Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1 | /****************************************************************************** |
| 2 | * |
| 3 | * Copyright (C) 2012 Ittiam Systems Pvt Ltd, Bangalore |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at: |
| 8 | * |
| 9 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | * |
| 17 | ******************************************************************************/ |
| 18 | /** |
| 19 | ******************************************************************************* |
| 20 | * @file |
| 21 | * ihevcd_utils.c |
| 22 | * |
| 23 | * @brief |
| 24 | * Contains miscellaneous utility functions such as init() etc |
| 25 | * |
| 26 | * @author |
| 27 | * Harish |
| 28 | * |
| 29 | * @par List of Functions: |
| 30 | * |
| 31 | * @remarks |
| 32 | * None |
| 33 | * |
| 34 | ******************************************************************************* |
| 35 | */ |
| 36 | /*****************************************************************************/ |
| 37 | /* File Includes */ |
| 38 | /*****************************************************************************/ |
| 39 | #include <stdio.h> |
| 40 | #include <stddef.h> |
| 41 | #include <stdlib.h> |
| 42 | #include <string.h> |
| 43 | #include <assert.h> |
| 44 | |
| 45 | #include "ihevc_typedefs.h" |
| 46 | #include "iv.h" |
| 47 | #include "ivd.h" |
| 48 | #include "ihevcd_cxa.h" |
| 49 | #include "ithread.h" |
| 50 | |
| 51 | #include "ihevc_defs.h" |
| 52 | #include "ihevc_debug.h" |
| 53 | #include "ihevc_defs.h" |
| 54 | #include "ihevc_error.h" |
| 55 | #include "ihevc_structs.h" |
| 56 | #include "ihevc_buf_mgr.h" |
| 57 | #include "ihevc_dpb_mgr.h" |
| 58 | #include "ihevc_macros.h" |
| 59 | #include "ihevc_platform_macros.h" |
| 60 | |
| 61 | #include "ihevc_common_tables.h" |
| 62 | #include "ihevc_buf_mgr.h" |
| 63 | #include "ihevc_disp_mgr.h" |
| 64 | #include "ihevc_cabac_tables.h" |
| 65 | |
| 66 | #include "ihevcd_defs.h" |
| 67 | |
| 68 | #include "ihevcd_function_selector.h" |
| 69 | #include "ihevcd_structs.h" |
| 70 | #include "ihevcd_error.h" |
| 71 | #include "ihevcd_nal.h" |
| 72 | #include "ihevcd_bitstream.h" |
| 73 | #include "ihevcd_utils.h" |
| 74 | #include "ihevcd_trace.h" |
| 75 | #include "ihevcd_process_slice.h" |
| 76 | #include "ihevcd_job_queue.h" |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 77 | #define MAX_DPB_PIC_BUF 6 |
| 78 | |
| 79 | /* Function declarations */ |
| 80 | mv_buf_t* ihevcd_mv_mgr_get_poc(buf_mgr_t *ps_mv_buf_mgr, UWORD32 abs_poc); |
| 81 | |
| 82 | /** |
| 83 | ******************************************************************************* |
| 84 | * |
| 85 | * @brief |
| 86 | * Used to get level index for a given level |
| 87 | * |
| 88 | * @par Description: |
| 89 | * Converts from level_idc (which is multiplied by 30) to an index that can be |
| 90 | * used as a lookup. Also used to ignore invalid levels like 2.2 , 3.2 etc |
| 91 | * |
| 92 | * @param[in] level |
| 93 | * Level of the stream |
| 94 | * |
| 95 | * @returns Level index for a given level |
| 96 | * |
| 97 | * @remarks |
| 98 | * |
| 99 | * |
| 100 | ******************************************************************************* |
| 101 | */ |
| 102 | WORD32 ihevcd_get_lvl_idx(WORD32 level) |
| 103 | { |
| 104 | WORD32 lvl_idx = 0; |
| 105 | |
| 106 | if(level < IHEVC_LEVEL_20) |
| 107 | { |
| 108 | lvl_idx = 0; |
| 109 | } |
| 110 | else if(level >= IHEVC_LEVEL_20 && level < IHEVC_LEVEL_21) |
| 111 | { |
| 112 | lvl_idx = 1; |
| 113 | } |
| 114 | else if(level >= IHEVC_LEVEL_21 && level < IHEVC_LEVEL_30) |
| 115 | { |
| 116 | lvl_idx = 2; |
| 117 | } |
| 118 | else if(level >= IHEVC_LEVEL_30 && level < IHEVC_LEVEL_31) |
| 119 | { |
| 120 | lvl_idx = 3; |
| 121 | } |
| 122 | else if(level >= IHEVC_LEVEL_31 && level < IHEVC_LEVEL_40) |
| 123 | { |
| 124 | lvl_idx = 4; |
| 125 | } |
| 126 | else if(level >= IHEVC_LEVEL_40 && level < IHEVC_LEVEL_41) |
| 127 | { |
| 128 | lvl_idx = 5; |
| 129 | } |
| 130 | else if(level >= IHEVC_LEVEL_41 && level < IHEVC_LEVEL_50) |
| 131 | { |
| 132 | lvl_idx = 6; |
| 133 | } |
| 134 | else if(level >= IHEVC_LEVEL_50 && level < IHEVC_LEVEL_51) |
| 135 | { |
| 136 | lvl_idx = 7; |
| 137 | } |
| 138 | else if(level >= IHEVC_LEVEL_51 && level < IHEVC_LEVEL_52) |
| 139 | { |
| 140 | lvl_idx = 8; |
| 141 | } |
| 142 | else if(level >= IHEVC_LEVEL_52 && level < IHEVC_LEVEL_60) |
| 143 | { |
| 144 | lvl_idx = 9; |
| 145 | } |
| 146 | else if(level >= IHEVC_LEVEL_60 && level < IHEVC_LEVEL_61) |
| 147 | { |
| 148 | lvl_idx = 10; |
| 149 | } |
| 150 | else if(level >= IHEVC_LEVEL_61 && level < IHEVC_LEVEL_62) |
| 151 | { |
| 152 | lvl_idx = 11; |
| 153 | } |
| 154 | else if(level >= IHEVC_LEVEL_62) |
| 155 | { |
| 156 | lvl_idx = 12; |
| 157 | } |
| 158 | |
| 159 | return (lvl_idx); |
| 160 | } |
| 161 | |
| 162 | /** |
| 163 | ******************************************************************************* |
| 164 | * |
| 165 | * @brief |
| 166 | * Used to get DPB size for a given level, and number of luma samples |
| 167 | * |
| 168 | * @par Description: |
| 169 | * For given width, height and level number of max_dpb_size is computed as per |
| 170 | * Annex A.4.1 |
| 171 | * |
| 172 | * @param[in] level |
| 173 | * Level of the stream |
| 174 | * |
| 175 | * @param[in] pic_size |
| 176 | * Width * Height |
| 177 | * |
| 178 | * @returns Number of buffers in DPB |
| 179 | * |
| 180 | * @remarks |
| 181 | * |
| 182 | * |
| 183 | ******************************************************************************* |
| 184 | */ |
| 185 | WORD32 ihevcd_get_dpb_size(WORD32 level, WORD32 pic_size) |
| 186 | { |
| 187 | |
| 188 | WORD32 max_luma_samples; |
| 189 | |
| 190 | WORD32 max_dpb_size; |
| 191 | WORD32 lvl_idx = ihevcd_get_lvl_idx(level); |
| 192 | max_luma_samples = gai4_ihevc_max_luma_pic_size[lvl_idx]; |
| 193 | |
| 194 | |
| 195 | |
| 196 | if(pic_size <= (max_luma_samples >> 2)) |
| 197 | { |
| 198 | max_dpb_size = MIN(4 * MAX_DPB_PIC_BUF, 16); |
| 199 | } |
| 200 | else if(pic_size <= (max_luma_samples >> 1)) |
| 201 | { |
| 202 | max_dpb_size = MIN(2 * MAX_DPB_PIC_BUF, 16); |
| 203 | } |
| 204 | else if(pic_size <= ((3 * max_luma_samples) >> 2)) |
| 205 | { |
| 206 | max_dpb_size = MIN((4 * MAX_DPB_PIC_BUF) / 3, 16); |
| 207 | } |
| 208 | else |
| 209 | { |
| 210 | max_dpb_size = MAX_DPB_PIC_BUF; |
| 211 | } |
| 212 | |
| 213 | return max_dpb_size; |
| 214 | } |
| 215 | /** |
| 216 | ******************************************************************************* |
| 217 | * |
| 218 | * @brief |
| 219 | * Used to get reference picture buffer size for a given level and |
| 220 | * and padding used |
| 221 | * |
| 222 | * @par Description: |
| 223 | * Used to get reference picture buffer size for a given level and padding used |
| 224 | * Each picture is padded on all four sides |
| 225 | * |
| 226 | * @param[in] pic_size |
| 227 | * Mumber of luma samples (Width * Height) |
| 228 | * |
| 229 | * @param[in] level |
| 230 | * Level |
| 231 | * |
| 232 | * @param[in] horz_pad |
| 233 | * Total padding used in horizontal direction |
| 234 | * |
| 235 | * @param[in] vert_pad |
| 236 | * Total padding used in vertical direction |
| 237 | * |
| 238 | * @returns Total picture buffer size |
| 239 | * |
| 240 | * @remarks |
| 241 | * |
| 242 | * |
| 243 | ******************************************************************************* |
| 244 | */ |
| 245 | WORD32 ihevcd_get_total_pic_buf_size(WORD32 pic_size, |
| 246 | WORD32 level, |
| 247 | WORD32 horz_pad, |
| 248 | WORD32 vert_pad, |
| 249 | WORD32 num_ref_frames, |
| 250 | WORD32 num_reorder_frames) |
| 251 | { |
| 252 | WORD32 size; |
| 253 | WORD32 num_luma_samples; |
| 254 | WORD32 lvl_idx; |
| 255 | WORD32 max_wd; |
| 256 | WORD32 max_dpb_size; |
| 257 | WORD32 num_samples; |
| 258 | WORD32 max_num_bufs; |
| 259 | WORD32 pad = MAX(horz_pad, vert_pad); |
| 260 | |
| 261 | |
| 262 | /* Get maximum number of buffers for the current picture size */ |
| 263 | max_dpb_size = ihevcd_get_dpb_size(level, pic_size); |
| 264 | |
| 265 | |
| 266 | max_num_bufs = (2 * max_dpb_size + 1); |
| 267 | /* If num_ref_frames and num_reorder_frmaes is specified |
| 268 | * Use minimum value |
| 269 | */ |
| 270 | max_num_bufs = MIN(max_num_bufs, (num_ref_frames + num_reorder_frames + 1)); |
| 271 | |
| 272 | /* Get level index */ |
| 273 | lvl_idx = ihevcd_get_lvl_idx(level); |
| 274 | |
| 275 | /* Maximum number of luma samples in a picture at given level */ |
| 276 | num_luma_samples = gai4_ihevc_max_luma_pic_size[lvl_idx]; |
| 277 | |
| 278 | /* Account for chroma */ |
| 279 | num_samples = num_luma_samples * 3 / 2; |
| 280 | |
| 281 | /* Maximum width of luma samples in a picture at given level */ |
| 282 | max_wd = gai4_ihevc_max_wd_ht[lvl_idx]; |
| 283 | |
| 284 | |
| 285 | /* Allocation is required for |
| 286 | * (Wd + horz_pad) * (Ht + vert_pad) * (2 * max_dpb_size + 1) |
| 287 | * |
| 288 | * Above expanded as |
| 289 | * ((Wd * Ht) + (horz_pad * vert_pad) + Wd * vert_pad + Ht * horz_pad) * (2 * max_dpb_size + 1) |
| 290 | * (Wd * Ht) * (2 * max_dpb_size + 1) + ((horz_pad * vert_pad) + Wd * vert_pad + Ht * horz_pad) * (2 * max_dpb_size + 1) |
| 291 | * Now max_dpb_size increases with smaller Wd and Ht, but Wd * ht * max_dpb_size will still be lesser or equal to max_wd * max_ht * dpb_size |
| 292 | * |
| 293 | * In the above equation (Wd * Ht) * (2 * max_dpb_size + 1) is accounted by using num_samples * (2 * max_dpb_size + 1) below |
| 294 | * |
| 295 | * For the padded area use MAX(horz_pad, vert_pad) as pad |
| 296 | * ((pad * pad) + pad * (Wd + Ht)) * (2 * max_dpb_size + 1) has to accounted from the above for padding |
| 297 | * |
| 298 | * Since Width and Height can change worst Wd + Ht is when One of the dimensions is max and other is min |
| 299 | * So use max_wd and min_ht |
| 300 | */ |
| 301 | |
| 302 | /* Number of bytes in reference pictures */ |
| 303 | size = num_samples * max_num_bufs; |
| 304 | |
| 305 | /* Account for padding area */ |
| 306 | size += ((pad * pad) + pad * (max_wd + max_wd)) * max_num_bufs; |
| 307 | |
| 308 | return size; |
| 309 | } |
| 310 | /** |
| 311 | ******************************************************************************* |
| 312 | * |
| 313 | * @brief |
| 314 | * Used to get MV bank size for a given number of luma samples |
| 315 | * |
| 316 | * @par Description: |
| 317 | * For given number of luma samples one MV bank size is computed |
| 318 | * Each MV bank includes pu_map and pu_t for all the min PUs(4x4) in a picture |
| 319 | * |
| 320 | * @param[in] num_luma_samples |
| 321 | * Max number of luma pixels in the frame |
| 322 | * |
| 323 | * @returns Total MV Bank size |
| 324 | * |
| 325 | * @remarks |
| 326 | * |
| 327 | * |
| 328 | ******************************************************************************* |
| 329 | */ |
| 330 | WORD32 ihevcd_get_pic_mv_bank_size(WORD32 num_luma_samples) |
| 331 | { |
| 332 | WORD32 size; |
| 333 | |
| 334 | WORD32 pic_size; |
| 335 | |
| 336 | WORD32 mv_bank_size; |
| 337 | WORD32 num_pu; |
| 338 | WORD32 num_ctb; |
| 339 | pic_size = num_luma_samples; |
| 340 | |
| 341 | |
| 342 | num_pu = pic_size / (MIN_PU_SIZE * MIN_PU_SIZE); |
| 343 | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
| 344 | |
| 345 | mv_bank_size = 0; |
| 346 | |
| 347 | /* Size for storing pu_t start index each CTB */ |
| 348 | /* One extra entry is needed to compute number of PUs in the last CTB */ |
| 349 | mv_bank_size += (num_ctb + 1) * sizeof(WORD32); |
| 350 | |
| 351 | /* Size for pu_map */ |
| 352 | mv_bank_size += num_pu; |
| 353 | |
| 354 | /* Size for storing pu_t for each PU */ |
| 355 | mv_bank_size += num_pu * sizeof(pu_t); |
| 356 | |
Harish Mahendrakar | df5f9ca | 2015-10-09 19:19:06 +0530 | [diff] [blame] | 357 | /* Size for storing slice_idx for each CTB */ |
| 358 | mv_bank_size += ALIGN4(num_ctb * sizeof(UWORD16)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 359 | |
| 360 | size = mv_bank_size; |
| 361 | return size; |
| 362 | } |
| 363 | /** |
| 364 | ******************************************************************************* |
| 365 | * |
| 366 | * @brief |
| 367 | * Used to get TU data size for a given number luma samples |
| 368 | * |
| 369 | * @par Description: |
| 370 | * For a given number of luma samples TU data size is computed |
| 371 | * Each TU data includes tu_map and tu_t and coeff data for all |
| 372 | * the min TUs(4x4) in given CTB |
| 373 | * |
| 374 | * @param[in] num_luma_samples |
| 375 | * Number of 64 x 64 CTBs for which TU data has to be allocated. |
| 376 | * |
| 377 | * @returns Total TU data size |
| 378 | * |
| 379 | * @remarks Assumption is num_luma_samples will be at least |
| 380 | * 64 x 64 to handle CTB of size 64 x 64. Can be frame size as well |
| 381 | * |
| 382 | ******************************************************************************* |
| 383 | */ |
| 384 | WORD32 ihevcd_get_tu_data_size(WORD32 num_luma_samples) |
| 385 | { |
| 386 | |
| 387 | |
| 388 | WORD32 tu_data_size; |
| 389 | WORD32 num_ctb; |
| 390 | WORD32 num_luma_tu, num_chroma_tu, num_tu; |
| 391 | num_ctb = num_luma_samples / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
| 392 | |
| 393 | num_luma_tu = num_luma_samples / (MIN_TU_SIZE * MIN_TU_SIZE); |
| 394 | num_chroma_tu = num_luma_tu >> 1; |
| 395 | |
| 396 | num_tu = num_luma_tu + num_chroma_tu; |
| 397 | tu_data_size = 0; |
| 398 | |
| 399 | /* Size for storing tu_t start index each CTB */ |
| 400 | /* One extra entry is needed to compute number of TUs in the last CTB */ |
| 401 | tu_data_size += (num_ctb + 1) * sizeof(WORD32); |
| 402 | |
| 403 | /* Size for storing tu map */ |
| 404 | tu_data_size += num_luma_tu * sizeof(UWORD8); |
| 405 | |
| 406 | /* Size for storing tu_t for each TU */ |
| 407 | tu_data_size += num_tu * sizeof(tu_t); |
| 408 | |
| 409 | /* Size for storing number of coded subblocks and scan_idx for each TU */ |
| 410 | tu_data_size += num_tu * (sizeof(WORD8) + sizeof(WORD8)); |
| 411 | |
| 412 | /* Size for storing coeff data for each TU */ |
| 413 | tu_data_size += num_tu * sizeof(tu_sblk_coeff_data_t); |
| 414 | |
| 415 | |
| 416 | return tu_data_size; |
| 417 | } |
| 418 | |
| 419 | |
| 420 | WORD32 ihevcd_nctb_cnt(codec_t *ps_codec, sps_t *ps_sps) |
| 421 | { |
| 422 | WORD32 nctb = 1; |
| 423 | UNUSED(ps_codec); |
| 424 | //TODO: Currently set to 1 |
| 425 | /* If CTB size is less than 32 x 32 then set nCTB as 4 */ |
| 426 | if(ps_sps->i1_log2_ctb_size < 5) |
| 427 | nctb = 1; |
| 428 | |
| 429 | return nctb; |
| 430 | } |
| 431 | |
| 432 | IHEVCD_ERROR_T ihevcd_get_tile_pos(pps_t *ps_pps, |
| 433 | sps_t *ps_sps, |
| 434 | WORD32 ctb_x, |
| 435 | WORD32 ctb_y, |
| 436 | WORD32 *pi4_ctb_tile_x, |
| 437 | WORD32 *pi4_ctb_tile_y, |
| 438 | WORD32 *pi4_tile_idx) |
| 439 | { |
| 440 | |
| 441 | tile_t *ps_tile_tmp; |
| 442 | WORD32 i; |
| 443 | WORD32 tile_row, tile_col; |
| 444 | |
| 445 | if(ctb_x < 0 || ctb_y < 0) |
| 446 | { |
| 447 | *pi4_ctb_tile_x = 0; |
| 448 | *pi4_ctb_tile_y = 0; |
| 449 | *pi4_tile_idx = 0; |
| 450 | |
| 451 | return (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
| 452 | } |
| 453 | |
| 454 | tile_row = 0; |
| 455 | tile_col = 0; |
| 456 | ps_tile_tmp = ps_pps->ps_tile; |
| 457 | if(0 == ps_pps->i1_tiles_enabled_flag) |
| 458 | { |
| 459 | *pi4_ctb_tile_x = ctb_x; |
| 460 | *pi4_ctb_tile_y = ctb_y; |
| 461 | *pi4_tile_idx = 0; |
| 462 | } |
| 463 | else |
| 464 | { |
| 465 | for(i = 0; i < ps_pps->i1_num_tile_columns; i++) |
| 466 | { |
| 467 | WORD16 next_tile_ctb_x; |
| 468 | ps_tile_tmp = ps_pps->ps_tile + i; //* ps_pps->i1_num_tile_rows; |
| 469 | if((ps_pps->i1_num_tile_columns - 1) == i) |
| 470 | { |
| 471 | next_tile_ctb_x = ps_sps->i2_pic_wd_in_ctb; |
| 472 | } |
| 473 | else |
| 474 | { |
| 475 | tile_t *ps_tile_next_tmp; |
| 476 | ps_tile_next_tmp = ps_pps->ps_tile + i + 1; |
| 477 | next_tile_ctb_x = ps_tile_next_tmp->u1_pos_x; |
| 478 | } |
| 479 | if((ctb_x >= ps_tile_tmp->u1_pos_x) && (ctb_x < next_tile_ctb_x)) |
| 480 | { |
| 481 | tile_col = i; |
| 482 | break; |
| 483 | } |
| 484 | } |
| 485 | *pi4_ctb_tile_x = ctb_x - ps_tile_tmp->u1_pos_x; |
| 486 | |
| 487 | for(i = 0; i < ps_pps->i1_num_tile_rows; i++) |
| 488 | { |
| 489 | WORD16 next_tile_ctb_y; |
| 490 | ps_tile_tmp = ps_pps->ps_tile + i * ps_pps->i1_num_tile_columns; |
| 491 | if((ps_pps->i1_num_tile_rows - 1) == i) |
| 492 | { |
| 493 | next_tile_ctb_y = ps_sps->i2_pic_ht_in_ctb; |
| 494 | } |
| 495 | else |
| 496 | { |
| 497 | tile_t *ps_tile_next_tmp; |
| 498 | ps_tile_next_tmp = ps_pps->ps_tile + ((i + 1) * ps_pps->i1_num_tile_columns); |
| 499 | next_tile_ctb_y = ps_tile_next_tmp->u1_pos_y; |
| 500 | } |
| 501 | if((ctb_y >= ps_tile_tmp->u1_pos_y) && (ctb_y < next_tile_ctb_y)) |
| 502 | { |
| 503 | tile_row = i; |
| 504 | break; |
| 505 | } |
| 506 | |
| 507 | } |
| 508 | *pi4_ctb_tile_y = ctb_y - ps_tile_tmp->u1_pos_y; |
| 509 | *pi4_tile_idx = tile_row * ps_pps->i1_num_tile_columns |
| 510 | + tile_col; |
| 511 | } |
| 512 | return (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
| 513 | } |
| 514 | /** |
| 515 | ******************************************************************************* |
| 516 | * |
| 517 | * @brief |
| 518 | * Function to initialize ps_pic_buf structs add pic buffers to |
| 519 | * buffer manager in case of non-shared mode |
| 520 | * |
| 521 | * @par Description: |
| 522 | * Function to initialize ps_pic_buf structs add pic buffers to |
| 523 | * buffer manager in case of non-shared mode |
| 524 | * To be called once per stream or for every reset |
| 525 | * |
| 526 | * @param[in] ps_codec |
| 527 | * Pointer to codec context |
| 528 | * |
| 529 | * @returns Error from IHEVCD_ERROR_T |
| 530 | * |
| 531 | * @remarks |
| 532 | * |
| 533 | * |
| 534 | ******************************************************************************* |
| 535 | */ |
| 536 | IHEVCD_ERROR_T ihevcd_pic_buf_mgr_add_bufs(codec_t *ps_codec) |
| 537 | { |
| 538 | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
| 539 | WORD32 i; |
| 540 | WORD32 max_dpb_size; |
| 541 | sps_t *ps_sps; |
| 542 | UWORD8 *pu1_buf; |
| 543 | pic_buf_t *ps_pic_buf; |
| 544 | WORD32 pic_buf_size_allocated; |
| 545 | |
| 546 | WORD32 max_num_bufs; |
| 547 | WORD32 pic_size; |
| 548 | WORD32 level; |
| 549 | |
| 550 | |
| 551 | /* Initialize MV Bank buffer manager */ |
| 552 | ps_sps = ps_codec->s_parse.ps_sps; |
| 553 | |
Ricardo Cerqueira | db8ddbb | 2015-12-15 23:16:20 +0000 | [diff] [blame] | 554 | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
| 555 | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 556 | |
| 557 | |
| 558 | /* Compute the number of MB Bank buffers needed */ |
| 559 | level = ps_codec->i4_init_level; |
| 560 | max_dpb_size = ihevcd_get_dpb_size(level, pic_size); |
| 561 | /* Allocate twice dpb size to handle worst case reorder without returning more |
| 562 | * than one output per call |
| 563 | */ |
| 564 | max_dpb_size *= 2; |
| 565 | /* Allocate one extra picture to handle current frame |
| 566 | * In case of asynchronous parsing and processing, number of buffers should increase here |
| 567 | * based on when parsing and processing threads are synchronized |
| 568 | */ |
| 569 | max_dpb_size++; |
| 570 | |
| 571 | /* If num_ref_frames and num_reorder_frmaes is specified |
| 572 | * Use minimum value |
| 573 | */ |
| 574 | max_num_bufs = MIN(max_dpb_size, (ps_codec->i4_init_num_ref + ps_codec->i4_init_num_reorder + 1)); |
| 575 | |
| 576 | |
| 577 | pu1_buf = (UWORD8 *)ps_codec->ps_pic_buf; |
| 578 | |
| 579 | ps_pic_buf = (pic_buf_t *)ps_codec->ps_pic_buf; |
| 580 | |
| 581 | pu1_buf += BUF_MGR_MAX_CNT * sizeof(pic_buf_t); |
| 582 | |
| 583 | /* In case of non-shared mode, add picture buffers to buffer manager |
| 584 | * In case of shared mode buffers are added in the run-time |
| 585 | */ |
| 586 | if(0 == ps_codec->i4_share_disp_buf) |
| 587 | { |
| 588 | WORD32 buf_ret; |
| 589 | WORD32 luma_samples; |
| 590 | WORD32 chroma_samples; |
| 591 | pic_buf_size_allocated = ps_codec->i4_total_pic_buf_size - |
| 592 | BUF_MGR_MAX_CNT * sizeof(pic_buf_t); |
| 593 | |
| 594 | luma_samples = (ps_codec->i4_strd) * |
| 595 | (ps_sps->i2_pic_height_in_luma_samples + PAD_HT); |
| 596 | |
| 597 | chroma_samples = luma_samples / 2; |
| 598 | |
| 599 | /* Try to add as many buffers as possible since memory is already allocated */ |
| 600 | /* If the number of buffers that can be added is less than max_num_bufs |
| 601 | * return with an error. |
| 602 | */ |
| 603 | for(i = 0; i < (2 * MAX_DPB_SIZE) + 1; i++) |
| 604 | { |
| 605 | pic_buf_size_allocated -= (luma_samples + chroma_samples); |
| 606 | |
| 607 | if(pic_buf_size_allocated < 0) |
| 608 | { |
| 609 | if(i < max_num_bufs) |
| 610 | { |
| 611 | ps_codec->s_parse.i4_error_code = IHEVCD_INSUFFICIENT_MEM_PICBUF; |
| 612 | return IHEVCD_INSUFFICIENT_MEM_PICBUF; |
| 613 | } |
| 614 | break; |
| 615 | } |
| 616 | |
| 617 | ps_pic_buf->pu1_luma = pu1_buf + ps_codec->i4_strd * PAD_TOP + PAD_LEFT; |
| 618 | pu1_buf += luma_samples; |
| 619 | |
| 620 | ps_pic_buf->pu1_chroma = pu1_buf + ps_codec->i4_strd * (PAD_TOP / 2) + PAD_LEFT; |
| 621 | pu1_buf += chroma_samples; |
| 622 | |
| 623 | buf_ret = ihevc_buf_mgr_add((buf_mgr_t *)ps_codec->pv_pic_buf_mgr, ps_pic_buf, i); |
| 624 | |
| 625 | if(0 != buf_ret) |
| 626 | { |
| 627 | ps_codec->s_parse.i4_error_code = IHEVCD_BUF_MGR_ERROR; |
| 628 | return IHEVCD_BUF_MGR_ERROR; |
| 629 | } |
| 630 | ps_pic_buf++; |
| 631 | } |
| 632 | } |
| 633 | |
| 634 | return ret; |
| 635 | } |
| 636 | /** |
| 637 | ******************************************************************************* |
| 638 | * |
| 639 | * @brief |
| 640 | * Function to add buffers to MV Bank buffer manager |
| 641 | * |
| 642 | * @par Description: |
| 643 | * Function to add buffers to MV Bank buffer manager |
| 644 | * To be called once per stream or for every reset |
| 645 | * |
| 646 | * @param[in] ps_codec |
| 647 | * Pointer to codec context |
| 648 | * |
| 649 | * @returns Error from IHEVCD_ERROR_T |
| 650 | * |
| 651 | * @remarks |
| 652 | * |
| 653 | * |
| 654 | ******************************************************************************* |
| 655 | */ |
| 656 | IHEVCD_ERROR_T ihevcd_mv_buf_mgr_add_bufs(codec_t *ps_codec) |
| 657 | { |
| 658 | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
| 659 | WORD32 i; |
| 660 | WORD32 max_dpb_size; |
| 661 | WORD32 mv_bank_size_allocated; |
| 662 | WORD32 pic_mv_bank_size; |
| 663 | WORD32 level; |
| 664 | sps_t *ps_sps; |
| 665 | UWORD8 *pu1_buf; |
| 666 | mv_buf_t *ps_mv_buf; |
| 667 | |
| 668 | |
| 669 | /* Initialize MV Bank buffer manager */ |
| 670 | ps_sps = ps_codec->s_parse.ps_sps; |
| 671 | |
| 672 | |
| 673 | /* Compute the number of MB Bank buffers needed */ |
| 674 | level = ps_codec->i4_init_level; |
| 675 | max_dpb_size = ihevcd_get_dpb_size(level, |
Ricardo Cerqueira | db8ddbb | 2015-12-15 23:16:20 +0000 | [diff] [blame] | 676 | ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
| 677 | ALIGN64(ps_sps->i2_pic_height_in_luma_samples)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 678 | |
| 679 | /* Allocate one extra MV Bank to handle current frame |
| 680 | * In case of asynchronous parsing and processing, number of buffers should increase here |
| 681 | * based on when parsing and processing threads are synchronized |
| 682 | */ |
| 683 | max_dpb_size++; |
| 684 | |
| 685 | pu1_buf = (UWORD8 *)ps_codec->pv_mv_bank_buf_base; |
| 686 | |
| 687 | ps_mv_buf = (mv_buf_t *)pu1_buf; |
| 688 | pu1_buf += BUF_MGR_MAX_CNT * sizeof(mv_buf_t); |
| 689 | ps_codec->ps_mv_buf = ps_mv_buf; |
| 690 | mv_bank_size_allocated = ps_codec->i4_total_mv_bank_size - BUF_MGR_MAX_CNT * sizeof(mv_buf_t); |
| 691 | |
| 692 | /* Compute MV bank size per picture */ |
Harish Mahendrakar | 769d4b3 | 2015-10-27 08:42:30 +0530 | [diff] [blame] | 693 | pic_mv_bank_size = ihevcd_get_pic_mv_bank_size(ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
| 694 | ALIGN64(ps_sps->i2_pic_height_in_luma_samples)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 695 | |
| 696 | for(i = 0; i < max_dpb_size; i++) |
| 697 | { |
| 698 | WORD32 buf_ret; |
| 699 | WORD32 num_pu; |
| 700 | WORD32 num_ctb; |
| 701 | WORD32 pic_size; |
| 702 | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
| 703 | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
| 704 | |
| 705 | |
| 706 | num_pu = pic_size / (MIN_PU_SIZE * MIN_PU_SIZE); |
| 707 | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
| 708 | |
| 709 | |
| 710 | mv_bank_size_allocated -= pic_mv_bank_size; |
| 711 | |
| 712 | if(mv_bank_size_allocated < 0) |
| 713 | { |
| 714 | ps_codec->s_parse.i4_error_code = IHEVCD_INSUFFICIENT_MEM_MVBANK; |
| 715 | return IHEVCD_INSUFFICIENT_MEM_MVBANK; |
| 716 | } |
| 717 | |
| 718 | ps_mv_buf->pu4_pic_pu_idx = (UWORD32 *)pu1_buf; |
| 719 | pu1_buf += (num_ctb + 1) * sizeof(WORD32); |
| 720 | |
| 721 | ps_mv_buf->pu1_pic_pu_map = pu1_buf; |
| 722 | pu1_buf += num_pu; |
| 723 | |
| 724 | ps_mv_buf->pu1_pic_slice_map = (UWORD16 *)pu1_buf; |
Harish Mahendrakar | df5f9ca | 2015-10-09 19:19:06 +0530 | [diff] [blame] | 725 | pu1_buf += ALIGN4(num_ctb * sizeof(UWORD16)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 726 | |
| 727 | ps_mv_buf->ps_pic_pu = (pu_t *)pu1_buf; |
Harish Mahendrakar | 769d4b3 | 2015-10-27 08:42:30 +0530 | [diff] [blame] | 728 | pu1_buf += num_pu * sizeof(pu_t); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 729 | |
| 730 | buf_ret = ihevc_buf_mgr_add((buf_mgr_t *)ps_codec->pv_mv_buf_mgr, ps_mv_buf, i); |
| 731 | |
| 732 | if(0 != buf_ret) |
| 733 | { |
| 734 | ps_codec->s_parse.i4_error_code = IHEVCD_BUF_MGR_ERROR; |
| 735 | return IHEVCD_BUF_MGR_ERROR; |
| 736 | } |
Harish Mahendrakar | 769d4b3 | 2015-10-27 08:42:30 +0530 | [diff] [blame] | 737 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 738 | ps_mv_buf++; |
| 739 | |
| 740 | } |
| 741 | return ret; |
| 742 | } |
| 743 | /** |
| 744 | ******************************************************************************* |
| 745 | * |
| 746 | * @brief |
| 747 | * Picture level initializations required during parsing |
| 748 | * |
| 749 | * @par Description: |
| 750 | * Initialize picture level context variables during parsing Initialize mv |
| 751 | * bank buffer manager in the first init call |
| 752 | * |
| 753 | * @param[in] ps_codec |
| 754 | * Pointer to codec context |
| 755 | * |
| 756 | * @returns Error from IHEVCD_ERROR_T |
| 757 | * |
| 758 | * @remarks |
| 759 | * |
| 760 | * |
| 761 | ******************************************************************************* |
| 762 | */ |
| 763 | IHEVCD_ERROR_T ihevcd_parse_pic_init(codec_t *ps_codec) |
| 764 | { |
| 765 | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
| 766 | mv_buf_t *ps_mv_buf; |
| 767 | sps_t *ps_sps; |
| 768 | WORD32 num_min_cu; |
| 769 | WORD32 cur_pic_buf_id; |
| 770 | WORD32 cur_mv_bank_buf_id; |
| 771 | pic_buf_t *ps_cur_pic; |
| 772 | slice_header_t *ps_slice_hdr; |
| 773 | UWORD8 *pu1_cur_pic_luma, *pu1_cur_pic_chroma; |
| 774 | WORD32 i; |
| 775 | |
| 776 | ps_codec->s_parse.i4_error_code = IHEVCD_SUCCESS; |
| 777 | ps_sps = ps_codec->s_parse.ps_sps; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 778 | ps_slice_hdr = ps_codec->s_parse.ps_slice_hdr; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 779 | /* If parse_pic_init is called, then slice data is present in the input bitstrea stream */ |
| 780 | ps_codec->i4_pic_present = 1; |
| 781 | |
| 782 | /* Memset picture level intra map and transquant bypass map to zero */ |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 783 | num_min_cu = ((ps_sps->i2_pic_height_in_luma_samples + 7) / 8) * ((ps_sps->i2_pic_width_in_luma_samples + 63) / 64); |
| 784 | memset(ps_codec->s_parse.pu1_pic_intra_flag, 0, num_min_cu); |
| 785 | memset(ps_codec->s_parse.pu1_pic_no_loop_filter_flag, 0, num_min_cu); |
| 786 | |
| 787 | |
| 788 | |
| 789 | if(0 == ps_codec->s_parse.i4_first_pic_init) |
| 790 | { |
| 791 | ret = ihevcd_mv_buf_mgr_add_bufs(ps_codec); |
| 792 | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
| 793 | |
| 794 | ret = ihevcd_pic_buf_mgr_add_bufs(ps_codec); |
| 795 | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
| 796 | |
| 797 | ps_codec->s_parse.i4_first_pic_init = 1; |
| 798 | } |
| 799 | |
| 800 | /* Initialize all the slice headers' slice addresses to zero */ |
| 801 | { |
| 802 | WORD32 slice_idx; |
| 803 | WORD32 slice_start_idx; |
| 804 | |
| 805 | slice_start_idx = ps_codec->i4_slice_error ? 2 : 1; |
| 806 | |
| 807 | for(slice_idx = slice_start_idx; slice_idx < MAX_SLICE_HDR_CNT; slice_idx++) |
| 808 | { |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 809 | slice_header_t *ps_slice_hdr_tmp = ps_codec->ps_slice_hdr_base + slice_idx; |
| 810 | ps_slice_hdr_tmp->i2_ctb_x = -1; |
| 811 | ps_slice_hdr_tmp->i2_ctb_y = -1; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 812 | |
| 813 | } |
| 814 | } |
| 815 | |
| 816 | /* Get free MV Bank to hold current picture's motion vector data */ |
| 817 | { |
| 818 | ps_mv_buf = (mv_buf_t *)ihevc_buf_mgr_get_next_free((buf_mgr_t *)ps_codec->pv_mv_buf_mgr, &cur_mv_bank_buf_id); |
| 819 | |
| 820 | /* If there are no free buffers then return with an error code. |
| 821 | * If the buffer is to be freed by another thread , change the |
| 822 | * following to call thread yield and wait for buffer to be freed |
| 823 | */ |
| 824 | if(NULL == ps_mv_buf) |
| 825 | { |
| 826 | ps_codec->s_parse.i4_error_code = IHEVCD_NO_FREE_MVBANK; |
| 827 | ps_codec->i4_error_code = IHEVCD_NO_FREE_MVBANK; |
| 828 | return IHEVCD_NO_FREE_MVBANK; |
| 829 | } |
| 830 | |
| 831 | ps_codec->s_parse.ps_cur_mv_buf = ps_mv_buf; |
| 832 | /* Set current ABS poc to ps_mv_buf, so that while freeing a reference buffer |
| 833 | * corresponding mv buffer can be found by looping through ps_codec->ps_mv_buf array |
| 834 | * and getting a buffer id to free |
| 835 | */ |
| 836 | ps_mv_buf->i4_abs_poc = ps_slice_hdr->i4_abs_pic_order_cnt; |
| 837 | } |
| 838 | |
| 839 | /* Get free picture buffer to hold current picture recon data */ |
| 840 | /* TODO: For asynchronous api the following initializations related to picture |
| 841 | * buffer should be moved to processing side |
| 842 | */ |
| 843 | { |
| 844 | |
| 845 | UWORD8 *pu1_buf; |
| 846 | ps_cur_pic = (pic_buf_t *)ihevc_buf_mgr_get_next_free((buf_mgr_t *)ps_codec->pv_pic_buf_mgr, &cur_pic_buf_id); |
| 847 | |
| 848 | /* If there are no free buffers then return with an error code. |
| 849 | * TODO: If the buffer is to be freed by another thread , change the |
| 850 | * following to call thread yield and wait for buffer to be freed |
| 851 | */ |
| 852 | if(NULL == ps_cur_pic) |
| 853 | { |
| 854 | ps_codec->s_parse.i4_error_code = IHEVCD_NO_FREE_PICBUF; |
| 855 | ps_codec->i4_error_code = IHEVCD_NO_FREE_PICBUF; |
| 856 | return IHEVCD_NO_FREE_PICBUF; |
| 857 | } |
| 858 | |
| 859 | /* Store input timestamp sent with input buffer */ |
| 860 | ps_cur_pic->u4_ts = ps_codec->u4_ts; |
| 861 | ps_cur_pic->i4_abs_poc = ps_slice_hdr->i4_abs_pic_order_cnt; |
| 862 | ps_cur_pic->i4_poc_lsb = ps_slice_hdr->i4_pic_order_cnt_lsb; |
| 863 | pu1_buf = ps_cur_pic->pu1_luma; |
| 864 | pu1_cur_pic_luma = pu1_buf; |
| 865 | |
| 866 | pu1_buf = ps_cur_pic->pu1_chroma; |
| 867 | |
| 868 | pu1_cur_pic_chroma = pu1_buf; |
| 869 | } |
| 870 | |
| 871 | if(0 == ps_codec->u4_pic_cnt) |
| 872 | { |
| 873 | memset(ps_cur_pic->pu1_luma, 128, (ps_sps->i2_pic_width_in_luma_samples + PAD_WD) * ps_sps->i2_pic_height_in_luma_samples); |
| 874 | memset(ps_cur_pic->pu1_chroma, 128, (ps_sps->i2_pic_width_in_luma_samples + PAD_WD) * ps_sps->i2_pic_height_in_luma_samples / 2); |
| 875 | } |
| 876 | |
| 877 | /* Fill the remaining entries of the reference lists with the nearest POC |
| 878 | * This is done to handle cases where there is a corruption in the reference index */ |
| 879 | { |
| 880 | pic_buf_t *ps_pic_buf_ref; |
| 881 | mv_buf_t *ps_mv_buf_ref; |
| 882 | WORD32 r_idx; |
| 883 | dpb_mgr_t *ps_dpb_mgr = (dpb_mgr_t *)ps_codec->pv_dpb_mgr; |
| 884 | buf_mgr_t *ps_mv_buf_mgr = (buf_mgr_t *)ps_codec->pv_mv_buf_mgr; |
| 885 | |
| 886 | ps_pic_buf_ref = ihevc_dpb_mgr_get_ref_by_nearest_poc(ps_dpb_mgr, ps_slice_hdr->i4_abs_pic_order_cnt); |
| 887 | if(NULL == ps_pic_buf_ref) |
| 888 | { |
Harish Mahendrakar | 769d4b3 | 2015-10-27 08:42:30 +0530 | [diff] [blame] | 889 | WORD32 size; |
| 890 | |
| 891 | WORD32 num_pu; |
| 892 | WORD32 num_ctb; |
| 893 | WORD32 pic_size; |
| 894 | /* In case current mv buffer itself is being used as reference mv buffer for colocated |
| 895 | * calculations, then memset all the buffers to zero. |
| 896 | */ |
| 897 | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
| 898 | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
| 899 | |
| 900 | num_pu = pic_size / (MIN_PU_SIZE * MIN_PU_SIZE); |
| 901 | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
| 902 | |
| 903 | memset(ps_mv_buf->l0_collocated_poc, 0, sizeof(ps_mv_buf->l0_collocated_poc)); |
| 904 | memset(ps_mv_buf->u1_l0_collocated_poc_lt, 0, sizeof(ps_mv_buf->u1_l0_collocated_poc_lt)); |
| 905 | memset(ps_mv_buf->l1_collocated_poc, 0, sizeof(ps_mv_buf->l1_collocated_poc)); |
| 906 | memset(ps_mv_buf->u1_l1_collocated_poc_lt, 0, sizeof(ps_mv_buf->u1_l1_collocated_poc_lt)); |
| 907 | |
| 908 | size = (num_ctb + 1) * sizeof(WORD32); |
| 909 | memset(ps_mv_buf->pu4_pic_pu_idx, 0, size); |
| 910 | |
| 911 | size = num_pu; |
| 912 | memset(ps_mv_buf->pu1_pic_pu_map, 0, size); |
| 913 | size = ALIGN4(num_ctb * sizeof(UWORD16)); |
| 914 | memset(ps_mv_buf->pu1_pic_slice_map, 0, size); |
| 915 | size = num_pu * sizeof(pu_t); |
| 916 | memset(ps_mv_buf->ps_pic_pu, 0, size); |
| 917 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 918 | ps_pic_buf_ref = ps_cur_pic; |
| 919 | ps_mv_buf_ref = ps_mv_buf; |
| 920 | } |
| 921 | else |
| 922 | { |
| 923 | ps_mv_buf_ref = ihevcd_mv_mgr_get_poc(ps_mv_buf_mgr, ps_pic_buf_ref->i4_abs_poc); |
| 924 | } |
| 925 | |
| 926 | for(r_idx = 0; r_idx < ps_slice_hdr->i1_num_ref_idx_l0_active; r_idx++) |
| 927 | { |
| 928 | if(NULL == ps_slice_hdr->as_ref_pic_list0[r_idx].pv_pic_buf) |
| 929 | { |
| 930 | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
| 931 | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
| 932 | } |
| 933 | } |
| 934 | |
| 935 | for(r_idx = ps_slice_hdr->i1_num_ref_idx_l0_active; r_idx < MAX_DPB_SIZE; r_idx++) |
| 936 | { |
| 937 | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
| 938 | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
| 939 | } |
| 940 | |
| 941 | for(r_idx = 0; r_idx < ps_slice_hdr->i1_num_ref_idx_l1_active; r_idx++) |
| 942 | { |
| 943 | if(NULL == ps_slice_hdr->as_ref_pic_list1[r_idx].pv_pic_buf) |
| 944 | { |
| 945 | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
| 946 | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
| 947 | } |
| 948 | } |
| 949 | |
| 950 | for(r_idx = ps_slice_hdr->i1_num_ref_idx_l1_active; r_idx < MAX_DPB_SIZE; r_idx++) |
| 951 | { |
| 952 | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
| 953 | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
| 954 | } |
| 955 | } |
| 956 | |
| 957 | |
| 958 | /* Reset the jobq to start of the jobq buffer */ |
| 959 | ihevcd_jobq_reset((jobq_t *)ps_codec->pv_proc_jobq); |
| 960 | |
| 961 | ps_codec->s_parse.i4_pic_pu_idx = 0; |
| 962 | ps_codec->s_parse.i4_pic_tu_idx = 0; |
| 963 | |
| 964 | ps_codec->s_parse.pu1_pic_pu_map = ps_mv_buf->pu1_pic_pu_map; |
| 965 | ps_codec->s_parse.ps_pic_pu = ps_mv_buf->ps_pic_pu; |
| 966 | ps_codec->s_parse.pu4_pic_pu_idx = ps_mv_buf->pu4_pic_pu_idx; |
| 967 | ps_codec->s_parse.pu1_slice_idx = (UWORD16 *)ps_mv_buf->pu1_pic_slice_map; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 968 | for(i = 0; i < MAX_PROCESS_THREADS; i++) |
| 969 | { |
| 970 | ps_codec->as_process[i].pu1_slice_idx = (UWORD16 *)ps_mv_buf->pu1_pic_slice_map; |
| 971 | } |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 972 | ps_codec->s_parse.pu1_pu_map = ps_codec->s_parse.pu1_pic_pu_map; |
| 973 | ps_codec->s_parse.ps_pu = ps_codec->s_parse.ps_pic_pu; |
| 974 | |
| 975 | { |
| 976 | UWORD8 *pu1_buf; |
| 977 | WORD32 ctb_luma_min_tu_cnt, ctb_chroma_min_tu_cnt, ctb_min_tu_cnt; |
| 978 | WORD32 pic_size; |
| 979 | WORD32 num_ctb; |
| 980 | |
Harish Mahendrakar | 769d4b3 | 2015-10-27 08:42:30 +0530 | [diff] [blame] | 981 | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
| 982 | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 983 | |
| 984 | ctb_luma_min_tu_cnt = pic_size / (MIN_TU_SIZE * MIN_TU_SIZE); |
| 985 | |
| 986 | ctb_chroma_min_tu_cnt = ctb_luma_min_tu_cnt >> 1; |
| 987 | |
| 988 | ctb_min_tu_cnt = ctb_luma_min_tu_cnt + ctb_chroma_min_tu_cnt; |
| 989 | |
| 990 | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 991 | pu1_buf = (UWORD8 *)ps_codec->pv_tu_data; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 992 | ps_codec->s_parse.pu4_pic_tu_idx = (UWORD32 *)pu1_buf; |
| 993 | pu1_buf += (num_ctb + 1) * sizeof(WORD32); |
| 994 | |
| 995 | ps_codec->s_parse.pu1_pic_tu_map = pu1_buf; |
| 996 | pu1_buf += ctb_min_tu_cnt; |
| 997 | |
| 998 | ps_codec->s_parse.ps_pic_tu = (tu_t *)pu1_buf; |
| 999 | pu1_buf += ctb_min_tu_cnt * sizeof(tu_t); |
| 1000 | |
| 1001 | ps_codec->s_parse.pv_pic_tu_coeff_data = pu1_buf; |
| 1002 | |
| 1003 | ps_codec->s_parse.pu1_tu_map = ps_codec->s_parse.pu1_pic_tu_map; |
| 1004 | ps_codec->s_parse.ps_tu = ps_codec->s_parse.ps_pic_tu; |
| 1005 | ps_codec->s_parse.pv_tu_coeff_data = ps_codec->s_parse.pv_pic_tu_coeff_data; |
| 1006 | } |
| 1007 | |
| 1008 | ps_codec->s_parse.s_bs_ctxt.ps_pic_pu = ps_codec->s_parse.ps_pic_pu; |
| 1009 | ps_codec->s_parse.s_bs_ctxt.pu4_pic_pu_idx = ps_codec->s_parse.pu4_pic_pu_idx; |
| 1010 | ps_codec->s_parse.s_bs_ctxt.pu4_pic_tu_idx = ps_codec->s_parse.pu4_pic_tu_idx; |
| 1011 | |
| 1012 | |
| 1013 | /* Set number of CTBs to be processed simultaneously */ |
| 1014 | ps_codec->i4_proc_nctb = ihevcd_nctb_cnt(ps_codec, ps_sps); |
| 1015 | |
| 1016 | /* Memset Parse Map and process map at the start of frame */ |
| 1017 | //TODO: In case of asynchronous API proc_map can not be set to zero here |
| 1018 | { |
| 1019 | WORD32 num_ctb; |
| 1020 | |
| 1021 | num_ctb = ps_sps->i4_pic_size_in_ctb; |
| 1022 | |
| 1023 | memset(ps_codec->pu1_parse_map, 0, num_ctb); |
| 1024 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1025 | memset(ps_codec->pu1_proc_map, 0, num_ctb); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1026 | } |
| 1027 | |
| 1028 | |
| 1029 | |
| 1030 | /* Initialize disp buf id to -1, this will be updated at the end of frame if there is |
| 1031 | * buffer to be displayed |
| 1032 | */ |
| 1033 | ps_codec->i4_disp_buf_id = -1; |
| 1034 | ps_codec->ps_disp_buf = NULL; |
| 1035 | |
| 1036 | ps_codec->i4_disable_deblk_pic = 0; |
| 1037 | ps_codec->i4_disable_sao_pic = 0; |
| 1038 | ps_codec->i4_fullpel_inter_pred = 0; |
| 1039 | ps_codec->i4_mv_frac_mask = 0x7FFFFFFF; |
| 1040 | |
| 1041 | /* If degrade is enabled, set the degrade flags appropriately */ |
| 1042 | if(ps_codec->i4_degrade_type && ps_codec->i4_degrade_pics) |
| 1043 | { |
| 1044 | WORD32 degrade_pic; |
| 1045 | ps_codec->i4_degrade_pic_cnt++; |
| 1046 | degrade_pic = 0; |
| 1047 | |
| 1048 | /* If degrade is to be done in all frames, then do not check further */ |
| 1049 | switch(ps_codec->i4_degrade_pics) |
| 1050 | { |
| 1051 | case 4: |
| 1052 | { |
| 1053 | degrade_pic = 1; |
| 1054 | break; |
| 1055 | } |
| 1056 | case 3: |
| 1057 | { |
| 1058 | if(ps_slice_hdr->i1_slice_type != ISLICE) |
| 1059 | degrade_pic = 1; |
| 1060 | |
| 1061 | break; |
| 1062 | } |
| 1063 | case 2: |
| 1064 | { |
| 1065 | |
| 1066 | /* If pic count hits non-degrade interval or it is an islice, then do not degrade */ |
| 1067 | if((ps_slice_hdr->i1_slice_type != ISLICE) && |
| 1068 | (ps_codec->i4_degrade_pic_cnt != ps_codec->i4_nondegrade_interval)) |
| 1069 | degrade_pic = 1; |
| 1070 | |
| 1071 | break; |
| 1072 | } |
| 1073 | case 1: |
| 1074 | { |
| 1075 | /* Check if the current picture is non-ref */ |
| 1076 | if((ps_slice_hdr->i1_nal_unit_type < NAL_BLA_W_LP) && |
| 1077 | (ps_slice_hdr->i1_nal_unit_type % 2 == 0)) |
| 1078 | { |
| 1079 | degrade_pic = 1; |
| 1080 | } |
| 1081 | break; |
| 1082 | } |
| 1083 | |
| 1084 | |
| 1085 | } |
| 1086 | if(degrade_pic) |
| 1087 | { |
| 1088 | if(ps_codec->i4_degrade_type & 0x1) |
| 1089 | ps_codec->i4_disable_sao_pic = 1; |
| 1090 | |
| 1091 | if(ps_codec->i4_degrade_type & 0x2) |
| 1092 | ps_codec->i4_disable_deblk_pic = 1; |
| 1093 | |
| 1094 | /* MC degrading is done only for non-ref pictures */ |
| 1095 | if((ps_slice_hdr->i1_nal_unit_type < NAL_BLA_W_LP) && |
| 1096 | (ps_slice_hdr->i1_nal_unit_type % 2 == 0)) |
| 1097 | { |
| 1098 | if(ps_codec->i4_degrade_type & 0x4) |
| 1099 | ps_codec->i4_mv_frac_mask = 0; |
| 1100 | |
| 1101 | if(ps_codec->i4_degrade_type & 0x8) |
| 1102 | ps_codec->i4_mv_frac_mask = 0; |
| 1103 | } |
| 1104 | } |
| 1105 | else |
| 1106 | ps_codec->i4_degrade_pic_cnt = 0; |
| 1107 | } |
| 1108 | |
| 1109 | |
| 1110 | { |
| 1111 | WORD32 i; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1112 | for(i = 0; i < MAX_PROCESS_THREADS; i++) |
| 1113 | { |
| 1114 | ps_codec->as_process[i].pu4_pic_pu_idx = ps_codec->s_parse.pu4_pic_pu_idx; |
| 1115 | ps_codec->as_process[i].ps_pic_pu = ps_codec->s_parse.ps_pic_pu; |
| 1116 | ps_codec->as_process[i].pu1_pic_pu_map = ps_codec->s_parse.pu1_pic_pu_map; |
| 1117 | ps_codec->as_process[i].pu4_pic_tu_idx = ps_codec->s_parse.pu4_pic_tu_idx; |
| 1118 | ps_codec->as_process[i].ps_pic_tu = ps_codec->s_parse.ps_pic_tu; |
| 1119 | ps_codec->as_process[i].pu1_pic_tu_map = ps_codec->s_parse.pu1_pic_tu_map; |
| 1120 | ps_codec->as_process[i].pv_pic_tu_coeff_data = ps_codec->s_parse.pv_pic_tu_coeff_data; |
| 1121 | ps_codec->as_process[i].i4_cur_mv_bank_buf_id = cur_mv_bank_buf_id; |
| 1122 | ps_codec->as_process[i].s_sao_ctxt.pu1_slice_idx = ps_codec->as_process[i].pu1_slice_idx; |
| 1123 | ps_codec->as_process[i].s_sao_ctxt.pu1_tile_idx = ps_codec->as_process[i].pu1_tile_idx; |
| 1124 | |
| 1125 | /* TODO: For asynchronous api the following initializations related to picture |
| 1126 | * buffer should be moved to processing side |
| 1127 | */ |
| 1128 | ps_codec->as_process[i].pu1_cur_pic_luma = pu1_cur_pic_luma; |
| 1129 | ps_codec->as_process[i].pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
| 1130 | ps_codec->as_process[i].ps_cur_pic = ps_cur_pic; |
| 1131 | ps_codec->as_process[i].i4_cur_pic_buf_id = cur_pic_buf_id; |
| 1132 | |
| 1133 | ps_codec->as_process[i].ps_out_buffer = ps_codec->ps_out_buffer; |
| 1134 | if(1 < ps_codec->i4_num_cores) |
| 1135 | { |
| 1136 | ps_codec->as_process[i].i4_check_parse_status = 1; |
| 1137 | ps_codec->as_process[i].i4_check_proc_status = 1; |
| 1138 | } |
| 1139 | else |
| 1140 | { |
| 1141 | ps_codec->as_process[i].i4_check_parse_status = 0; |
| 1142 | ps_codec->as_process[i].i4_check_proc_status = 0; |
| 1143 | } |
| 1144 | ps_codec->as_process[i].pu1_pic_intra_flag = ps_codec->s_parse.pu1_pic_intra_flag; |
| 1145 | ps_codec->as_process[i].pu1_pic_no_loop_filter_flag = ps_codec->s_parse.pu1_pic_no_loop_filter_flag; |
| 1146 | ps_codec->as_process[i].i4_init_done = 0; |
| 1147 | |
| 1148 | ps_codec->as_process[i].s_bs_ctxt.pu4_pic_tu_idx = ps_codec->as_process[i].pu4_pic_tu_idx; |
| 1149 | ps_codec->as_process[i].s_bs_ctxt.pu4_pic_pu_idx = ps_codec->as_process[i].pu4_pic_pu_idx; |
| 1150 | ps_codec->as_process[i].s_bs_ctxt.ps_pic_pu = ps_codec->as_process[i].ps_pic_pu; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1151 | ps_codec->as_process[i].s_deblk_ctxt.pu1_pic_no_loop_filter_flag = ps_codec->s_parse.pu1_pic_no_loop_filter_flag; |
| 1152 | ps_codec->as_process[i].s_deblk_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
| 1153 | ps_codec->as_process[i].s_deblk_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1154 | ps_codec->as_process[i].s_sao_ctxt.pu1_pic_no_loop_filter_flag = ps_codec->s_parse.pu1_pic_no_loop_filter_flag; |
| 1155 | ps_codec->as_process[i].s_sao_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
| 1156 | ps_codec->as_process[i].s_sao_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1157 | if(i < (ps_codec->i4_num_cores - 1)) |
| 1158 | { |
| 1159 | ithread_create(ps_codec->apv_process_thread_handle[i], NULL, |
| 1160 | (void *)ihevcd_process_thread, |
| 1161 | (void *)&ps_codec->as_process[i]); |
| 1162 | ps_codec->ai4_process_thread_created[i] = 1; |
| 1163 | } |
| 1164 | else |
| 1165 | { |
| 1166 | ps_codec->ai4_process_thread_created[i] = 0; |
| 1167 | } |
| 1168 | |
| 1169 | } |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1170 | ps_codec->s_parse.s_deblk_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
| 1171 | ps_codec->s_parse.s_deblk_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
| 1172 | |
| 1173 | ps_codec->s_parse.s_sao_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
| 1174 | ps_codec->s_parse.s_sao_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
| 1175 | } |
| 1176 | /* Since any input bitstream buffer that contains slice data will be sent to output(even in |
| 1177 | * case of error, this buffer is added to display queue and next buffer in the display queue |
| 1178 | * will be returned as the display buffer. |
| 1179 | * Note: If format conversion (or frame copy) is used and is scheduled |
| 1180 | * in a different thread then it has to check if the processing for the current row is complete before |
| 1181 | * it copies/converts a given row. In case of low delay or in case of B pictures, current frame being decoded has to be |
| 1182 | * returned, which requires a status check to ensure that the current row is reconstructed before copying. |
| 1183 | */ |
| 1184 | /* Add current picture to display manager */ |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1185 | { |
| 1186 | WORD32 abs_poc; |
| 1187 | slice_header_t *ps_slice_hdr; |
| 1188 | ps_slice_hdr = ps_codec->s_parse.ps_slice_hdr; |
| 1189 | abs_poc = ps_slice_hdr->i4_abs_pic_order_cnt; |
| 1190 | ihevc_disp_mgr_add((disp_mgr_t *)ps_codec->pv_disp_buf_mgr, |
| 1191 | ps_codec->as_process[0].i4_cur_pic_buf_id, |
| 1192 | abs_poc, |
| 1193 | ps_codec->as_process[0].ps_cur_pic); |
| 1194 | } |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1195 | ps_codec->ps_disp_buf = NULL; |
| 1196 | /* Get picture to be displayed if number of pictures decoded is more than max allowed reorder */ |
| 1197 | /* Since the current will be decoded, check is fore >= instead of > */ |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1198 | if(((WORD32)(ps_codec->u4_pic_cnt - ps_codec->u4_disp_cnt) >= ps_sps->ai1_sps_max_num_reorder_pics[ps_sps->i1_sps_max_sub_layers - 1]) || |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1199 | ((WORD32)(ps_codec->u4_pic_cnt - ps_codec->u4_disp_cnt) >= ps_codec->i4_init_num_reorder)) |
| 1200 | |
| 1201 | { |
| 1202 | ps_codec->ps_disp_buf = (pic_buf_t *)ihevc_disp_mgr_get((disp_mgr_t *)ps_codec->pv_disp_buf_mgr, &ps_codec->i4_disp_buf_id); |
| 1203 | ps_codec->u4_disp_cnt++; |
| 1204 | } |
| 1205 | |
| 1206 | ps_codec->s_fmt_conv.i4_cur_row = 0; |
| 1207 | /* Set number of rows to be processed at a time */ |
| 1208 | ps_codec->s_fmt_conv.i4_num_rows = 4; |
| 1209 | |
| 1210 | if(ps_codec->u4_enable_fmt_conv_ahead && (ps_codec->i4_num_cores > 1)) |
| 1211 | { |
| 1212 | process_ctxt_t *ps_proc; |
| 1213 | |
| 1214 | /* i4_num_cores - 1 contexts are currently being used by other threads */ |
| 1215 | ps_proc = &ps_codec->as_process[ps_codec->i4_num_cores - 1]; |
| 1216 | |
| 1217 | /* If the frame being decoded and displayed are different, schedule format conversion jobs |
| 1218 | * this will keep the proc threads busy and lets parse thread decode few CTBs ahead |
| 1219 | * If the frame being decoded and displayed are same, then format conversion is scheduled later. |
| 1220 | */ |
| 1221 | if((ps_codec->ps_disp_buf) && (ps_codec->i4_disp_buf_id != ps_proc->i4_cur_pic_buf_id) && |
| 1222 | ((0 == ps_codec->i4_share_disp_buf) || (IV_YUV_420P == ps_codec->e_chroma_fmt))) |
| 1223 | { |
| 1224 | |
| 1225 | for(i = 0; i < ps_sps->i2_pic_ht_in_ctb; i++) |
| 1226 | { |
| 1227 | proc_job_t s_job; |
| 1228 | IHEVCD_ERROR_T ret; |
| 1229 | s_job.i4_cmd = CMD_FMTCONV; |
| 1230 | s_job.i2_ctb_cnt = 0; |
| 1231 | s_job.i2_ctb_x = 0; |
| 1232 | s_job.i2_ctb_y = i; |
| 1233 | s_job.i2_slice_idx = 0; |
| 1234 | s_job.i4_tu_coeff_data_ofst = 0; |
| 1235 | ret = ihevcd_jobq_queue((jobq_t *)ps_codec->s_parse.pv_proc_jobq, |
| 1236 | &s_job, sizeof(proc_job_t), 1); |
| 1237 | if(ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS) |
| 1238 | return ret; |
| 1239 | } |
| 1240 | } |
| 1241 | } |
| 1242 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1243 | |
| 1244 | return ret; |
| 1245 | } |
| 1246 | |
| 1247 | |