| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2014 The WebRTC project authors. All Rights Reserved. |
| 3 | * |
| 4 | * Use of this source code is governed by a BSD-style license |
| 5 | * that can be found in the LICENSE file in the root of the source |
| 6 | * tree. An additional intellectual property rights grant can be found |
| 7 | * in the file PATENTS. All contributing project authors may |
| 8 | * be found in the AUTHORS file in the root of the source tree. |
| 9 | */ |
| 10 | |
| Mirko Bonadei | 92ea95e | 2017-09-15 06:47:31 +0200 | [diff] [blame] | 11 | #include "modules/audio_processing/utility/ooura_fft.h" |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 12 | |
| Mirko Bonadei | 92ea95e | 2017-09-15 06:47:31 +0200 | [diff] [blame] | 13 | #include "modules/audio_processing/utility/ooura_fft_tables_common.h" |
| Mirko Bonadei | 7120742 | 2017-09-15 13:58:09 +0200 | [diff] [blame] | 14 | #include "typedefs.h" // NOLINT(build/include) |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 15 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 16 | namespace webrtc { |
| 17 | |
| 18 | #if defined(MIPS_FPU_LE) |
| 19 | void bitrv2_128_mips(float* a) { |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 20 | // n is 128 |
| 21 | float xr, xi, yr, yi; |
| 22 | |
| 23 | xr = a[8]; |
| 24 | xi = a[9]; |
| 25 | yr = a[16]; |
| 26 | yi = a[17]; |
| 27 | a[8] = yr; |
| 28 | a[9] = yi; |
| 29 | a[16] = xr; |
| 30 | a[17] = xi; |
| 31 | |
| 32 | xr = a[64]; |
| 33 | xi = a[65]; |
| 34 | yr = a[2]; |
| 35 | yi = a[3]; |
| 36 | a[64] = yr; |
| 37 | a[65] = yi; |
| 38 | a[2] = xr; |
| 39 | a[3] = xi; |
| 40 | |
| 41 | xr = a[72]; |
| 42 | xi = a[73]; |
| 43 | yr = a[18]; |
| 44 | yi = a[19]; |
| 45 | a[72] = yr; |
| 46 | a[73] = yi; |
| 47 | a[18] = xr; |
| 48 | a[19] = xi; |
| 49 | |
| 50 | xr = a[80]; |
| 51 | xi = a[81]; |
| 52 | yr = a[10]; |
| 53 | yi = a[11]; |
| 54 | a[80] = yr; |
| 55 | a[81] = yi; |
| 56 | a[10] = xr; |
| 57 | a[11] = xi; |
| 58 | |
| 59 | xr = a[88]; |
| 60 | xi = a[89]; |
| 61 | yr = a[26]; |
| 62 | yi = a[27]; |
| 63 | a[88] = yr; |
| 64 | a[89] = yi; |
| 65 | a[26] = xr; |
| 66 | a[27] = xi; |
| 67 | |
| 68 | xr = a[74]; |
| 69 | xi = a[75]; |
| 70 | yr = a[82]; |
| 71 | yi = a[83]; |
| 72 | a[74] = yr; |
| 73 | a[75] = yi; |
| 74 | a[82] = xr; |
| 75 | a[83] = xi; |
| 76 | |
| 77 | xr = a[32]; |
| 78 | xi = a[33]; |
| 79 | yr = a[4]; |
| 80 | yi = a[5]; |
| 81 | a[32] = yr; |
| 82 | a[33] = yi; |
| 83 | a[4] = xr; |
| 84 | a[5] = xi; |
| 85 | |
| 86 | xr = a[40]; |
| 87 | xi = a[41]; |
| 88 | yr = a[20]; |
| 89 | yi = a[21]; |
| 90 | a[40] = yr; |
| 91 | a[41] = yi; |
| 92 | a[20] = xr; |
| 93 | a[21] = xi; |
| 94 | |
| 95 | xr = a[48]; |
| 96 | xi = a[49]; |
| 97 | yr = a[12]; |
| 98 | yi = a[13]; |
| 99 | a[48] = yr; |
| 100 | a[49] = yi; |
| 101 | a[12] = xr; |
| 102 | a[13] = xi; |
| 103 | |
| 104 | xr = a[56]; |
| 105 | xi = a[57]; |
| 106 | yr = a[28]; |
| 107 | yi = a[29]; |
| 108 | a[56] = yr; |
| 109 | a[57] = yi; |
| 110 | a[28] = xr; |
| 111 | a[29] = xi; |
| 112 | |
| 113 | xr = a[34]; |
| 114 | xi = a[35]; |
| 115 | yr = a[68]; |
| 116 | yi = a[69]; |
| 117 | a[34] = yr; |
| 118 | a[35] = yi; |
| 119 | a[68] = xr; |
| 120 | a[69] = xi; |
| 121 | |
| 122 | xr = a[42]; |
| 123 | xi = a[43]; |
| 124 | yr = a[84]; |
| 125 | yi = a[85]; |
| 126 | a[42] = yr; |
| 127 | a[43] = yi; |
| 128 | a[84] = xr; |
| 129 | a[85] = xi; |
| 130 | |
| 131 | xr = a[50]; |
| 132 | xi = a[51]; |
| 133 | yr = a[76]; |
| 134 | yi = a[77]; |
| 135 | a[50] = yr; |
| 136 | a[51] = yi; |
| 137 | a[76] = xr; |
| 138 | a[77] = xi; |
| 139 | |
| 140 | xr = a[58]; |
| 141 | xi = a[59]; |
| 142 | yr = a[92]; |
| 143 | yi = a[93]; |
| 144 | a[58] = yr; |
| 145 | a[59] = yi; |
| 146 | a[92] = xr; |
| 147 | a[93] = xi; |
| 148 | |
| 149 | xr = a[44]; |
| 150 | xi = a[45]; |
| 151 | yr = a[52]; |
| 152 | yi = a[53]; |
| 153 | a[44] = yr; |
| 154 | a[45] = yi; |
| 155 | a[52] = xr; |
| 156 | a[53] = xi; |
| 157 | |
| 158 | xr = a[96]; |
| 159 | xi = a[97]; |
| 160 | yr = a[6]; |
| 161 | yi = a[7]; |
| 162 | a[96] = yr; |
| 163 | a[97] = yi; |
| 164 | a[6] = xr; |
| 165 | a[7] = xi; |
| 166 | |
| 167 | xr = a[104]; |
| 168 | xi = a[105]; |
| 169 | yr = a[22]; |
| 170 | yi = a[23]; |
| 171 | a[104] = yr; |
| 172 | a[105] = yi; |
| 173 | a[22] = xr; |
| 174 | a[23] = xi; |
| 175 | |
| 176 | xr = a[112]; |
| 177 | xi = a[113]; |
| 178 | yr = a[14]; |
| 179 | yi = a[15]; |
| 180 | a[112] = yr; |
| 181 | a[113] = yi; |
| 182 | a[14] = xr; |
| 183 | a[15] = xi; |
| 184 | |
| 185 | xr = a[120]; |
| 186 | xi = a[121]; |
| 187 | yr = a[30]; |
| 188 | yi = a[31]; |
| 189 | a[120] = yr; |
| 190 | a[121] = yi; |
| 191 | a[30] = xr; |
| 192 | a[31] = xi; |
| 193 | |
| 194 | xr = a[98]; |
| 195 | xi = a[99]; |
| 196 | yr = a[70]; |
| 197 | yi = a[71]; |
| 198 | a[98] = yr; |
| 199 | a[99] = yi; |
| 200 | a[70] = xr; |
| 201 | a[71] = xi; |
| 202 | |
| 203 | xr = a[106]; |
| 204 | xi = a[107]; |
| 205 | yr = a[86]; |
| 206 | yi = a[87]; |
| 207 | a[106] = yr; |
| 208 | a[107] = yi; |
| 209 | a[86] = xr; |
| 210 | a[87] = xi; |
| 211 | |
| 212 | xr = a[114]; |
| 213 | xi = a[115]; |
| 214 | yr = a[78]; |
| 215 | yi = a[79]; |
| 216 | a[114] = yr; |
| 217 | a[115] = yi; |
| 218 | a[78] = xr; |
| 219 | a[79] = xi; |
| 220 | |
| 221 | xr = a[122]; |
| 222 | xi = a[123]; |
| 223 | yr = a[94]; |
| 224 | yi = a[95]; |
| 225 | a[122] = yr; |
| 226 | a[123] = yi; |
| 227 | a[94] = xr; |
| 228 | a[95] = xi; |
| 229 | |
| 230 | xr = a[100]; |
| 231 | xi = a[101]; |
| 232 | yr = a[38]; |
| 233 | yi = a[39]; |
| 234 | a[100] = yr; |
| 235 | a[101] = yi; |
| 236 | a[38] = xr; |
| 237 | a[39] = xi; |
| 238 | |
| 239 | xr = a[108]; |
| 240 | xi = a[109]; |
| 241 | yr = a[54]; |
| 242 | yi = a[55]; |
| 243 | a[108] = yr; |
| 244 | a[109] = yi; |
| 245 | a[54] = xr; |
| 246 | a[55] = xi; |
| 247 | |
| 248 | xr = a[116]; |
| 249 | xi = a[117]; |
| 250 | yr = a[46]; |
| 251 | yi = a[47]; |
| 252 | a[116] = yr; |
| 253 | a[117] = yi; |
| 254 | a[46] = xr; |
| 255 | a[47] = xi; |
| 256 | |
| 257 | xr = a[124]; |
| 258 | xi = a[125]; |
| 259 | yr = a[62]; |
| 260 | yi = a[63]; |
| 261 | a[124] = yr; |
| 262 | a[125] = yi; |
| 263 | a[62] = xr; |
| 264 | a[63] = xi; |
| 265 | |
| 266 | xr = a[110]; |
| 267 | xi = a[111]; |
| 268 | yr = a[118]; |
| 269 | yi = a[119]; |
| 270 | a[110] = yr; |
| 271 | a[111] = yi; |
| 272 | a[118] = xr; |
| 273 | a[119] = xi; |
| 274 | } |
| 275 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 276 | void cft1st_128_mips(float* a) { |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 277 | float f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11, f12, f13, f14; |
| 278 | int a_ptr, p1_rdft, p2_rdft, count; |
| andrew@webrtc.org | db75a66 | 2014-09-30 15:17:50 +0000 | [diff] [blame] | 279 | const float* first = rdft_wk3ri_first; |
| 280 | const float* second = rdft_wk3ri_second; |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 281 | |
| 282 | __asm __volatile ( |
| 283 | ".set push \n\t" |
| 284 | ".set noreorder \n\t" |
| 285 | // first 8 |
| 286 | "lwc1 %[f0], 0(%[a]) \n\t" |
| 287 | "lwc1 %[f1], 4(%[a]) \n\t" |
| 288 | "lwc1 %[f2], 8(%[a]) \n\t" |
| 289 | "lwc1 %[f3], 12(%[a]) \n\t" |
| 290 | "lwc1 %[f4], 16(%[a]) \n\t" |
| 291 | "lwc1 %[f5], 20(%[a]) \n\t" |
| 292 | "lwc1 %[f6], 24(%[a]) \n\t" |
| 293 | "lwc1 %[f7], 28(%[a]) \n\t" |
| 294 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 295 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 296 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 297 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 298 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 299 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 300 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 301 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 302 | "add.s %[f7], %[f8], %[f2] \n\t" |
| 303 | "sub.s %[f8], %[f8], %[f2] \n\t" |
| 304 | "sub.s %[f2], %[f1], %[f4] \n\t" |
| 305 | "add.s %[f1], %[f1], %[f4] \n\t" |
| 306 | "add.s %[f4], %[f6], %[f3] \n\t" |
| 307 | "sub.s %[f6], %[f6], %[f3] \n\t" |
| 308 | "sub.s %[f3], %[f0], %[f5] \n\t" |
| 309 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 310 | "swc1 %[f7], 0(%[a]) \n\t" |
| 311 | "swc1 %[f8], 16(%[a]) \n\t" |
| 312 | "swc1 %[f2], 28(%[a]) \n\t" |
| 313 | "swc1 %[f1], 12(%[a]) \n\t" |
| 314 | "swc1 %[f4], 4(%[a]) \n\t" |
| 315 | "swc1 %[f6], 20(%[a]) \n\t" |
| 316 | "swc1 %[f3], 8(%[a]) \n\t" |
| 317 | "swc1 %[f0], 24(%[a]) \n\t" |
| 318 | // second 8 |
| 319 | "lwc1 %[f0], 32(%[a]) \n\t" |
| 320 | "lwc1 %[f1], 36(%[a]) \n\t" |
| 321 | "lwc1 %[f2], 40(%[a]) \n\t" |
| 322 | "lwc1 %[f3], 44(%[a]) \n\t" |
| 323 | "lwc1 %[f4], 48(%[a]) \n\t" |
| 324 | "lwc1 %[f5], 52(%[a]) \n\t" |
| 325 | "lwc1 %[f6], 56(%[a]) \n\t" |
| 326 | "lwc1 %[f7], 60(%[a]) \n\t" |
| 327 | "add.s %[f8], %[f4], %[f6] \n\t" |
| 328 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 329 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 330 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 331 | "add.s %[f3], %[f0], %[f2] \n\t" |
| 332 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 333 | "add.s %[f2], %[f5], %[f7] \n\t" |
| 334 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 335 | "add.s %[f7], %[f4], %[f1] \n\t" |
| 336 | "sub.s %[f4], %[f4], %[f1] \n\t" |
| 337 | "add.s %[f1], %[f3], %[f8] \n\t" |
| 338 | "sub.s %[f3], %[f3], %[f8] \n\t" |
| 339 | "sub.s %[f8], %[f0], %[f5] \n\t" |
| 340 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 341 | "add.s %[f5], %[f6], %[f2] \n\t" |
| 342 | "sub.s %[f6], %[f2], %[f6] \n\t" |
| 343 | "lwc1 %[f9], 8(%[rdft_w]) \n\t" |
| 344 | "sub.s %[f2], %[f8], %[f7] \n\t" |
| 345 | "add.s %[f8], %[f8], %[f7] \n\t" |
| 346 | "sub.s %[f7], %[f4], %[f0] \n\t" |
| 347 | "add.s %[f4], %[f4], %[f0] \n\t" |
| 348 | // prepare for loop |
| 349 | "addiu %[a_ptr], %[a], 64 \n\t" |
| 350 | "addiu %[p1_rdft], %[rdft_w], 8 \n\t" |
| 351 | "addiu %[p2_rdft], %[rdft_w], 16 \n\t" |
| 352 | "addiu %[count], $zero, 7 \n\t" |
| 353 | // finish second 8 |
| 354 | "mul.s %[f2], %[f9], %[f2] \n\t" |
| 355 | "mul.s %[f8], %[f9], %[f8] \n\t" |
| 356 | "mul.s %[f7], %[f9], %[f7] \n\t" |
| 357 | "mul.s %[f4], %[f9], %[f4] \n\t" |
| 358 | "swc1 %[f1], 32(%[a]) \n\t" |
| 359 | "swc1 %[f3], 52(%[a]) \n\t" |
| 360 | "swc1 %[f5], 36(%[a]) \n\t" |
| 361 | "swc1 %[f6], 48(%[a]) \n\t" |
| 362 | "swc1 %[f2], 40(%[a]) \n\t" |
| 363 | "swc1 %[f8], 44(%[a]) \n\t" |
| 364 | "swc1 %[f7], 56(%[a]) \n\t" |
| 365 | "swc1 %[f4], 60(%[a]) \n\t" |
| 366 | // loop |
| 367 | "1: \n\t" |
| 368 | "lwc1 %[f0], 0(%[a_ptr]) \n\t" |
| 369 | "lwc1 %[f1], 4(%[a_ptr]) \n\t" |
| 370 | "lwc1 %[f2], 8(%[a_ptr]) \n\t" |
| 371 | "lwc1 %[f3], 12(%[a_ptr]) \n\t" |
| 372 | "lwc1 %[f4], 16(%[a_ptr]) \n\t" |
| 373 | "lwc1 %[f5], 20(%[a_ptr]) \n\t" |
| 374 | "lwc1 %[f6], 24(%[a_ptr]) \n\t" |
| 375 | "lwc1 %[f7], 28(%[a_ptr]) \n\t" |
| 376 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 377 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 378 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 379 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 380 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 381 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 382 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 383 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 384 | "lwc1 %[f10], 4(%[p1_rdft]) \n\t" |
| 385 | "lwc1 %[f11], 0(%[p2_rdft]) \n\t" |
| 386 | "lwc1 %[f12], 4(%[p2_rdft]) \n\t" |
| 387 | "lwc1 %[f13], 8(%[first]) \n\t" |
| 388 | "lwc1 %[f14], 12(%[first]) \n\t" |
| 389 | "add.s %[f7], %[f8], %[f2] \n\t" |
| 390 | "sub.s %[f8], %[f8], %[f2] \n\t" |
| 391 | "add.s %[f2], %[f6], %[f3] \n\t" |
| 392 | "sub.s %[f6], %[f6], %[f3] \n\t" |
| 393 | "add.s %[f3], %[f0], %[f5] \n\t" |
| 394 | "sub.s %[f0], %[f0], %[f5] \n\t" |
| 395 | "add.s %[f5], %[f1], %[f4] \n\t" |
| 396 | "sub.s %[f1], %[f1], %[f4] \n\t" |
| 397 | "swc1 %[f7], 0(%[a_ptr]) \n\t" |
| 398 | "swc1 %[f2], 4(%[a_ptr]) \n\t" |
| 399 | "mul.s %[f4], %[f9], %[f8] \n\t" |
| 400 | #if defined(MIPS32_R2_LE) |
| 401 | "mul.s %[f8], %[f10], %[f8] \n\t" |
| 402 | "mul.s %[f7], %[f11], %[f0] \n\t" |
| 403 | "mul.s %[f0], %[f12], %[f0] \n\t" |
| 404 | "mul.s %[f2], %[f13], %[f3] \n\t" |
| 405 | "mul.s %[f3], %[f14], %[f3] \n\t" |
| 406 | "nmsub.s %[f4], %[f4], %[f10], %[f6] \n\t" |
| 407 | "madd.s %[f8], %[f8], %[f9], %[f6] \n\t" |
| 408 | "nmsub.s %[f7], %[f7], %[f12], %[f5] \n\t" |
| 409 | "madd.s %[f0], %[f0], %[f11], %[f5] \n\t" |
| 410 | "nmsub.s %[f2], %[f2], %[f14], %[f1] \n\t" |
| 411 | "madd.s %[f3], %[f3], %[f13], %[f1] \n\t" |
| 412 | #else |
| 413 | "mul.s %[f7], %[f10], %[f6] \n\t" |
| 414 | "mul.s %[f6], %[f9], %[f6] \n\t" |
| 415 | "mul.s %[f8], %[f10], %[f8] \n\t" |
| 416 | "mul.s %[f2], %[f11], %[f0] \n\t" |
| 417 | "mul.s %[f11], %[f11], %[f5] \n\t" |
| 418 | "mul.s %[f5], %[f12], %[f5] \n\t" |
| 419 | "mul.s %[f0], %[f12], %[f0] \n\t" |
| 420 | "mul.s %[f12], %[f13], %[f3] \n\t" |
| 421 | "mul.s %[f13], %[f13], %[f1] \n\t" |
| 422 | "mul.s %[f1], %[f14], %[f1] \n\t" |
| 423 | "mul.s %[f3], %[f14], %[f3] \n\t" |
| 424 | "sub.s %[f4], %[f4], %[f7] \n\t" |
| 425 | "add.s %[f8], %[f6], %[f8] \n\t" |
| 426 | "sub.s %[f7], %[f2], %[f5] \n\t" |
| 427 | "add.s %[f0], %[f11], %[f0] \n\t" |
| 428 | "sub.s %[f2], %[f12], %[f1] \n\t" |
| 429 | "add.s %[f3], %[f13], %[f3] \n\t" |
| 430 | #endif |
| 431 | "swc1 %[f4], 16(%[a_ptr]) \n\t" |
| 432 | "swc1 %[f8], 20(%[a_ptr]) \n\t" |
| 433 | "swc1 %[f7], 8(%[a_ptr]) \n\t" |
| 434 | "swc1 %[f0], 12(%[a_ptr]) \n\t" |
| 435 | "swc1 %[f2], 24(%[a_ptr]) \n\t" |
| 436 | "swc1 %[f3], 28(%[a_ptr]) \n\t" |
| 437 | "lwc1 %[f0], 32(%[a_ptr]) \n\t" |
| 438 | "lwc1 %[f1], 36(%[a_ptr]) \n\t" |
| 439 | "lwc1 %[f2], 40(%[a_ptr]) \n\t" |
| 440 | "lwc1 %[f3], 44(%[a_ptr]) \n\t" |
| 441 | "lwc1 %[f4], 48(%[a_ptr]) \n\t" |
| 442 | "lwc1 %[f5], 52(%[a_ptr]) \n\t" |
| 443 | "lwc1 %[f6], 56(%[a_ptr]) \n\t" |
| 444 | "lwc1 %[f7], 60(%[a_ptr]) \n\t" |
| 445 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 446 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 447 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 448 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 449 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 450 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 451 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 452 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 453 | "lwc1 %[f11], 8(%[p2_rdft]) \n\t" |
| 454 | "lwc1 %[f12], 12(%[p2_rdft]) \n\t" |
| 455 | "lwc1 %[f13], 8(%[second]) \n\t" |
| 456 | "lwc1 %[f14], 12(%[second]) \n\t" |
| 457 | "add.s %[f7], %[f8], %[f2] \n\t" |
| 458 | "sub.s %[f8], %[f2], %[f8] \n\t" |
| 459 | "add.s %[f2], %[f6], %[f3] \n\t" |
| 460 | "sub.s %[f6], %[f3], %[f6] \n\t" |
| 461 | "add.s %[f3], %[f0], %[f5] \n\t" |
| 462 | "sub.s %[f0], %[f0], %[f5] \n\t" |
| 463 | "add.s %[f5], %[f1], %[f4] \n\t" |
| 464 | "sub.s %[f1], %[f1], %[f4] \n\t" |
| 465 | "swc1 %[f7], 32(%[a_ptr]) \n\t" |
| 466 | "swc1 %[f2], 36(%[a_ptr]) \n\t" |
| 467 | "mul.s %[f4], %[f10], %[f8] \n\t" |
| 468 | #if defined(MIPS32_R2_LE) |
| 469 | "mul.s %[f10], %[f10], %[f6] \n\t" |
| 470 | "mul.s %[f7], %[f11], %[f0] \n\t" |
| 471 | "mul.s %[f11], %[f11], %[f5] \n\t" |
| 472 | "mul.s %[f2], %[f13], %[f3] \n\t" |
| 473 | "mul.s %[f13], %[f13], %[f1] \n\t" |
| 474 | "madd.s %[f4], %[f4], %[f9], %[f6] \n\t" |
| 475 | "nmsub.s %[f10], %[f10], %[f9], %[f8] \n\t" |
| 476 | "nmsub.s %[f7], %[f7], %[f12], %[f5] \n\t" |
| 477 | "madd.s %[f11], %[f11], %[f12], %[f0] \n\t" |
| 478 | "nmsub.s %[f2], %[f2], %[f14], %[f1] \n\t" |
| 479 | "madd.s %[f13], %[f13], %[f14], %[f3] \n\t" |
| 480 | #else |
| 481 | "mul.s %[f2], %[f9], %[f6] \n\t" |
| 482 | "mul.s %[f10], %[f10], %[f6] \n\t" |
| 483 | "mul.s %[f9], %[f9], %[f8] \n\t" |
| 484 | "mul.s %[f7], %[f11], %[f0] \n\t" |
| 485 | "mul.s %[f8], %[f12], %[f5] \n\t" |
| 486 | "mul.s %[f11], %[f11], %[f5] \n\t" |
| 487 | "mul.s %[f12], %[f12], %[f0] \n\t" |
| 488 | "mul.s %[f5], %[f13], %[f3] \n\t" |
| 489 | "mul.s %[f0], %[f14], %[f1] \n\t" |
| 490 | "mul.s %[f13], %[f13], %[f1] \n\t" |
| 491 | "mul.s %[f14], %[f14], %[f3] \n\t" |
| 492 | "add.s %[f4], %[f4], %[f2] \n\t" |
| 493 | "sub.s %[f10], %[f10], %[f9] \n\t" |
| 494 | "sub.s %[f7], %[f7], %[f8] \n\t" |
| 495 | "add.s %[f11], %[f11], %[f12] \n\t" |
| 496 | "sub.s %[f2], %[f5], %[f0] \n\t" |
| 497 | "add.s %[f13], %[f13], %[f14] \n\t" |
| 498 | #endif |
| 499 | "swc1 %[f4], 48(%[a_ptr]) \n\t" |
| 500 | "swc1 %[f10], 52(%[a_ptr]) \n\t" |
| 501 | "swc1 %[f7], 40(%[a_ptr]) \n\t" |
| 502 | "swc1 %[f11], 44(%[a_ptr]) \n\t" |
| 503 | "swc1 %[f2], 56(%[a_ptr]) \n\t" |
| 504 | "swc1 %[f13], 60(%[a_ptr]) \n\t" |
| 505 | "addiu %[count], %[count], -1 \n\t" |
| 506 | "lwc1 %[f9], 8(%[p1_rdft]) \n\t" |
| 507 | "addiu %[a_ptr], %[a_ptr], 64 \n\t" |
| 508 | "addiu %[p1_rdft], %[p1_rdft], 8 \n\t" |
| 509 | "addiu %[p2_rdft], %[p2_rdft], 16 \n\t" |
| 510 | "addiu %[first], %[first], 8 \n\t" |
| 511 | "bgtz %[count], 1b \n\t" |
| 512 | " addiu %[second], %[second], 8 \n\t" |
| 513 | ".set pop \n\t" |
| 514 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 515 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| 516 | [f8] "=&f" (f8), [f9] "=&f" (f9), [f10] "=&f" (f10), [f11] "=&f" (f11), |
| 517 | [f12] "=&f" (f12), [f13] "=&f" (f13), [f14] "=&f" (f14), |
| 518 | [a_ptr] "=&r" (a_ptr), [p1_rdft] "=&r" (p1_rdft), [first] "+r" (first), |
| 519 | [p2_rdft] "=&r" (p2_rdft), [count] "=&r" (count), [second] "+r" (second) |
| 520 | : [a] "r" (a), [rdft_w] "r" (rdft_w) |
| 521 | : "memory" |
| 522 | ); |
| 523 | } |
| 524 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 525 | void cftmdl_128_mips(float* a) { |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 526 | float f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11, f12, f13, f14; |
| 527 | int tmp_a, count; |
| 528 | __asm __volatile ( |
| 529 | ".set push \n\t" |
| 530 | ".set noreorder \n\t" |
| 531 | "addiu %[tmp_a], %[a], 0 \n\t" |
| 532 | "addiu %[count], $zero, 4 \n\t" |
| 533 | "1: \n\t" |
| 534 | "addiu %[count], %[count], -1 \n\t" |
| 535 | "lwc1 %[f0], 0(%[tmp_a]) \n\t" |
| 536 | "lwc1 %[f2], 32(%[tmp_a]) \n\t" |
| 537 | "lwc1 %[f4], 64(%[tmp_a]) \n\t" |
| 538 | "lwc1 %[f6], 96(%[tmp_a]) \n\t" |
| 539 | "lwc1 %[f1], 4(%[tmp_a]) \n\t" |
| 540 | "lwc1 %[f3], 36(%[tmp_a]) \n\t" |
| 541 | "lwc1 %[f5], 68(%[tmp_a]) \n\t" |
| 542 | "lwc1 %[f7], 100(%[tmp_a]) \n\t" |
| 543 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 544 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 545 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 546 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 547 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 548 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 549 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 550 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 551 | "add.s %[f7], %[f8], %[f2] \n\t" |
| 552 | "sub.s %[f8], %[f8], %[f2] \n\t" |
| 553 | "add.s %[f2], %[f1], %[f4] \n\t" |
| 554 | "sub.s %[f1], %[f1], %[f4] \n\t" |
| 555 | "add.s %[f4], %[f6], %[f3] \n\t" |
| 556 | "sub.s %[f6], %[f6], %[f3] \n\t" |
| 557 | "sub.s %[f3], %[f0], %[f5] \n\t" |
| 558 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 559 | "swc1 %[f7], 0(%[tmp_a]) \n\t" |
| 560 | "swc1 %[f8], 64(%[tmp_a]) \n\t" |
| 561 | "swc1 %[f2], 36(%[tmp_a]) \n\t" |
| 562 | "swc1 %[f1], 100(%[tmp_a]) \n\t" |
| 563 | "swc1 %[f4], 4(%[tmp_a]) \n\t" |
| 564 | "swc1 %[f6], 68(%[tmp_a]) \n\t" |
| 565 | "swc1 %[f3], 32(%[tmp_a]) \n\t" |
| 566 | "swc1 %[f0], 96(%[tmp_a]) \n\t" |
| 567 | "bgtz %[count], 1b \n\t" |
| 568 | " addiu %[tmp_a], %[tmp_a], 8 \n\t" |
| 569 | ".set pop \n\t" |
| 570 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 571 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| 572 | [f8] "=&f" (f8), [tmp_a] "=&r" (tmp_a), [count] "=&r" (count) |
| 573 | : [a] "r" (a) |
| 574 | : "memory" |
| 575 | ); |
| 576 | f9 = rdft_w[2]; |
| 577 | __asm __volatile ( |
| 578 | ".set push \n\t" |
| 579 | ".set noreorder \n\t" |
| 580 | "addiu %[tmp_a], %[a], 128 \n\t" |
| 581 | "addiu %[count], $zero, 4 \n\t" |
| 582 | "1: \n\t" |
| 583 | "addiu %[count], %[count], -1 \n\t" |
| 584 | "lwc1 %[f0], 0(%[tmp_a]) \n\t" |
| 585 | "lwc1 %[f2], 32(%[tmp_a]) \n\t" |
| 586 | "lwc1 %[f5], 68(%[tmp_a]) \n\t" |
| 587 | "lwc1 %[f7], 100(%[tmp_a]) \n\t" |
| 588 | "lwc1 %[f1], 4(%[tmp_a]) \n\t" |
| 589 | "lwc1 %[f3], 36(%[tmp_a]) \n\t" |
| 590 | "lwc1 %[f4], 64(%[tmp_a]) \n\t" |
| 591 | "lwc1 %[f6], 96(%[tmp_a]) \n\t" |
| 592 | "sub.s %[f8], %[f0], %[f2] \n\t" |
| 593 | "add.s %[f0], %[f0], %[f2] \n\t" |
| 594 | "sub.s %[f2], %[f5], %[f7] \n\t" |
| 595 | "add.s %[f5], %[f5], %[f7] \n\t" |
| 596 | "sub.s %[f7], %[f1], %[f3] \n\t" |
| 597 | "add.s %[f1], %[f1], %[f3] \n\t" |
| 598 | "sub.s %[f3], %[f4], %[f6] \n\t" |
| 599 | "add.s %[f4], %[f4], %[f6] \n\t" |
| 600 | "sub.s %[f6], %[f8], %[f2] \n\t" |
| 601 | "add.s %[f8], %[f8], %[f2] \n\t" |
| 602 | "add.s %[f2], %[f5], %[f1] \n\t" |
| 603 | "sub.s %[f5], %[f5], %[f1] \n\t" |
| 604 | "add.s %[f1], %[f3], %[f7] \n\t" |
| 605 | "sub.s %[f3], %[f3], %[f7] \n\t" |
| 606 | "add.s %[f7], %[f0], %[f4] \n\t" |
| 607 | "sub.s %[f0], %[f0], %[f4] \n\t" |
| 608 | "sub.s %[f4], %[f6], %[f1] \n\t" |
| 609 | "add.s %[f6], %[f6], %[f1] \n\t" |
| 610 | "sub.s %[f1], %[f3], %[f8] \n\t" |
| 611 | "add.s %[f3], %[f3], %[f8] \n\t" |
| 612 | "mul.s %[f4], %[f4], %[f9] \n\t" |
| 613 | "mul.s %[f6], %[f6], %[f9] \n\t" |
| 614 | "mul.s %[f1], %[f1], %[f9] \n\t" |
| 615 | "mul.s %[f3], %[f3], %[f9] \n\t" |
| 616 | "swc1 %[f7], 0(%[tmp_a]) \n\t" |
| 617 | "swc1 %[f2], 4(%[tmp_a]) \n\t" |
| 618 | "swc1 %[f5], 64(%[tmp_a]) \n\t" |
| 619 | "swc1 %[f0], 68(%[tmp_a]) \n\t" |
| 620 | "swc1 %[f4], 32(%[tmp_a]) \n\t" |
| 621 | "swc1 %[f6], 36(%[tmp_a]) \n\t" |
| 622 | "swc1 %[f1], 96(%[tmp_a]) \n\t" |
| 623 | "swc1 %[f3], 100(%[tmp_a]) \n\t" |
| 624 | "bgtz %[count], 1b \n\t" |
| 625 | " addiu %[tmp_a], %[tmp_a], 8 \n\t" |
| 626 | ".set pop \n\t" |
| 627 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 628 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| 629 | [f8] "=&f" (f8), [tmp_a] "=&r" (tmp_a), [count] "=&r" (count) |
| 630 | : [a] "r" (a), [f9] "f" (f9) |
| 631 | : "memory" |
| 632 | ); |
| 633 | f10 = rdft_w[3]; |
| 634 | f11 = rdft_w[4]; |
| 635 | f12 = rdft_w[5]; |
| 636 | f13 = rdft_wk3ri_first[2]; |
| 637 | f14 = rdft_wk3ri_first[3]; |
| 638 | |
| 639 | __asm __volatile ( |
| 640 | ".set push \n\t" |
| 641 | ".set noreorder \n\t" |
| 642 | "addiu %[tmp_a], %[a], 256 \n\t" |
| 643 | "addiu %[count], $zero, 4 \n\t" |
| 644 | "1: \n\t" |
| 645 | "addiu %[count], %[count], -1 \n\t" |
| 646 | "lwc1 %[f0], 0(%[tmp_a]) \n\t" |
| 647 | "lwc1 %[f2], 32(%[tmp_a]) \n\t" |
| 648 | "lwc1 %[f4], 64(%[tmp_a]) \n\t" |
| 649 | "lwc1 %[f6], 96(%[tmp_a]) \n\t" |
| 650 | "lwc1 %[f1], 4(%[tmp_a]) \n\t" |
| 651 | "lwc1 %[f3], 36(%[tmp_a]) \n\t" |
| 652 | "lwc1 %[f5], 68(%[tmp_a]) \n\t" |
| 653 | "lwc1 %[f7], 100(%[tmp_a]) \n\t" |
| 654 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 655 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 656 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 657 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 658 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 659 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 660 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 661 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 662 | "sub.s %[f7], %[f8], %[f2] \n\t" |
| 663 | "add.s %[f8], %[f8], %[f2] \n\t" |
| 664 | "add.s %[f2], %[f1], %[f4] \n\t" |
| 665 | "sub.s %[f1], %[f1], %[f4] \n\t" |
| 666 | "sub.s %[f4], %[f6], %[f3] \n\t" |
| 667 | "add.s %[f6], %[f6], %[f3] \n\t" |
| 668 | "sub.s %[f3], %[f0], %[f5] \n\t" |
| 669 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 670 | "swc1 %[f8], 0(%[tmp_a]) \n\t" |
| 671 | "swc1 %[f6], 4(%[tmp_a]) \n\t" |
| 672 | "mul.s %[f5], %[f9], %[f7] \n\t" |
| 673 | #if defined(MIPS32_R2_LE) |
| 674 | "mul.s %[f7], %[f10], %[f7] \n\t" |
| 675 | "mul.s %[f8], %[f11], %[f3] \n\t" |
| 676 | "mul.s %[f3], %[f12], %[f3] \n\t" |
| 677 | "mul.s %[f6], %[f13], %[f0] \n\t" |
| 678 | "mul.s %[f0], %[f14], %[f0] \n\t" |
| 679 | "nmsub.s %[f5], %[f5], %[f10], %[f4] \n\t" |
| 680 | "madd.s %[f7], %[f7], %[f9], %[f4] \n\t" |
| 681 | "nmsub.s %[f8], %[f8], %[f12], %[f2] \n\t" |
| 682 | "madd.s %[f3], %[f3], %[f11], %[f2] \n\t" |
| 683 | "nmsub.s %[f6], %[f6], %[f14], %[f1] \n\t" |
| 684 | "madd.s %[f0], %[f0], %[f13], %[f1] \n\t" |
| 685 | "swc1 %[f5], 64(%[tmp_a]) \n\t" |
| 686 | "swc1 %[f7], 68(%[tmp_a]) \n\t" |
| 687 | #else |
| 688 | "mul.s %[f8], %[f10], %[f4] \n\t" |
| 689 | "mul.s %[f4], %[f9], %[f4] \n\t" |
| 690 | "mul.s %[f7], %[f10], %[f7] \n\t" |
| 691 | "mul.s %[f6], %[f11], %[f3] \n\t" |
| 692 | "mul.s %[f3], %[f12], %[f3] \n\t" |
| 693 | "sub.s %[f5], %[f5], %[f8] \n\t" |
| 694 | "mul.s %[f8], %[f12], %[f2] \n\t" |
| 695 | "mul.s %[f2], %[f11], %[f2] \n\t" |
| 696 | "add.s %[f7], %[f4], %[f7] \n\t" |
| 697 | "mul.s %[f4], %[f13], %[f0] \n\t" |
| 698 | "mul.s %[f0], %[f14], %[f0] \n\t" |
| 699 | "sub.s %[f8], %[f6], %[f8] \n\t" |
| 700 | "mul.s %[f6], %[f14], %[f1] \n\t" |
| 701 | "mul.s %[f1], %[f13], %[f1] \n\t" |
| 702 | "add.s %[f3], %[f2], %[f3] \n\t" |
| 703 | "swc1 %[f5], 64(%[tmp_a]) \n\t" |
| 704 | "swc1 %[f7], 68(%[tmp_a]) \n\t" |
| 705 | "sub.s %[f6], %[f4], %[f6] \n\t" |
| 706 | "add.s %[f0], %[f1], %[f0] \n\t" |
| 707 | #endif |
| 708 | "swc1 %[f8], 32(%[tmp_a]) \n\t" |
| 709 | "swc1 %[f3], 36(%[tmp_a]) \n\t" |
| 710 | "swc1 %[f6], 96(%[tmp_a]) \n\t" |
| 711 | "swc1 %[f0], 100(%[tmp_a]) \n\t" |
| 712 | "bgtz %[count], 1b \n\t" |
| 713 | " addiu %[tmp_a], %[tmp_a], 8 \n\t" |
| 714 | ".set pop \n\t" |
| 715 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 716 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| 717 | [f8] "=&f" (f8), [tmp_a] "=&r" (tmp_a), [count] "=&r" (count) |
| 718 | : [a] "r" (a), [f9] "f" (f9), [f10] "f" (f10), [f11] "f" (f11), |
| 719 | [f12] "f" (f12), [f13] "f" (f13), [f14] "f" (f14) |
| 720 | : "memory" |
| 721 | ); |
| 722 | f11 = rdft_w[6]; |
| 723 | f12 = rdft_w[7]; |
| 724 | f13 = rdft_wk3ri_second[2]; |
| 725 | f14 = rdft_wk3ri_second[3]; |
| 726 | __asm __volatile ( |
| 727 | ".set push \n\t" |
| 728 | ".set noreorder \n\t" |
| 729 | "addiu %[tmp_a], %[a], 384 \n\t" |
| 730 | "addiu %[count], $zero, 4 \n\t" |
| 731 | "1: \n\t" |
| 732 | "addiu %[count], %[count], -1 \n\t" |
| 733 | "lwc1 %[f0], 0(%[tmp_a]) \n\t" |
| 734 | "lwc1 %[f1], 4(%[tmp_a]) \n\t" |
| 735 | "lwc1 %[f2], 32(%[tmp_a]) \n\t" |
| 736 | "lwc1 %[f3], 36(%[tmp_a]) \n\t" |
| 737 | "lwc1 %[f4], 64(%[tmp_a]) \n\t" |
| 738 | "lwc1 %[f5], 68(%[tmp_a]) \n\t" |
| 739 | "lwc1 %[f6], 96(%[tmp_a]) \n\t" |
| 740 | "lwc1 %[f7], 100(%[tmp_a]) \n\t" |
| 741 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 742 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 743 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 744 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 745 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 746 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 747 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 748 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 749 | "sub.s %[f7], %[f2], %[f8] \n\t" |
| 750 | "add.s %[f2], %[f2], %[f8] \n\t" |
| 751 | "add.s %[f8], %[f1], %[f4] \n\t" |
| 752 | "sub.s %[f1], %[f1], %[f4] \n\t" |
| 753 | "sub.s %[f4], %[f3], %[f6] \n\t" |
| 754 | "add.s %[f3], %[f3], %[f6] \n\t" |
| 755 | "sub.s %[f6], %[f0], %[f5] \n\t" |
| 756 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 757 | "swc1 %[f2], 0(%[tmp_a]) \n\t" |
| 758 | "swc1 %[f3], 4(%[tmp_a]) \n\t" |
| 759 | "mul.s %[f5], %[f10], %[f7] \n\t" |
| 760 | #if defined(MIPS32_R2_LE) |
| 761 | "mul.s %[f7], %[f9], %[f7] \n\t" |
| 762 | "mul.s %[f2], %[f12], %[f8] \n\t" |
| 763 | "mul.s %[f8], %[f11], %[f8] \n\t" |
| 764 | "mul.s %[f3], %[f14], %[f1] \n\t" |
| 765 | "mul.s %[f1], %[f13], %[f1] \n\t" |
| 766 | "madd.s %[f5], %[f5], %[f9], %[f4] \n\t" |
| 767 | "msub.s %[f7], %[f7], %[f10], %[f4] \n\t" |
| 768 | "msub.s %[f2], %[f2], %[f11], %[f6] \n\t" |
| 769 | "madd.s %[f8], %[f8], %[f12], %[f6] \n\t" |
| 770 | "msub.s %[f3], %[f3], %[f13], %[f0] \n\t" |
| 771 | "madd.s %[f1], %[f1], %[f14], %[f0] \n\t" |
| 772 | "swc1 %[f5], 64(%[tmp_a]) \n\t" |
| 773 | "swc1 %[f7], 68(%[tmp_a]) \n\t" |
| 774 | #else |
| 775 | "mul.s %[f2], %[f9], %[f4] \n\t" |
| 776 | "mul.s %[f4], %[f10], %[f4] \n\t" |
| 777 | "mul.s %[f7], %[f9], %[f7] \n\t" |
| 778 | "mul.s %[f3], %[f11], %[f6] \n\t" |
| 779 | "mul.s %[f6], %[f12], %[f6] \n\t" |
| 780 | "add.s %[f5], %[f5], %[f2] \n\t" |
| 781 | "sub.s %[f7], %[f4], %[f7] \n\t" |
| 782 | "mul.s %[f2], %[f12], %[f8] \n\t" |
| 783 | "mul.s %[f8], %[f11], %[f8] \n\t" |
| 784 | "mul.s %[f4], %[f14], %[f1] \n\t" |
| 785 | "mul.s %[f1], %[f13], %[f1] \n\t" |
| 786 | "sub.s %[f2], %[f3], %[f2] \n\t" |
| 787 | "mul.s %[f3], %[f13], %[f0] \n\t" |
| 788 | "mul.s %[f0], %[f14], %[f0] \n\t" |
| 789 | "add.s %[f8], %[f8], %[f6] \n\t" |
| 790 | "swc1 %[f5], 64(%[tmp_a]) \n\t" |
| 791 | "swc1 %[f7], 68(%[tmp_a]) \n\t" |
| 792 | "sub.s %[f3], %[f3], %[f4] \n\t" |
| 793 | "add.s %[f1], %[f1], %[f0] \n\t" |
| 794 | #endif |
| 795 | "swc1 %[f2], 32(%[tmp_a]) \n\t" |
| 796 | "swc1 %[f8], 36(%[tmp_a]) \n\t" |
| 797 | "swc1 %[f3], 96(%[tmp_a]) \n\t" |
| 798 | "swc1 %[f1], 100(%[tmp_a]) \n\t" |
| 799 | "bgtz %[count], 1b \n\t" |
| 800 | " addiu %[tmp_a], %[tmp_a], 8 \n\t" |
| 801 | ".set pop \n\t" |
| 802 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 803 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| 804 | [f8] "=&f" (f8), [tmp_a] "=&r" (tmp_a), [count] "=&r" (count) |
| 805 | : [a] "r" (a), [f9] "f" (f9), [f10] "f" (f10), [f11] "f" (f11), |
| 806 | [f12] "f" (f12), [f13] "f" (f13), [f14] "f" (f14) |
| 807 | : "memory" |
| 808 | ); |
| 809 | } |
| 810 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 811 | void cftfsub_128_mips(float* a) { |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 812 | float f0, f1, f2, f3, f4, f5, f6, f7, f8; |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 813 | int tmp_a, count; |
| 814 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 815 | cft1st_128_mips(a); |
| 816 | cftmdl_128_mips(a); |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 817 | |
| 818 | __asm __volatile ( |
| 819 | ".set push \n\t" |
| 820 | ".set noreorder \n\t" |
| 821 | "addiu %[tmp_a], %[a], 0 \n\t" |
| 822 | "addiu %[count], $zero, 16 \n\t" |
| 823 | "1: \n\t" |
| 824 | "addiu %[count], %[count], -1 \n\t" |
| 825 | "lwc1 %[f0], 0(%[tmp_a]) \n\t" |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 826 | "lwc1 %[f2], 128(%[tmp_a]) \n\t" |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 827 | "lwc1 %[f4], 256(%[tmp_a]) \n\t" |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 828 | "lwc1 %[f6], 384(%[tmp_a]) \n\t" |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 829 | "lwc1 %[f1], 4(%[tmp_a]) \n\t" |
| 830 | "lwc1 %[f3], 132(%[tmp_a]) \n\t" |
| 831 | "lwc1 %[f5], 260(%[tmp_a]) \n\t" |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 832 | "lwc1 %[f7], 388(%[tmp_a]) \n\t" |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 833 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 834 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 835 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 836 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 837 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 838 | "sub.s %[f1], %[f1], %[f3] \n\t" |
| 839 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 840 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 841 | "add.s %[f7], %[f8], %[f2] \n\t" |
| 842 | "sub.s %[f8], %[f8], %[f2] \n\t" |
| 843 | "add.s %[f2], %[f1], %[f4] \n\t" |
| 844 | "sub.s %[f1], %[f1], %[f4] \n\t" |
| 845 | "add.s %[f4], %[f6], %[f3] \n\t" |
| 846 | "sub.s %[f6], %[f6], %[f3] \n\t" |
| 847 | "sub.s %[f3], %[f0], %[f5] \n\t" |
| 848 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 849 | "swc1 %[f7], 0(%[tmp_a]) \n\t" |
| 850 | "swc1 %[f8], 256(%[tmp_a]) \n\t" |
| 851 | "swc1 %[f2], 132(%[tmp_a]) \n\t" |
| 852 | "swc1 %[f1], 388(%[tmp_a]) \n\t" |
| 853 | "swc1 %[f4], 4(%[tmp_a]) \n\t" |
| 854 | "swc1 %[f6], 260(%[tmp_a]) \n\t" |
| 855 | "swc1 %[f3], 128(%[tmp_a]) \n\t" |
| 856 | "swc1 %[f0], 384(%[tmp_a]) \n\t" |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 857 | "bgtz %[count], 1b \n\t" |
| 858 | " addiu %[tmp_a], %[tmp_a], 8 \n\t" |
| 859 | ".set pop \n\t" |
| 860 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 861 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 862 | [f8] "=&f" (f8), [tmp_a] "=&r" (tmp_a), |
| andrew@webrtc.org | c0907ef | 2014-02-21 00:13:31 +0000 | [diff] [blame] | 863 | [count] "=&r" (count) |
| 864 | : [a] "r" (a) |
| 865 | : "memory" |
| 866 | ); |
| 867 | } |
| 868 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 869 | void cftbsub_128_mips(float* a) { |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 870 | float f0, f1, f2, f3, f4, f5, f6, f7, f8; |
| 871 | int tmp_a, count; |
| 872 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 873 | cft1st_128_mips(a); |
| 874 | cftmdl_128_mips(a); |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 875 | |
| 876 | __asm __volatile ( |
| 877 | ".set push \n\t" |
| 878 | ".set noreorder \n\t" |
| 879 | "addiu %[tmp_a], %[a], 0 \n\t" |
| 880 | "addiu %[count], $zero, 16 \n\t" |
| 881 | "1: \n\t" |
| 882 | "addiu %[count], %[count], -1 \n\t" |
| 883 | "lwc1 %[f0], 0(%[tmp_a]) \n\t" |
| 884 | "lwc1 %[f2], 128(%[tmp_a]) \n\t" |
| 885 | "lwc1 %[f4], 256(%[tmp_a]) \n\t" |
| 886 | "lwc1 %[f6], 384(%[tmp_a]) \n\t" |
| 887 | "lwc1 %[f1], 4(%[tmp_a]) \n\t" |
| 888 | "lwc1 %[f3], 132(%[tmp_a]) \n\t" |
| 889 | "lwc1 %[f5], 260(%[tmp_a]) \n\t" |
| 890 | "lwc1 %[f7], 388(%[tmp_a]) \n\t" |
| 891 | "add.s %[f8], %[f0], %[f2] \n\t" |
| 892 | "sub.s %[f0], %[f0], %[f2] \n\t" |
| 893 | "add.s %[f2], %[f4], %[f6] \n\t" |
| 894 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 895 | "add.s %[f6], %[f1], %[f3] \n\t" |
| 896 | "sub.s %[f1], %[f3], %[f1] \n\t" |
| 897 | "add.s %[f3], %[f5], %[f7] \n\t" |
| 898 | "sub.s %[f5], %[f5], %[f7] \n\t" |
| 899 | "add.s %[f7], %[f8], %[f2] \n\t" |
| 900 | "sub.s %[f8], %[f8], %[f2] \n\t" |
| 901 | "sub.s %[f2], %[f1], %[f4] \n\t" |
| 902 | "add.s %[f1], %[f1], %[f4] \n\t" |
| 903 | "add.s %[f4], %[f3], %[f6] \n\t" |
| 904 | "sub.s %[f6], %[f3], %[f6] \n\t" |
| 905 | "sub.s %[f3], %[f0], %[f5] \n\t" |
| 906 | "add.s %[f0], %[f0], %[f5] \n\t" |
| 907 | "neg.s %[f4], %[f4] \n\t" |
| 908 | "swc1 %[f7], 0(%[tmp_a]) \n\t" |
| 909 | "swc1 %[f8], 256(%[tmp_a]) \n\t" |
| 910 | "swc1 %[f2], 132(%[tmp_a]) \n\t" |
| 911 | "swc1 %[f1], 388(%[tmp_a]) \n\t" |
| 912 | "swc1 %[f6], 260(%[tmp_a]) \n\t" |
| 913 | "swc1 %[f3], 128(%[tmp_a]) \n\t" |
| 914 | "swc1 %[f0], 384(%[tmp_a]) \n\t" |
| 915 | "swc1 %[f4], 4(%[tmp_a]) \n\t" |
| 916 | "bgtz %[count], 1b \n\t" |
| 917 | " addiu %[tmp_a], %[tmp_a], 8 \n\t" |
| 918 | ".set pop \n\t" |
| 919 | : [f0] "=&f" (f0), [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), |
| 920 | [f4] "=&f" (f4), [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), |
| 921 | [f8] "=&f" (f8), [tmp_a] "=&r" (tmp_a), [count] "=&r" (count) |
| 922 | : [a] "r" (a) |
| 923 | : "memory" |
| 924 | ); |
| 925 | } |
| 926 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 927 | void rftfsub_128_mips(float* a) { |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 928 | const float* c = rdft_w + 32; |
| 929 | const float f0 = 0.5f; |
| 930 | float* a1 = &a[2]; |
| 931 | float* a2 = &a[126]; |
| 932 | const float* c1 = &c[1]; |
| 933 | const float* c2 = &c[31]; |
| 934 | float f1, f2, f3 ,f4, f5, f6, f7, f8, f9, f10, f11, f12, f13, f14, f15; |
| 935 | int count; |
| 936 | |
| 937 | __asm __volatile ( |
| 938 | ".set push \n\t" |
| 939 | ".set noreorder \n\t" |
| 940 | "lwc1 %[f6], 0(%[c2]) \n\t" |
| 941 | "lwc1 %[f1], 0(%[a1]) \n\t" |
| 942 | "lwc1 %[f2], 0(%[a2]) \n\t" |
| 943 | "lwc1 %[f3], 4(%[a1]) \n\t" |
| 944 | "lwc1 %[f4], 4(%[a2]) \n\t" |
| 945 | "lwc1 %[f5], 0(%[c1]) \n\t" |
| 946 | "sub.s %[f6], %[f0], %[f6] \n\t" |
| 947 | "sub.s %[f7], %[f1], %[f2] \n\t" |
| 948 | "add.s %[f8], %[f3], %[f4] \n\t" |
| 949 | "addiu %[count], $zero, 15 \n\t" |
| 950 | "mul.s %[f9], %[f6], %[f7] \n\t" |
| 951 | "mul.s %[f6], %[f6], %[f8] \n\t" |
| 952 | #if !defined(MIPS32_R2_LE) |
| 953 | "mul.s %[f8], %[f5], %[f8] \n\t" |
| 954 | "mul.s %[f5], %[f5], %[f7] \n\t" |
| 955 | "sub.s %[f9], %[f9], %[f8] \n\t" |
| 956 | "add.s %[f6], %[f6], %[f5] \n\t" |
| 957 | #else |
| 958 | "nmsub.s %[f9], %[f9], %[f5], %[f8] \n\t" |
| 959 | "madd.s %[f6], %[f6], %[f5], %[f7] \n\t" |
| 960 | #endif |
| 961 | "sub.s %[f1], %[f1], %[f9] \n\t" |
| 962 | "add.s %[f2], %[f2], %[f9] \n\t" |
| 963 | "sub.s %[f3], %[f3], %[f6] \n\t" |
| 964 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 965 | "swc1 %[f1], 0(%[a1]) \n\t" |
| 966 | "swc1 %[f2], 0(%[a2]) \n\t" |
| 967 | "swc1 %[f3], 4(%[a1]) \n\t" |
| 968 | "swc1 %[f4], 4(%[a2]) \n\t" |
| 969 | "addiu %[a1], %[a1], 8 \n\t" |
| 970 | "addiu %[a2], %[a2], -8 \n\t" |
| 971 | "addiu %[c1], %[c1], 4 \n\t" |
| 972 | "addiu %[c2], %[c2], -4 \n\t" |
| 973 | "1: \n\t" |
| 974 | "lwc1 %[f6], 0(%[c2]) \n\t" |
| 975 | "lwc1 %[f1], 0(%[a1]) \n\t" |
| 976 | "lwc1 %[f2], 0(%[a2]) \n\t" |
| 977 | "lwc1 %[f3], 4(%[a1]) \n\t" |
| 978 | "lwc1 %[f4], 4(%[a2]) \n\t" |
| 979 | "lwc1 %[f5], 0(%[c1]) \n\t" |
| 980 | "sub.s %[f6], %[f0], %[f6] \n\t" |
| 981 | "sub.s %[f7], %[f1], %[f2] \n\t" |
| 982 | "add.s %[f8], %[f3], %[f4] \n\t" |
| 983 | "lwc1 %[f10], -4(%[c2]) \n\t" |
| 984 | "lwc1 %[f11], 8(%[a1]) \n\t" |
| 985 | "lwc1 %[f12], -8(%[a2]) \n\t" |
| 986 | "mul.s %[f9], %[f6], %[f7] \n\t" |
| 987 | "mul.s %[f6], %[f6], %[f8] \n\t" |
| 988 | #if !defined(MIPS32_R2_LE) |
| 989 | "mul.s %[f8], %[f5], %[f8] \n\t" |
| 990 | "mul.s %[f5], %[f5], %[f7] \n\t" |
| 991 | "lwc1 %[f13], 12(%[a1]) \n\t" |
| 992 | "lwc1 %[f14], -4(%[a2]) \n\t" |
| 993 | "lwc1 %[f15], 4(%[c1]) \n\t" |
| 994 | "sub.s %[f9], %[f9], %[f8] \n\t" |
| 995 | "add.s %[f6], %[f6], %[f5] \n\t" |
| 996 | #else |
| 997 | "lwc1 %[f13], 12(%[a1]) \n\t" |
| 998 | "lwc1 %[f14], -4(%[a2]) \n\t" |
| 999 | "lwc1 %[f15], 4(%[c1]) \n\t" |
| 1000 | "nmsub.s %[f9], %[f9], %[f5], %[f8] \n\t" |
| 1001 | "madd.s %[f6], %[f6], %[f5], %[f7] \n\t" |
| 1002 | #endif |
| 1003 | "sub.s %[f10], %[f0], %[f10] \n\t" |
| 1004 | "sub.s %[f5], %[f11], %[f12] \n\t" |
| 1005 | "add.s %[f7], %[f13], %[f14] \n\t" |
| 1006 | "sub.s %[f1], %[f1], %[f9] \n\t" |
| 1007 | "add.s %[f2], %[f2], %[f9] \n\t" |
| 1008 | "sub.s %[f3], %[f3], %[f6] \n\t" |
| 1009 | "mul.s %[f8], %[f10], %[f5] \n\t" |
| 1010 | "mul.s %[f10], %[f10], %[f7] \n\t" |
| 1011 | #if !defined(MIPS32_R2_LE) |
| 1012 | "mul.s %[f9], %[f15], %[f7] \n\t" |
| 1013 | "mul.s %[f15], %[f15], %[f5] \n\t" |
| 1014 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 1015 | "swc1 %[f1], 0(%[a1]) \n\t" |
| 1016 | "swc1 %[f2], 0(%[a2]) \n\t" |
| 1017 | "sub.s %[f8], %[f8], %[f9] \n\t" |
| 1018 | "add.s %[f10], %[f10], %[f15] \n\t" |
| 1019 | #else |
| 1020 | "swc1 %[f1], 0(%[a1]) \n\t" |
| 1021 | "swc1 %[f2], 0(%[a2]) \n\t" |
| 1022 | "sub.s %[f4], %[f4], %[f6] \n\t" |
| 1023 | "nmsub.s %[f8], %[f8], %[f15], %[f7] \n\t" |
| 1024 | "madd.s %[f10], %[f10], %[f15], %[f5] \n\t" |
| 1025 | #endif |
| 1026 | "swc1 %[f3], 4(%[a1]) \n\t" |
| 1027 | "swc1 %[f4], 4(%[a2]) \n\t" |
| 1028 | "sub.s %[f11], %[f11], %[f8] \n\t" |
| 1029 | "add.s %[f12], %[f12], %[f8] \n\t" |
| 1030 | "sub.s %[f13], %[f13], %[f10] \n\t" |
| 1031 | "sub.s %[f14], %[f14], %[f10] \n\t" |
| 1032 | "addiu %[c2], %[c2], -8 \n\t" |
| 1033 | "addiu %[c1], %[c1], 8 \n\t" |
| 1034 | "swc1 %[f11], 8(%[a1]) \n\t" |
| 1035 | "swc1 %[f12], -8(%[a2]) \n\t" |
| 1036 | "swc1 %[f13], 12(%[a1]) \n\t" |
| 1037 | "swc1 %[f14], -4(%[a2]) \n\t" |
| 1038 | "addiu %[a1], %[a1], 16 \n\t" |
| 1039 | "addiu %[count], %[count], -1 \n\t" |
| 1040 | "bgtz %[count], 1b \n\t" |
| 1041 | " addiu %[a2], %[a2], -16 \n\t" |
| 1042 | ".set pop \n\t" |
| 1043 | : [a1] "+r" (a1), [a2] "+r" (a2), [c1] "+r" (c1), [c2] "+r" (c2), |
| 1044 | [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), [f4] "=&f" (f4), |
| 1045 | [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), [f8] "=&f" (f8), |
| 1046 | [f9] "=&f" (f9), [f10] "=&f" (f10), [f11] "=&f" (f11), [f12] "=&f" (f12), |
| 1047 | [f13] "=&f" (f13), [f14] "=&f" (f14), [f15] "=&f" (f15), |
| 1048 | [count] "=&r" (count) |
| 1049 | : [f0] "f" (f0) |
| 1050 | : "memory" |
| 1051 | ); |
| 1052 | } |
| 1053 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 1054 | void rftbsub_128_mips(float* a) { |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 1055 | const float *c = rdft_w + 32; |
| 1056 | const float f0 = 0.5f; |
| 1057 | float* a1 = &a[2]; |
| 1058 | float* a2 = &a[126]; |
| 1059 | const float* c1 = &c[1]; |
| 1060 | const float* c2 = &c[31]; |
| 1061 | float f1, f2, f3 ,f4, f5, f6, f7, f8, f9, f10, f11, f12, f13, f14, f15; |
| 1062 | int count; |
| 1063 | |
| 1064 | a[1] = -a[1]; |
| 1065 | a[65] = -a[65]; |
| 1066 | |
| 1067 | __asm __volatile ( |
| 1068 | ".set push \n\t" |
| 1069 | ".set noreorder \n\t" |
| 1070 | "lwc1 %[f6], 0(%[c2]) \n\t" |
| 1071 | "lwc1 %[f1], 0(%[a1]) \n\t" |
| 1072 | "lwc1 %[f2], 0(%[a2]) \n\t" |
| 1073 | "lwc1 %[f3], 4(%[a1]) \n\t" |
| 1074 | "lwc1 %[f4], 4(%[a2]) \n\t" |
| 1075 | "lwc1 %[f5], 0(%[c1]) \n\t" |
| 1076 | "sub.s %[f6], %[f0], %[f6] \n\t" |
| 1077 | "sub.s %[f7], %[f1], %[f2] \n\t" |
| 1078 | "add.s %[f8], %[f3], %[f4] \n\t" |
| 1079 | "addiu %[count], $zero, 15 \n\t" |
| 1080 | "mul.s %[f9], %[f6], %[f7] \n\t" |
| 1081 | "mul.s %[f6], %[f6], %[f8] \n\t" |
| 1082 | #if !defined(MIPS32_R2_LE) |
| 1083 | "mul.s %[f8], %[f5], %[f8] \n\t" |
| 1084 | "mul.s %[f5], %[f5], %[f7] \n\t" |
| 1085 | "add.s %[f9], %[f9], %[f8] \n\t" |
| 1086 | "sub.s %[f6], %[f6], %[f5] \n\t" |
| 1087 | #else |
| 1088 | "madd.s %[f9], %[f9], %[f5], %[f8] \n\t" |
| 1089 | "nmsub.s %[f6], %[f6], %[f5], %[f7] \n\t" |
| 1090 | #endif |
| 1091 | "sub.s %[f1], %[f1], %[f9] \n\t" |
| 1092 | "add.s %[f2], %[f2], %[f9] \n\t" |
| 1093 | "sub.s %[f3], %[f6], %[f3] \n\t" |
| 1094 | "sub.s %[f4], %[f6], %[f4] \n\t" |
| 1095 | "swc1 %[f1], 0(%[a1]) \n\t" |
| 1096 | "swc1 %[f2], 0(%[a2]) \n\t" |
| 1097 | "swc1 %[f3], 4(%[a1]) \n\t" |
| 1098 | "swc1 %[f4], 4(%[a2]) \n\t" |
| 1099 | "addiu %[a1], %[a1], 8 \n\t" |
| 1100 | "addiu %[a2], %[a2], -8 \n\t" |
| 1101 | "addiu %[c1], %[c1], 4 \n\t" |
| 1102 | "addiu %[c2], %[c2], -4 \n\t" |
| 1103 | "1: \n\t" |
| 1104 | "lwc1 %[f6], 0(%[c2]) \n\t" |
| 1105 | "lwc1 %[f1], 0(%[a1]) \n\t" |
| 1106 | "lwc1 %[f2], 0(%[a2]) \n\t" |
| 1107 | "lwc1 %[f3], 4(%[a1]) \n\t" |
| 1108 | "lwc1 %[f4], 4(%[a2]) \n\t" |
| 1109 | "lwc1 %[f5], 0(%[c1]) \n\t" |
| 1110 | "sub.s %[f6], %[f0], %[f6] \n\t" |
| 1111 | "sub.s %[f7], %[f1], %[f2] \n\t" |
| 1112 | "add.s %[f8], %[f3], %[f4] \n\t" |
| 1113 | "lwc1 %[f10], -4(%[c2]) \n\t" |
| 1114 | "lwc1 %[f11], 8(%[a1]) \n\t" |
| 1115 | "lwc1 %[f12], -8(%[a2]) \n\t" |
| 1116 | "mul.s %[f9], %[f6], %[f7] \n\t" |
| 1117 | "mul.s %[f6], %[f6], %[f8] \n\t" |
| 1118 | #if !defined(MIPS32_R2_LE) |
| 1119 | "mul.s %[f8], %[f5], %[f8] \n\t" |
| 1120 | "mul.s %[f5], %[f5], %[f7] \n\t" |
| 1121 | "lwc1 %[f13], 12(%[a1]) \n\t" |
| 1122 | "lwc1 %[f14], -4(%[a2]) \n\t" |
| 1123 | "lwc1 %[f15], 4(%[c1]) \n\t" |
| 1124 | "add.s %[f9], %[f9], %[f8] \n\t" |
| 1125 | "sub.s %[f6], %[f6], %[f5] \n\t" |
| 1126 | #else |
| 1127 | "lwc1 %[f13], 12(%[a1]) \n\t" |
| 1128 | "lwc1 %[f14], -4(%[a2]) \n\t" |
| 1129 | "lwc1 %[f15], 4(%[c1]) \n\t" |
| 1130 | "madd.s %[f9], %[f9], %[f5], %[f8] \n\t" |
| 1131 | "nmsub.s %[f6], %[f6], %[f5], %[f7] \n\t" |
| 1132 | #endif |
| 1133 | "sub.s %[f10], %[f0], %[f10] \n\t" |
| 1134 | "sub.s %[f5], %[f11], %[f12] \n\t" |
| 1135 | "add.s %[f7], %[f13], %[f14] \n\t" |
| 1136 | "sub.s %[f1], %[f1], %[f9] \n\t" |
| 1137 | "add.s %[f2], %[f2], %[f9] \n\t" |
| 1138 | "sub.s %[f3], %[f6], %[f3] \n\t" |
| 1139 | "mul.s %[f8], %[f10], %[f5] \n\t" |
| 1140 | "mul.s %[f10], %[f10], %[f7] \n\t" |
| 1141 | #if !defined(MIPS32_R2_LE) |
| 1142 | "mul.s %[f9], %[f15], %[f7] \n\t" |
| 1143 | "mul.s %[f15], %[f15], %[f5] \n\t" |
| 1144 | "sub.s %[f4], %[f6], %[f4] \n\t" |
| 1145 | "swc1 %[f1], 0(%[a1]) \n\t" |
| 1146 | "swc1 %[f2], 0(%[a2]) \n\t" |
| 1147 | "add.s %[f8], %[f8], %[f9] \n\t" |
| 1148 | "sub.s %[f10], %[f10], %[f15] \n\t" |
| 1149 | #else |
| 1150 | "swc1 %[f1], 0(%[a1]) \n\t" |
| 1151 | "swc1 %[f2], 0(%[a2]) \n\t" |
| 1152 | "sub.s %[f4], %[f6], %[f4] \n\t" |
| 1153 | "madd.s %[f8], %[f8], %[f15], %[f7] \n\t" |
| 1154 | "nmsub.s %[f10], %[f10], %[f15], %[f5] \n\t" |
| 1155 | #endif |
| 1156 | "swc1 %[f3], 4(%[a1]) \n\t" |
| 1157 | "swc1 %[f4], 4(%[a2]) \n\t" |
| 1158 | "sub.s %[f11], %[f11], %[f8] \n\t" |
| 1159 | "add.s %[f12], %[f12], %[f8] \n\t" |
| 1160 | "sub.s %[f13], %[f10], %[f13] \n\t" |
| 1161 | "sub.s %[f14], %[f10], %[f14] \n\t" |
| 1162 | "addiu %[c2], %[c2], -8 \n\t" |
| 1163 | "addiu %[c1], %[c1], 8 \n\t" |
| 1164 | "swc1 %[f11], 8(%[a1]) \n\t" |
| 1165 | "swc1 %[f12], -8(%[a2]) \n\t" |
| 1166 | "swc1 %[f13], 12(%[a1]) \n\t" |
| 1167 | "swc1 %[f14], -4(%[a2]) \n\t" |
| 1168 | "addiu %[a1], %[a1], 16 \n\t" |
| 1169 | "addiu %[count], %[count], -1 \n\t" |
| 1170 | "bgtz %[count], 1b \n\t" |
| 1171 | " addiu %[a2], %[a2], -16 \n\t" |
| 1172 | ".set pop \n\t" |
| 1173 | : [a1] "+r" (a1), [a2] "+r" (a2), [c1] "+r" (c1), [c2] "+r" (c2), |
| 1174 | [f1] "=&f" (f1), [f2] "=&f" (f2), [f3] "=&f" (f3), [f4] "=&f" (f4), |
| 1175 | [f5] "=&f" (f5), [f6] "=&f" (f6), [f7] "=&f" (f7), [f8] "=&f" (f8), |
| 1176 | [f9] "=&f" (f9), [f10] "=&f" (f10), [f11] "=&f" (f11), [f12] "=&f" (f12), |
| 1177 | [f13] "=&f" (f13), [f14] "=&f" (f14), [f15] "=&f" (f15), |
| 1178 | [count] "=&r" (count) |
| 1179 | : [f0] "f" (f0) |
| 1180 | : "memory" |
| 1181 | ); |
| 1182 | } |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 1183 | #endif |
| andrew@webrtc.org | a22485e | 2014-08-29 17:51:28 +0000 | [diff] [blame] | 1184 | |
| peah | 81b9291 | 2016-10-06 06:46:20 -0700 | [diff] [blame] | 1185 | } // namespace webrtc |