Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 5361 | serge | 1 | /* |
| 2 | * Copyright 2000-2013 Intel Corporation All Rights Reserved |
||
| 3 | * |
||
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
||
| 5 | * you may not use this file except in compliance with the License. |
||
| 6 | * You may obtain a copy of the License at |
||
| 7 | * |
||
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
||
| 9 | * |
||
| 10 | * Unless required by applicable law or agreed to in writing, software |
||
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
||
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
||
| 13 | * See the License for the specific language governing permissions and |
||
| 14 | * limitations under the License. |
||
| 15 | * Authors: |
||
| 16 | * Zhao Yakui |
||
| 17 | */ |
||
| 18 | |||
| 19 | |||
| 20 | |||
| 21 | // Module name: common.inc |
||
| 22 | // |
||
| 23 | // Common header file for all Video-Processing kernels |
||
| 24 | // |
||
| 25 | |||
| 26 | .default_execution_size (16) |
||
| 27 | .default_register_type :ub |
||
| 28 | |||
| 29 | .reg_count_total 128 |
||
| 30 | .reg_count_payload 7 |
||
| 31 | |||
| 32 | //========== Common constants ========== |
||
| 33 | |||
| 34 | |||
| 35 | //========== Macros ========== |
||
| 36 | |||
| 37 | |||
| 38 | //Fast Jump, For more details see "Set_Layer_N.asm" |
||
| 39 | |||
| 40 | |||
| 41 | //========== Defines ==================== |
||
| 42 | |||
| 43 | //========== Static Parameters (Common To All) ========== |
||
| 44 | //r1 |
||
| 45 | |||
| 46 | |||
| 47 | //r2 |
||
| 48 | |||
| 49 | // e.g. byte0 byte1 byte2 |
||
| 50 | // YUYV 0 1 3 |
||
| 51 | // YVYU 0 3 1 |
||
| 52 | |||
| 53 | //Color Pipe (IECP) parameters |
||
| 54 | |||
| 55 | |||
| 56 | //ByteCopy |
||
| 57 | |||
| 58 | |||
| 59 | //r4 |
||
| 60 | |||
| 61 | // e.g. byte0 byte1 byte2 |
||
| 62 | // YUYV 0 1 3 |
||
| 63 | // YVYU 0 3 1 |
||
| 64 | |||
| 65 | |||
| 66 | //========== Inline parameters (Common To All) =========== |
||
| 67 | |||
| 68 | |||
| 69 | //============== Binding Index Table=========== |
||
| 70 | //Common between DNDI and DNUV |
||
| 71 | |||
| 72 | |||
| 73 | //================= Common Message Descriptor ===== |
||
| 74 | // Message descriptor for thread spawning |
||
| 75 | // Message Descriptors |
||
| 76 | // = 000 0001 (min message len 1 ) 0,0000 (resp len 0 -add later) |
||
| 77 | // 0000,0000,0000 |
||
| 78 | // 0001(Spawn a root thread),0001 (Root thread spawn thread) |
||
| 79 | // = 0x02000011 |
||
| 80 | // Thread Spawner Message Descriptor |
||
| 81 | |||
| 82 | |||
| 83 | // Message descriptor for atomic operation add |
||
| 84 | // Message Descriptors |
||
| 85 | // = 000 0110 (min message len 6 ) 0,0000 (resp len 0 -add later) |
||
| 86 | // 1(header present)001,10(typed atomic operation)0(return enabled)0(slot group, low 8 bits),0111 (AOP_Add) |
||
| 87 | // 0000,0000 (Binding table index, added later) |
||
| 88 | // = 0x02000011 |
||
| 89 | |||
| 90 | // Atomic Operation Add Message Descriptor |
||
| 91 | |||
| 92 | |||
| 93 | // Message descriptor for dataport media write |
||
| 94 | // Message Descriptors |
||
| 95 | // = 000 0001 (min message len 1 - add later) 00000 (resp len 0) |
||
| 96 | // 1 (header present 1) 0 1010 (media block write) 000000 |
||
| 97 | // 00000000 (binding table index - set later) |
||
| 98 | // = 0x020A8000 |
||
| 99 | |||
| 100 | |||
| 101 | // Message Length defines |
||
| 102 | |||
| 103 | |||
| 104 | // Response Length defines |
||
| 105 | |||
| 106 | |||
| 107 | // Block Width and Height Size defines |
||
| 108 | |||
| 109 | |||
| 110 | // Extended Message Descriptors |
||
| 111 | |||
| 112 | |||
| 113 | // Common message descriptors: |
||
| 114 | |||
| 115 | |||
| 116 | //===================== Math Function Control =================================== |
||
| 117 | |||
| 118 | |||
| 119 | //============ Message Registers =============== |
||
| 120 | // buf4 starts from r28 |
||
| 121 | |||
| 122 | |||
| 123 | //#define mMSGHDR_EOT r43 // Dummy Message Register for EOT |
||
| 124 | |||
| 125 | |||
| 126 | .declare mubMSGPAYLOAD Base=r30 ElementSize=1 SrcRegion=<16;16,1> Type=ub |
||
| 127 | .declare muwMSGPAYLOAD Base=r30 ElementSize=2 SrcRegion=<16;16,1> Type=uw |
||
| 128 | .declare mudMSGPAYLOAD Base=r30 ElementSize=4 SrcRegion=<8;8,1> Type=ud |
||
| 129 | .declare mfMSGPAYLOAD Base=r30 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 130 | |||
| 131 | //=================== End of thread instruction =========================== |
||
| 132 | |||
| 133 | |||
| 134 | //=====================Pointers Used===================================== |
||
| 135 | |||
| 136 | |||
| 137 | //======================================================================= |
||
| 138 | |||
| 139 | |||
| 140 | //r9-r17 |
||
| 141 | // Define temp space for any usages |
||
| 142 | |||
| 143 | |||
| 144 | // Common Buffers |
||
| 145 | |||
| 146 | |||
| 147 | // temp space for rotation |
||
| 148 | |||
| 149 | .declare fROBUF Base=r9.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=f |
||
| 150 | |||
| 151 | .declare udROBUF Base=r9.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 152 | |||
| 153 | .declare uwROBUF Base=r9.0 ElementSize=2 SrcRegion=<16;16,1> DstRegion=<1> Type=uw |
||
| 154 | |||
| 155 | .declare ubROBUF Base=r9.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 156 | |||
| 157 | .declare ub4ROBUF Base=r9.0 ElementSize=1 SrcRegion=<32;8,4> DstRegion=<4> Type=ub |
||
| 158 | |||
| 159 | |||
| 160 | // End of common.inc |
||
| 161 | |||
| 162 | |||
| 163 | // Module name: YUV_to_RGB.asm |
||
| 164 | // |
||
| 165 | // Convert YUV to RGB, handle it by 16x4 block |
||
| 166 | // |
||
| 167 | |||
| 168 | |||
| 169 | // Description: Includes all definitions explicit to Fast Composite. |
||
| 170 | |||
| 171 | |||
| 172 | //========== GRF partition ========== |
||
| 173 | // r0 header : r0 (1 GRF) |
||
| 174 | // Static parameters : r1 - r6 (6 GRFS) |
||
| 175 | // Inline parameters : r7 - r8 (2 GRFs) |
||
| 176 | // MSGSRC : r27 (1 GRF) |
||
| 177 | //=================================== |
||
| 178 | |||
| 179 | //Interface: |
||
| 180 | //========== Static Parameters (Explicit To Fast Composite) ========== |
||
| 181 | //r1 |
||
| 182 | //CSC Set 0 |
||
| 183 | |||
| 184 | |||
| 185 | .declare udCSC_CURBE Base=r1.0 ElementSize=4 Type=ud |
||
| 186 | |||
| 187 | //Constant alpha |
||
| 188 | |||
| 189 | |||
| 190 | //r2 |
||
| 191 | |||
| 192 | |||
| 193 | // Gen7 AVS WA |
||
| 194 | |||
| 195 | |||
| 196 | // WiDi Definitions |
||
| 197 | |||
| 198 | |||
| 199 | //Colorfill |
||
| 200 | |||
| 201 | |||
| 202 | // 0: 0-degree, 1: 90, 2: 180, 3: 270-degree, clockwise. |
||
| 203 | |||
| 204 | .declare ubCOLOR_PIXEL_VAL Base=r2.20 ElementSize=1 SrcRegion=<0;1,0> DstRegion=<1> Type=ub |
||
| 205 | |||
| 206 | //r3 |
||
| 207 | //Normalised Ratio of Horizontal step size with main video for all layers |
||
| 208 | |||
| 209 | |||
| 210 | //Normalised Ratio of Horizontal step size with main video for all layers becomes |
||
| 211 | //Normalised Horizontal step size for all layers in VP_Setup.asm |
||
| 212 | |||
| 213 | |||
| 214 | //r4 |
||
| 215 | //Normalised Vertical step size for all layers |
||
| 216 | |||
| 217 | |||
| 218 | //r5 |
||
| 219 | //Normalised Vertical Frame Origin for all layers |
||
| 220 | |||
| 221 | |||
| 222 | //r6 |
||
| 223 | //Normalised Horizontal Frame Origin for all layers |
||
| 224 | |||
| 225 | |||
| 226 | //========== Inline Parameters (Explicit To Fast Composite) ========== |
||
| 227 | |||
| 228 | |||
| 229 | //Main video Step X |
||
| 230 | |||
| 231 | |||
| 232 | //====================== Binding table (Explicit To Fast Composite)========================================= |
||
| 233 | |||
| 234 | |||
| 235 | //Used by Interlaced Scaling Kernels |
||
| 236 | |||
| 237 | |||
| 238 | //========== Sampler State Table Index (Explicit To Fast Composite)========== |
||
| 239 | //Sampler Index for AVS/IEF messages |
||
| 240 | |||
| 241 | |||
| 242 | //Sampler Index for SIMD16 sampler messages |
||
| 243 | |||
| 244 | |||
| 245 | //============================================================================= |
||
| 246 | |||
| 247 | .declare fBUFFER_0 Base=r64.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=f |
||
| 248 | .declare fBUFFER_1 Base=r80.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=f |
||
| 249 | .declare fBUFFER_2 Base=r96.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=f |
||
| 250 | .declare fBUFFER_3 Base=r112.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=f |
||
| 251 | |||
| 252 | .declare bBUFFER_0 Base=r64.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 253 | .declare bBUFFER_1 Base=r80.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 254 | .declare bBUFFER_2 Base=r96.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 255 | .declare bBUFFER_3 Base=r112.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 256 | |||
| 257 | //Pointer to mask reg |
||
| 258 | |||
| 259 | |||
| 260 | //r18 |
||
| 261 | |||
| 262 | |||
| 263 | //Always keep Cannel Pointers and Offsets in same GRF, so that we can use |
||
| 264 | // NODDCLR, NODDCHK flags. -rT |
||
| 265 | |||
| 266 | |||
| 267 | .declare udCSC_COEFF_0 Base=r18.0 ElementSize=4 Type=ud // 1 GRF |
||
| 268 | |||
| 269 | //r19 |
||
| 270 | |||
| 271 | |||
| 272 | .declare udCSC_COEFF_1 Base=r19.0 ElementSize=4 Type=ud // 1 GRF |
||
| 273 | |||
| 274 | |||
| 275 | //r20 |
||
| 276 | |||
| 277 | .declare uwALPHA_MASK_REG_TEMP Base=r20.0 ElementSize=2 SrcRegion=<16;16,1> Type=uw // 1 GRF |
||
| 278 | |||
| 279 | //r21 |
||
| 280 | |||
| 281 | .declare uwALPHA_MASK_REG Base=r21.0 ElementSize=2 SrcRegion=<16;16,1> Type=uw // 1 GRF |
||
| 282 | |||
| 283 | //r22 |
||
| 284 | |||
| 285 | |||
| 286 | //Always keep Cannel Pointers and Offsets in same GRF, so that we can use |
||
| 287 | // NODDCLR, NODDCHK flags. -rT |
||
| 288 | |||
| 289 | |||
| 290 | //Keep fORIGIN_X_NLAS, fY_OFFSET_2ND_BLOCK, fSTEP_X_NLAS, pMSGDSC_COPY, ubCONST_ALPHA_COPY as |
||
| 291 | //sub registers of same GRF to enable using NODDCLR NODDCHK. -rT |
||
| 292 | |||
| 293 | //r23 |
||
| 294 | |||
| 295 | |||
| 296 | //Lumakey |
||
| 297 | |||
| 298 | |||
| 299 | //r24 |
||
| 300 | |||
| 301 | |||
| 302 | //r25 |
||
| 303 | |||
| 304 | |||
| 305 | //r26 |
||
| 306 | |||
| 307 | |||
| 308 | //defines to generate LABELS during compile time. |
||
| 309 | |||
| 310 | |||
| 311 | //Msg payload buffers; upto 4 full-size messages can be written |
||
| 312 | |||
| 313 | //Unnecessary to use the MSGPayLoad, So it is temporiarily used for conversion of YUV->RGB |
||
| 314 | |||
| 315 | .declare fBUFFER_R Base=r28.0 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 316 | .declare fBUFFER_G Base=r30.0 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 317 | .declare fBUFFER_B Base=r32.0 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 318 | |||
| 319 | .declare fBUFFER_Y Base=r36.0 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 320 | .declare fBUFFER_U Base=r38.0 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 321 | .declare fBUFFER_V Base=r40.0 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 322 | |||
| 323 | |||
| 324 | .declare wTempR Base=r42.0 ElementSize=2 Type=w |
||
| 325 | .declare wTempG Base=r44.0 ElementSize=2 Type=w |
||
| 326 | .declare wTempB Base=r46.0 ElementSize=2 Type=w |
||
| 327 | |||
| 328 | .declare ubTempR Base=r42.0 ElementSize=1 Type=ub |
||
| 329 | .declare ubTempG Base=r44.0 ElementSize=1 Type=ub |
||
| 330 | .declare ubTempB Base=r46.0 ElementSize=1 Type=ub |
||
| 331 | |||
| 332 | // the r17 register (nTEMP0) is originally defined from "Common.inc" |
||
| 333 | // instead of re-defining a nTEMP0 here, we use "SAVE_RGB" suffix for its naming |
||
| 334 | |||
| 335 | .declare wTemp0 Base=r17.0 ElementSize=2 Type=uw |
||
| 336 | |||
| 337 | |||
| 338 | //_SAVE_INC_ |
||
| 339 | // NTSC standard |
||
| 340 | // R = Clamp ( 1.164(Y-16/255) + 1.596(Cr-128/255)) |
||
| 341 | // G = Clamp ( 1.164(Y-16/255) - 0.813(Cr-128/255) - 0.392(Cb-128/255)) |
||
| 342 | // B = Clamp ( 1.164(Y-16/255) + 2.017(Cb-128/255)) |
||
| 343 | // ITU-R conversion, Now we are using ITU-R conversion |
||
| 344 | // R = clip( Y + 1.402*(Cr-128)) // ITU-R |
||
| 345 | // G = clip( Y - 0.344*(Cb-128) - 0.714*(Cr-128)) |
||
| 346 | // B = clip( Y + 1.772*(Cb-128)) |
||
| 347 | |||
| 348 | // At the save module we have all 8 address sub-registers available. |
||
| 349 | // So we will use PING-PONG type of scheme to save the data using |
||
| 350 | // pointers pBUF_CHNL_TOP_8x4 and pBUF_CHNL_BOT_8x4. This will help |
||
| 351 | // reduce dependency. - rT |
||
| 352 | |||
| 353 | //wBUFF_CHNL_PTR points to either buffer 0 or buffer 4. |
||
| 354 | //Add appropriate offsets to get pointers for all buffers (1,2,3 or 5). |
||
| 355 | //Offsets are zero for buffer 0 and buffer 4. |
||
| 356 | //Y/U/V is also stored as R/G/B for the internal purpose |
||
| 357 | //for BUFFER_0 |
||
| 358 | mov (4) a0.0<1>:uw r22.0<4;4,1>:uw |
||
| 359 | //the first line in the block 0 |
||
| 360 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 1]<16;8,2>:ub |
||
| 361 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 17]<16;8,2>:ub |
||
| 362 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 1]<16;8,2>:ub |
||
| 363 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 17]<16;8,2>:ub |
||
| 364 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 1]<16;8,2>:ub |
||
| 365 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 17]<16;8,2>:ub |
||
| 366 | |||
| 367 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 368 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 369 | |||
| 370 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 371 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 372 | |||
| 373 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 374 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 375 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 376 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 377 | |||
| 378 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 379 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 380 | |||
| 381 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 382 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 383 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 384 | |||
| 385 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 386 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 387 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 388 | |||
| 389 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 390 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 391 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 392 | |||
| 393 | mov (16) r[a0.1,1]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 394 | mov (16) r[a0.2,1]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 395 | mov (16) r[a0.0,1]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 396 | |||
| 397 | //the second line in the block 0 |
||
| 398 | |||
| 399 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 33]<16;8,2>:ub |
||
| 400 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 49]<16;8,2>:ub |
||
| 401 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 33]<16;8,2>:ub |
||
| 402 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 49]<16;8,2>:ub |
||
| 403 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 33]<16;8,2>:ub |
||
| 404 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 49]<16;8,2>:ub |
||
| 405 | |||
| 406 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 407 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 408 | |||
| 409 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 410 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 411 | |||
| 412 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 413 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 414 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 415 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 416 | |||
| 417 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 418 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 419 | |||
| 420 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 421 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 422 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 423 | |||
| 424 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 425 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 426 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 427 | |||
| 428 | |||
| 429 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 430 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 431 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 432 | mov (16) r[a0.1,33]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 433 | mov (16) r[a0.2,33]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 434 | mov (16) r[a0.0,33]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 435 | |||
| 436 | //the third line in the block 0 |
||
| 437 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 65]<16;8,2>:ub |
||
| 438 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 81]<16;8,2>:ub |
||
| 439 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 65]<16;8,2>:ub |
||
| 440 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 81]<16;8,2>:ub |
||
| 441 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 65]<16;8,2>:ub |
||
| 442 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 81]<16;8,2>:ub |
||
| 443 | |||
| 444 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 445 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 446 | |||
| 447 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 448 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 449 | |||
| 450 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 451 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 452 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 453 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 454 | |||
| 455 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 456 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 457 | |||
| 458 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 459 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 460 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 461 | |||
| 462 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 463 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 464 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 465 | |||
| 466 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 467 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 468 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 469 | mov (16) r[a0.1,65]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 470 | mov (16) r[a0.2,65]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 471 | mov (16) r[a0.0,65]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 472 | |||
| 473 | //the fourth line in the block 0 |
||
| 474 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 97]<16;8,2>:ub |
||
| 475 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 113]<16;8,2>:ub |
||
| 476 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 97]<16;8,2>:ub |
||
| 477 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 113]<16;8,2>:ub |
||
| 478 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 97]<16;8,2>:ub |
||
| 479 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 113]<16;8,2>:ub |
||
| 480 | |||
| 481 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 482 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 483 | |||
| 484 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 485 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 486 | |||
| 487 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 488 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 489 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 490 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 491 | |||
| 492 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 493 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 494 | |||
| 495 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 496 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 497 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 498 | |||
| 499 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 500 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 501 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 502 | |||
| 503 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 504 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 505 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 506 | mov (16) r[a0.1,97]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 507 | mov (16) r[a0.2,97]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 508 | mov (16) r[a0.0,97]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 509 | |||
| 510 | |||
| 511 | //for BUFFER_1 |
||
| 512 | add (4) a0.0<1>:uw r22.0<4;4,1>:uw 512:uw |
||
| 513 | //the first line in the block 1 |
||
| 514 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 1]<16;8,2>:ub |
||
| 515 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 17]<16;8,2>:ub |
||
| 516 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 1]<16;8,2>:ub |
||
| 517 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 17]<16;8,2>:ub |
||
| 518 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 1]<16;8,2>:ub |
||
| 519 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 17]<16;8,2>:ub |
||
| 520 | |||
| 521 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 522 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 523 | |||
| 524 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 525 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 526 | |||
| 527 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 528 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 529 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 530 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 531 | |||
| 532 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 533 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 534 | |||
| 535 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 536 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 537 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 538 | |||
| 539 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 540 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 541 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 542 | |||
| 543 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 544 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 545 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 546 | |||
| 547 | mov (16) r[a0.1,1]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 548 | mov (16) r[a0.2,1]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 549 | mov (16) r[a0.0,1]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 550 | |||
| 551 | //the second line in the block 1 |
||
| 552 | |||
| 553 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 33]<16;8,2>:ub |
||
| 554 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 49]<16;8,2>:ub |
||
| 555 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 33]<16;8,2>:ub |
||
| 556 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 49]<16;8,2>:ub |
||
| 557 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 33]<16;8,2>:ub |
||
| 558 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 49]<16;8,2>:ub |
||
| 559 | |||
| 560 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 561 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 562 | |||
| 563 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 564 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 565 | |||
| 566 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 567 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 568 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 569 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 570 | |||
| 571 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 572 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 573 | |||
| 574 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 575 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 576 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 577 | |||
| 578 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 579 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 580 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 581 | |||
| 582 | |||
| 583 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 584 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 585 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 586 | mov (16) r[a0.1,33]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 587 | mov (16) r[a0.2,33]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 588 | mov (16) r[a0.0,33]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 589 | |||
| 590 | //the third line in the block 1 |
||
| 591 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 65]<16;8,2>:ub |
||
| 592 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 81]<16;8,2>:ub |
||
| 593 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 65]<16;8,2>:ub |
||
| 594 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 81]<16;8,2>:ub |
||
| 595 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 65]<16;8,2>:ub |
||
| 596 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 81]<16;8,2>:ub |
||
| 597 | |||
| 598 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 599 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 600 | |||
| 601 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 602 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 603 | |||
| 604 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 605 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 606 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 607 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 608 | |||
| 609 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 610 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 611 | |||
| 612 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 613 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 614 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 615 | |||
| 616 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 617 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 618 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 619 | |||
| 620 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 621 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 622 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 623 | mov (16) r[a0.1,65]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 624 | mov (16) r[a0.2,65]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 625 | mov (16) r[a0.0,65]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 626 | |||
| 627 | //the fourth line in the block 1 |
||
| 628 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 97]<16;8,2>:ub |
||
| 629 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 113]<16;8,2>:ub |
||
| 630 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 97]<16;8,2>:ub |
||
| 631 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 113]<16;8,2>:ub |
||
| 632 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 97]<16;8,2>:ub |
||
| 633 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 113]<16;8,2>:ub |
||
| 634 | |||
| 635 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 636 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 637 | |||
| 638 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 639 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 640 | |||
| 641 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 642 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 643 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 644 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 645 | |||
| 646 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 647 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 648 | |||
| 649 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 650 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 651 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 652 | |||
| 653 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 654 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 655 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 656 | |||
| 657 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 658 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 659 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 660 | mov (16) r[a0.1,97]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 661 | mov (16) r[a0.2,97]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 662 | mov (16) r[a0.0,97]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 663 | |||
| 664 | |||
| 665 | //for BUFFER_2 |
||
| 666 | add (4) a0.0<1>:uw r22.0<4;4,1>:uw 1024:uw |
||
| 667 | //the first line in the block 2 |
||
| 668 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 1]<16;8,2>:ub |
||
| 669 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 17]<16;8,2>:ub |
||
| 670 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 1]<16;8,2>:ub |
||
| 671 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 17]<16;8,2>:ub |
||
| 672 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 1]<16;8,2>:ub |
||
| 673 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 17]<16;8,2>:ub |
||
| 674 | |||
| 675 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 676 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 677 | |||
| 678 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 679 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 680 | |||
| 681 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 682 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 683 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 684 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 685 | |||
| 686 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 687 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 688 | |||
| 689 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 690 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 691 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 692 | |||
| 693 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 694 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 695 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 696 | |||
| 697 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 698 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 699 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 700 | |||
| 701 | mov (16) r[a0.1,1]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 702 | mov (16) r[a0.2,1]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 703 | mov (16) r[a0.0,1]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 704 | |||
| 705 | //the second line in the block 2 |
||
| 706 | |||
| 707 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 33]<16;8,2>:ub |
||
| 708 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 49]<16;8,2>:ub |
||
| 709 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 33]<16;8,2>:ub |
||
| 710 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 49]<16;8,2>:ub |
||
| 711 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 33]<16;8,2>:ub |
||
| 712 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 49]<16;8,2>:ub |
||
| 713 | |||
| 714 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 715 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 716 | |||
| 717 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 718 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 719 | |||
| 720 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 721 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 722 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 723 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 724 | |||
| 725 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 726 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 727 | |||
| 728 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 729 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 730 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 731 | |||
| 732 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 733 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 734 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 735 | |||
| 736 | |||
| 737 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 738 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 739 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 740 | mov (16) r[a0.1,33]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 741 | mov (16) r[a0.2,33]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 742 | mov (16) r[a0.0,33]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 743 | |||
| 744 | //the third line in the block 2 |
||
| 745 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 65]<16;8,2>:ub |
||
| 746 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 81]<16;8,2>:ub |
||
| 747 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 65]<16;8,2>:ub |
||
| 748 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 81]<16;8,2>:ub |
||
| 749 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 65]<16;8,2>:ub |
||
| 750 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 81]<16;8,2>:ub |
||
| 751 | |||
| 752 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 753 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 754 | |||
| 755 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 756 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 757 | |||
| 758 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 759 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 760 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 761 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 762 | |||
| 763 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 764 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 765 | |||
| 766 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 767 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 768 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 769 | |||
| 770 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 771 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 772 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 773 | |||
| 774 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 775 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 776 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 777 | mov (16) r[a0.1,65]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 778 | mov (16) r[a0.2,65]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 779 | mov (16) r[a0.0,65]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 780 | |||
| 781 | //the fourth line in the block 2 |
||
| 782 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 97]<16;8,2>:ub |
||
| 783 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 113]<16;8,2>:ub |
||
| 784 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 97]<16;8,2>:ub |
||
| 785 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 113]<16;8,2>:ub |
||
| 786 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 97]<16;8,2>:ub |
||
| 787 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 113]<16;8,2>:ub |
||
| 788 | |||
| 789 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 790 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 791 | |||
| 792 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 793 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 794 | |||
| 795 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 796 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 797 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 798 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 799 | |||
| 800 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 801 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 802 | |||
| 803 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 804 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 805 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 806 | |||
| 807 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 808 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 809 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 810 | |||
| 811 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 812 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 813 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 814 | mov (16) r[a0.1,97]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 815 | mov (16) r[a0.2,97]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 816 | mov (16) r[a0.0,97]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 817 | |||
| 818 | |||
| 819 | //for BUFFER_3 |
||
| 820 | add (4) a0.0<1>:uw r22.0<4;4,1>:uw 1536:uw |
||
| 821 | //the first line in the block 3 |
||
| 822 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 1]<16;8,2>:ub |
||
| 823 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 17]<16;8,2>:ub |
||
| 824 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 1]<16;8,2>:ub |
||
| 825 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 17]<16;8,2>:ub |
||
| 826 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 1]<16;8,2>:ub |
||
| 827 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 17]<16;8,2>:ub |
||
| 828 | |||
| 829 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 830 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 831 | |||
| 832 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 833 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 834 | |||
| 835 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 836 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 837 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 838 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 839 | |||
| 840 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 841 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 842 | |||
| 843 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 844 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 845 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 846 | |||
| 847 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 848 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 849 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 850 | |||
| 851 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 852 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 853 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 854 | |||
| 855 | mov (16) r[a0.1,1]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 856 | mov (16) r[a0.2,1]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 857 | mov (16) r[a0.0,1]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 858 | |||
| 859 | //the second line in the block 3 |
||
| 860 | |||
| 861 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 33]<16;8,2>:ub |
||
| 862 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 49]<16;8,2>:ub |
||
| 863 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 33]<16;8,2>:ub |
||
| 864 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 49]<16;8,2>:ub |
||
| 865 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 33]<16;8,2>:ub |
||
| 866 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 49]<16;8,2>:ub |
||
| 867 | |||
| 868 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 869 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 870 | |||
| 871 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 872 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 873 | |||
| 874 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 875 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 876 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 877 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 878 | |||
| 879 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 880 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 881 | |||
| 882 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 883 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 884 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 885 | |||
| 886 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 887 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 888 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 889 | |||
| 890 | |||
| 891 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 892 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 893 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 894 | mov (16) r[a0.1,33]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 895 | mov (16) r[a0.2,33]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 896 | mov (16) r[a0.0,33]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 897 | |||
| 898 | //the third line in the block 3 |
||
| 899 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 65]<16;8,2>:ub |
||
| 900 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 81]<16;8,2>:ub |
||
| 901 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 65]<16;8,2>:ub |
||
| 902 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 81]<16;8,2>:ub |
||
| 903 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 65]<16;8,2>:ub |
||
| 904 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 81]<16;8,2>:ub |
||
| 905 | |||
| 906 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 907 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 908 | |||
| 909 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 910 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 911 | |||
| 912 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 913 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 914 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 915 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 916 | |||
| 917 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 918 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 919 | |||
| 920 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 921 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 922 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 923 | |||
| 924 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 925 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 926 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 927 | |||
| 928 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 929 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 930 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 931 | mov (16) r[a0.1,65]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 932 | mov (16) r[a0.2,65]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 933 | mov (16) r[a0.0,65]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 934 | |||
| 935 | //the fourth line in the block 3 |
||
| 936 | mov (8) fBUFFER_Y(0, 0)<1> r[a0.1, 97]<16;8,2>:ub |
||
| 937 | mov (8) fBUFFER_Y(1, 0)<1> r[a0.1, 113]<16;8,2>:ub |
||
| 938 | mov (8) fBUFFER_U(0, 0)<1> r[a0.2, 97]<16;8,2>:ub |
||
| 939 | mov (8) fBUFFER_U(1, 0)<1> r[a0.2, 113]<16;8,2>:ub |
||
| 940 | mov (8) fBUFFER_V(0, 0)<1> r[a0.0, 97]<16;8,2>:ub |
||
| 941 | mov (8) fBUFFER_V(1, 0)<1> r[a0.0, 113]<16;8,2>:ub |
||
| 942 | |||
| 943 | add (16) fBUFFER_U(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> -128.0f |
||
| 944 | add (16) fBUFFER_V(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -128.0f |
||
| 945 | |||
| 946 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 947 | mac (16) fBUFFER_R(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 1.402f |
||
| 948 | |||
| 949 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 950 | mac (8) acc0.0<1>:f fBUFFER_U(0, 0)<8;8,1> -0.344f |
||
| 951 | mac (8) acc1.0<1>:f fBUFFER_U(1, 0)<8;8,1> -0.344f |
||
| 952 | mac (16) fBUFFER_G(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> -0.714f |
||
| 953 | |||
| 954 | mov (16) acc0.0<1>:f fBUFFER_Y(0, 0)<8;8,1> |
||
| 955 | mac (16) fBUFFER_B(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 1.772f |
||
| 956 | |||
| 957 | mul.sat (16) fBUFFER_Y(0, 0)<1> fBUFFER_R(0, 0)<8;8,1> 0.0039216f |
||
| 958 | mul.sat (16) fBUFFER_U(0, 0)<1> fBUFFER_G(0, 0)<8;8,1> 0.0039216f |
||
| 959 | mul.sat (16) fBUFFER_V(0, 0)<1> fBUFFER_B(0, 0)<8;8,1> 0.0039216f |
||
| 960 | |||
| 961 | mul (16) fBUFFER_R(0, 0)<1> fBUFFER_Y(0, 0)<8;8,1> 255.0f |
||
| 962 | mul (16) fBUFFER_G(0, 0)<1> fBUFFER_U(0, 0)<8;8,1> 255.0f |
||
| 963 | mul (16) fBUFFER_B(0, 0)<1> fBUFFER_V(0, 0)<8;8,1> 255.0f |
||
| 964 | |||
| 965 | mov (16) wTempR(0, 0)<2> fBUFFER_R(0, 0)<8;8,1> |
||
| 966 | mov (16) wTempG(0, 0)<2> fBUFFER_G(0, 0)<8;8,1> |
||
| 967 | mov (16) wTempB(0, 0)<2> fBUFFER_B(0, 0)<8;8,1> |
||
| 968 | mov (16) r[a0.1,97]<2>:ub ubTempR(0, 0)<32;8,4> |
||
| 969 | mov (16) r[a0.2,97]<2>:ub ubTempG(0, 0)<32;8,4> |
||
| 970 | mov (16) r[a0.0,97]<2>:ub ubTempB(0, 0)<32;8,4> |
||
| 971 |