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_process_slice.c |
| 22 | * |
| 23 | * @brief |
| 24 | * Contains functions for processing slice data |
| 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_structs.h" |
| 55 | #include "ihevc_macros.h" |
| 56 | #include "ihevc_platform_macros.h" |
| 57 | #include "ihevc_cabac_tables.h" |
| 58 | #include "ihevc_padding.h" |
| 59 | #include "ihevc_iquant_itrans_recon.h" |
| 60 | #include "ihevc_chroma_iquant_itrans_recon.h" |
| 61 | #include "ihevc_recon.h" |
| 62 | #include "ihevc_chroma_recon.h" |
| 63 | #include "ihevc_iquant_recon.h" |
| 64 | #include "ihevc_chroma_iquant_recon.h" |
| 65 | #include "ihevc_intra_pred.h" |
| 66 | |
| 67 | #include "ihevc_error.h" |
| 68 | #include "ihevc_common_tables.h" |
| 69 | #include "ihevc_quant_tables.h" |
| 70 | #include "ihevcd_common_tables.h" |
| 71 | |
| 72 | #include "ihevcd_profile.h" |
| 73 | #include "ihevcd_trace.h" |
| 74 | #include "ihevcd_defs.h" |
| 75 | #include "ihevcd_function_selector.h" |
| 76 | #include "ihevcd_structs.h" |
| 77 | #include "ihevcd_error.h" |
| 78 | #include "ihevcd_nal.h" |
| 79 | #include "ihevcd_bitstream.h" |
| 80 | #include "ihevcd_job_queue.h" |
| 81 | #include "ihevcd_utils.h" |
| 82 | #include "ihevcd_debug.h" |
| 83 | #include "ihevcd_get_mv.h" |
| 84 | #include "ihevcd_inter_pred.h" |
| 85 | #include "ihevcd_iquant_itrans_recon_ctb.h" |
| 86 | #include "ihevcd_boundary_strength.h" |
| 87 | #include "ihevcd_deblk.h" |
| 88 | #include "ihevcd_fmt_conv.h" |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 89 | #include "ihevcd_sao.h" |
| 90 | #include "ihevcd_profile.h" |
| 91 | |
| 92 | IHEVCD_ERROR_T ihevcd_fmt_conv(codec_t *ps_codec, |
| 93 | process_ctxt_t *ps_proc, |
| 94 | UWORD8 *pu1_y_dst, |
| 95 | UWORD8 *pu1_u_dst, |
| 96 | UWORD8 *pu1_v_dst, |
| 97 | WORD32 cur_row, |
| 98 | WORD32 num_rows); |
| 99 | |
| 100 | typedef enum |
| 101 | { |
| 102 | PROC_ALL, |
| 103 | PROC_INTER_PRED, |
| 104 | PROC_RECON, |
| 105 | PROC_DEBLK, |
| 106 | PROC_SAO |
| 107 | }proc_type_t; |
| 108 | |
| 109 | void ihevcd_proc_map_check(process_ctxt_t *ps_proc, proc_type_t proc_type, WORD32 nctb) |
| 110 | { |
| 111 | tile_t *ps_tile = ps_proc->ps_tile; |
| 112 | sps_t *ps_sps = ps_proc->ps_sps; |
| 113 | pps_t *ps_pps = ps_proc->ps_pps; |
| 114 | codec_t *ps_codec = ps_proc->ps_codec; |
| 115 | WORD32 idx; |
| 116 | WORD32 nop_cnt; |
| 117 | WORD32 bit_pos = proc_type; |
| 118 | WORD32 bit_mask = (1 << bit_pos); |
| 119 | |
| 120 | if(ps_proc->i4_check_proc_status) |
| 121 | { |
| 122 | nop_cnt = PROC_NOP_CNT; |
| 123 | while(1) |
| 124 | { |
| 125 | volatile UWORD8 *pu1_buf; |
| 126 | volatile WORD32 status; |
| 127 | status = 1; |
| 128 | /* Check if all dependencies for the next nCTBs are met */ |
| 129 | { |
| 130 | WORD32 x_pos; |
| 131 | |
| 132 | { |
| 133 | /* Check if the top right of next nCTBs are processed */ |
| 134 | if(ps_proc->i4_ctb_y > 0) |
| 135 | { |
| 136 | x_pos = (ps_proc->i4_ctb_tile_x + nctb); |
| 137 | idx = MIN(x_pos, (ps_tile->u2_wd - 1)); |
| 138 | |
| 139 | /* Check if top-right CTB for the last CTB in nCTB is within the tile */ |
| 140 | { |
| 141 | idx += ps_tile->u1_pos_x; |
| 142 | idx += ((ps_proc->i4_ctb_y - 1) |
| 143 | * ps_sps->i2_pic_wd_in_ctb); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 144 | pu1_buf = (ps_codec->pu1_proc_map + idx); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 145 | status = *pu1_buf & bit_mask; |
| 146 | } |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | /* If tiles are enabled, then test left and top-left as well */ |
| 151 | ps_pps = ps_proc->ps_pps; |
| 152 | if(ps_pps->i1_tiles_enabled_flag) |
| 153 | { |
| 154 | /*Check if left ctb is processed*/ |
| 155 | if((ps_proc->i4_ctb_x > 0) && ((0 != status))) |
| 156 | { |
| 157 | x_pos = ps_tile->u1_pos_x + ps_proc->i4_ctb_tile_x - 1; |
| 158 | idx = x_pos + (ps_proc->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 159 | pu1_buf = (ps_codec->pu1_proc_map + idx); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 160 | status = *pu1_buf & bit_mask; |
| 161 | } |
| 162 | |
| 163 | /*Check if top left ctb is processed*/ |
| 164 | if((ps_proc->i4_ctb_x > 0) && (0 != status) && (ps_proc->i4_ctb_y > 0)) |
| 165 | { |
| 166 | x_pos = ps_tile->u1_pos_x + ps_proc->i4_ctb_tile_x - 1; |
| 167 | idx = x_pos + ((ps_proc->i4_ctb_y - 1) * ps_sps->i2_pic_wd_in_ctb); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 168 | pu1_buf = (ps_codec->pu1_proc_map + idx); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 169 | status = *pu1_buf & bit_mask; |
| 170 | } |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | if(status) |
| 175 | break; |
| 176 | |
| 177 | /* if dependencies are not met, then wait for few cycles. |
| 178 | * Even after few iterations, if the dependencies are not met then yield |
| 179 | */ |
| 180 | if(nop_cnt > 0) |
| 181 | { |
| 182 | NOP(128); |
| 183 | nop_cnt -= 128; |
| 184 | } |
| 185 | else |
| 186 | { |
| 187 | nop_cnt = PROC_NOP_CNT; |
| 188 | ithread_yield(); |
| 189 | //NOP(128 * 16); |
| 190 | } |
| 191 | } |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | void ihevcd_proc_map_update(process_ctxt_t *ps_proc, proc_type_t proc_type, WORD32 nctb) |
| 196 | { |
| 197 | codec_t *ps_codec = ps_proc->ps_codec; |
| 198 | WORD32 i, idx; |
| 199 | WORD32 bit_pos = proc_type; |
| 200 | WORD32 bit_mask = (1 << bit_pos); |
| 201 | |
| 202 | /* Update the current CTBs processing status */ |
| 203 | if(ps_proc->i4_check_proc_status) |
| 204 | { |
| 205 | for(i = 0; i < nctb; i++) |
| 206 | { |
| 207 | sps_t *ps_sps = ps_proc->ps_sps; |
| 208 | UWORD8 *pu1_buf; |
| 209 | idx = (ps_proc->i4_ctb_x + i); |
| 210 | idx += ((ps_proc->i4_ctb_y) * ps_sps->i2_pic_wd_in_ctb); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 211 | pu1_buf = (ps_codec->pu1_proc_map + idx); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 212 | *pu1_buf = *pu1_buf | bit_mask; |
| 213 | } |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | |
| 218 | void ihevcd_slice_hdr_update(process_ctxt_t *ps_proc) |
| 219 | { |
| 220 | |
| 221 | /* Slice x and y are initialized in proc_init. But initialize slice x and y count here |
| 222 | * if a new slice begins at the middle of a row since proc_init is invoked only at the beginning of each row */ |
| 223 | if(!((ps_proc->i4_ctb_x == 0) && (ps_proc->i4_ctb_y == 0))) |
| 224 | { |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 225 | slice_header_t *ps_slice_hdr_next = ps_proc->ps_codec->ps_slice_hdr_base + ((ps_proc->i4_cur_slice_idx + 1) & (MAX_SLICE_HDR_CNT - 1)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 226 | |
| 227 | if((ps_slice_hdr_next->i2_ctb_x == ps_proc->i4_ctb_x) |
| 228 | && (ps_slice_hdr_next->i2_ctb_y == ps_proc->i4_ctb_y)) |
| 229 | { |
| 230 | if(0 == ps_slice_hdr_next->i1_dependent_slice_flag) |
| 231 | { |
| 232 | ps_proc->i4_ctb_slice_x = 0; |
| 233 | ps_proc->i4_ctb_slice_y = 0; |
| 234 | } |
| 235 | |
| 236 | ps_proc->i4_cur_slice_idx++; |
| 237 | ps_proc->ps_slice_hdr = ps_slice_hdr_next; |
| 238 | } |
| 239 | |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | void ihevcd_ctb_pos_update(process_ctxt_t *ps_proc, WORD32 nctb) |
| 244 | { |
| 245 | WORD32 tile_start_ctb_idx, slice_start_ctb_idx; |
| 246 | slice_header_t *ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 247 | tile_t *ps_tile = ps_proc->ps_tile; |
| 248 | sps_t *ps_sps = ps_proc->ps_sps; |
| 249 | |
| 250 | /* Update x and y positions */ |
| 251 | ps_proc->i4_ctb_tile_x += nctb; |
| 252 | ps_proc->i4_ctb_x += nctb; |
| 253 | |
| 254 | ps_proc->i4_ctb_slice_x += nctb; |
| 255 | /*If tile are enabled, then handle the tile & slice counters differently*/ |
| 256 | if(ps_proc->ps_pps->i1_tiles_enabled_flag) |
| 257 | { |
| 258 | /* Update slice counters*/ |
| 259 | slice_start_ctb_idx = ps_slice_hdr->i2_ctb_x + (ps_slice_hdr->i2_ctb_y * ps_sps->i2_pic_wd_in_ctb); |
| 260 | tile_start_ctb_idx = ps_tile->u1_pos_x + (ps_tile->u1_pos_y * ps_sps->i2_pic_wd_in_ctb); |
| 261 | /* |
| 262 | * There can be 2 cases where slice counters must be handled differently. |
| 263 | * 1 - Multiple tiles span across a single/one of the many slice. |
| 264 | * 2 - Multiple slices span across a single/one of the many tiles. |
| 265 | */ |
| 266 | |
| 267 | /*Case 1 */ |
| 268 | if(slice_start_ctb_idx < tile_start_ctb_idx) |
| 269 | { |
| 270 | /*End of tile row*/ |
| 271 | if(ps_proc->i4_ctb_x > ps_slice_hdr->i2_ctb_x) |
| 272 | { |
| 273 | if(ps_proc->i4_ctb_slice_x >= (ps_tile->u2_wd + ps_tile->u1_pos_x)) |
| 274 | { |
| 275 | ps_proc->i4_ctb_slice_y++; |
| 276 | ps_proc->i4_ctb_slice_x = ps_proc->i4_ctb_slice_x |
| 277 | - ps_tile->u2_wd; |
| 278 | } |
| 279 | } |
| 280 | else |
| 281 | { |
| 282 | WORD32 temp_stride = (ps_sps->i2_pic_wd_in_ctb - ps_slice_hdr->i2_ctb_x); |
| 283 | if(ps_proc->i4_ctb_slice_x >= (temp_stride + ps_tile->u2_wd + ps_tile->u1_pos_x)) |
| 284 | { |
| 285 | ps_proc->i4_ctb_slice_y++; |
| 286 | ps_proc->i4_ctb_slice_x = ps_proc->i4_ctb_slice_x |
| 287 | - ps_tile->u2_wd; |
| 288 | } |
| 289 | } |
| 290 | } |
| 291 | /*Case 2*/ |
| 292 | else if(ps_proc->i4_ctb_slice_x >= (ps_tile->u2_wd)) |
| 293 | { |
| 294 | /*End of tile row*/ |
| 295 | ps_proc->i4_ctb_slice_y++; |
| 296 | ps_proc->i4_ctb_slice_x = 0; |
| 297 | } |
| 298 | } |
| 299 | else |
| 300 | { |
| 301 | if(ps_proc->i4_ctb_slice_x >= ps_tile->u2_wd) |
| 302 | { |
| 303 | ps_proc->i4_ctb_slice_y++; |
| 304 | ps_proc->i4_ctb_slice_x = ps_proc->i4_ctb_slice_x |
| 305 | - ps_tile->u2_wd; |
| 306 | } |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | void ihevcd_ctb_avail_update(process_ctxt_t *ps_proc) |
| 311 | { |
| 312 | slice_header_t *ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 313 | sps_t *ps_sps = ps_proc->ps_sps; |
| 314 | tile_t *ps_tile_prev; |
| 315 | tile_t *ps_tile = ps_proc->ps_tile; |
| 316 | WORD32 cur_pu_idx; |
| 317 | WORD32 tile_start_ctb_idx, slice_start_ctb_idx; |
| 318 | WORD16 i2_wd_in_ctb; |
| 319 | WORD32 continuous_tiles = 0; |
| 320 | WORD32 cur_ctb_idx; |
| 321 | WORD32 check_tile_wd; |
| 322 | |
| 323 | if((0 != ps_tile->u1_pos_x) && (0 != ps_tile->u1_pos_y)) |
| 324 | { |
| 325 | ps_tile_prev = ps_tile - 1; |
| 326 | } |
| 327 | else |
| 328 | { |
| 329 | ps_tile_prev = ps_tile; |
| 330 | } |
| 331 | |
| 332 | |
| 333 | check_tile_wd = ps_slice_hdr->i2_ctb_x + ps_tile_prev->u2_wd; |
| 334 | if(!(((check_tile_wd >= ps_sps->i2_pic_wd_in_ctb) && (check_tile_wd % ps_sps->i2_pic_wd_in_ctb == ps_tile->u1_pos_x)) |
| 335 | || ((ps_slice_hdr->i2_ctb_x == ps_tile->u1_pos_x)))) |
| 336 | { |
| 337 | continuous_tiles = 1; |
| 338 | } |
| 339 | |
| 340 | slice_start_ctb_idx = ps_slice_hdr->i2_ctb_x + (ps_slice_hdr->i2_ctb_y * ps_sps->i2_pic_wd_in_ctb); |
| 341 | tile_start_ctb_idx = ps_tile->u1_pos_x + (ps_tile->u1_pos_y * ps_sps->i2_pic_wd_in_ctb); |
| 342 | |
| 343 | if((slice_start_ctb_idx < tile_start_ctb_idx) && (continuous_tiles)) |
| 344 | { |
| 345 | //Slices span across multiple tiles. |
| 346 | i2_wd_in_ctb = ps_sps->i2_pic_wd_in_ctb; |
| 347 | } |
| 348 | else |
| 349 | { |
| 350 | i2_wd_in_ctb = ps_tile->u2_wd; |
| 351 | } |
| 352 | cur_ctb_idx = ps_proc->i4_ctb_x |
| 353 | + ps_proc->i4_ctb_y * (ps_sps->i2_pic_wd_in_ctb); |
| 354 | |
| 355 | /* Ctb level availability */ |
| 356 | /* Bottom left will not be available at a CTB level, no need to pass this */ |
| 357 | ps_proc->u1_top_ctb_avail = 1; |
| 358 | ps_proc->u1_left_ctb_avail = 1; |
| 359 | ps_proc->u1_top_lt_ctb_avail = 1; |
| 360 | ps_proc->u1_top_rt_ctb_avail = 1; |
| 361 | /* slice and tile boundaries */ |
| 362 | |
| 363 | if((0 == ps_proc->i4_ctb_y) || (0 == ps_proc->i4_ctb_tile_y)) |
| 364 | { |
| 365 | ps_proc->u1_top_ctb_avail = 0; |
| 366 | ps_proc->u1_top_lt_ctb_avail = 0; |
| 367 | ps_proc->u1_top_rt_ctb_avail = 0; |
| 368 | } |
| 369 | |
| 370 | if((0 == ps_proc->i4_ctb_x) || (0 == ps_proc->i4_ctb_tile_x)) |
| 371 | { |
| 372 | ps_proc->u1_left_ctb_avail = 0; |
| 373 | ps_proc->u1_top_lt_ctb_avail = 0; |
| 374 | if((0 == ps_proc->i4_ctb_slice_y) || (0 == ps_proc->i4_ctb_tile_y)) |
| 375 | { |
| 376 | ps_proc->u1_top_ctb_avail = 0; |
| 377 | if((i2_wd_in_ctb - 1) != ps_proc->i4_ctb_slice_x) |
| 378 | { |
| 379 | ps_proc->u1_top_rt_ctb_avail = 0; |
| 380 | } |
| 381 | } |
| 382 | } |
| 383 | /*For slices not beginning at start of a ctb row*/ |
| 384 | else if(ps_proc->i4_ctb_x > 0) |
| 385 | { |
| 386 | if((0 == ps_proc->i4_ctb_slice_y) || (0 == ps_proc->i4_ctb_tile_y)) |
| 387 | { |
| 388 | ps_proc->u1_top_ctb_avail = 0; |
| 389 | ps_proc->u1_top_lt_ctb_avail = 0; |
| 390 | if(0 == ps_proc->i4_ctb_slice_x) |
| 391 | { |
| 392 | ps_proc->u1_left_ctb_avail = 0; |
| 393 | } |
| 394 | if((i2_wd_in_ctb - 1) != ps_proc->i4_ctb_slice_x) |
| 395 | { |
| 396 | ps_proc->u1_top_rt_ctb_avail = 0; |
| 397 | } |
| 398 | } |
| 399 | else if((1 == ps_proc->i4_ctb_slice_y) && (0 == ps_proc->i4_ctb_slice_x)) |
| 400 | { |
| 401 | ps_proc->u1_top_lt_ctb_avail = 0; |
| 402 | } |
| 403 | } |
| 404 | |
| 405 | if((ps_proc->i4_ctb_x == (ps_sps->i2_pic_wd_in_ctb - 1)) || ((ps_tile->u2_wd - 1) == ps_proc->i4_ctb_tile_x)) |
| 406 | { |
| 407 | ps_proc->u1_top_rt_ctb_avail = 0; |
| 408 | } |
| 409 | |
| 410 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 411 | { |
| 412 | WORD32 next_ctb_idx; |
| 413 | next_ctb_idx = cur_ctb_idx + 1; |
| 414 | |
| 415 | if(ps_tile->u2_wd == (ps_proc->i4_ctb_tile_x + 1)) |
| 416 | { |
| 417 | if((ps_proc->i4_ctb_tile_y + 1) == ps_tile->u2_ht) |
| 418 | { |
| 419 | //Last tile |
| 420 | if(((ps_proc->i4_ctb_tile_y + 1 + ps_tile->u1_pos_y) == ps_sps->i2_pic_ht_in_ctb) && ((ps_proc->i4_ctb_tile_x + 1 + ps_tile->u1_pos_x) == ps_sps->i2_pic_wd_in_ctb)) |
| 421 | { |
| 422 | next_ctb_idx = cur_ctb_idx + 1; |
| 423 | } |
| 424 | else //Not last tile, but new tile |
| 425 | { |
| 426 | tile_t *ps_tile_next = ps_proc->ps_tile + 1; |
| 427 | next_ctb_idx = ps_tile_next->u1_pos_x + (ps_tile_next->u1_pos_y * ps_sps->i2_pic_wd_in_ctb); |
| 428 | } |
| 429 | } |
| 430 | else //End of each tile row |
| 431 | { |
| 432 | next_ctb_idx = ((ps_tile->u1_pos_y + ps_proc->i4_ctb_tile_y + 1) * ps_sps->i2_pic_wd_in_ctb) + ps_tile->u1_pos_x; |
| 433 | } |
| 434 | } |
| 435 | ps_proc->i4_next_pu_ctb_cnt = next_ctb_idx; |
| 436 | ps_proc->i4_ctb_pu_cnt = |
| 437 | ps_proc->pu4_pic_pu_idx[next_ctb_idx] |
| 438 | - ps_proc->pu4_pic_pu_idx[cur_ctb_idx]; |
| 439 | cur_pu_idx = ps_proc->pu4_pic_pu_idx[cur_ctb_idx]; |
| 440 | ps_proc->i4_ctb_start_pu_idx = cur_pu_idx; |
| 441 | ps_proc->ps_pu = &ps_proc->ps_pic_pu[cur_pu_idx]; |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | void ihevcd_update_ctb_tu_cnt(process_ctxt_t *ps_proc) |
| 446 | { |
| 447 | sps_t *ps_sps = ps_proc->ps_sps; |
| 448 | codec_t *ps_codec = ps_proc->ps_codec; |
| 449 | WORD32 cur_ctb_idx; |
| 450 | |
| 451 | cur_ctb_idx = ps_proc->i4_ctb_x |
| 452 | + ps_proc->i4_ctb_y * (ps_sps->i2_pic_wd_in_ctb); |
| 453 | |
| 454 | { |
| 455 | tile_t *ps_tile; |
| 456 | WORD32 next_ctb_tu_idx; |
| 457 | ps_tile = ps_proc->ps_tile; |
| 458 | |
| 459 | |
| 460 | if(1 == ps_codec->i4_num_cores) |
| 461 | { |
| 462 | next_ctb_tu_idx = cur_ctb_idx % RESET_TU_BUF_NCTB + 1; |
| 463 | if(ps_tile->u2_wd == (ps_proc->i4_ctb_tile_x + 1)) |
| 464 | { |
| 465 | if((ps_proc->i4_ctb_tile_y + 1) == ps_tile->u2_ht) |
| 466 | { |
| 467 | //Last tile |
| 468 | if(((ps_proc->i4_ctb_tile_y + 1 + ps_tile->u1_pos_y) == ps_sps->i2_pic_ht_in_ctb) && ((ps_proc->i4_ctb_tile_x + 1 + ps_tile->u1_pos_x) == ps_sps->i2_pic_wd_in_ctb)) |
| 469 | { |
| 470 | next_ctb_tu_idx = (cur_ctb_idx % RESET_TU_BUF_NCTB) + 1; |
| 471 | } |
| 472 | else //Not last tile, but new tile |
| 473 | { |
| 474 | tile_t *ps_tile_next = ps_proc->ps_tile + 1; |
| 475 | next_ctb_tu_idx = ps_tile_next->u1_pos_x + (ps_tile_next->u1_pos_y * ps_sps->i2_pic_wd_in_ctb); |
| 476 | } |
| 477 | } |
| 478 | else //End of each tile row |
| 479 | { |
| 480 | next_ctb_tu_idx = ((ps_tile->u1_pos_y + ps_proc->i4_ctb_tile_y + 1) * ps_sps->i2_pic_wd_in_ctb) + ps_tile->u1_pos_x; |
| 481 | } |
| 482 | } |
| 483 | ps_proc->i4_next_tu_ctb_cnt = next_ctb_tu_idx; |
| 484 | ps_proc->i4_ctb_tu_cnt = ps_proc->pu4_pic_tu_idx[next_ctb_tu_idx] - ps_proc->pu4_pic_tu_idx[cur_ctb_idx % RESET_TU_BUF_NCTB]; |
| 485 | } |
| 486 | else |
| 487 | { |
| 488 | next_ctb_tu_idx = cur_ctb_idx + 1; |
| 489 | if(ps_tile->u2_wd == (ps_proc->i4_ctb_tile_x + 1)) |
| 490 | { |
| 491 | if((ps_proc->i4_ctb_tile_y + 1) == ps_tile->u2_ht) |
| 492 | { |
| 493 | //Last tile |
| 494 | if(((ps_proc->i4_ctb_tile_y + 1 + ps_tile->u1_pos_y) == ps_sps->i2_pic_ht_in_ctb) && ((ps_proc->i4_ctb_tile_x + 1 + ps_tile->u1_pos_x) == ps_sps->i2_pic_wd_in_ctb)) |
| 495 | { |
| 496 | next_ctb_tu_idx = (cur_ctb_idx % RESET_TU_BUF_NCTB) + 1; |
| 497 | } |
| 498 | else //Not last tile, but new tile |
| 499 | { |
| 500 | tile_t *ps_tile_next = ps_proc->ps_tile + 1; |
| 501 | next_ctb_tu_idx = ps_tile_next->u1_pos_x + (ps_tile_next->u1_pos_y * ps_sps->i2_pic_wd_in_ctb); |
| 502 | } |
| 503 | } |
| 504 | else //End of each tile row |
| 505 | { |
| 506 | next_ctb_tu_idx = ((ps_tile->u1_pos_y + ps_proc->i4_ctb_tile_y + 1) * ps_sps->i2_pic_wd_in_ctb) + ps_tile->u1_pos_x; |
| 507 | } |
| 508 | } |
| 509 | ps_proc->i4_next_tu_ctb_cnt = next_ctb_tu_idx; |
| 510 | ps_proc->i4_ctb_tu_cnt = ps_proc->pu4_pic_tu_idx[next_ctb_tu_idx] - |
| 511 | ps_proc->pu4_pic_tu_idx[cur_ctb_idx]; |
| 512 | } |
| 513 | } |
| 514 | } |
| 515 | |
| 516 | IHEVCD_ERROR_T ihevcd_process(process_ctxt_t *ps_proc) |
| 517 | { |
| 518 | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
| 519 | codec_t *ps_codec; |
| 520 | sps_t *ps_sps = ps_proc->ps_sps; |
| 521 | |
| 522 | WORD32 nctb; |
| 523 | WORD32 i; |
| 524 | WORD32 idx; |
| 525 | WORD32 nop_cnt; |
| 526 | WORD32 num_minpu_in_ctb; |
| 527 | WORD32 cur_slice_idx, cur_ctb_tile_x, cur_ctb_slice_x, cur_ctb_tile_y, cur_ctb_slice_y; |
| 528 | WORD32 nxt_ctb_slice_y, nxt_ctb_slice_x; |
| 529 | tu_t *ps_tu_cur, *ps_tu_nxt; |
| 530 | UWORD8 *pu1_pu_map_cur, *pu1_pu_map_nxt; |
| 531 | WORD32 num_ctb, num_ctb_tmp; |
| 532 | proc_type_t proc_type; |
| 533 | |
| 534 | |
| 535 | WORD32 ctb_size = 1 << ps_sps->i1_log2_ctb_size; |
| 536 | |
| 537 | PROFILE_DISABLE_PROCESS_CTB(); |
| 538 | |
| 539 | ps_codec = ps_proc->ps_codec; |
| 540 | num_minpu_in_ctb = (ctb_size / MIN_PU_SIZE) * (ctb_size / MIN_PU_SIZE); |
| 541 | |
| 542 | nctb = MIN(ps_codec->i4_proc_nctb, ps_proc->i4_ctb_cnt); |
| 543 | nctb = MIN(nctb, (ps_proc->ps_tile->u2_wd - ps_proc->i4_ctb_tile_x)); |
| 544 | |
| 545 | if(ps_proc->i4_cur_slice_idx > (MAX_SLICE_HDR_CNT - 2 * ps_sps->i2_pic_wd_in_ctb)) |
| 546 | { |
| 547 | num_ctb = 1; |
| 548 | } |
| 549 | else |
| 550 | { |
| 551 | num_ctb = ps_proc->i4_nctb; |
| 552 | } |
| 553 | nxt_ctb_slice_y = ps_proc->i4_ctb_slice_y; |
| 554 | nxt_ctb_slice_x = ps_proc->i4_ctb_slice_x; |
| 555 | pu1_pu_map_nxt = ps_proc->pu1_pu_map; |
| 556 | ps_tu_nxt = ps_proc->ps_tu; |
| 557 | |
| 558 | while(ps_proc->i4_ctb_cnt) |
| 559 | { |
| 560 | ps_proc->i4_ctb_slice_y = nxt_ctb_slice_y; |
| 561 | ps_proc->i4_ctb_slice_x = nxt_ctb_slice_x; |
| 562 | ps_proc->pu1_pu_map = pu1_pu_map_nxt; |
| 563 | ps_proc->ps_tu = ps_tu_nxt; |
| 564 | |
| 565 | cur_ctb_tile_x = ps_proc->i4_ctb_tile_x; |
| 566 | cur_ctb_tile_y = ps_proc->i4_ctb_tile_y; |
| 567 | cur_ctb_slice_x = ps_proc->i4_ctb_slice_x; |
| 568 | cur_ctb_slice_y = ps_proc->i4_ctb_slice_y; |
| 569 | cur_slice_idx = ps_proc->i4_cur_slice_idx; |
| 570 | ps_tu_cur = ps_proc->ps_tu; |
| 571 | pu1_pu_map_cur = ps_proc->pu1_pu_map; |
| 572 | proc_type = PROC_INTER_PRED; |
| 573 | |
| 574 | if(ps_proc->i4_ctb_cnt < num_ctb) |
| 575 | { |
| 576 | num_ctb = ps_proc->i4_ctb_cnt; |
| 577 | } |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 578 | num_ctb_tmp = num_ctb; |
| 579 | |
| 580 | while(num_ctb_tmp) |
| 581 | { |
| 582 | slice_header_t *ps_slice_hdr; |
| 583 | tile_t *ps_tile = ps_proc->ps_tile; |
| 584 | |
| 585 | /* Waiting for Parsing to be done*/ |
| 586 | { |
| 587 | |
| 588 | |
| 589 | nop_cnt = PROC_NOP_CNT; |
| 590 | if(ps_proc->i4_check_parse_status || ps_proc->i4_check_proc_status) |
| 591 | { |
| 592 | while(1) |
| 593 | { |
| 594 | volatile UWORD8 *pu1_buf; |
| 595 | volatile WORD32 status; |
| 596 | status = 1; |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 597 | /* Check if all dependencies for the next nCTBs are met */ |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 598 | /* Check if the next nCTBs are parsed */ |
| 599 | if(ps_proc->i4_check_parse_status) |
| 600 | { |
| 601 | idx = (ps_proc->i4_ctb_x + nctb - 1); |
| 602 | idx += (ps_proc->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb); |
| 603 | pu1_buf = (ps_codec->pu1_parse_map + idx); |
| 604 | status = *pu1_buf; |
| 605 | } |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 606 | |
| 607 | if(status) |
| 608 | break; |
| 609 | |
| 610 | /* if dependencies are not met, then wait for few cycles. |
| 611 | * Even after few iterations, if the dependencies are not met then yield |
| 612 | */ |
| 613 | if(nop_cnt > 0) |
| 614 | { |
| 615 | NOP(128); |
| 616 | nop_cnt -= 128; |
| 617 | } |
| 618 | else |
| 619 | { |
| 620 | nop_cnt = PROC_NOP_CNT; |
| 621 | ithread_yield(); |
| 622 | } |
| 623 | } |
| 624 | } |
| 625 | } |
| 626 | |
| 627 | /* Check proc map to ensure dependencies for recon are met */ |
| 628 | ihevcd_proc_map_check(ps_proc, proc_type, nctb); |
| 629 | |
| 630 | ihevcd_slice_hdr_update(ps_proc); |
| 631 | ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 632 | |
| 633 | //ihevcd_mv_prediction(); |
| 634 | //ihevcd_lvl_unpack(); |
| 635 | //ihevcd_inter_iq_it_recon(); |
| 636 | //Following does prediction, iq, it and recon on a TU by TU basis for intra TUs |
| 637 | //ihevcd_intra_process(); |
| 638 | //ihevcd_ctb_boundary_strength_islice(ps_proc, ctb_size); |
| 639 | //ihevcd_deblk_ctb(ps_proc); |
| 640 | |
| 641 | /* iq,it recon of Intra TU */ |
| 642 | { |
| 643 | UWORD32 *pu4_ctb_top_pu_idx, *pu4_ctb_left_pu_idx, *pu4_ctb_top_left_pu_idx; |
| 644 | WORD32 cur_ctb_idx; |
| 645 | |
| 646 | ihevcd_ctb_avail_update(ps_proc); |
| 647 | |
| 648 | #if DEBUG_DUMP_FRAME_BUFFERS_INFO |
| 649 | au1_pic_avail_ctb_flags[ps_proc->i4_ctb_x + ps_proc->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb] = |
| 650 | ((ps_proc->u1_top_ctb_avail << 3) | (ps_proc->u1_left_ctb_avail << 2) | (ps_proc->u1_top_lt_ctb_avail << 1) | (ps_proc->u1_top_rt_ctb_avail)); |
| 651 | au4_pic_ctb_slice_xy[ps_proc->i4_ctb_x + ps_proc->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb] = |
| 652 | (((UWORD16)ps_proc->i4_ctb_slice_x << 16) | ((UWORD16)ps_proc->i4_ctb_slice_y << 16)); |
| 653 | #endif |
| 654 | |
| 655 | /*************************************************/ |
| 656 | /**************** MV pred **********************/ |
| 657 | /*************************************************/ |
| 658 | if(PSLICE == ps_slice_hdr->i1_slice_type |
| 659 | || BSLICE == ps_slice_hdr->i1_slice_type) |
| 660 | { |
| 661 | mv_ctxt_t s_mv_ctxt; |
| 662 | |
| 663 | pu4_ctb_top_pu_idx = ps_proc->pu4_pic_pu_idx_top |
| 664 | + (ps_proc->i4_ctb_x * ctb_size / MIN_PU_SIZE); |
| 665 | pu4_ctb_left_pu_idx = ps_proc->pu4_pic_pu_idx_left; |
| 666 | pu4_ctb_top_left_pu_idx = &ps_proc->u4_ctb_top_left_pu_idx; |
| 667 | |
| 668 | /* Initializing s_mv_ctxt */ |
| 669 | if(ps_codec->i4_num_cores > MV_PRED_NUM_CORES_THRESHOLD) |
| 670 | { |
| 671 | s_mv_ctxt.ps_pps = ps_proc->ps_pps; |
| 672 | s_mv_ctxt.ps_sps = ps_proc->ps_sps; |
| 673 | s_mv_ctxt.ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 674 | s_mv_ctxt.i4_ctb_x = ps_proc->i4_ctb_x; |
| 675 | s_mv_ctxt.i4_ctb_y = ps_proc->i4_ctb_y; |
| 676 | s_mv_ctxt.ps_pu = ps_proc->ps_pu; |
| 677 | s_mv_ctxt.ps_pic_pu = ps_proc->ps_pic_pu; |
| 678 | s_mv_ctxt.ps_tile = ps_tile; |
| 679 | s_mv_ctxt.pu4_pic_pu_idx_map = ps_proc->pu4_pic_pu_idx_map; |
| 680 | s_mv_ctxt.pu4_pic_pu_idx = ps_proc->pu4_pic_pu_idx; |
| 681 | s_mv_ctxt.pu1_pic_pu_map = ps_proc->pu1_pic_pu_map; |
| 682 | s_mv_ctxt.i4_ctb_pu_cnt = ps_proc->i4_ctb_pu_cnt; |
| 683 | s_mv_ctxt.i4_ctb_start_pu_idx = ps_proc->i4_ctb_start_pu_idx; |
| 684 | s_mv_ctxt.u1_top_ctb_avail = ps_proc->u1_top_ctb_avail; |
| 685 | s_mv_ctxt.u1_top_rt_ctb_avail = ps_proc->u1_top_rt_ctb_avail; |
| 686 | s_mv_ctxt.u1_top_lt_ctb_avail = ps_proc->u1_top_lt_ctb_avail; |
| 687 | s_mv_ctxt.u1_left_ctb_avail = ps_proc->u1_left_ctb_avail; |
| 688 | |
| 689 | ihevcd_get_mv_ctb(&s_mv_ctxt, pu4_ctb_top_pu_idx, |
| 690 | pu4_ctb_left_pu_idx, pu4_ctb_top_left_pu_idx); |
| 691 | } |
| 692 | |
| 693 | ihevcd_inter_pred_ctb(ps_proc); |
| 694 | } |
| 695 | else if(ps_codec->i4_num_cores > MV_PRED_NUM_CORES_THRESHOLD) |
| 696 | { |
| 697 | WORD32 next_ctb_idx, num_pu_per_ctb, ctb_start_pu_idx, pu_cnt; |
| 698 | pu_t *ps_pu; |
| 699 | WORD32 num_minpu_in_ctb = (ctb_size / MIN_PU_SIZE) * (ctb_size / MIN_PU_SIZE); |
| 700 | UWORD8 *pu1_pic_pu_map_ctb = ps_proc->pu1_pic_pu_map + |
| 701 | (ps_proc->i4_ctb_x + ps_proc->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb) * num_minpu_in_ctb; |
| 702 | WORD32 row, col; |
| 703 | UWORD32 *pu4_nbr_pu_idx = ps_proc->pu4_pic_pu_idx_map; |
| 704 | WORD32 nbr_pu_idx_strd = MAX_CTB_SIZE / MIN_PU_SIZE + 2; |
Naveen Kumar P | 41f9419 | 2017-06-22 15:36:50 +0530 | [diff] [blame] | 705 | WORD32 ctb_size_in_min_pu = (ctb_size / MIN_PU_SIZE); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 706 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 707 | /* Neighbor PU idx update inside CTB */ |
| 708 | /* 1byte per 4x4. Indicates the PU idx that 4x4 block belongs to */ |
| 709 | |
| 710 | cur_ctb_idx = ps_proc->i4_ctb_x |
| 711 | + ps_proc->i4_ctb_y * (ps_sps->i2_pic_wd_in_ctb); |
| 712 | next_ctb_idx = ps_proc->i4_next_pu_ctb_cnt; |
| 713 | num_pu_per_ctb = ps_proc->pu4_pic_pu_idx[next_ctb_idx] |
| 714 | - ps_proc->pu4_pic_pu_idx[cur_ctb_idx]; |
| 715 | ctb_start_pu_idx = ps_proc->pu4_pic_pu_idx[cur_ctb_idx]; |
| 716 | ps_pu = &ps_proc->ps_pic_pu[ctb_start_pu_idx]; |
| 717 | |
| 718 | for(pu_cnt = 0; pu_cnt < num_pu_per_ctb; pu_cnt++, ps_pu++) |
| 719 | { |
| 720 | UWORD32 cur_pu_idx; |
| 721 | WORD32 pu_ht = (ps_pu->b4_ht + 1) << 2; |
| 722 | WORD32 pu_wd = (ps_pu->b4_wd + 1) << 2; |
| 723 | |
| 724 | cur_pu_idx = ctb_start_pu_idx + pu_cnt; |
| 725 | |
| 726 | for(row = 0; row < pu_ht / MIN_PU_SIZE; row++) |
| 727 | for(col = 0; col < pu_wd / MIN_PU_SIZE; col++) |
| 728 | pu4_nbr_pu_idx[(1 + ps_pu->b4_pos_x + col) |
| 729 | + (1 + ps_pu->b4_pos_y + row) |
| 730 | * nbr_pu_idx_strd] = |
| 731 | cur_pu_idx; |
| 732 | } |
| 733 | |
| 734 | /* Updating Top and Left pointers */ |
| 735 | { |
| 736 | WORD32 rows_remaining = ps_sps->i2_pic_height_in_luma_samples |
| 737 | - (ps_proc->i4_ctb_y << ps_sps->i1_log2_ctb_size); |
| 738 | WORD32 ctb_size_left = MIN(ctb_size, rows_remaining); |
| 739 | |
| 740 | /* Top Left */ |
| 741 | /* saving top left before updating top ptr, as updating top ptr will overwrite the top left for the next ctb */ |
| 742 | ps_proc->u4_ctb_top_left_pu_idx = ps_proc->pu4_pic_pu_idx_top[((ps_proc->i4_ctb_x + 1) * ctb_size / MIN_PU_SIZE) - 1]; |
| 743 | for(i = 0; i < ctb_size / MIN_PU_SIZE; i++) |
| 744 | { |
| 745 | /* Left */ |
| 746 | /* Last column of au4_nbr_pu_idx */ |
| 747 | ps_proc->pu4_pic_pu_idx_left[i] = |
| 748 | pu4_nbr_pu_idx[(ctb_size / MIN_PU_SIZE) + (i + 1) * nbr_pu_idx_strd]; |
| 749 | /* Top */ |
| 750 | /* Last row of au4_nbr_pu_idx */ |
| 751 | ps_proc->pu4_pic_pu_idx_top[(ps_proc->i4_ctb_x * ctb_size / MIN_PU_SIZE) + i] = |
| 752 | pu4_nbr_pu_idx[(ctb_size_left / MIN_PU_SIZE) * nbr_pu_idx_strd + i + 1]; |
| 753 | |
| 754 | } |
Naveen Kumar P | 41f9419 | 2017-06-22 15:36:50 +0530 | [diff] [blame] | 755 | |
| 756 | /* Updating the CTB level PU idx (Used for collocated MV pred)*/ |
| 757 | { |
| 758 | WORD32 ctb_row, ctb_col, index_pic_map, index_nbr_map; |
| 759 | WORD32 first_pu_of_ctb; |
| 760 | first_pu_of_ctb = pu4_nbr_pu_idx[1 + nbr_pu_idx_strd]; |
| 761 | |
| 762 | index_pic_map = 0 * ctb_size_in_min_pu + 0; |
| 763 | index_nbr_map = (0 + 1) * nbr_pu_idx_strd + (0 + 1); |
| 764 | |
| 765 | for(ctb_row = 0; ctb_row < ctb_size_in_min_pu; ctb_row++) |
| 766 | { |
| 767 | for(ctb_col = 0; ctb_col < ctb_size_in_min_pu; ctb_col++) |
| 768 | { |
| 769 | pu1_pic_pu_map_ctb[index_pic_map + ctb_col] = pu4_nbr_pu_idx[index_nbr_map + ctb_col] |
| 770 | - first_pu_of_ctb; |
| 771 | } |
| 772 | index_pic_map += ctb_size_in_min_pu; |
| 773 | index_nbr_map += nbr_pu_idx_strd; |
| 774 | } |
| 775 | } |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 776 | } |
| 777 | } |
| 778 | } |
| 779 | |
| 780 | if(ps_proc->ps_pps->i1_tiles_enabled_flag) |
| 781 | { |
| 782 | /*Update the tile index buffer with tile information for the current ctb*/ |
| 783 | UWORD16 *pu1_tile_idx = ps_proc->pu1_tile_idx; |
| 784 | pu1_tile_idx[(ps_proc->i4_ctb_x + (ps_proc->i4_ctb_y * ps_sps->i2_pic_wd_in_ctb))] |
| 785 | = ps_proc->i4_cur_tile_idx; |
| 786 | } |
| 787 | |
| 788 | /*************************************************/ |
| 789 | /*********** BS, QP and Deblocking **************/ |
| 790 | /*************************************************/ |
| 791 | /* Boundary strength call has to be after IQ IT recon since QP population needs ps_proc->i4_qp_const_inc_ctb flag */ |
| 792 | |
| 793 | { |
| 794 | slice_header_t *ps_slice_hdr; |
| 795 | ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 796 | |
| 797 | |
| 798 | /* Check if deblock is disabled for the current slice or if it is disabled for the current picture |
| 799 | * because of disable deblock api |
| 800 | */ |
| 801 | if(0 == ps_codec->i4_disable_deblk_pic) |
| 802 | { |
| 803 | if(ps_codec->i4_num_cores > MV_PRED_NUM_CORES_THRESHOLD) |
| 804 | { |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 805 | /* Boundary strength calculation is done irrespective of whether deblocking is disabled |
| 806 | * in the slice or not, to handle deblocking slice boundaries */ |
| 807 | if((0 == ps_codec->i4_slice_error)) |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 808 | { |
| 809 | ihevcd_update_ctb_tu_cnt(ps_proc); |
| 810 | ps_proc->s_bs_ctxt.ps_pps = ps_proc->ps_pps; |
| 811 | ps_proc->s_bs_ctxt.ps_sps = ps_proc->ps_sps; |
| 812 | ps_proc->s_bs_ctxt.ps_codec = ps_proc->ps_codec; |
| 813 | ps_proc->s_bs_ctxt.i4_ctb_tu_cnt = ps_proc->i4_ctb_tu_cnt; |
| 814 | ps_proc->s_bs_ctxt.i4_ctb_x = ps_proc->i4_ctb_x; |
| 815 | ps_proc->s_bs_ctxt.i4_ctb_y = ps_proc->i4_ctb_y; |
| 816 | ps_proc->s_bs_ctxt.i4_ctb_tile_x = ps_proc->i4_ctb_tile_x; |
| 817 | ps_proc->s_bs_ctxt.i4_ctb_tile_y = ps_proc->i4_ctb_tile_y; |
| 818 | ps_proc->s_bs_ctxt.i4_ctb_slice_x = ps_proc->i4_ctb_slice_x; |
| 819 | ps_proc->s_bs_ctxt.i4_ctb_slice_y = ps_proc->i4_ctb_slice_y; |
| 820 | ps_proc->s_bs_ctxt.ps_tu = ps_proc->ps_tu; |
| 821 | ps_proc->s_bs_ctxt.ps_pu = ps_proc->ps_pu; |
| 822 | ps_proc->s_bs_ctxt.pu4_pic_pu_idx_map = ps_proc->pu4_pic_pu_idx_map; |
| 823 | ps_proc->s_bs_ctxt.i4_next_pu_ctb_cnt = ps_proc->i4_next_pu_ctb_cnt; |
| 824 | ps_proc->s_bs_ctxt.i4_next_tu_ctb_cnt = ps_proc->i4_next_tu_ctb_cnt; |
| 825 | ps_proc->s_bs_ctxt.pu1_slice_idx = ps_proc->pu1_slice_idx; |
| 826 | ps_proc->s_bs_ctxt.ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 827 | ps_proc->s_bs_ctxt.ps_tile = ps_proc->ps_tile; |
| 828 | |
| 829 | if(ISLICE == ps_slice_hdr->i1_slice_type) |
| 830 | { |
| 831 | ihevcd_ctb_boundary_strength_islice(&ps_proc->s_bs_ctxt); |
| 832 | } |
| 833 | else |
| 834 | { |
| 835 | ihevcd_ctb_boundary_strength_pbslice(&ps_proc->s_bs_ctxt); |
| 836 | } |
| 837 | } |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 838 | |
| 839 | /* Boundary strength is set to zero if deblocking is disabled for the current slice */ |
| 840 | if((0 != ps_slice_hdr->i1_slice_disable_deblocking_filter_flag)) |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 841 | { |
Harish Mahendrakar | 707042f | 2014-06-04 10:31:48 -0700 | [diff] [blame] | 842 | WORD32 bs_strd = (ps_sps->i2_pic_wd_in_ctb + 1) * (ctb_size * ctb_size / 8 / 16); |
| 843 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 844 | UWORD32 *pu4_vert_bs = (UWORD32 *)((UWORD8 *)ps_proc->s_bs_ctxt.pu4_pic_vert_bs + |
| 845 | ps_proc->i4_ctb_x * (ctb_size * ctb_size / 8 / 16) + |
Harish Mahendrakar | 707042f | 2014-06-04 10:31:48 -0700 | [diff] [blame] | 846 | ps_proc->i4_ctb_y * bs_strd); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 847 | UWORD32 *pu4_horz_bs = (UWORD32 *)((UWORD8 *)ps_proc->s_bs_ctxt.pu4_pic_horz_bs + |
| 848 | ps_proc->i4_ctb_x * (ctb_size * ctb_size / 8 / 16) + |
Harish Mahendrakar | 707042f | 2014-06-04 10:31:48 -0700 | [diff] [blame] | 849 | ps_proc->i4_ctb_y * bs_strd); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 850 | |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 851 | memset(pu4_vert_bs, 0, (ctb_size / 8) * (ctb_size / 4) / 8 * 2); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 852 | memset(pu4_horz_bs, 0, (ctb_size / 8) * (ctb_size / 4) / 8 * 2); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 853 | } |
| 854 | } |
| 855 | } |
| 856 | } |
| 857 | |
| 858 | /* Per CTB update the following */ |
| 859 | { |
| 860 | WORD32 cur_ctb_idx = ps_proc->i4_ctb_x |
| 861 | + ps_proc->i4_ctb_y * (ps_sps->i2_pic_wd_in_ctb); |
| 862 | cur_ctb_idx++; |
| 863 | |
| 864 | ps_proc->pu1_pu_map += nctb * num_minpu_in_ctb; |
| 865 | ps_proc->ps_tu += ps_proc->i4_ctb_tu_cnt; |
| 866 | if((1 == ps_codec->i4_num_cores) && |
| 867 | (0 == cur_ctb_idx % RESET_TU_BUF_NCTB)) |
| 868 | { |
| 869 | ps_proc->ps_tu = ps_proc->ps_pic_tu; |
| 870 | } |
| 871 | ps_proc->ps_pu += ps_proc->i4_ctb_pu_cnt; |
| 872 | } |
| 873 | |
| 874 | /* Update proc map for recon*/ |
| 875 | ihevcd_proc_map_update(ps_proc, proc_type, nctb); |
| 876 | |
| 877 | num_ctb_tmp -= nctb; |
| 878 | ihevcd_ctb_pos_update(ps_proc, nctb); |
| 879 | |
| 880 | } |
| 881 | |
| 882 | if(cur_slice_idx != ps_proc->i4_cur_slice_idx) |
| 883 | { |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 884 | ps_proc->ps_slice_hdr = ps_codec->ps_slice_hdr_base + ((cur_slice_idx)&(MAX_SLICE_HDR_CNT - 1)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 885 | ps_proc->i4_cur_slice_idx = cur_slice_idx; |
| 886 | } |
| 887 | /* Restore the saved variables */ |
| 888 | num_ctb_tmp = num_ctb; |
| 889 | ps_proc->i4_ctb_x -= num_ctb; |
| 890 | ps_proc->i4_ctb_tile_x = cur_ctb_tile_x; |
| 891 | ps_proc->i4_ctb_slice_x = cur_ctb_slice_x; |
| 892 | ps_proc->i4_ctb_tile_y = cur_ctb_tile_y; |
| 893 | ps_proc->i4_ctb_slice_y = cur_ctb_slice_y; |
| 894 | ps_proc->pu1_pu_map = pu1_pu_map_cur; |
| 895 | ps_proc->ps_tu = ps_tu_cur; |
| 896 | proc_type = PROC_RECON; |
| 897 | |
| 898 | while(num_ctb_tmp) |
| 899 | { |
| 900 | |
| 901 | /* Check proc map to ensure dependencies for recon are met */ |
| 902 | ihevcd_proc_map_check(ps_proc, proc_type, nctb); |
| 903 | |
| 904 | ihevcd_slice_hdr_update(ps_proc); |
| 905 | |
| 906 | { |
| 907 | |
| 908 | ihevcd_ctb_avail_update(ps_proc); |
| 909 | |
| 910 | /*************************************************/ |
| 911 | /**************** IQ IT RECON *******************/ |
| 912 | /*************************************************/ |
| 913 | |
| 914 | ihevcd_update_ctb_tu_cnt(ps_proc); |
| 915 | |
| 916 | /* When scaling matrix is not to be used(scaling_list_enable_flag is zero in SPS), |
| 917 | * default value of 16 has to be used. Since the value is same for all sizes, |
| 918 | * same table is used for all cases. |
| 919 | */ |
| 920 | if(0 == ps_sps->i1_scaling_list_enable_flag) |
| 921 | { |
| 922 | ps_proc->api2_dequant_intra_matrix[0] = |
| 923 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 924 | ps_proc->api2_dequant_intra_matrix[1] = |
| 925 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 926 | ps_proc->api2_dequant_intra_matrix[2] = |
| 927 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 928 | ps_proc->api2_dequant_intra_matrix[3] = |
| 929 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 930 | |
| 931 | ps_proc->api2_dequant_inter_matrix[0] = |
| 932 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 933 | ps_proc->api2_dequant_inter_matrix[1] = |
| 934 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 935 | ps_proc->api2_dequant_inter_matrix[2] = |
| 936 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 937 | ps_proc->api2_dequant_inter_matrix[3] = |
| 938 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 939 | } |
| 940 | else |
| 941 | { |
| 942 | if(0 == ps_sps->i1_sps_scaling_list_data_present_flag) |
| 943 | { |
| 944 | ps_proc->api2_dequant_intra_matrix[0] = |
| 945 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 946 | ps_proc->api2_dequant_intra_matrix[1] = |
| 947 | (WORD16 *)gi2_intra_default_scale_mat_8x8; |
| 948 | ps_proc->api2_dequant_intra_matrix[2] = |
| 949 | (WORD16 *)gi2_intra_default_scale_mat_16x16; |
| 950 | ps_proc->api2_dequant_intra_matrix[3] = |
| 951 | (WORD16 *)gi2_intra_default_scale_mat_32x32; |
| 952 | |
| 953 | ps_proc->api2_dequant_inter_matrix[0] = |
| 954 | (WORD16 *)gi2_flat_scale_mat_32x32; |
| 955 | ps_proc->api2_dequant_inter_matrix[1] = |
| 956 | (WORD16 *)gi2_inter_default_scale_mat_8x8; |
| 957 | ps_proc->api2_dequant_inter_matrix[2] = |
| 958 | (WORD16 *)gi2_inter_default_scale_mat_16x16; |
| 959 | ps_proc->api2_dequant_inter_matrix[3] = |
| 960 | (WORD16 *)gi2_inter_default_scale_mat_32x32; |
| 961 | } |
| 962 | /*TODO: Add support for custom scaling matrices */ |
| 963 | } |
| 964 | |
| 965 | |
| 966 | /* CTB Level pointers */ |
| 967 | ps_proc->pu1_cur_ctb_luma = ps_proc->pu1_cur_pic_luma |
| 968 | + (ps_proc->i4_ctb_x * ctb_size |
| 969 | + ps_proc->i4_ctb_y * ctb_size |
| 970 | * ps_codec->i4_strd); |
| 971 | ps_proc->pu1_cur_ctb_chroma = ps_proc->pu1_cur_pic_chroma |
| 972 | + ps_proc->i4_ctb_x * ctb_size |
| 973 | + (ps_proc->i4_ctb_y * ctb_size * ps_codec->i4_strd / 2); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 974 | |
| 975 | ihevcd_iquant_itrans_recon_ctb(ps_proc); |
| 976 | } |
| 977 | |
| 978 | /* Per CTB update the following */ |
| 979 | { |
| 980 | WORD32 cur_ctb_idx = ps_proc->i4_ctb_x |
| 981 | + ps_proc->i4_ctb_y * (ps_sps->i2_pic_wd_in_ctb); |
| 982 | cur_ctb_idx++; |
| 983 | |
| 984 | ps_proc->pu1_pu_map += nctb * num_minpu_in_ctb; |
| 985 | ps_proc->ps_tu += ps_proc->i4_ctb_tu_cnt; |
| 986 | if((1 == ps_codec->i4_num_cores) && |
| 987 | (0 == cur_ctb_idx % RESET_TU_BUF_NCTB)) |
| 988 | { |
| 989 | ps_proc->ps_tu = ps_proc->ps_pic_tu; |
| 990 | } |
| 991 | ps_proc->ps_pu += ps_proc->i4_ctb_pu_cnt; |
| 992 | } |
| 993 | |
| 994 | |
| 995 | /* Update proc map for recon*/ |
| 996 | ihevcd_proc_map_update(ps_proc, proc_type, nctb); |
| 997 | |
| 998 | num_ctb_tmp -= nctb; |
| 999 | ihevcd_ctb_pos_update(ps_proc, nctb); |
| 1000 | } |
| 1001 | |
| 1002 | if(cur_slice_idx != ps_proc->i4_cur_slice_idx) |
| 1003 | { |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1004 | ps_proc->ps_slice_hdr = ps_codec->ps_slice_hdr_base + ((cur_slice_idx)&(MAX_SLICE_HDR_CNT - 1)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1005 | ps_proc->i4_cur_slice_idx = cur_slice_idx; |
| 1006 | } |
| 1007 | /* Restore the saved variables */ |
| 1008 | num_ctb_tmp = num_ctb; |
| 1009 | ps_proc->i4_ctb_x -= num_ctb; |
| 1010 | ps_proc->i4_ctb_tile_x = cur_ctb_tile_x; |
| 1011 | ps_proc->i4_ctb_slice_x = cur_ctb_slice_x; |
| 1012 | ps_proc->i4_ctb_tile_y = cur_ctb_tile_y; |
| 1013 | ps_proc->i4_ctb_slice_y = cur_ctb_slice_y; |
| 1014 | pu1_pu_map_nxt = ps_proc->pu1_pu_map; |
| 1015 | ps_tu_nxt = ps_proc->ps_tu; |
| 1016 | ps_proc->pu1_pu_map = pu1_pu_map_cur; |
| 1017 | ps_proc->ps_tu = ps_tu_cur; |
| 1018 | proc_type = PROC_DEBLK; |
| 1019 | |
| 1020 | while(num_ctb_tmp) |
| 1021 | { |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 1022 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1023 | |
| 1024 | /* Check proc map to ensure dependencies for deblk are met */ |
| 1025 | ihevcd_proc_map_check(ps_proc, proc_type, nctb); |
| 1026 | |
| 1027 | ihevcd_slice_hdr_update(ps_proc); |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 1028 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1029 | |
| 1030 | if(((0 == FRAME_ILF_PAD || ps_codec->i4_num_cores != 1)) && |
| 1031 | (0 == ps_codec->i4_disable_deblk_pic)) |
| 1032 | { |
| 1033 | WORD32 i4_is_last_ctb_x = 0; |
| 1034 | WORD32 i4_is_last_ctb_y = 0; |
| 1035 | |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 1036 | |
| 1037 | /* Deblocking is done irrespective of whether it is disabled in the slice or not, |
| 1038 | * to handle deblocking the slice boundaries */ |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1039 | { |
| 1040 | ps_proc->s_deblk_ctxt.ps_pps = ps_proc->ps_pps; |
| 1041 | ps_proc->s_deblk_ctxt.ps_sps = ps_proc->ps_sps; |
| 1042 | ps_proc->s_deblk_ctxt.ps_codec = ps_proc->ps_codec; |
| 1043 | ps_proc->s_deblk_ctxt.ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 1044 | ps_proc->s_deblk_ctxt.i4_ctb_x = ps_proc->i4_ctb_x; |
| 1045 | ps_proc->s_deblk_ctxt.i4_ctb_y = ps_proc->i4_ctb_y; |
| 1046 | ps_proc->s_deblk_ctxt.pu1_slice_idx = ps_proc->pu1_slice_idx; |
| 1047 | ps_proc->s_deblk_ctxt.is_chroma_yuv420sp_vu = (ps_codec->e_ref_chroma_fmt == IV_YUV_420SP_VU); |
| 1048 | |
| 1049 | /* Populating Current CTB's no_loop_filter flags */ |
| 1050 | { |
| 1051 | WORD32 row; |
| 1052 | WORD32 log2_ctb_size = ps_sps->i1_log2_ctb_size; |
| 1053 | |
| 1054 | /* Loop filter strd in units of num bits */ |
| 1055 | WORD32 loop_filter_strd = ((ps_sps->i2_pic_width_in_luma_samples + 63) >> 6) << 3; |
| 1056 | /* Bit position is the current 8x8 bit offset wrt pic_no_loop_filter |
| 1057 | * bit_pos has to be a WOR32 so that when it is negative, the downshift still retains it to be a negative value */ |
| 1058 | WORD32 bit_pos = ((ps_proc->i4_ctb_y << (log2_ctb_size - 3)) - 1) * loop_filter_strd + (ps_proc->i4_ctb_x << (log2_ctb_size - 3)) - 1; |
| 1059 | |
| 1060 | for(row = 0; row < (ctb_size >> 3) + 1; row++) |
| 1061 | { |
| 1062 | /* Go to the corresponding byte - read 32 bits and downshift */ |
| 1063 | ps_proc->s_deblk_ctxt.au2_ctb_no_loop_filter_flag[row] = (*(UWORD32 *)(ps_proc->pu1_pic_no_loop_filter_flag + (bit_pos >> 3))) >> (bit_pos & 7); |
| 1064 | bit_pos += loop_filter_strd; |
| 1065 | } |
| 1066 | } |
| 1067 | |
| 1068 | ihevcd_deblk_ctb(&ps_proc->s_deblk_ctxt, i4_is_last_ctb_x, i4_is_last_ctb_y); |
| 1069 | |
| 1070 | /* If the last CTB in the row was a complete CTB then deblocking has to be called from remaining pixels, since deblocking |
| 1071 | * is applied on a shifted CTB structure |
| 1072 | */ |
| 1073 | if(ps_proc->i4_ctb_x == ps_sps->i2_pic_wd_in_ctb - 1) |
| 1074 | { |
| 1075 | WORD32 i4_is_last_ctb_x = 1; |
| 1076 | WORD32 i4_is_last_ctb_y = 0; |
| 1077 | |
| 1078 | WORD32 last_x_pos; |
| 1079 | last_x_pos = (ps_sps->i2_pic_wd_in_ctb << ps_sps->i1_log2_ctb_size); |
| 1080 | if(last_x_pos == ps_sps->i2_pic_width_in_luma_samples) |
| 1081 | { |
| 1082 | ihevcd_deblk_ctb(&ps_proc->s_deblk_ctxt, i4_is_last_ctb_x, i4_is_last_ctb_y); |
| 1083 | } |
| 1084 | } |
| 1085 | |
| 1086 | |
| 1087 | /* If the last CTB in the column was a complete CTB then deblocking has to be called from remaining pixels, since deblocking |
| 1088 | * is applied on a shifted CTB structure |
| 1089 | */ |
| 1090 | if(ps_proc->i4_ctb_y == ps_sps->i2_pic_ht_in_ctb - 1) |
| 1091 | { |
| 1092 | WORD32 i4_is_last_ctb_x = 0; |
| 1093 | WORD32 i4_is_last_ctb_y = 1; |
| 1094 | WORD32 last_y_pos; |
| 1095 | last_y_pos = (ps_sps->i2_pic_ht_in_ctb << ps_sps->i1_log2_ctb_size); |
| 1096 | if(last_y_pos == ps_sps->i2_pic_height_in_luma_samples) |
| 1097 | { |
| 1098 | ihevcd_deblk_ctb(&ps_proc->s_deblk_ctxt, i4_is_last_ctb_x, i4_is_last_ctb_y); |
| 1099 | } |
| 1100 | } |
| 1101 | } |
| 1102 | } |
| 1103 | |
| 1104 | /* Update proc map for deblk*/ |
| 1105 | ihevcd_proc_map_update(ps_proc, proc_type, nctb); |
| 1106 | |
| 1107 | num_ctb_tmp -= nctb; |
| 1108 | ihevcd_ctb_pos_update(ps_proc, nctb); |
| 1109 | } |
| 1110 | |
| 1111 | if(cur_slice_idx != ps_proc->i4_cur_slice_idx) |
| 1112 | { |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1113 | ps_proc->ps_slice_hdr = ps_codec->ps_slice_hdr_base + ((cur_slice_idx)&(MAX_SLICE_HDR_CNT - 1)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1114 | ps_proc->i4_cur_slice_idx = cur_slice_idx; |
| 1115 | } |
| 1116 | /* Restore the saved variables */ |
| 1117 | num_ctb_tmp = num_ctb; |
| 1118 | ps_proc->i4_ctb_x -= num_ctb; |
| 1119 | ps_proc->i4_ctb_tile_x = cur_ctb_tile_x; |
| 1120 | ps_proc->i4_ctb_tile_y = cur_ctb_tile_y; |
| 1121 | ps_proc->pu1_pu_map = pu1_pu_map_cur; |
| 1122 | ps_proc->ps_tu = ps_tu_cur; |
| 1123 | nxt_ctb_slice_y = ps_proc->i4_ctb_slice_y; |
| 1124 | nxt_ctb_slice_x = ps_proc->i4_ctb_slice_x; |
| 1125 | ps_proc->i4_ctb_slice_y = cur_ctb_slice_y; |
| 1126 | ps_proc->i4_ctb_slice_x = cur_ctb_slice_x; |
| 1127 | proc_type = PROC_SAO; |
| 1128 | |
| 1129 | while(num_ctb_tmp) |
| 1130 | { |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 1131 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1132 | |
| 1133 | /* Check proc map to ensure dependencies for SAO are met */ |
| 1134 | ihevcd_proc_map_check(ps_proc, proc_type, nctb); |
| 1135 | |
| 1136 | ihevcd_slice_hdr_update(ps_proc); |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 1137 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1138 | |
| 1139 | if(0 == FRAME_ILF_PAD || ps_codec->i4_num_cores != 1) |
| 1140 | { |
Naveen Kumar Ponnusamy | 221aacc | 2014-10-28 11:23:59 +0530 | [diff] [blame] | 1141 | /* SAO is done even when it is disabled in the current slice, because |
| 1142 | * it is performed on a shifted CTB and the neighbor CTBs can belong |
| 1143 | * to different slices with SAO enabled */ |
| 1144 | if(0 == ps_codec->i4_disable_sao_pic) |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1145 | { |
| 1146 | ps_proc->s_sao_ctxt.ps_pps = ps_proc->ps_pps; |
| 1147 | ps_proc->s_sao_ctxt.ps_sps = ps_proc->ps_sps; |
| 1148 | ps_proc->s_sao_ctxt.ps_tile = ps_proc->ps_tile; |
| 1149 | ps_proc->s_sao_ctxt.ps_codec = ps_proc->ps_codec; |
| 1150 | ps_proc->s_sao_ctxt.ps_slice_hdr = ps_proc->ps_slice_hdr; |
| 1151 | ps_proc->s_sao_ctxt.i4_cur_slice_idx = ps_proc->i4_cur_slice_idx; |
| 1152 | |
| 1153 | |
| 1154 | #if SAO_PROCESS_SHIFT_CTB |
| 1155 | ps_proc->s_sao_ctxt.i4_ctb_x = ps_proc->i4_ctb_x; |
| 1156 | ps_proc->s_sao_ctxt.i4_ctb_y = ps_proc->i4_ctb_y; |
| 1157 | ps_proc->s_sao_ctxt.is_chroma_yuv420sp_vu = (ps_codec->e_ref_chroma_fmt == IV_YUV_420SP_VU); |
| 1158 | |
| 1159 | ihevcd_sao_shift_ctb(&ps_proc->s_sao_ctxt); |
| 1160 | #else |
| 1161 | if(ps_proc->i4_ctb_x > 1 && ps_proc->i4_ctb_y > 0) |
| 1162 | { |
| 1163 | ps_proc->s_sao_ctxt.i4_ctb_x = ps_proc->i4_ctb_x - 2; |
| 1164 | ps_proc->s_sao_ctxt.i4_ctb_y = ps_proc->i4_ctb_y - 1; |
| 1165 | |
| 1166 | ihevcd_sao_ctb(&ps_proc->s_sao_ctxt); |
| 1167 | } |
| 1168 | |
| 1169 | if(ps_sps->i2_pic_wd_in_ctb - 1 == ps_proc->i4_ctb_x && ps_proc->i4_ctb_y > 0) |
| 1170 | { |
| 1171 | ps_proc->s_sao_ctxt.i4_ctb_x = ps_proc->i4_ctb_x - 1; |
| 1172 | ps_proc->s_sao_ctxt.i4_ctb_y = ps_proc->i4_ctb_y - 1; |
| 1173 | |
| 1174 | ihevcd_sao_ctb(&ps_proc->s_sao_ctxt); |
| 1175 | |
| 1176 | ps_proc->s_sao_ctxt.i4_ctb_x = ps_proc->i4_ctb_x; |
| 1177 | ps_proc->s_sao_ctxt.i4_ctb_y = ps_proc->i4_ctb_y - 1; |
| 1178 | |
| 1179 | ihevcd_sao_ctb(&ps_proc->s_sao_ctxt); |
| 1180 | |
| 1181 | if(ps_sps->i2_pic_ht_in_ctb - 1 == ps_proc->i4_ctb_y) |
| 1182 | { |
| 1183 | WORD32 i4_ctb_x; |
| 1184 | ps_proc->s_sao_ctxt.i4_ctb_y = ps_proc->i4_ctb_y; |
| 1185 | for(i4_ctb_x = 0; i4_ctb_x < ps_sps->i2_pic_wd_in_ctb; i4_ctb_x++) |
| 1186 | { |
| 1187 | ps_proc->s_sao_ctxt.i4_ctb_x = i4_ctb_x; |
| 1188 | ihevcd_sao_ctb(&ps_proc->s_sao_ctxt); |
| 1189 | } |
| 1190 | } |
| 1191 | } |
| 1192 | #endif |
| 1193 | } |
| 1194 | |
| 1195 | |
| 1196 | /* Call padding if required */ |
| 1197 | { |
| 1198 | #if SAO_PROCESS_SHIFT_CTB |
| 1199 | |
| 1200 | if(0 == ps_proc->i4_ctb_x) |
| 1201 | { |
| 1202 | WORD32 pad_ht_luma; |
| 1203 | WORD32 pad_ht_chroma; |
| 1204 | |
| 1205 | ps_proc->pu1_cur_ctb_luma = ps_proc->pu1_cur_pic_luma |
| 1206 | + (ps_proc->i4_ctb_x * ctb_size |
| 1207 | + ps_proc->i4_ctb_y * ctb_size |
| 1208 | * ps_codec->i4_strd); |
| 1209 | ps_proc->pu1_cur_ctb_chroma = ps_proc->pu1_cur_pic_chroma |
| 1210 | + ps_proc->i4_ctb_x * ctb_size |
| 1211 | + (ps_proc->i4_ctb_y * ctb_size * ps_codec->i4_strd / 2); |
| 1212 | |
| 1213 | pad_ht_luma = ctb_size; |
| 1214 | pad_ht_luma += (ps_sps->i2_pic_ht_in_ctb - 1) == ps_proc->i4_ctb_y ? 8 : 0; |
| 1215 | pad_ht_chroma = ctb_size / 2; |
| 1216 | /* Pad left after 1st CTB is processed */ |
| 1217 | ps_codec->s_func_selector.ihevc_pad_left_luma_fptr(ps_proc->pu1_cur_ctb_luma - 8 * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_luma, PAD_LEFT); |
| 1218 | ps_codec->s_func_selector.ihevc_pad_left_chroma_fptr(ps_proc->pu1_cur_ctb_chroma - 16 * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_chroma, PAD_LEFT); |
| 1219 | } |
| 1220 | |
| 1221 | if((ps_sps->i2_pic_wd_in_ctb - 1) == ps_proc->i4_ctb_x) |
| 1222 | { |
| 1223 | WORD32 pad_ht_luma; |
| 1224 | WORD32 pad_ht_chroma; |
| 1225 | WORD32 cols_remaining = ps_sps->i2_pic_width_in_luma_samples - (ps_proc->i4_ctb_x << ps_sps->i1_log2_ctb_size); |
| 1226 | |
| 1227 | ps_proc->pu1_cur_ctb_luma = ps_proc->pu1_cur_pic_luma |
| 1228 | + (ps_proc->i4_ctb_x * ctb_size |
| 1229 | + ps_proc->i4_ctb_y * ctb_size |
| 1230 | * ps_codec->i4_strd); |
| 1231 | ps_proc->pu1_cur_ctb_chroma = ps_proc->pu1_cur_pic_chroma |
| 1232 | + ps_proc->i4_ctb_x * ctb_size |
| 1233 | + (ps_proc->i4_ctb_y * ctb_size * ps_codec->i4_strd / 2); |
| 1234 | |
| 1235 | pad_ht_luma = ctb_size; |
| 1236 | pad_ht_chroma = ctb_size / 2; |
| 1237 | if((ps_sps->i2_pic_ht_in_ctb - 1) == ps_proc->i4_ctb_y) |
| 1238 | { |
| 1239 | pad_ht_luma += 8; |
| 1240 | pad_ht_chroma += 16; |
| 1241 | ps_codec->s_func_selector.ihevc_pad_left_chroma_fptr(ps_proc->pu1_cur_pic_chroma + (ps_sps->i2_pic_height_in_luma_samples / 2 - 16) * ps_codec->i4_strd, |
| 1242 | ps_codec->i4_strd, 16, PAD_LEFT); |
| 1243 | } |
| 1244 | /* Pad right after last CTB in the current row is processed */ |
| 1245 | ps_codec->s_func_selector.ihevc_pad_right_luma_fptr(ps_proc->pu1_cur_ctb_luma + cols_remaining - 8 * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_luma, PAD_RIGHT); |
| 1246 | ps_codec->s_func_selector.ihevc_pad_right_chroma_fptr(ps_proc->pu1_cur_ctb_chroma + cols_remaining - 16 * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_chroma, PAD_RIGHT); |
| 1247 | |
| 1248 | if((ps_sps->i2_pic_ht_in_ctb - 1) == ps_proc->i4_ctb_y) |
| 1249 | { |
| 1250 | UWORD8 *pu1_buf; |
| 1251 | /* Since SAO is shifted by 8x8, chroma padding can not be done till second row is processed */ |
| 1252 | /* Hence moving top padding to to end of frame, Moving it to second row also results in problems when there is only one row */ |
| 1253 | /* Pad top after padding left and right for current rows after processing 1st CTB row */ |
| 1254 | ihevc_pad_top(ps_proc->pu1_cur_pic_luma - PAD_LEFT, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_TOP); |
| 1255 | ihevc_pad_top(ps_proc->pu1_cur_pic_chroma - PAD_LEFT, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_TOP / 2); |
| 1256 | |
| 1257 | pu1_buf = ps_proc->pu1_cur_pic_luma + ps_codec->i4_strd * ps_sps->i2_pic_height_in_luma_samples - PAD_LEFT; |
| 1258 | /* Pad top after padding left and right for current rows after processing 1st CTB row */ |
| 1259 | ihevc_pad_bottom(pu1_buf, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_BOT); |
| 1260 | |
| 1261 | pu1_buf = ps_proc->pu1_cur_pic_chroma + ps_codec->i4_strd * (ps_sps->i2_pic_height_in_luma_samples / 2) - PAD_LEFT; |
| 1262 | ihevc_pad_bottom(pu1_buf, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_BOT / 2); |
| 1263 | } |
| 1264 | } |
| 1265 | #else |
| 1266 | if(ps_proc->i4_ctb_y > 1) |
| 1267 | { |
| 1268 | if(0 == ps_proc->i4_ctb_x) |
| 1269 | { |
| 1270 | WORD32 pad_ht_luma; |
| 1271 | WORD32 pad_ht_chroma; |
| 1272 | |
| 1273 | pad_ht_luma = ctb_size; |
| 1274 | pad_ht_chroma = ctb_size / 2; |
| 1275 | /* Pad left after 1st CTB is processed */ |
| 1276 | ps_codec->s_func_selector.ihevc_pad_left_luma_fptr(ps_proc->pu1_cur_ctb_luma - 2 * ctb_size * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_luma, PAD_LEFT); |
| 1277 | ps_codec->s_func_selector.ihevc_pad_left_chroma_fptr(ps_proc->pu1_cur_ctb_chroma - ctb_size * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_chroma, PAD_LEFT); |
| 1278 | } |
| 1279 | else if((ps_sps->i2_pic_wd_in_ctb - 1) == ps_proc->i4_ctb_x) |
| 1280 | { |
| 1281 | WORD32 pad_ht_luma; |
| 1282 | WORD32 pad_ht_chroma; |
| 1283 | WORD32 cols_remaining = ps_sps->i2_pic_width_in_luma_samples - (ps_proc->i4_ctb_x << ps_sps->i1_log2_ctb_size); |
| 1284 | |
| 1285 | pad_ht_luma = ((ps_sps->i2_pic_ht_in_ctb - 1) == ps_proc->i4_ctb_y) ? 3 * ctb_size : ctb_size; |
| 1286 | pad_ht_chroma = ((ps_sps->i2_pic_ht_in_ctb - 1) == ps_proc->i4_ctb_y) ? 3 * ctb_size / 2 : ctb_size / 2; |
| 1287 | /* Pad right after last CTB in the current row is processed */ |
| 1288 | ps_codec->s_func_selector.ihevc_pad_right_luma_fptr(ps_proc->pu1_cur_ctb_luma + cols_remaining - 2 * ctb_size * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_luma, PAD_RIGHT); |
| 1289 | ps_codec->s_func_selector.ihevc_pad_right_chroma_fptr(ps_proc->pu1_cur_ctb_chroma + cols_remaining - ctb_size * ps_codec->i4_strd, ps_codec->i4_strd, pad_ht_chroma, PAD_RIGHT); |
| 1290 | |
| 1291 | if((ps_sps->i2_pic_ht_in_ctb - 1) == ps_proc->i4_ctb_y) |
| 1292 | { |
| 1293 | UWORD8 *pu1_buf; |
| 1294 | WORD32 pad_ht_luma; |
| 1295 | WORD32 pad_ht_chroma; |
| 1296 | |
| 1297 | pad_ht_luma = 2 * ctb_size; |
| 1298 | pad_ht_chroma = ctb_size; |
| 1299 | |
| 1300 | ps_codec->s_func_selector.ihevc_pad_left_luma_fptr(ps_proc->pu1_cur_pic_luma + ps_codec->i4_strd * (ps_sps->i2_pic_height_in_luma_samples - 2 * ctb_size), |
| 1301 | ps_codec->i4_strd, pad_ht_luma, PAD_LEFT); |
| 1302 | ps_codec->s_func_selector.ihevc_pad_left_chroma_fptr(ps_proc->pu1_cur_pic_chroma + ps_codec->i4_strd * (ps_sps->i2_pic_height_in_luma_samples / 2 - ctb_size), |
| 1303 | ps_codec->i4_strd, pad_ht_chroma, PAD_LEFT); |
| 1304 | |
| 1305 | /* Since SAO is shifted by 8x8, chroma padding can not be done till second row is processed */ |
| 1306 | /* Hence moving top padding to to end of frame, Moving it to second row also results in problems when there is only one row */ |
| 1307 | /* Pad top after padding left and right for current rows after processing 1st CTB row */ |
| 1308 | ihevc_pad_top(ps_proc->pu1_cur_pic_luma - PAD_LEFT, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_TOP); |
| 1309 | ihevc_pad_top(ps_proc->pu1_cur_pic_chroma - PAD_LEFT, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_TOP / 2); |
| 1310 | |
| 1311 | pu1_buf = ps_proc->pu1_cur_pic_luma + ps_codec->i4_strd * ps_sps->i2_pic_height_in_luma_samples - PAD_LEFT; |
| 1312 | /* Pad top after padding left and right for current rows after processing 1st CTB row */ |
| 1313 | ihevc_pad_bottom(pu1_buf, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_BOT); |
| 1314 | |
| 1315 | pu1_buf = ps_proc->pu1_cur_pic_chroma + ps_codec->i4_strd * (ps_sps->i2_pic_height_in_luma_samples / 2) - PAD_LEFT; |
| 1316 | ihevc_pad_bottom(pu1_buf, ps_codec->i4_strd, ps_sps->i2_pic_width_in_luma_samples + PAD_WD, PAD_BOT / 2); |
| 1317 | } |
| 1318 | } |
| 1319 | } |
| 1320 | #endif |
| 1321 | } |
| 1322 | } |
| 1323 | |
| 1324 | |
| 1325 | /* Update proc map for SAO*/ |
| 1326 | ihevcd_proc_map_update(ps_proc, proc_type, nctb); |
| 1327 | /* Update proc map for Completion of CTB*/ |
| 1328 | ihevcd_proc_map_update(ps_proc, PROC_ALL, nctb); |
| 1329 | { |
| 1330 | tile_t *ps_tile; |
| 1331 | |
| 1332 | ps_tile = ps_proc->ps_tile; |
| 1333 | num_ctb_tmp -= nctb; |
| 1334 | |
| 1335 | ps_proc->i4_ctb_tile_x += nctb; |
| 1336 | ps_proc->i4_ctb_x += nctb; |
| 1337 | |
| 1338 | ps_proc->i4_ctb_slice_x += nctb; |
| 1339 | |
| 1340 | |
| 1341 | /* Update tile counters */ |
| 1342 | if(ps_proc->i4_ctb_tile_x >= (ps_tile->u2_wd)) |
| 1343 | { |
| 1344 | /*End of tile row*/ |
| 1345 | ps_proc->i4_ctb_tile_x = 0; |
| 1346 | ps_proc->i4_ctb_x = ps_tile->u1_pos_x; |
| 1347 | |
| 1348 | ps_proc->i4_ctb_tile_y++; |
| 1349 | ps_proc->i4_ctb_y++; |
| 1350 | if(ps_proc->i4_ctb_tile_y == ps_tile->u2_ht) |
| 1351 | { |
| 1352 | /* Reached End of Tile */ |
| 1353 | ps_proc->i4_ctb_tile_y = 0; |
| 1354 | ps_proc->i4_ctb_tile_x = 0; |
| 1355 | ps_proc->ps_tile++; |
| 1356 | //End of picture |
| 1357 | if(!((ps_tile->u2_ht + ps_tile->u1_pos_y == ps_sps->i2_pic_ht_in_ctb) && (ps_tile->u2_wd + ps_tile->u1_pos_x == ps_sps->i2_pic_wd_in_ctb))) |
| 1358 | { |
| 1359 | ps_tile = ps_proc->ps_tile; |
| 1360 | ps_proc->i4_ctb_x = ps_tile->u1_pos_x; |
| 1361 | ps_proc->i4_ctb_y = ps_tile->u1_pos_y; |
| 1362 | |
| 1363 | } |
| 1364 | } |
| 1365 | } |
| 1366 | } |
| 1367 | } |
| 1368 | |
| 1369 | ps_proc->i4_ctb_cnt -= num_ctb; |
| 1370 | } |
| 1371 | return ret; |
| 1372 | } |
| 1373 | |
| 1374 | void ihevcd_init_proc_ctxt(process_ctxt_t *ps_proc, WORD32 tu_coeff_data_ofst) |
| 1375 | { |
| 1376 | codec_t *ps_codec; |
| 1377 | slice_header_t *ps_slice_hdr; |
| 1378 | pps_t *ps_pps; |
| 1379 | sps_t *ps_sps; |
| 1380 | tile_t *ps_tile, *ps_tile_prev; |
| 1381 | WORD32 tile_idx; |
| 1382 | WORD32 ctb_size; |
| 1383 | WORD32 num_minpu_in_ctb; |
| 1384 | WORD32 num_ctb_in_row; |
| 1385 | WORD32 ctb_addr; |
| 1386 | WORD32 i4_wd_in_ctb; |
| 1387 | WORD32 tile_start_ctb_idx; |
| 1388 | WORD32 slice_start_ctb_idx; |
| 1389 | WORD32 check_tile_wd; |
| 1390 | WORD32 continuous_tiles = 0; //Refers to tiles that are continuous, within a slice, horizontally |
| 1391 | |
| 1392 | ps_codec = ps_proc->ps_codec; |
| 1393 | |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1394 | ps_slice_hdr = ps_codec->ps_slice_hdr_base + ((ps_proc->i4_cur_slice_idx) & (MAX_SLICE_HDR_CNT - 1)); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1395 | ps_proc->ps_slice_hdr = ps_slice_hdr; |
| 1396 | ps_proc->ps_pps = ps_codec->ps_pps_base + ps_slice_hdr->i1_pps_id; |
| 1397 | ps_pps = ps_proc->ps_pps; |
| 1398 | ps_proc->ps_sps = ps_codec->ps_sps_base + ps_pps->i1_sps_id; |
| 1399 | ps_sps = ps_proc->ps_sps; |
| 1400 | ps_proc->i4_init_done = 1; |
| 1401 | ctb_size = 1 << ps_sps->i1_log2_ctb_size; |
| 1402 | num_minpu_in_ctb = (ctb_size / MIN_PU_SIZE) * (ctb_size / MIN_PU_SIZE); |
| 1403 | num_ctb_in_row = ps_sps->i2_pic_wd_in_ctb; |
| 1404 | |
| 1405 | ps_proc->s_sao_ctxt.pu1_slice_idx = ps_proc->pu1_slice_idx; |
| 1406 | |
| 1407 | ihevcd_get_tile_pos(ps_pps, ps_sps, ps_proc->i4_ctb_x, ps_proc->i4_ctb_y, |
| 1408 | &ps_proc->i4_ctb_tile_x, &ps_proc->i4_ctb_tile_y, |
| 1409 | &tile_idx); |
| 1410 | |
| 1411 | ps_proc->ps_tile = ps_pps->ps_tile + tile_idx; |
| 1412 | ps_proc->i4_cur_tile_idx = tile_idx; |
| 1413 | ps_tile = ps_proc->ps_tile; |
| 1414 | |
| 1415 | if(ps_pps->i1_tiles_enabled_flag) |
| 1416 | { |
| 1417 | if(tile_idx) |
| 1418 | ps_tile_prev = ps_tile - 1; |
| 1419 | else |
| 1420 | ps_tile_prev = ps_tile; |
| 1421 | |
| 1422 | slice_start_ctb_idx = ps_slice_hdr->i2_ctb_x + (ps_slice_hdr->i2_ctb_y * ps_sps->i2_pic_wd_in_ctb); |
| 1423 | tile_start_ctb_idx = ps_tile->u1_pos_x + (ps_tile->u1_pos_y * ps_sps->i2_pic_wd_in_ctb); |
| 1424 | |
| 1425 | /*Check if |
| 1426 | * 1. Last tile that ends in frame boundary and 1st tile in next row belongs to same slice |
| 1427 | * 1.1. If it does, check if the slice that has these tiles spans across the frame row. |
| 1428 | * 2. Vertical tiles are present within a slice */ |
| 1429 | if(((ps_slice_hdr->i2_ctb_x == ps_tile->u1_pos_x) && (ps_slice_hdr->i2_ctb_y != ps_tile->u1_pos_y))) |
| 1430 | { |
| 1431 | continuous_tiles = 1; |
| 1432 | } |
| 1433 | else |
| 1434 | { |
| 1435 | check_tile_wd = ps_slice_hdr->i2_ctb_x + ps_tile_prev->u2_wd; |
| 1436 | if(!(((check_tile_wd >= ps_sps->i2_pic_wd_in_ctb) && (check_tile_wd % ps_sps->i2_pic_wd_in_ctb == ps_tile->u1_pos_x)) |
| 1437 | || ((ps_slice_hdr->i2_ctb_x == ps_tile->u1_pos_x)))) |
| 1438 | { |
| 1439 | continuous_tiles = 1; |
| 1440 | } |
| 1441 | } |
| 1442 | |
| 1443 | { |
| 1444 | WORD32 i2_independent_ctb_x = ps_slice_hdr->i2_independent_ctb_x; |
| 1445 | WORD32 i2_independent_ctb_y = ps_slice_hdr->i2_independent_ctb_y; |
| 1446 | |
| 1447 | /* Handles cases where |
| 1448 | * 1. Slices begin at the start of each tile |
| 1449 | * 2. Tiles lie in the same slice row.i.e, starting tile_x > slice_x, but tile_y == slice_y |
| 1450 | * */ |
| 1451 | if(ps_proc->i4_ctb_x >= i2_independent_ctb_x) |
| 1452 | { |
| 1453 | ps_proc->i4_ctb_slice_x = ps_proc->i4_ctb_x - i2_independent_ctb_x; |
| 1454 | } |
| 1455 | else |
| 1456 | { |
| 1457 | /* Indicates multiple tiles in a slice case where |
| 1458 | * The new tile belongs to an older slice that started in the previous rows-not the present row |
| 1459 | * & (tile_y > slice_y and tile_x < slice_x) |
| 1460 | */ |
| 1461 | if((slice_start_ctb_idx < tile_start_ctb_idx) && (continuous_tiles)) |
| 1462 | { |
| 1463 | i4_wd_in_ctb = ps_sps->i2_pic_wd_in_ctb; |
| 1464 | } |
| 1465 | /* Indicates many-tiles-in-one-slice case, for slices that end without spanning the frame width*/ |
| 1466 | else |
| 1467 | { |
| 1468 | i4_wd_in_ctb = ps_tile->u2_wd; |
| 1469 | } |
| 1470 | |
| 1471 | if(continuous_tiles) |
| 1472 | { |
| 1473 | ps_proc->i4_ctb_slice_x = i4_wd_in_ctb |
| 1474 | - (i2_independent_ctb_x - ps_proc->i4_ctb_x); |
| 1475 | } |
| 1476 | else |
| 1477 | { |
| 1478 | ps_proc->i4_ctb_slice_x = ps_proc->i4_ctb_x - ps_tile->u1_pos_x; |
| 1479 | } |
| 1480 | } |
| 1481 | /* Initialize ctb slice y to zero and at the start of slice row initialize it |
| 1482 | to difference between ctb_y and slice's start ctb y */ |
| 1483 | |
| 1484 | ps_proc->i4_ctb_slice_y = ps_proc->i4_ctb_y - i2_independent_ctb_y; |
| 1485 | |
| 1486 | /*If beginning of tile, check if slice counters are set correctly*/ |
| 1487 | if((0 == ps_proc->i4_ctb_tile_x) && (0 == ps_proc->i4_ctb_tile_y)) |
| 1488 | { |
| 1489 | if(ps_slice_hdr->i1_dependent_slice_flag) |
| 1490 | { |
| 1491 | ps_proc->i4_ctb_slice_x = 0; |
| 1492 | ps_proc->i4_ctb_slice_y = 0; |
| 1493 | } |
| 1494 | /*For slices that span across multiple tiles*/ |
| 1495 | else if(slice_start_ctb_idx < tile_start_ctb_idx) |
| 1496 | { |
| 1497 | ps_proc->i4_ctb_slice_y = ps_tile->u1_pos_y - i2_independent_ctb_y; |
| 1498 | /* Two Cases |
| 1499 | * 1 - slice spans across frame-width- but dose not start from 1st column |
| 1500 | * 2 - Slice spans across multiple tiles anywhere is a frame |
| 1501 | */ |
| 1502 | /*TODO:In a multiple slice clip, if an independent slice span across more than 2 tiles in a row, it is not supported*/ |
| 1503 | if(continuous_tiles) //Case 2-implemented for slices that span not more than 2 tiles |
| 1504 | { |
| 1505 | if(i2_independent_ctb_y <= ps_tile->u1_pos_y) |
| 1506 | { |
| 1507 | //Check if ctb x is before or after |
| 1508 | if(i2_independent_ctb_x > ps_tile->u1_pos_x) |
| 1509 | { |
| 1510 | ps_proc->i4_ctb_slice_y -= 1; |
| 1511 | } |
| 1512 | } |
| 1513 | } |
| 1514 | } |
| 1515 | } |
| 1516 | //Slice starts from a column which is not the starting tile-column, but is within the tile |
| 1517 | if(((i2_independent_ctb_x - ps_tile->u1_pos_x) != 0) && ((ps_proc->i4_ctb_slice_y != 0)) |
| 1518 | && ((i2_independent_ctb_x >= ps_tile->u1_pos_x) && (i2_independent_ctb_x < ps_tile->u1_pos_x + ps_tile->u2_wd))) |
| 1519 | { |
| 1520 | ps_proc->i4_ctb_slice_y -= 1; |
| 1521 | } |
| 1522 | } |
| 1523 | } |
| 1524 | else |
| 1525 | { |
| 1526 | WORD32 i2_independent_ctb_x = ps_slice_hdr->i2_independent_ctb_x; |
| 1527 | WORD32 i2_independent_ctb_y = ps_slice_hdr->i2_independent_ctb_y; |
| 1528 | |
| 1529 | |
| 1530 | { |
| 1531 | ps_proc->i4_ctb_slice_x = ps_proc->i4_ctb_x - i2_independent_ctb_x; |
| 1532 | ps_proc->i4_ctb_slice_y = ps_proc->i4_ctb_y - i2_independent_ctb_y; |
| 1533 | if(ps_proc->i4_ctb_slice_x < 0) |
| 1534 | { |
| 1535 | ps_proc->i4_ctb_slice_x += ps_sps->i2_pic_wd_in_ctb; |
| 1536 | ps_proc->i4_ctb_slice_y -= 1; |
| 1537 | } |
| 1538 | |
| 1539 | /* Initialize ctb slice y to zero and at the start of slice row initialize it |
| 1540 | to difference between ctb_y and slice's start ctb y */ |
| 1541 | } |
| 1542 | } |
| 1543 | |
| 1544 | /* Compute TU offset for the current CTB set */ |
| 1545 | { |
| 1546 | |
| 1547 | WORD32 ctb_luma_min_tu_cnt; |
| 1548 | WORD32 ctb_addr; |
| 1549 | |
| 1550 | ctb_addr = ps_proc->i4_ctb_y * num_ctb_in_row + ps_proc->i4_ctb_x; |
| 1551 | |
| 1552 | ctb_luma_min_tu_cnt = (1 << ps_sps->i1_log2_ctb_size) / MIN_TU_SIZE; |
| 1553 | ctb_luma_min_tu_cnt *= ctb_luma_min_tu_cnt; |
| 1554 | |
| 1555 | ps_proc->pu1_tu_map = ps_proc->pu1_pic_tu_map |
| 1556 | + ctb_luma_min_tu_cnt * ctb_addr; |
| 1557 | if(1 == ps_codec->i4_num_cores) |
| 1558 | { |
| 1559 | ps_proc->ps_tu = ps_proc->ps_pic_tu + ps_proc->pu4_pic_tu_idx[ctb_addr % RESET_TU_BUF_NCTB]; |
| 1560 | } |
| 1561 | else |
| 1562 | { |
| 1563 | ps_proc->ps_tu = ps_proc->ps_pic_tu + ps_proc->pu4_pic_tu_idx[ctb_addr]; |
| 1564 | } |
| 1565 | ps_proc->pv_tu_coeff_data = (UWORD8 *)ps_proc->pv_pic_tu_coeff_data |
| 1566 | + tu_coeff_data_ofst; |
| 1567 | |
| 1568 | } |
| 1569 | |
| 1570 | /* Compute PU related elements for the current CTB set */ |
| 1571 | { |
| 1572 | WORD32 pu_idx; |
| 1573 | ctb_addr = ps_proc->i4_ctb_y * num_ctb_in_row + ps_proc->i4_ctb_x; |
| 1574 | pu_idx = ps_proc->pu4_pic_pu_idx[ctb_addr]; |
| 1575 | ps_proc->pu1_pu_map = ps_proc->pu1_pic_pu_map |
| 1576 | + ctb_addr * num_minpu_in_ctb; |
| 1577 | ps_proc->ps_pu = ps_proc->ps_pic_pu + pu_idx; |
| 1578 | } |
| 1579 | |
| 1580 | /* Number of ctbs processed in one loop of process function */ |
| 1581 | { |
| 1582 | ps_proc->i4_nctb = MIN(ps_codec->u4_nctb, ps_tile->u2_wd); |
| 1583 | } |
| 1584 | |
| 1585 | } |
| 1586 | void ihevcd_process_thread(process_ctxt_t *ps_proc) |
| 1587 | { |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1588 | { |
| 1589 | ithread_set_affinity(ps_proc->i4_id + 1); |
| 1590 | } |
| 1591 | while(1) |
| 1592 | { |
| 1593 | IHEVCD_ERROR_T ret; |
| 1594 | proc_job_t s_job; |
| 1595 | |
| 1596 | ret = ihevcd_jobq_dequeue((jobq_t *)ps_proc->pv_proc_jobq, &s_job, |
| 1597 | sizeof(proc_job_t), 1); |
| 1598 | if((IHEVCD_ERROR_T)IHEVCD_SUCCESS != ret) |
| 1599 | break; |
| 1600 | |
| 1601 | ps_proc->i4_ctb_cnt = s_job.i2_ctb_cnt; |
| 1602 | ps_proc->i4_ctb_x = s_job.i2_ctb_x; |
| 1603 | ps_proc->i4_ctb_y = s_job.i2_ctb_y; |
| 1604 | ps_proc->i4_cur_slice_idx = s_job.i2_slice_idx; |
| 1605 | |
| 1606 | |
| 1607 | |
| 1608 | if(CMD_PROCESS == s_job.i4_cmd) |
| 1609 | { |
| 1610 | ihevcd_init_proc_ctxt(ps_proc, s_job.i4_tu_coeff_data_ofst); |
Harish Mahendrakar | 0d8951c | 2014-05-16 10:31:13 -0700 | [diff] [blame] | 1611 | ihevcd_process(ps_proc); |
| 1612 | } |
| 1613 | else if(CMD_FMTCONV == s_job.i4_cmd) |
| 1614 | { |
| 1615 | sps_t *ps_sps; |
| 1616 | codec_t *ps_codec; |
| 1617 | ivd_out_bufdesc_t *ps_out_buffer; |
| 1618 | WORD32 num_rows; |
| 1619 | |
| 1620 | if(0 == ps_proc->i4_init_done) |
| 1621 | { |
| 1622 | ihevcd_init_proc_ctxt(ps_proc, 0); |
| 1623 | } |
| 1624 | ps_sps = ps_proc->ps_sps; |
| 1625 | ps_codec = ps_proc->ps_codec; |
| 1626 | ps_out_buffer = ps_proc->ps_out_buffer; |
| 1627 | num_rows = 1 << ps_sps->i1_log2_ctb_size; |
| 1628 | |
| 1629 | num_rows = MIN(num_rows, (ps_codec->i4_disp_ht - (s_job.i2_ctb_y << ps_sps->i1_log2_ctb_size))); |
| 1630 | |
| 1631 | if(num_rows < 0) |
| 1632 | num_rows = 0; |
| 1633 | |
| 1634 | ihevcd_fmt_conv(ps_proc->ps_codec, ps_proc, ps_out_buffer->pu1_bufs[0], ps_out_buffer->pu1_bufs[1], ps_out_buffer->pu1_bufs[2], |
| 1635 | s_job.i2_ctb_y << ps_sps->i1_log2_ctb_size, num_rows); |
| 1636 | } |
| 1637 | } |
| 1638 | //ithread_exit(0); |
| 1639 | return; |
| 1640 | } |
| 1641 | |