | threadpool /* threadpool */); |
| 2990 | if (status != xnn_status_success) { |
| 2991 | std::cerr << "failed to setup operation #76" << std::endl; |
| 2992 | return ExecutionPlan(); |
| 2993 | } |
| 2994 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 2995 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2996 | op77, |
| 2997 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 2998 | v77.data() /* input */, v78.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 2999 | threadpool /* threadpool */); |
| 3000 | if (status != xnn_status_success) { |
| 3001 | std::cerr << "failed to setup operation #77" << std::endl; |
| 3002 | return ExecutionPlan(); |
| 3003 | } |
| 3004 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3005 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3006 | op78, |
| 3007 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3008 | v78.data() /* input */, v79.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3009 | threadpool /* threadpool */); |
| 3010 | if (status != xnn_status_success) { |
| 3011 | std::cerr << "failed to setup operation #78" << std::endl; |
| 3012 | return ExecutionPlan(); |
| 3013 | } |
| 3014 | |
| 3015 | { |
| 3016 | const size_t a_shape[] = { 1, 7, 7, 576 }; |
| 3017 | const size_t b_shape[] = { 1, 1, 1, 576 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3018 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3019 | op79, |
| 3020 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3021 | v76.data() /* a */, v79.data() /* b */, v80.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3022 | threadpool /* threadpool */); |
| 3023 | } |
| 3024 | if (status != xnn_status_success) { |
| 3025 | std::cerr << "failed to setup operation #79" << std::endl; |
| 3026 | return ExecutionPlan(); |
| 3027 | } |
| 3028 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3029 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3030 | op80, |
| 3031 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3032 | v80.data() /* input */, v81.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3033 | threadpool /* threadpool */); |
| 3034 | if (status != xnn_status_success) { |
| 3035 | std::cerr << "failed to setup operation #80" << std::endl; |
| 3036 | return ExecutionPlan(); |
| 3037 | } |
| 3038 | |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 3039 | { |
| 3040 | const size_t a_shape[] = { 1, 7, 7, 96 }; |
| 3041 | const size_t b_shape[] = { 1, 7, 7, 96 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3042 | status = xnn_setup_add_nd_f16( |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 3043 | op81, |
| 3044 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3045 | v81.data() /* a */, v72.data() /* b */, v82.data() /* output */, |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 3046 | threadpool /* threadpool */); |
| 3047 | } |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3048 | if (status != xnn_status_success) { |
| 3049 | std::cerr << "failed to setup operation #81" << std::endl; |
| 3050 | return ExecutionPlan(); |
| 3051 | } |
| 3052 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3053 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3054 | op82, |
| 3055 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3056 | v82.data() /* input */, v83.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3057 | threadpool /* threadpool */); |
| 3058 | if (status != xnn_status_success) { |
| 3059 | std::cerr << "failed to setup operation #82" << std::endl; |
| 3060 | return ExecutionPlan(); |
| 3061 | } |
| 3062 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3063 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3064 | op83, |
| 3065 | 49 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3066 | v83.data() /* input */, v84.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3067 | threadpool /* threadpool */); |
| 3068 | if (status != xnn_status_success) { |
| 3069 | std::cerr << "failed to setup operation #83" << std::endl; |
| 3070 | return ExecutionPlan(); |
| 3071 | } |
| 3072 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3073 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3074 | op84, |
| 3075 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3076 | v84.data() /* input */, v85.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3077 | threadpool /* threadpool */); |
| 3078 | if (status != xnn_status_success) { |
| 3079 | std::cerr << "failed to setup operation #84" << std::endl; |
| 3080 | return ExecutionPlan(); |
| 3081 | } |
| 3082 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3083 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3084 | op85, |
| 3085 | 49 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3086 | v85.data() /* input */, v86.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3087 | threadpool /* threadpool */); |
| 3088 | if (status != xnn_status_success) { |
| 3089 | std::cerr << "failed to setup operation #85" << std::endl; |
| 3090 | return ExecutionPlan(); |
| 3091 | } |
| 3092 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3093 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3094 | op86, |
| 3095 | 1 /* batch size */, 49 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3096 | v86.data() /* input */, v87.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3097 | threadpool /* threadpool */); |
| 3098 | if (status != xnn_status_success) { |
| 3099 | std::cerr << "failed to setup operation #86" << std::endl; |
| 3100 | return ExecutionPlan(); |
| 3101 | } |
| 3102 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3103 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3104 | op87, |
| 3105 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3106 | v87.data() /* input */, v88.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3107 | threadpool /* threadpool */); |
| 3108 | if (status != xnn_status_success) { |
| 3109 | std::cerr << "failed to setup operation #87" << std::endl; |
| 3110 | return ExecutionPlan(); |
| 3111 | } |
| 3112 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3113 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3114 | op88, |
| 3115 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3116 | v88.data() /* input */, v89.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3117 | threadpool /* threadpool */); |
| 3118 | if (status != xnn_status_success) { |
| 3119 | std::cerr << "failed to setup operation #88" << std::endl; |
| 3120 | return ExecutionPlan(); |
| 3121 | } |
| 3122 | |
| 3123 | { |
| 3124 | const size_t a_shape[] = { 1, 7, 7, 576 }; |
| 3125 | const size_t b_shape[] = { 1, 1, 1, 576 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3126 | status = xnn_setup_multiply_nd_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3127 | op89, |
| 3128 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3129 | v86.data() /* a */, v89.data() /* b */, v90.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3130 | threadpool /* threadpool */); |
| 3131 | } |
| 3132 | if (status != xnn_status_success) { |
| 3133 | std::cerr << "failed to setup operation #89" << std::endl; |
| 3134 | return ExecutionPlan(); |
| 3135 | } |
| 3136 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3137 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3138 | op90, |
| 3139 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3140 | v90.data() /* input */, v91.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3141 | threadpool /* threadpool */); |
| 3142 | if (status != xnn_status_success) { |
| 3143 | std::cerr << "failed to setup operation #90" << std::endl; |
| 3144 | return ExecutionPlan(); |
| 3145 | } |
| 3146 | |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 3147 | { |
| 3148 | const size_t a_shape[] = { 1, 7, 7, 96 }; |
| 3149 | const size_t b_shape[] = { 1, 7, 7, 96 }; |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3150 | status = xnn_setup_add_nd_f16( |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 3151 | op91, |
| 3152 | 4, a_shape, 4, b_shape, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3153 | v91.data() /* a */, v82.data() /* b */, v92.data() /* output */, |
Marat Dukhan | b2164a3 | 2020-06-04 19:05:27 -0700 | [diff] [blame] | 3154 | threadpool /* threadpool */); |
| 3155 | } |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3156 | if (status != xnn_status_success) { |
| 3157 | std::cerr << "failed to setup operation #91" << std::endl; |
| 3158 | return ExecutionPlan(); |
| 3159 | } |
| 3160 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3161 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3162 | op92, |
| 3163 | 1 /* batch size */, 7 /* input height */, 7 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3164 | v92.data() /* input */, v93.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3165 | threadpool /* threadpool */); |
| 3166 | if (status != xnn_status_success) { |
| 3167 | std::cerr << "failed to setup operation #92" << std::endl; |
| 3168 | return ExecutionPlan(); |
| 3169 | } |
| 3170 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3171 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3172 | op93, |
| 3173 | 49 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3174 | v93.data() /* input */, v94.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3175 | threadpool /* threadpool */); |
| 3176 | if (status != xnn_status_success) { |
| 3177 | std::cerr << "failed to setup operation #93" << std::endl; |
| 3178 | return ExecutionPlan(); |
| 3179 | } |
| 3180 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3181 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3182 | op94, |
| 3183 | 1 /* batch size */, 49 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3184 | v94.data() /* input */, v95.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3185 | threadpool /* threadpool */); |
| 3186 | if (status != xnn_status_success) { |
| 3187 | std::cerr << "failed to setup operation #94" << std::endl; |
| 3188 | return ExecutionPlan(); |
| 3189 | } |
| 3190 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3191 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3192 | op95, |
| 3193 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3194 | v95.data() /* input */, v96.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3195 | threadpool /* threadpool */); |
| 3196 | if (status != xnn_status_success) { |
| 3197 | std::cerr << "failed to setup operation #95" << std::endl; |
| 3198 | return ExecutionPlan(); |
| 3199 | } |
| 3200 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3201 | status = xnn_setup_hardswish_nc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3202 | op96, |
| 3203 | 1 /* batch size */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3204 | v96.data() /* input */, v97.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3205 | threadpool /* threadpool */); |
| 3206 | if (status != xnn_status_success) { |
| 3207 | std::cerr << "failed to setup operation #96" << std::endl; |
| 3208 | return ExecutionPlan(); |
| 3209 | } |
| 3210 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3211 | status = xnn_setup_global_average_pooling_nwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3212 | op97, |
| 3213 | 1 /* batch size */, 1 /* width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3214 | v97.data() /* input */, v98.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3215 | threadpool /* threadpool */); |
| 3216 | if (status != xnn_status_success) { |
| 3217 | std::cerr << "failed to setup operation #97" << std::endl; |
| 3218 | return ExecutionPlan(); |
| 3219 | } |
| 3220 | |
Marat Dukhan | 270a2c4 | 2020-06-26 16:45:52 -0700 | [diff] [blame] | 3221 | status = xnn_setup_convolution2d_nhwc_f16( |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3222 | op98, |
| 3223 | 1 /* batch size */, 1 /* input height */, 1 /* input width */, |
Marat Dukhan | 52e061d | 2021-03-09 00:02:44 -0800 | [diff] [blame] | 3224 | v98.data() /* input */, v99.data() /* output */, |
Marat Dukhan | c08cdf5 | 2019-12-09 09:17:51 -0800 | [diff] [blame] | 3225 | threadpool /* threadpool */); |
| 3226 | if (status != xnn_status_success) { |
| 3227 | std::cerr << "failed to setup operation #98" << std::endl; |
| 3228 | return ExecutionPlan(); |
| 3229 | } |
| 3230 | |
| 3231 | #pragma clang diagnostic push |
| 3232 | #pragma clang diagnostic ignored "-Wpessimizing-move" |
| 3233 | return operators; |
| 3234 | #pragma clang diagnostic pop |
| 3235 | } |
| 3236 | |
| 3237 | } // namespace models |