Go to most recent revision | Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 3769 | Serge | 1 | /* |
| 2 | * Copyright 2000-2011 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 | */ |
||
| 16 | // 44 // Total instruction count |
||
| 17 | // 1 // Total kernel count |
||
| 18 | |||
| 19 | .kernel PL3_DN_PL3 |
||
| 20 | .code |
||
| 21 | |||
| 22 | |||
| 23 | |||
| 24 | // FileName: DN_PL_Core.asm |
||
| 25 | // Author: Vivek Kumar |
||
| 26 | // Description: Tasks for DN only case (16x8 block) for planar format |
||
| 27 | |||
| 28 | |||
| 29 | |||
| 30 | // FileName: DN.asm |
||
| 31 | // Author: Vivek Kumar |
||
| 32 | // Description: Tasks for DN only case (16x8 block) |
||
| 33 | |||
| 34 | |||
| 35 | |||
| 36 | |||
| 37 | // Module name: common.inc |
||
| 38 | // |
||
| 39 | // Common header file for all Video-Processing kernels |
||
| 40 | // |
||
| 41 | |||
| 42 | .default_execution_size (16) |
||
| 43 | .default_register_type :ub |
||
| 44 | |||
| 45 | .reg_count_total 128 |
||
| 46 | .reg_count_payload 7 |
||
| 47 | |||
| 48 | //========== Common constants ========== |
||
| 49 | |||
| 50 | |||
| 51 | //========== Macros ========== |
||
| 52 | |||
| 53 | |||
| 54 | //Fast Jump, For more details see "Set_Layer_N.asm" |
||
| 55 | |||
| 56 | |||
| 57 | //========== Defines ==================== |
||
| 58 | |||
| 59 | //========== Static Parameters (Common To All) ========== |
||
| 60 | //r1 |
||
| 61 | |||
| 62 | |||
| 63 | //r2 |
||
| 64 | |||
| 65 | // e.g. byte0 byte1 byte2 |
||
| 66 | // YUYV 0 1 3 |
||
| 67 | // YVYU 0 3 1 |
||
| 68 | |||
| 69 | //Color Pipe (IECP) parameters |
||
| 70 | |||
| 71 | |||
| 72 | //ByteCopy |
||
| 73 | |||
| 74 | |||
| 75 | //r4 |
||
| 76 | |||
| 77 | // e.g. byte0 byte1 byte2 |
||
| 78 | // YUYV 0 1 3 |
||
| 79 | // YVYU 0 3 1 |
||
| 80 | |||
| 81 | |||
| 82 | //========== Inline parameters (Common To All) =========== |
||
| 83 | |||
| 84 | |||
| 85 | //============== Binding Index Table=========== |
||
| 86 | //Common between DNDI and DNUV |
||
| 87 | |||
| 88 | |||
| 89 | //================= Common Message Descriptor ===== |
||
| 90 | // Message descriptor for thread spawning |
||
| 91 | // Message Descriptors |
||
| 92 | // = 000 0001 (min message len 1 ) 0,0000 (resp len 0 -add later) |
||
| 93 | // 0000,0000,0000 |
||
| 94 | // 0001(Spawn a root thread),0001 (Root thread spawn thread) |
||
| 95 | // = 0x02000011 |
||
| 96 | // Thread Spawner Message Descriptor |
||
| 97 | |||
| 98 | |||
| 99 | // Message descriptor for atomic operation add |
||
| 100 | // Message Descriptors |
||
| 101 | // = 000 0110 (min message len 6 ) 0,0000 (resp len 0 -add later) |
||
| 102 | // 1(header present)001,10(typed atomic operation)0(return enabled)0(slot group, low 8 bits),0111 (AOP_Add) |
||
| 103 | // 0000,0000 (Binding table index, added later) |
||
| 104 | // = 0x02000011 |
||
| 105 | |||
| 106 | // Atomic Operation Add Message Descriptor |
||
| 107 | |||
| 108 | |||
| 109 | // Message descriptor for dataport media write |
||
| 110 | // Message Descriptors |
||
| 111 | // = 000 0001 (min message len 1 - add later) 00000 (resp len 0) |
||
| 112 | // 1 (header present 1) 0 1010 (media block write) 000000 |
||
| 113 | // 00000000 (binding table index - set later) |
||
| 114 | // = 0x020A8000 |
||
| 115 | |||
| 116 | |||
| 117 | // Message Length defines |
||
| 118 | |||
| 119 | |||
| 120 | // Response Length defines |
||
| 121 | |||
| 122 | |||
| 123 | // Block Width and Height Size defines |
||
| 124 | |||
| 125 | |||
| 126 | // Extended Message Descriptors |
||
| 127 | |||
| 128 | |||
| 129 | // Common message descriptors: |
||
| 130 | |||
| 131 | |||
| 132 | //===================== Math Function Control =================================== |
||
| 133 | |||
| 134 | |||
| 135 | //============ Message Registers =============== |
||
| 136 | // buf4 starts from r28 |
||
| 137 | |||
| 138 | |||
| 139 | //#define mMSGHDR_EOT r43 // Dummy Message Register for EOT |
||
| 140 | |||
| 141 | |||
| 142 | .declare mubMSGPAYLOAD Base=r30 ElementSize=1 SrcRegion=<16;16,1> Type=ub |
||
| 143 | .declare muwMSGPAYLOAD Base=r30 ElementSize=2 SrcRegion=<16;16,1> Type=uw |
||
| 144 | .declare mudMSGPAYLOAD Base=r30 ElementSize=4 SrcRegion=<8;8,1> Type=ud |
||
| 145 | .declare mfMSGPAYLOAD Base=r30 ElementSize=4 SrcRegion=<8;8,1> Type=f |
||
| 146 | |||
| 147 | //=================== End of thread instruction =========================== |
||
| 148 | |||
| 149 | |||
| 150 | //=====================Pointers Used===================================== |
||
| 151 | |||
| 152 | |||
| 153 | //======================================================================= |
||
| 154 | |||
| 155 | |||
| 156 | //r9-r17 |
||
| 157 | // Define temp space for any usages |
||
| 158 | |||
| 159 | |||
| 160 | // Common Buffers |
||
| 161 | |||
| 162 | |||
| 163 | // temp space for rotation |
||
| 164 | |||
| 165 | .declare fROBUF Base=r9.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=f |
||
| 166 | |||
| 167 | .declare udROBUF Base=r9.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 168 | |||
| 169 | .declare uwROBUF Base=r9.0 ElementSize=2 SrcRegion=<16;16,1> DstRegion=<1> Type=uw |
||
| 170 | |||
| 171 | .declare ubROBUF Base=r9.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 172 | |||
| 173 | .declare ub4ROBUF Base=r9.0 ElementSize=1 SrcRegion=<32;8,4> DstRegion=<4> Type=ub |
||
| 174 | |||
| 175 | |||
| 176 | // End of common.inc |
||
| 177 | |||
| 178 | |||
| 179 | // FileName: DNDI.inc |
||
| 180 | // Author: Vivek Kumar |
||
| 181 | // Description: Include file for DN, DI and DNDI |
||
| 182 | // Inputs: DI_ENABLE, DN_ENABLE, DN_PLANAR, DN_PACKED |
||
| 183 | |||
| 184 | |||
| 185 | |||
| 186 | |||
| 187 | // End of common.inc |
||
| 188 | |||
| 189 | |||
| 190 | //Interface: |
||
| 191 | //Static Parameters: |
||
| 192 | //r1 |
||
| 193 | |||
| 194 | |||
| 195 | //====================== Binding table (Explicit To DNDI)========================================= |
||
| 196 | |||
| 197 | |||
| 198 | .declare mudMSGHDR_DNDI Base=r18 ElementSize=4 Type=ud |
||
| 199 | .declare mdMSGHDR_DNDI Base=r18 ElementSize=4 Type=d |
||
| 200 | .declare mwMSGHDR_DNDI Base=r18 ElementSize=2 Type=w |
||
| 201 | |||
| 202 | |||
| 203 | .declare mudMSGHDR_STMM Base=r20 ElementSize=4 Type=ud |
||
| 204 | |||
| 205 | |||
| 206 | .declare mudMSGHDR_HIST Base=r22 ElementSize=4 Type=ud |
||
| 207 | |||
| 208 | |||
| 209 | .declare mudMSGHDR_ENC_STATS Base=r24 ElementSize=4 Type=ud |
||
| 210 | .declare muwMSGHDR_ENC_STATS Base=r24 ElementSize=2 Type=uw |
||
| 211 | .declare mubMSGHDR_ENC_STATS Base=r24 ElementSize=1 Type=ub |
||
| 212 | |||
| 213 | |||
| 214 | .declare mudMSGHDR_DN_OUT Base=r31.0 ElementSize=4 Type=ud |
||
| 215 | .declare mdMSGHDR_DN_OUT Base=r31.0 ElementSize=4 Type=d |
||
| 216 | .declare mubMSGHDR_DN_OUT Base=r31.0 ElementSize=1 Type=ub |
||
| 217 | |||
| 218 | |||
| 219 | .declare mudMSGHDR_UVCOPY Base=r36 ElementSize=4 Type=ud |
||
| 220 | .declare mdMSGHDR_UVCOPY Base=r36 ElementSize=4 Type=d |
||
| 221 | .declare mudMSGHDR_UCOPY Base=r36 ElementSize=4 Type=ud |
||
| 222 | .declare mudMSGHDR_VCOPY Base=r38 ElementSize=4 Type=ud |
||
| 223 | |||
| 224 | |||
| 225 | .declare mudMSGHDR_DI_OUT1 Base=r18.0 ElementSize=4 Type=ud |
||
| 226 | .declare mubMSGHDR_DI_OUT1 Base=r18.0 ElementSize=1 Type=ub |
||
| 227 | |||
| 228 | |||
| 229 | .declare mudMSGHDR_DI_OUT2 Base=r23.0 ElementSize=4 Type=ud |
||
| 230 | .declare mubMSGHDR_DI_OUT2 Base=r23.0 ElementSize=1 Type=ub |
||
| 231 | |||
| 232 | //r45 |
||
| 233 | //Use r45 as message header, so no need to "mov" the data. |
||
| 234 | |||
| 235 | .declare mudDN_Y_OUT Base=r45.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 236 | |||
| 237 | // Message response (Denoised & DI-ed pixels & statistics); Use buffer 5 |
||
| 238 | .declare udDNDI_RESP Base=r46.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 239 | .declare uwDNDI_RESP Base=r46.0 ElementSize=2 SrcRegion=<16;16,1> DstRegion=<1> Type=uw |
||
| 240 | .declare ubDNDI_RESP Base=r46.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 241 | |||
| 242 | // Message response (UV Copy); Use buffer 5 |
||
| 243 | .declare udDNDI_UV_RESP Base=r58.0 ElementSize=4 SrcRegion=<8;8,1> DstRegion=<1> Type=ud |
||
| 244 | .declare ubDNDI_UV_RESP Base=r58.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub |
||
| 245 | |||
| 246 | //Temp GRFs: For 42X to 422 Conversion |
||
| 247 | .declare uwDNDI_UVCOPY_TEMP Base=r10.0 ElementSize=2 SrcRegion=<16;16,1> DstRegion=<1> Type=uw //8 GRFs |
||
| 248 | .declare ubDNDI_UVCOPY_TEMP Base=r10.0 ElementSize=1 SrcRegion=<16;16,1> DstRegion=<1> Type=ub //8 GRFs |
||
| 249 | //--------------------------------------------------------------------------- |
||
| 250 | // Message descriptors |
||
| 251 | //--------------------------------------------------------------------------- |
||
| 252 | // Extended message descriptor |
||
| 253 | // Message descriptor for sampler read |
||
| 254 | // = 000 0010 (message len 2) 00000 (resp len - set later, 12 or 5 or 11) |
||
| 255 | // 1 (header present 1) 0 11 (SIMD32/64 mode) |
||
| 256 | // 1000 (message type) 0000 (DI state index) |
||
| 257 | // 00000000 (binding table index - set later) |
||
| 258 | // = 0x040b8000 |
||
| 259 | |||
| 260 | |||
| 261 | // Attention: The Message Length is The Number of GRFs with Data Only, without the Header |
||
| 262 | |||
| 263 | |||
| 264 | //--------------------------------------------------------------------------- |
||
| 265 | // VDI Return Data format |
||
| 266 | //--------------------------------------------------------------------------- |
||
| 267 | // Defines for DI enabled |
||
| 268 | |||
| 269 | |||
| 270 | // Defines for DI disabled |
||
| 271 | |||
| 272 | |||
| 273 | |||
| 274 | // FileName: DNDI_Command.asm |
||
| 275 | // Author: Vivek Kumar |
||
| 276 | // Description: Sends a message to the VDI to process one DN (16x8) or DNDI (16x4) block |
||
| 277 | |||
| 278 | // Prepare the DNDI send command |
||
| 279 | mov (8) mudMSGHDR_DNDI(0)<1> r0.0<8;8,1>:ud // message header |
||
| 280 | mov (1) mwMSGHDR_DNDI(1,4)<1> r7.0<0;1,0>:w { NoDDClr } // horizontal origin // Do we need to add offset here? -vK |
||
| 281 | mov (1) mwMSGHDR_DNDI(1,12)<1> r7.1<0;1,0>:w { NoDDChk } // vertical origin // Can these 2 be combined? - vK |
||
| 282 | |||
| 283 | send (8) udDNDI_RESP(0)<1> r18 0x2 0x45E8003:ud |
||
| 284 | |||
| 285 | // On Gen6, with VDI walker, use the XY pair returned rather than programmed above |
||
| 286 | // VDI_RETURNED_XY is ordered XY in case of walker enables and the same as programmed in case of walker disabled |
||
| 287 | mov (2) r7.0<1>:w uwDNDI_RESP(4,14)<2;2,1> // horizontal/Vertial origin in W.14 and W.15 |
||
| 288 | |||
| 289 | |||
| 290 | |||
| 291 | // FileName: DN_Hist_Save.asm |
||
| 292 | // Author: Vivek Kumar |
||
| 293 | // Description: Saves DN history data to statistics surface |
||
| 294 | |||
| 295 | // Write denoise history to memory |
||
| 296 | mov (8) r27<1>:ud r0.0<8;8,1>:ud // message header |
||
| 297 | |||
| 298 | |||
| 299 | mov (2) mudMSGHDR_HIST(1)<1> udDNDI_RESP(4,0)<2;2,1> // Move denoise history to MRF (4x2) |
||
| 300 | |||
| 301 | |||
| 302 | shr (2) r27.0<1>:ud r7.0<2;2,1>:w 2:w // X,Y origin / 4 |
||
| 303 | add (1) r27.0<1>:ud r27.0<0;1,0>:ud r1.12<0;1,0>:uw { NoDDClr } // Add pitch to X origin |
||
| 304 | mov (1) r27.2<1>:ud 0x10003:ud { NoDDChk } // block width and height |
||
| 305 | |||
| 306 | mov (8) mudMSGHDR_HIST(0)<1> r27.0<8;8,1>:ud |
||
| 307 | send (8) null<1>:d r22 0x5 0x40A8021:ud |
||
| 308 | |||
| 309 | |||
| 310 | |||
| 311 | // FileName: DNDI_Enc_Stats_Save.asm |
||
| 312 | // Author: Vivek Kumar |
||
| 313 | // Description: Saves Encoder Statistics data to statistics surface in case of DI enabled (for 16x4 block) |
||
| 314 | |||
| 315 | // Write encoder statistics to memory |
||
| 316 | //Currently enable this only on Gen6 validation |
||
| 317 | mov (8) mudMSGHDR_ENC_STATS(1)<1> 0x0:ud // Init payload MRF |
||
| 318 | mov (8) mudMSGHDR_ENC_STATS(0)<1> r0.0<8;8,1>:ud // message header |
||
| 319 | |||
| 320 | shr (1) mudMSGHDR_ENC_STATS(0,0)<1> r7.0<0;1,0>:w 1:w { NoDDClr } //enable the flag after testing on si { NoDDClr } // X origin / 2 |
||
| 321 | mul (1) acc0.1<1>:ud r7.1<0;1,0>:w 3:w // Y origin * 3 |
||
| 322 | shr (1) mudMSGHDR_ENC_STATS(0,1)<1> acc0.1<0;1,0>:ud 2:w { NoDDClr, NoDDChk } //enable the flag after testing on si { NoDDClr, NoDDChk } // Y origin * 3/4 |
||
| 323 | mov (1) mudMSGHDR_ENC_STATS(0,2)<1> 0x50003:ud { NoDDChk } //enable the flag after testing on si { NoDDChk } // block width and height (8x3) |
||
| 324 | add (2) mudMSGHDR_ENC_STATS(0,0)<1> mudMSGHDR_ENC_STATS(0,0)<2;2,1> r1.12<2;2,1>:uw // Add pitch to X,Y origin |
||
| 325 | |||
| 326 | |||
| 327 | //Data block for Encoder Statistics |
||
| 328 | //---------------------------------------------------- |
||
| 329 | //| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | Bytes |
||
| 330 | //---------------------------------------------------- |
||
| 331 | //| BNE | X | X | X | X | |
||
| 332 | //---------------------------------------------------- |
||
| 333 | //| X | SVCM | X | |
||
| 334 | //---------------------------------------------------- |
||
| 335 | //| SHCM | STAD | X | |
||
| 336 | //---------------------------------------------------- |
||
| 337 | //| X | X | |
||
| 338 | //---------------------------------------------------- |
||
| 339 | //| X | SVCM | X | |
||
| 340 | //---------------------------------------------------- |
||
| 341 | //| SHCM | STAD | X | |
||
| 342 | //---------------------------------------------------- |
||
| 343 | mov (1) mubMSGHDR_ENC_STATS(1,0)<1> ubDNDI_RESP(4,8)<0;1,0> { NoDDClr } // Move encoder statistics to MRF |
||
| 344 | mov (1) muwMSGHDR_ENC_STATS(1,3)<1> uwDNDI_RESP(4,11)<0;1,0> { NoDDClr, NoDDChk } // Move encoder statistics to MRF |
||
| 345 | mov (2) muwMSGHDR_ENC_STATS(1,4)<1> uwDNDI_RESP(4,12)<2;2,1> { NoDDClr, NoDDChk } // Move encoder statistics to MRF |
||
| 346 | mov (1) muwMSGHDR_ENC_STATS(1,9)<1> uwDNDI_RESP(4,8)<0;1,0> { NoDDClr, NoDDChk } // Move encoder statistics to MRF |
||
| 347 | mov (2) muwMSGHDR_ENC_STATS(1,10)<1> uwDNDI_RESP(4,9)<2;2,1> { NoDDChk } // Move encoder statistics to MRF |
||
| 348 | |||
| 349 | |||
| 350 | send (8) null<1>:d r24 0x5 0x40A8021:ud |
||
| 351 | |||
| 352 | |||
| 353 | |||
| 354 | // FileName: DN_Load_UV_IMC3_16x8.asm |
||
| 355 | // Author: Vivek Kumar |
||
| 356 | // Description: Read UV for 16x8 block through DATAPORT |
||
| 357 | |||
| 358 | |||
| 359 | |||
| 360 | // FileName: UVCopy_Load_16x8.asm |
||
| 361 | // Author: Vivek Kumar |
||
| 362 | // Description: Read UV for 16x8 block through DATAPORT |
||
| 363 | |||
| 364 | //CHANGE : Read extra UV data to convert to 422. -rT |
||
| 365 | //we are reading extra data in ALL cases irrespective of whether upsampling is reqd or not later on, to keep things simple. |
||
| 366 | |||
| 367 | |||
| 368 | add (2) r27.0<1>:d r7.0<2;2,1>:w r4.4<2;2,1>:w // Source Y Block origin |
||
| 369 | asr (2) r27.0<1>:d r27.0<2;2,1>:d 1:w { NoDDClr } // U/V block origin should be half of Y's |
||
| 370 | mov (1) r27.2<1>:ud 0x40007:ud { NoDDChk } // U/V block width and height (8x5) |
||
| 371 | mov (8) r36<1>:ud r27.0<8;8,1>:ud |
||
| 372 | mov (8) r38<1>:ud r27.0<8;8,1>:ud |
||
| 373 | send (8) udDNDI_UV_RESP(0)<1> r36 0x4 0x2290001:ud |
||
| 374 | send (8) udDNDI_UV_RESP(2)<1> r38 0x4 0x2290002:ud |
||
| 375 | |||
| 376 | //Update Header for Save |
||
| 377 | mov (1) mudMSGHDR_UCOPY(0,2)<1> 0x30007:ud // U block width and height (8x4) |
||
| 378 | mov (1) mudMSGHDR_VCOPY(0,2)<1> 0x30007:ud // V block width and height (8x4) |
||
| 379 | |||
| 380 | |||
| 381 | |||
| 382 | // FileName: DN_Save_Y_16x8.asm |
||
| 383 | // Author: Vivek Kumar |
||
| 384 | // Description: Save one 16x8 blocks of Y channel of DN output for reference |
||
| 385 | |||
| 386 | |||
| 387 | mov (8) mudDN_Y_OUT(0,0)<1> r0<8;8,1>:ud // message header |
||
| 388 | mov (2) mudDN_Y_OUT(0,0)<1> r7.0<2;2,1>:w { NoDDClr } // X origin |
||
| 389 | mov (1) mudDN_Y_OUT(0,2)<1> 0x7000F:ud { NoDDChk } // block width and height (16x8) |
||
| 390 | |||
| 391 | //send out data through data port |
||
| 392 | send (8) null<1>:d mudDN_Y_OUT 0x5 0xA0A8018:ud |
||
| 393 | |||
| 394 | |||
| 395 | |||
| 396 | // FileName: DN_Save_UV_IMC3_16x8.asm |
||
| 397 | // Author: Vivek Kumar |
||
| 398 | // Description: Save UV for 16x8 block through DATAPORT |
||
| 399 | |||
| 400 | |||
| 401 | |||
| 402 | // FileName: UVCopy_Save_16x8.asm |
||
| 403 | // Author: Vivek Kumar |
||
| 404 | // Description: Save UV for 16x8 block through DATAPORT |
||
| 405 | |||
| 406 | |||
| 407 | //Reuse the header from Load component |
||
| 408 | //Header is modified at the end of load - to be usable for save. |
||
| 409 | |||
| 410 | |||
| 411 | mov (8) mudMSGHDR_UCOPY(1)<1> udDNDI_UV_RESP(0)<8;8,1> |
||
| 412 | mov (8) mudMSGHDR_VCOPY(1)<1> udDNDI_UV_RESP(2)<8;8,1> |
||
| 413 | send (4) null<1>:d r36 0x5 0x40A8019:ud |
||
| 414 | send (4) null<1>:d r38 0x5 0x40A801A:ud |
||
| 415 | |||
| 416 | |||
| 417 | |||
| 418 | //End of Thread message |
||
| 419 | |||
| 420 | mov (8) r127<1>:ud r0.0<8;8,1>:ud |
||
| 421 | send (1) null<1>:d r127 0x27 0x02000010 |
||
| 422 | |||
| 423 | |||
| 424 | .end_code |
||
| 425 | .end_kernel |