/src/libhevc/decoder/ihevcd_utils.c
Line | Count | Source |
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" |
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 | 0 | { |
104 | 0 | WORD32 lvl_idx = 0; |
105 | |
|
106 | 0 | if(level < IHEVC_LEVEL_20) |
107 | 0 | { |
108 | 0 | lvl_idx = 0; |
109 | 0 | } |
110 | 0 | else if(level >= IHEVC_LEVEL_20 && level < IHEVC_LEVEL_21) |
111 | 0 | { |
112 | 0 | lvl_idx = 1; |
113 | 0 | } |
114 | 0 | else if(level >= IHEVC_LEVEL_21 && level < IHEVC_LEVEL_30) |
115 | 0 | { |
116 | 0 | lvl_idx = 2; |
117 | 0 | } |
118 | 0 | else if(level >= IHEVC_LEVEL_30 && level < IHEVC_LEVEL_31) |
119 | 0 | { |
120 | 0 | lvl_idx = 3; |
121 | 0 | } |
122 | 0 | else if(level >= IHEVC_LEVEL_31 && level < IHEVC_LEVEL_40) |
123 | 0 | { |
124 | 0 | lvl_idx = 4; |
125 | 0 | } |
126 | 0 | else if(level >= IHEVC_LEVEL_40 && level < IHEVC_LEVEL_41) |
127 | 0 | { |
128 | 0 | lvl_idx = 5; |
129 | 0 | } |
130 | 0 | else if(level >= IHEVC_LEVEL_41 && level < IHEVC_LEVEL_50) |
131 | 0 | { |
132 | 0 | lvl_idx = 6; |
133 | 0 | } |
134 | 0 | else if(level >= IHEVC_LEVEL_50 && level < IHEVC_LEVEL_51) |
135 | 0 | { |
136 | 0 | lvl_idx = 7; |
137 | 0 | } |
138 | 0 | else if(level >= IHEVC_LEVEL_51 && level < IHEVC_LEVEL_52) |
139 | 0 | { |
140 | 0 | lvl_idx = 8; |
141 | 0 | } |
142 | 0 | else if(level >= IHEVC_LEVEL_52 && level < IHEVC_LEVEL_60) |
143 | 0 | { |
144 | 0 | lvl_idx = 9; |
145 | 0 | } |
146 | 0 | else if(level >= IHEVC_LEVEL_60 && level < IHEVC_LEVEL_61) |
147 | 0 | { |
148 | 0 | lvl_idx = 10; |
149 | 0 | } |
150 | 0 | else if(level >= IHEVC_LEVEL_61 && level < IHEVC_LEVEL_62) |
151 | 0 | { |
152 | 0 | lvl_idx = 11; |
153 | 0 | } |
154 | 0 | else if(level >= IHEVC_LEVEL_62) |
155 | 0 | { |
156 | 0 | lvl_idx = 12; |
157 | 0 | } |
158 | |
|
159 | 0 | return (lvl_idx); |
160 | 0 | } |
161 | | |
162 | | /** |
163 | | ******************************************************************************* |
164 | | * |
165 | | * @brief |
166 | | * Used to get reference picture buffer size for a given level and |
167 | | * and padding used |
168 | | * |
169 | | * @par Description: |
170 | | * Used to get reference picture buffer size for a given level and padding used |
171 | | * Each picture is padded on all four sides |
172 | | * |
173 | | * @param[in] pic_size |
174 | | * Mumber of luma samples (Width * Height) |
175 | | * |
176 | | * @param[in] level |
177 | | * Level |
178 | | * |
179 | | * @param[in] horz_pad |
180 | | * Total padding used in horizontal direction |
181 | | * |
182 | | * @param[in] vert_pad |
183 | | * Total padding used in vertical direction |
184 | | * |
185 | | * @returns Total picture buffer size |
186 | | * |
187 | | * @remarks |
188 | | * |
189 | | * |
190 | | ******************************************************************************* |
191 | | */ |
192 | | WORD32 ihevcd_get_total_pic_buf_size(codec_t *ps_codec, |
193 | | WORD32 wd, |
194 | | WORD32 ht) |
195 | 2.26k | { |
196 | 2.26k | WORD32 size; |
197 | 2.26k | WORD32 num_luma_samples; |
198 | 2.26k | WORD32 max_dpb_size; |
199 | 2.26k | WORD32 num_samples; |
200 | | |
201 | | |
202 | 2.26k | sps_t *ps_sps = (ps_codec->s_parse.ps_sps_base + ps_codec->i4_sps_id); |
203 | | |
204 | | /* Get maximum number of buffers for the current picture size */ |
205 | 2.26k | max_dpb_size = ps_sps->ai1_sps_max_dec_pic_buffering[ps_sps->i1_sps_max_sub_layers - 1]; |
206 | | |
207 | 2.26k | if(ps_codec->e_frm_out_mode != IVD_DECODE_FRAME_OUT) |
208 | 2.26k | max_dpb_size += ps_sps->ai1_sps_max_num_reorder_pics[ps_sps->i1_sps_max_sub_layers - 1]; |
209 | | |
210 | 2.26k | max_dpb_size++; |
211 | | /* Allocation is required for |
212 | | * (Wd + horz_pad) * (Ht + vert_pad) * (2 * max_dpb_size + 1) |
213 | | */ |
214 | | |
215 | | /* Account for padding area */ |
216 | 2.26k | num_luma_samples = (wd + PAD_WD) * (ht + PAD_HT); |
217 | | |
218 | | /* Account for chroma */ |
219 | 2.26k | if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV444) |
220 | 0 | { |
221 | 0 | num_samples = num_luma_samples * 3; |
222 | 0 | } |
223 | 2.26k | else if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV422) |
224 | 0 | { |
225 | 0 | num_samples = num_luma_samples * 2; |
226 | 0 | } |
227 | 2.26k | else if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) |
228 | 2.26k | { |
229 | 2.26k | num_samples = num_luma_samples * 3 / 2; |
230 | 2.26k | } |
231 | 0 | else |
232 | 0 | { |
233 | 0 | num_samples = num_luma_samples; |
234 | 0 | } |
235 | | |
236 | | /* Number of bytes in reference pictures */ |
237 | 2.26k | size = num_samples * max_dpb_size; |
238 | | |
239 | | |
240 | 2.26k | return size; |
241 | 2.26k | } |
242 | | /** |
243 | | ******************************************************************************* |
244 | | * |
245 | | * @brief |
246 | | * Used to get MV bank size for a given number of luma samples |
247 | | * |
248 | | * @par Description: |
249 | | * For given number of luma samples one MV bank size is computed |
250 | | * Each MV bank includes pu_map and pu_t for all the min PUs(4x4) in a picture |
251 | | * |
252 | | * @param[in] num_luma_samples |
253 | | * Max number of luma pixels in the frame |
254 | | * |
255 | | * @returns Total MV Bank size |
256 | | * |
257 | | * @remarks |
258 | | * |
259 | | * |
260 | | ******************************************************************************* |
261 | | */ |
262 | | WORD32 ihevcd_get_pic_mv_bank_size(WORD32 num_luma_samples) |
263 | 4.15k | { |
264 | 4.15k | WORD32 size; |
265 | | |
266 | 4.15k | WORD32 pic_size; |
267 | | |
268 | 4.15k | WORD32 mv_bank_size; |
269 | 4.15k | WORD32 num_pu; |
270 | 4.15k | WORD32 num_ctb; |
271 | 4.15k | pic_size = num_luma_samples; |
272 | | |
273 | | |
274 | 4.15k | num_pu = pic_size / (MIN_PU_SIZE * MIN_PU_SIZE); |
275 | 4.15k | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
276 | | |
277 | 4.15k | mv_bank_size = 0; |
278 | | |
279 | | /* Size for storing pu_t start index each CTB */ |
280 | | /* One extra entry is needed to compute number of PUs in the last CTB */ |
281 | 4.15k | mv_bank_size += (num_ctb + 1) * sizeof(WORD32); |
282 | | |
283 | | /* Size for pu_map */ |
284 | 4.15k | mv_bank_size += num_pu; |
285 | | |
286 | | /* Size for storing pu_t for each PU */ |
287 | 4.15k | mv_bank_size += num_pu * sizeof(pu_t); |
288 | | |
289 | | /* Size for storing slice_idx for each CTB */ |
290 | 4.15k | mv_bank_size += ALIGN4(num_ctb * sizeof(UWORD16)); |
291 | | |
292 | 4.15k | size = mv_bank_size; |
293 | 4.15k | return size; |
294 | 4.15k | } |
295 | | /** |
296 | | ******************************************************************************* |
297 | | * |
298 | | * @brief |
299 | | * Used to get TU data size for a given number luma samples |
300 | | * |
301 | | * @par Description: |
302 | | * For a given number of luma samples TU data size is computed |
303 | | * Each TU data includes tu_map and tu_t and coeff data for all |
304 | | * the min TUs(4x4) in given CTB |
305 | | * |
306 | | * @param[in] num_luma_samples |
307 | | * Number of 64 x 64 CTBs for which TU data has to be allocated. |
308 | | * |
309 | | * @returns Total TU data size |
310 | | * |
311 | | * @remarks Assumption is num_luma_samples will be at least |
312 | | * 64 x 64 to handle CTB of size 64 x 64. Can be frame size as well |
313 | | * |
314 | | ******************************************************************************* |
315 | | */ |
316 | | WORD32 ihevcd_get_tu_data_size(codec_t *ps_codec, WORD32 num_luma_samples) |
317 | 2.26k | { |
318 | | |
319 | 2.26k | sps_t *ps_sps = (ps_codec->s_parse.ps_sps_base + ps_codec->i4_sps_id); |
320 | 2.26k | WORD32 tu_data_size; |
321 | 2.26k | WORD32 num_ctb; |
322 | 2.26k | WORD32 num_luma_tu, num_chroma_tu, num_tu; |
323 | 2.26k | num_ctb = num_luma_samples / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
324 | | |
325 | 2.26k | num_luma_tu = num_luma_samples / (MIN_TU_SIZE * MIN_TU_SIZE); |
326 | | |
327 | 2.26k | if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV444) |
328 | 0 | { |
329 | 0 | num_chroma_tu = num_luma_tu << 1; |
330 | 0 | } |
331 | 2.26k | else if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV422) |
332 | 0 | { |
333 | 0 | num_chroma_tu = num_luma_tu; |
334 | 0 | } |
335 | 2.26k | else if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) |
336 | 2.26k | { |
337 | 2.26k | num_chroma_tu = num_luma_tu >> 1; |
338 | 2.26k | } |
339 | 0 | else |
340 | 0 | { |
341 | 0 | num_chroma_tu = 0; |
342 | 0 | } |
343 | | |
344 | 2.26k | num_tu = num_luma_tu + num_chroma_tu; |
345 | 2.26k | tu_data_size = 0; |
346 | | |
347 | | /* Size for storing tu_t start index each CTB */ |
348 | | /* One extra entry is needed to compute number of TUs in the last CTB */ |
349 | 2.26k | tu_data_size += (num_ctb + 1) * sizeof(WORD32); |
350 | | |
351 | | /* Size for storing tu map */ |
352 | 2.26k | tu_data_size += num_luma_tu * sizeof(UWORD8); |
353 | | |
354 | | /* Size for storing tu_t for each TU */ |
355 | 2.26k | tu_data_size += num_tu * sizeof(tu_t); |
356 | | |
357 | | /* Size for storing number of coded subblocks and scan_idx for each TU */ |
358 | 2.26k | tu_data_size += num_tu * (sizeof(WORD8) + sizeof(WORD8)); |
359 | | |
360 | | /* Size for storing coeff data for each TU */ |
361 | 2.26k | tu_data_size += num_tu * sizeof(tu_sblk_coeff_data_t); |
362 | | |
363 | | |
364 | 2.26k | return tu_data_size; |
365 | 2.26k | } |
366 | | |
367 | | |
368 | | WORD32 ihevcd_nctb_cnt(codec_t *ps_codec, sps_t *ps_sps) |
369 | 10.9k | { |
370 | 10.9k | WORD32 nctb = 1; |
371 | 10.9k | UNUSED(ps_codec); |
372 | | //TODO: Currently set to 1 |
373 | | /* If CTB size is less than 32 x 32 then set nCTB as 4 */ |
374 | 10.9k | if(ps_sps->i1_log2_ctb_size < 5) |
375 | 506 | nctb = 1; |
376 | | |
377 | 10.9k | return nctb; |
378 | 10.9k | } |
379 | | |
380 | | IHEVCD_ERROR_T ihevcd_get_tile_pos(pps_t *ps_pps, |
381 | | sps_t *ps_sps, |
382 | | WORD32 ctb_x, |
383 | | WORD32 ctb_y, |
384 | | WORD32 *pi4_ctb_tile_x, |
385 | | WORD32 *pi4_ctb_tile_y, |
386 | | WORD32 *pi4_tile_idx) |
387 | 269k | { |
388 | | |
389 | 269k | tile_t *ps_tile_tmp; |
390 | 269k | WORD32 i; |
391 | 269k | WORD32 tile_row, tile_col; |
392 | | |
393 | 269k | if(ctb_x < 0 || ctb_y < 0) |
394 | 0 | { |
395 | 0 | *pi4_ctb_tile_x = 0; |
396 | 0 | *pi4_ctb_tile_y = 0; |
397 | 0 | *pi4_tile_idx = 0; |
398 | |
|
399 | 0 | return (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
400 | 0 | } |
401 | | |
402 | 269k | tile_row = 0; |
403 | 269k | tile_col = 0; |
404 | 269k | ps_tile_tmp = ps_pps->ps_tile; |
405 | 269k | if(0 == ps_pps->i1_tiles_enabled_flag) |
406 | 151k | { |
407 | 151k | *pi4_ctb_tile_x = ctb_x; |
408 | 151k | *pi4_ctb_tile_y = ctb_y; |
409 | 151k | *pi4_tile_idx = 0; |
410 | 151k | } |
411 | 117k | else |
412 | 117k | { |
413 | 230k | for(i = 0; i < ps_pps->i1_num_tile_columns; i++) |
414 | 230k | { |
415 | 230k | WORD16 next_tile_ctb_x; |
416 | 230k | ps_tile_tmp = ps_pps->ps_tile + i; //* ps_pps->i1_num_tile_rows; |
417 | 230k | if((ps_pps->i1_num_tile_columns - 1) == i) |
418 | 44.8k | { |
419 | 44.8k | next_tile_ctb_x = ps_sps->i2_pic_wd_in_ctb; |
420 | 44.8k | } |
421 | 185k | else |
422 | 185k | { |
423 | 185k | tile_t *ps_tile_next_tmp; |
424 | 185k | ps_tile_next_tmp = ps_pps->ps_tile + i + 1; |
425 | 185k | next_tile_ctb_x = ps_tile_next_tmp->u1_pos_x; |
426 | 185k | } |
427 | 230k | if((ctb_x >= ps_tile_tmp->u1_pos_x) && (ctb_x < next_tile_ctb_x)) |
428 | 117k | { |
429 | 117k | tile_col = i; |
430 | 117k | break; |
431 | 117k | } |
432 | 230k | } |
433 | 117k | *pi4_ctb_tile_x = ctb_x - ps_tile_tmp->u1_pos_x; |
434 | | |
435 | 229k | for(i = 0; i < ps_pps->i1_num_tile_rows; i++) |
436 | 229k | { |
437 | 229k | WORD16 next_tile_ctb_y; |
438 | 229k | ps_tile_tmp = ps_pps->ps_tile + i * ps_pps->i1_num_tile_columns; |
439 | 229k | if((ps_pps->i1_num_tile_rows - 1) == i) |
440 | 47.8k | { |
441 | 47.8k | next_tile_ctb_y = ps_sps->i2_pic_ht_in_ctb; |
442 | 47.8k | } |
443 | 181k | else |
444 | 181k | { |
445 | 181k | tile_t *ps_tile_next_tmp; |
446 | 181k | ps_tile_next_tmp = ps_pps->ps_tile + ((i + 1) * ps_pps->i1_num_tile_columns); |
447 | 181k | next_tile_ctb_y = ps_tile_next_tmp->u1_pos_y; |
448 | 181k | } |
449 | 229k | if((ctb_y >= ps_tile_tmp->u1_pos_y) && (ctb_y < next_tile_ctb_y)) |
450 | 117k | { |
451 | 117k | tile_row = i; |
452 | 117k | break; |
453 | 117k | } |
454 | | |
455 | 229k | } |
456 | 117k | *pi4_ctb_tile_y = ctb_y - ps_tile_tmp->u1_pos_y; |
457 | 117k | *pi4_tile_idx = tile_row * ps_pps->i1_num_tile_columns |
458 | 117k | + tile_col; |
459 | 117k | } |
460 | 269k | return (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
461 | 269k | } |
462 | | /** |
463 | | ******************************************************************************* |
464 | | * |
465 | | * @brief |
466 | | * Function to initialize ps_pic_buf structs add pic buffers to |
467 | | * buffer manager in case of non-shared mode |
468 | | * |
469 | | * @par Description: |
470 | | * Function to initialize ps_pic_buf structs add pic buffers to |
471 | | * buffer manager in case of non-shared mode |
472 | | * To be called once per stream or for every reset |
473 | | * |
474 | | * @param[in] ps_codec |
475 | | * Pointer to codec context |
476 | | * |
477 | | * @returns Error from IHEVCD_ERROR_T |
478 | | * |
479 | | * @remarks |
480 | | * |
481 | | * |
482 | | ******************************************************************************* |
483 | | */ |
484 | | IHEVCD_ERROR_T ihevcd_pic_buf_mgr_add_bufs(codec_t *ps_codec) |
485 | 1.89k | { |
486 | 1.89k | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
487 | 1.89k | WORD32 i; |
488 | 1.89k | WORD32 max_dpb_size; |
489 | 1.89k | sps_t *ps_sps; |
490 | 1.89k | UWORD8 *pu1_buf; |
491 | 1.89k | pic_buf_t *ps_pic_buf; |
492 | 1.89k | WORD32 pic_buf_size_allocated; |
493 | 1.89k | WORD32 h_samp_factor, v_samp_factor; |
494 | 1.89k | WORD32 chroma_pixel_strd = 2; |
495 | | |
496 | | |
497 | | /* Initialize Pic buffer manager */ |
498 | 1.89k | ps_sps = ps_codec->s_parse.ps_sps; |
499 | | |
500 | 1.89k | h_samp_factor = (CHROMA_FMT_IDC_YUV444 == ps_sps->i1_chroma_format_idc) ? 1 : 2; |
501 | 1.89k | v_samp_factor = (CHROMA_FMT_IDC_YUV420 == ps_sps->i1_chroma_format_idc) ? 2 : 1; |
502 | | |
503 | | /* Compute the number of Pic buffers needed */ |
504 | 1.89k | max_dpb_size = ps_sps->ai1_sps_max_dec_pic_buffering[ps_sps->i1_sps_max_sub_layers - 1]; |
505 | | |
506 | 1.89k | if(ps_codec->e_frm_out_mode != IVD_DECODE_FRAME_OUT) |
507 | 1.89k | max_dpb_size += ps_sps->ai1_sps_max_num_reorder_pics[ps_sps->i1_sps_max_sub_layers - 1]; |
508 | | |
509 | | /* Allocate one extra picture to handle current frame |
510 | | * In case of asynchronous parsing and processing, number of buffers should increase here |
511 | | * based on when parsing and processing threads are synchronized |
512 | | */ |
513 | 1.89k | max_dpb_size++; |
514 | | |
515 | | |
516 | 1.89k | pu1_buf = (UWORD8 *)ps_codec->pu1_ref_pic_buf_base; |
517 | | |
518 | 1.89k | ps_pic_buf = (pic_buf_t *)ps_codec->ps_pic_buf; |
519 | | |
520 | | /* In case of non-shared mode, add picture buffers to buffer manager |
521 | | * In case of shared mode buffers are added in the run-time |
522 | | */ |
523 | 1.89k | if(0 == ps_codec->i4_share_disp_buf) |
524 | 1.89k | { |
525 | 1.89k | WORD32 buf_ret; |
526 | 1.89k | WORD32 luma_samples; |
527 | 1.89k | WORD32 chroma_samples; |
528 | 1.89k | pic_buf_size_allocated = ps_codec->i4_total_pic_buf_size; |
529 | | |
530 | 1.89k | luma_samples = (ps_codec->i4_strd) * |
531 | 1.89k | (ps_sps->i2_pic_height_in_luma_samples + PAD_HT); |
532 | | |
533 | 1.89k | if(CHROMA_FMT_IDC_MONOCHROME == ps_sps->i1_chroma_format_idc) |
534 | 0 | { |
535 | 0 | chroma_samples = 0; |
536 | 0 | } |
537 | 1.89k | else |
538 | 1.89k | { |
539 | 1.89k | chroma_samples = luma_samples * 2 / (h_samp_factor * v_samp_factor); |
540 | 1.89k | } |
541 | | |
542 | | /* Try to add as many buffers as possible since memory is already allocated */ |
543 | | /* If the number of buffers that can be added is less than max_num_bufs |
544 | | * return with an error. |
545 | | */ |
546 | 19.1k | for(i = 0; i < max_dpb_size; i++) |
547 | 17.2k | { |
548 | 17.2k | pic_buf_size_allocated -= (luma_samples + chroma_samples); |
549 | | |
550 | 17.2k | if(pic_buf_size_allocated < 0) |
551 | 0 | { |
552 | 0 | ps_codec->s_parse.i4_error_code = IHEVCD_INSUFFICIENT_MEM_PICBUF; |
553 | 0 | return IHEVCD_INSUFFICIENT_MEM_PICBUF; |
554 | 0 | } |
555 | | |
556 | 17.2k | ps_pic_buf->pu1_luma = pu1_buf + ps_codec->i4_strd * PAD_TOP + PAD_LEFT; |
557 | 17.2k | pu1_buf += luma_samples; |
558 | | |
559 | 17.2k | if(chroma_samples) |
560 | 17.2k | { |
561 | 17.2k | ps_pic_buf->pu1_chroma = pu1_buf |
562 | 17.2k | + (ps_codec->i4_strd * chroma_pixel_strd / h_samp_factor) * (PAD_TOP / v_samp_factor) |
563 | 17.2k | + (PAD_LEFT * chroma_pixel_strd / h_samp_factor); |
564 | 17.2k | pu1_buf += chroma_samples; |
565 | 17.2k | } |
566 | 0 | else |
567 | 0 | { |
568 | 0 | ps_pic_buf->pu1_chroma = NULL; |
569 | 0 | } |
570 | | |
571 | | /* Pad boundary pixels (one pixel on all sides) */ |
572 | | /* This ensures SAO does not read uninitialized pixels */ |
573 | | /* Note these are not used in actual processing */ |
574 | 17.2k | { |
575 | 17.2k | UWORD8 *pu1_buf; |
576 | 17.2k | WORD32 strd, wd, ht; |
577 | 17.2k | WORD32 i; |
578 | 17.2k | strd = ps_codec->i4_strd; |
579 | 17.2k | wd = ps_codec->i4_wd; |
580 | 17.2k | ht = ps_codec->i4_ht; |
581 | | |
582 | 17.2k | pu1_buf = ps_pic_buf->pu1_luma; |
583 | 3.48M | for(i = 0; i < ht; i++) |
584 | 3.47M | { |
585 | 3.47M | pu1_buf[-1] = 0; |
586 | 3.47M | pu1_buf[wd] = 0; |
587 | 3.47M | pu1_buf += strd; |
588 | 3.47M | } |
589 | 17.2k | pu1_buf = ps_pic_buf->pu1_luma; |
590 | 17.2k | memset(pu1_buf - strd - 1, 0, wd + 2); |
591 | | |
592 | 17.2k | pu1_buf += strd * ht; |
593 | 17.2k | memset(pu1_buf - 1, 0, wd + 2); |
594 | | |
595 | 17.2k | if(ps_pic_buf->pu1_chroma) |
596 | 17.2k | { |
597 | 17.2k | pu1_buf = ps_pic_buf->pu1_chroma; |
598 | 17.2k | ht /= v_samp_factor; |
599 | 17.2k | WORD32 chroma_strd_scale = chroma_pixel_strd / h_samp_factor; |
600 | 1.75M | for(i = 0; i < ht; i++) |
601 | 1.73M | { |
602 | 1.73M | pu1_buf[-1] = 0; |
603 | 1.73M | pu1_buf[-2] = 0; |
604 | 1.73M | pu1_buf[wd * chroma_strd_scale] = 0; |
605 | 1.73M | pu1_buf[wd * chroma_strd_scale + 1] = 0; |
606 | 1.73M | pu1_buf += (strd * chroma_strd_scale); |
607 | 1.73M | } |
608 | 17.2k | pu1_buf = ps_pic_buf->pu1_chroma; |
609 | 17.2k | memset(pu1_buf - (strd * chroma_strd_scale) - 2, 0, wd * chroma_strd_scale + 4); |
610 | | |
611 | 17.2k | pu1_buf += (strd * chroma_strd_scale) * ht; |
612 | 17.2k | memset(pu1_buf - 2, 0, wd * chroma_strd_scale + 4); |
613 | 17.2k | } |
614 | 17.2k | } |
615 | | |
616 | 17.2k | buf_ret = ihevc_buf_mgr_add((buf_mgr_t *)ps_codec->pv_pic_buf_mgr, ps_pic_buf, i); |
617 | | |
618 | | |
619 | 17.2k | if(0 != buf_ret) |
620 | 0 | { |
621 | 0 | ps_codec->s_parse.i4_error_code = IHEVCD_BUF_MGR_ERROR; |
622 | 0 | return IHEVCD_BUF_MGR_ERROR; |
623 | 0 | } |
624 | 17.2k | ps_pic_buf++; |
625 | 17.2k | } |
626 | 1.89k | } |
627 | 0 | else |
628 | 0 | { |
629 | | /* In case of shared mode, buffers are added without adjusting for padding. |
630 | | Update luma and chroma pointers here to account for padding as per stride. |
631 | | In some cases stride might not be available when set_display_frame is called. |
632 | | Hence updated luma and chroma pointers here */ |
633 | |
|
634 | 0 | for(i = 0; i < BUF_MGR_MAX_CNT; i++) |
635 | 0 | { |
636 | 0 | ps_pic_buf = ihevc_buf_mgr_get_buf((buf_mgr_t *)ps_codec->pv_pic_buf_mgr, i); |
637 | 0 | if((NULL == ps_pic_buf) || |
638 | 0 | (NULL == ps_pic_buf->pu1_luma) || |
639 | 0 | (NULL == ps_pic_buf->pu1_chroma)) |
640 | 0 | { |
641 | 0 | break; |
642 | 0 | } |
643 | 0 | ps_pic_buf->pu1_luma += ps_codec->i4_strd * PAD_TOP + PAD_LEFT; |
644 | 0 | ps_pic_buf->pu1_chroma += (ps_codec->i4_strd * chroma_pixel_strd / h_samp_factor) * (PAD_TOP / v_samp_factor) |
645 | 0 | + (PAD_LEFT * chroma_pixel_strd / h_samp_factor); |
646 | 0 | } |
647 | 0 | } |
648 | | |
649 | 1.89k | return ret; |
650 | 1.89k | } |
651 | | /** |
652 | | ******************************************************************************* |
653 | | * |
654 | | * @brief |
655 | | * Function to add buffers to MV Bank buffer manager |
656 | | * |
657 | | * @par Description: |
658 | | * Function to add buffers to MV Bank buffer manager |
659 | | * To be called once per stream or for every reset |
660 | | * |
661 | | * @param[in] ps_codec |
662 | | * Pointer to codec context |
663 | | * |
664 | | * @returns Error from IHEVCD_ERROR_T |
665 | | * |
666 | | * @remarks |
667 | | * |
668 | | * |
669 | | ******************************************************************************* |
670 | | */ |
671 | | IHEVCD_ERROR_T ihevcd_mv_buf_mgr_add_bufs(codec_t *ps_codec) |
672 | 1.89k | { |
673 | 1.89k | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
674 | 1.89k | WORD32 i; |
675 | 1.89k | WORD32 max_dpb_size; |
676 | 1.89k | WORD32 mv_bank_size_allocated; |
677 | 1.89k | WORD32 pic_mv_bank_size; |
678 | | |
679 | 1.89k | sps_t *ps_sps; |
680 | 1.89k | UWORD8 *pu1_buf; |
681 | 1.89k | mv_buf_t *ps_mv_buf; |
682 | | |
683 | | |
684 | | /* Initialize MV Bank buffer manager */ |
685 | 1.89k | ps_sps = ps_codec->s_parse.ps_sps; |
686 | | |
687 | | |
688 | | /* Compute the number of MV Bank buffers needed */ |
689 | 1.89k | max_dpb_size = ps_sps->ai1_sps_max_dec_pic_buffering[ps_sps->i1_sps_max_sub_layers - 1]; |
690 | | |
691 | | /* Allocate one extra MV Bank to handle current frame |
692 | | * In case of asynchronous parsing and processing, number of buffers should increase here |
693 | | * based on when parsing and processing threads are synchronized |
694 | | */ |
695 | 1.89k | max_dpb_size++; |
696 | | |
697 | 1.89k | ps_codec->i4_max_dpb_size = max_dpb_size; |
698 | | |
699 | 1.89k | pu1_buf = (UWORD8 *)ps_codec->pv_mv_bank_buf_base; |
700 | | |
701 | 1.89k | ps_mv_buf = (mv_buf_t *)pu1_buf; |
702 | 1.89k | pu1_buf += max_dpb_size * sizeof(mv_buf_t); |
703 | 1.89k | ps_codec->ps_mv_buf = ps_mv_buf; |
704 | 1.89k | mv_bank_size_allocated = ps_codec->i4_total_mv_bank_size - max_dpb_size * sizeof(mv_buf_t); |
705 | | |
706 | | /* Compute MV bank size per picture */ |
707 | 1.89k | pic_mv_bank_size = ihevcd_get_pic_mv_bank_size(ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
708 | 1.89k | ALIGN64(ps_sps->i2_pic_height_in_luma_samples)); |
709 | | |
710 | 17.2k | for(i = 0; i < max_dpb_size; i++) |
711 | 15.3k | { |
712 | 15.3k | WORD32 buf_ret; |
713 | 15.3k | WORD32 num_pu; |
714 | 15.3k | WORD32 num_ctb; |
715 | 15.3k | WORD32 pic_size; |
716 | 15.3k | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
717 | 15.3k | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
718 | | |
719 | | |
720 | 15.3k | num_pu = pic_size / (MIN_PU_SIZE * MIN_PU_SIZE); |
721 | 15.3k | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
722 | | |
723 | | |
724 | 15.3k | mv_bank_size_allocated -= pic_mv_bank_size; |
725 | | |
726 | 15.3k | if(mv_bank_size_allocated < 0) |
727 | 0 | { |
728 | 0 | ps_codec->s_parse.i4_error_code = IHEVCD_INSUFFICIENT_MEM_MVBANK; |
729 | 0 | return IHEVCD_INSUFFICIENT_MEM_MVBANK; |
730 | 0 | } |
731 | | |
732 | 15.3k | ps_mv_buf->pu4_pic_pu_idx = (UWORD32 *)pu1_buf; |
733 | 15.3k | pu1_buf += (num_ctb + 1) * sizeof(WORD32); |
734 | | |
735 | 15.3k | ps_mv_buf->pu1_pic_pu_map = pu1_buf; |
736 | 15.3k | pu1_buf += num_pu; |
737 | | |
738 | 15.3k | ps_mv_buf->pu1_pic_slice_map = (UWORD16 *)pu1_buf; |
739 | 15.3k | pu1_buf += ALIGN4(num_ctb * sizeof(UWORD16)); |
740 | | |
741 | 15.3k | ps_mv_buf->ps_pic_pu = (pu_t *)pu1_buf; |
742 | 15.3k | pu1_buf += num_pu * sizeof(pu_t); |
743 | | |
744 | 15.3k | buf_ret = ihevc_buf_mgr_add((buf_mgr_t *)ps_codec->pv_mv_buf_mgr, ps_mv_buf, i); |
745 | | |
746 | 15.3k | if(0 != buf_ret) |
747 | 0 | { |
748 | 0 | ps_codec->s_parse.i4_error_code = IHEVCD_BUF_MGR_ERROR; |
749 | 0 | return IHEVCD_BUF_MGR_ERROR; |
750 | 0 | } |
751 | | |
752 | 15.3k | ps_mv_buf++; |
753 | | |
754 | 15.3k | } |
755 | 1.89k | return ret; |
756 | 1.89k | } |
757 | | /** |
758 | | ******************************************************************************* |
759 | | * |
760 | | * @brief |
761 | | * Output buffer check |
762 | | * |
763 | | * @par Description: |
764 | | * Check for the number of buffers and buffer sizes of output buffer |
765 | | * |
766 | | * @param[in] ps_codec |
767 | | * Pointer to codec context |
768 | | * |
769 | | * @returns Error from IHEVCD_ERROR_T |
770 | | * |
771 | | * @remarks |
772 | | * |
773 | | * |
774 | | ******************************************************************************* |
775 | | */ |
776 | | IHEVCD_ERROR_T ihevcd_check_out_buf_size(codec_t *ps_codec) |
777 | 11.2k | { |
778 | 11.2k | ivd_out_bufdesc_t *ps_out_buffer = ps_codec->ps_out_buffer; |
779 | 11.2k | UWORD32 au4_min_out_buf_size[IVD_VIDDEC_MAX_IO_BUFFERS]; |
780 | 11.2k | UWORD32 u4_min_num_out_bufs = 0, i; |
781 | 11.2k | UWORD32 wd, ht; |
782 | | |
783 | 11.2k | if(0 == ps_codec->i4_share_disp_buf) |
784 | 11.2k | { |
785 | 11.2k | wd = ps_codec->i4_disp_wd; |
786 | 11.2k | ht = ps_codec->i4_disp_ht; |
787 | 11.2k | } |
788 | 0 | else |
789 | 0 | { |
790 | | /* In case of shared mode, do not check validity of ps_codec->ps_out_buffer */ |
791 | 0 | return (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
792 | 0 | } |
793 | | |
794 | 11.2k | if(ps_codec->i4_disp_strd > (WORD32)wd) |
795 | 0 | wd = ps_codec->i4_disp_strd; |
796 | | |
797 | 11.2k | if(ps_codec->e_chroma_fmt == IV_YUV_420P) |
798 | 3.47k | u4_min_num_out_bufs = MIN_OUT_BUFS_420; |
799 | 7.82k | else if(ps_codec->e_chroma_fmt == IV_YUV_444P) |
800 | 0 | u4_min_num_out_bufs = MIN_OUT_BUFS_444; |
801 | 7.82k | else if((ps_codec->e_chroma_fmt == IV_YUV_420SP_UV) |
802 | 4.96k | || (ps_codec->e_chroma_fmt == IV_YUV_420SP_VU)) |
803 | 4.83k | u4_min_num_out_bufs = MIN_OUT_BUFS_420SP; |
804 | 2.98k | else if(ps_codec->e_chroma_fmt == IV_GRAY) |
805 | 2.98k | u4_min_num_out_bufs = MIN_OUT_BUFS_GRAY; |
806 | | |
807 | 11.2k | if(ps_codec->e_chroma_fmt == IV_YUV_420P) |
808 | 3.47k | { |
809 | 3.47k | au4_min_out_buf_size[0] = (wd * ht); |
810 | 3.47k | au4_min_out_buf_size[1] = (wd * ht) >> 2; |
811 | 3.47k | au4_min_out_buf_size[2] = (wd * ht) >> 2; |
812 | 3.47k | } |
813 | 7.82k | else if(ps_codec->e_chroma_fmt == IV_YUV_444P) |
814 | 0 | { |
815 | 0 | au4_min_out_buf_size[0] = (wd * ht); |
816 | 0 | au4_min_out_buf_size[1] = (wd * ht); |
817 | 0 | au4_min_out_buf_size[2] = (wd * ht); |
818 | 0 | } |
819 | 7.82k | else if((ps_codec->e_chroma_fmt == IV_YUV_420SP_UV) |
820 | 4.96k | || (ps_codec->e_chroma_fmt == IV_YUV_420SP_VU)) |
821 | 4.83k | { |
822 | 4.83k | au4_min_out_buf_size[0] = (wd * ht); |
823 | 4.83k | au4_min_out_buf_size[1] = (wd * ht) >> 1; |
824 | 4.83k | au4_min_out_buf_size[2] = 0; |
825 | 4.83k | } |
826 | 2.98k | else if(ps_codec->e_chroma_fmt == IV_GRAY) |
827 | 2.98k | { |
828 | 2.98k | au4_min_out_buf_size[0] = (wd * ht); |
829 | 2.98k | au4_min_out_buf_size[1] = 0; |
830 | 2.98k | au4_min_out_buf_size[2] = 0; |
831 | 2.98k | } |
832 | | |
833 | | |
834 | 11.2k | if(ps_out_buffer->u4_num_bufs < u4_min_num_out_bufs) |
835 | 0 | { |
836 | 0 | return (IHEVCD_ERROR_T)IV_FAIL; |
837 | 0 | } |
838 | | |
839 | 33.6k | for (i = 0 ; i < u4_min_num_out_bufs; i++) |
840 | 22.7k | { |
841 | 22.7k | if(ps_out_buffer->u4_min_out_buf_size[i] < au4_min_out_buf_size[i]) |
842 | 369 | { |
843 | 369 | return (IHEVCD_ERROR_T)IV_FAIL; |
844 | 369 | } |
845 | 22.7k | } |
846 | | |
847 | 10.9k | return (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
848 | 11.2k | } |
849 | | |
850 | | /** |
851 | | ******************************************************************************* |
852 | | * |
853 | | * @brief |
854 | | * Picture level initializations required during parsing |
855 | | * |
856 | | * @par Description: |
857 | | * Initialize picture level context variables during parsing Initialize mv |
858 | | * bank buffer manager in the first init call |
859 | | * |
860 | | * @param[in] ps_codec |
861 | | * Pointer to codec context |
862 | | * |
863 | | * @returns Error from IHEVCD_ERROR_T |
864 | | * |
865 | | * @remarks |
866 | | * |
867 | | * |
868 | | ******************************************************************************* |
869 | | */ |
870 | | IHEVCD_ERROR_T ihevcd_parse_pic_init(codec_t *ps_codec) |
871 | 11.2k | { |
872 | 11.2k | IHEVCD_ERROR_T ret = (IHEVCD_ERROR_T)IHEVCD_SUCCESS; |
873 | 11.2k | mv_buf_t *ps_mv_buf; |
874 | 11.2k | sps_t *ps_sps; |
875 | 11.2k | WORD32 num_min_cu; |
876 | 11.2k | WORD32 cur_pic_buf_id; |
877 | 11.2k | WORD32 cur_mv_bank_buf_id; |
878 | 11.2k | pic_buf_t *ps_cur_pic; |
879 | 11.2k | slice_header_t *ps_slice_hdr; |
880 | 11.2k | UWORD8 *pu1_cur_pic_luma, *pu1_cur_pic_chroma; |
881 | 11.2k | WORD32 h_samp_factor, v_samp_factor; |
882 | 11.2k | WORD32 chroma_pixel_strd = 2; |
883 | 11.2k | WORD32 i; |
884 | | |
885 | 11.2k | ps_codec->s_parse.i4_error_code = IHEVCD_SUCCESS; |
886 | 11.2k | ps_sps = ps_codec->s_parse.ps_sps; |
887 | 11.2k | ps_slice_hdr = ps_codec->s_parse.ps_slice_hdr; |
888 | | |
889 | | /* Memset picture level intra map and transquant bypass map to zero */ |
890 | 11.2k | num_min_cu = ((ps_sps->i2_pic_height_in_luma_samples + 7) / 8) * ((ps_sps->i2_pic_width_in_luma_samples + 63) / 64); |
891 | 11.2k | memset(ps_codec->s_parse.pu1_pic_intra_flag, 0, num_min_cu); |
892 | 11.2k | memset(ps_codec->s_parse.pu1_pic_no_loop_filter_flag, 0, num_min_cu); |
893 | | |
894 | 11.2k | h_samp_factor = (CHROMA_FMT_IDC_YUV444 == ps_sps->i1_chroma_format_idc) ? 1 : 2; |
895 | 11.2k | v_samp_factor = (CHROMA_FMT_IDC_YUV420 == ps_sps->i1_chroma_format_idc) ? 2 : 1; |
896 | | |
897 | 11.2k | if(0 == ps_codec->s_parse.i4_first_pic_init) |
898 | 1.89k | { |
899 | 1.89k | ret = ihevcd_mv_buf_mgr_add_bufs(ps_codec); |
900 | 1.89k | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
901 | | |
902 | 1.89k | ret = ihevcd_pic_buf_mgr_add_bufs(ps_codec); |
903 | 1.89k | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
904 | | |
905 | 1.89k | ps_codec->s_parse.i4_first_pic_init = 1; |
906 | 1.89k | } |
907 | | |
908 | | /* Output buffer check */ |
909 | 11.2k | ret = ihevcd_check_out_buf_size(ps_codec); |
910 | 11.2k | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
911 | | |
912 | | /* Initialize all the slice headers' slice addresses to zero */ |
913 | 10.9k | { |
914 | 10.9k | WORD32 slice_idx; |
915 | 10.9k | WORD32 slice_start_idx; |
916 | | |
917 | 10.9k | slice_start_idx = ps_codec->i4_slice_error ? 2 : 1; |
918 | | |
919 | 2.79M | for(slice_idx = slice_start_idx; slice_idx < MAX_SLICE_HDR_CNT; slice_idx++) |
920 | 2.78M | { |
921 | 2.78M | slice_header_t *ps_slice_hdr_tmp = ps_codec->ps_slice_hdr_base + slice_idx; |
922 | 2.78M | ps_slice_hdr_tmp->i2_ctb_x = -1; |
923 | 2.78M | ps_slice_hdr_tmp->i2_ctb_y = -1; |
924 | | |
925 | 2.78M | } |
926 | 10.9k | } |
927 | | |
928 | | /* Get free MV Bank to hold current picture's motion vector data */ |
929 | 10.9k | { |
930 | 10.9k | 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); |
931 | | |
932 | | /* If there are no free buffers then return with an error code. |
933 | | * If the buffer is to be freed by another thread , change the |
934 | | * following to call thread yield and wait for buffer to be freed |
935 | | */ |
936 | 10.9k | if(NULL == ps_mv_buf) |
937 | 6 | { |
938 | 6 | ps_codec->s_parse.i4_error_code = IHEVCD_NO_FREE_MVBANK; |
939 | 6 | ps_codec->i4_error_code = IHEVCD_NO_FREE_MVBANK; |
940 | 6 | return IHEVCD_NO_FREE_MVBANK; |
941 | 6 | } |
942 | | |
943 | 10.9k | ps_codec->s_parse.ps_cur_mv_buf = ps_mv_buf; |
944 | | /* Set current ABS poc to ps_mv_buf, so that while freeing a reference buffer |
945 | | * corresponding mv buffer can be found by looping through ps_codec->ps_mv_buf array |
946 | | * and getting a buffer id to free |
947 | | */ |
948 | 10.9k | ps_mv_buf->i4_abs_poc = ps_slice_hdr->i4_abs_pic_order_cnt; |
949 | 10.9k | } |
950 | | |
951 | | /* Get free picture buffer to hold current picture recon data */ |
952 | | /* TODO: For asynchronous api the following initializations related to picture |
953 | | * buffer should be moved to processing side |
954 | | */ |
955 | 0 | { |
956 | | |
957 | 10.9k | UWORD8 *pu1_buf; |
958 | 10.9k | 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); |
959 | | |
960 | | /* If there are no free buffers then return with an error code. |
961 | | * TODO: If the buffer is to be freed by another thread , change the |
962 | | * following to call thread yield and wait for buffer to be freed |
963 | | */ |
964 | 10.9k | if(NULL == ps_cur_pic) |
965 | 0 | { |
966 | 0 | ps_codec->s_parse.i4_error_code = IHEVCD_NO_FREE_PICBUF; |
967 | 0 | ps_codec->i4_error_code = IHEVCD_NO_FREE_PICBUF; |
968 | 0 | return IHEVCD_NO_FREE_PICBUF; |
969 | 0 | } |
970 | | |
971 | | /* Store input timestamp sent with input buffer */ |
972 | 10.9k | ps_cur_pic->u4_ts = ps_codec->u4_ts; |
973 | 10.9k | ps_cur_pic->i4_abs_poc = ps_slice_hdr->i4_abs_pic_order_cnt; |
974 | 10.9k | ps_cur_pic->i4_poc_lsb = ps_slice_hdr->i4_pic_order_cnt_lsb; |
975 | 10.9k | pu1_buf = ps_cur_pic->pu1_luma; |
976 | 10.9k | pu1_cur_pic_luma = pu1_buf; |
977 | | |
978 | 10.9k | pu1_buf = ps_cur_pic->pu1_chroma; |
979 | 10.9k | pu1_cur_pic_chroma = pu1_buf; |
980 | | |
981 | 10.9k | #ifndef DISABLE_SEI |
982 | 10.9k | ps_cur_pic->s_sei_params.i1_sei_parameters_present_flag = 0; |
983 | 10.9k | if(ps_codec->s_parse.s_sei_params.i1_sei_parameters_present_flag) |
984 | 210 | { |
985 | 210 | sei_params_t *ps_sei = &ps_codec->s_parse.s_sei_params; |
986 | 210 | ps_cur_pic->s_sei_params = ps_codec->s_parse.s_sei_params; |
987 | | |
988 | | /* Once sei_params is copied to pic_buf, |
989 | | * mark sei_params in s_parse as not present, |
990 | | * this ensures that future frames do not use this data again. |
991 | | */ |
992 | 210 | ps_sei->i1_sei_parameters_present_flag = 0; |
993 | 210 | ps_sei->i1_user_data_registered_present_flag = 0; |
994 | 210 | ps_sei->i1_aud_present_flag = 0; |
995 | 210 | ps_sei->i1_time_code_present_flag = 0; |
996 | 210 | ps_sei->i1_buf_period_params_present_flag = 0; |
997 | 210 | ps_sei->i1_pic_timing_params_present_flag = 0; |
998 | 210 | ps_sei->i1_recovery_point_params_present_flag = 0; |
999 | 210 | ps_sei->i1_active_parameter_set = 0; |
1000 | 210 | ps_sei->i4_sei_mastering_disp_colour_vol_params_present_flags = 0; |
1001 | 210 | } |
1002 | 10.9k | #endif |
1003 | 10.9k | } |
1004 | | |
1005 | 10.9k | if(0 == ps_codec->u4_pic_cnt) |
1006 | 1.59k | { |
1007 | 1.59k | 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); |
1008 | 1.59k | if(ps_sps->i1_chroma_format_idc != CHROMA_FMT_IDC_MONOCHROME) |
1009 | 1.59k | { |
1010 | 1.59k | memset(ps_cur_pic->pu1_chroma, |
1011 | 1.59k | 128, |
1012 | 1.59k | (((ps_sps->i2_pic_width_in_luma_samples + PAD_WD) * (chroma_pixel_strd / h_samp_factor)) |
1013 | 1.59k | * ps_sps->i2_pic_height_in_luma_samples / v_samp_factor)); |
1014 | 1.59k | } |
1015 | 1.59k | } |
1016 | | |
1017 | | /* Fill the remaining entries of the reference lists with the nearest POC |
1018 | | * This is done to handle cases where there is a corruption in the reference index */ |
1019 | 10.9k | { |
1020 | 10.9k | pic_buf_t *ps_pic_buf_ref; |
1021 | 10.9k | mv_buf_t *ps_mv_buf_ref; |
1022 | 10.9k | WORD32 r_idx; |
1023 | 10.9k | dpb_mgr_t *ps_dpb_mgr = (dpb_mgr_t *)ps_codec->pv_dpb_mgr; |
1024 | 10.9k | buf_mgr_t *ps_mv_buf_mgr = (buf_mgr_t *)ps_codec->pv_mv_buf_mgr; |
1025 | | |
1026 | 10.9k | ps_pic_buf_ref = ihevc_dpb_mgr_get_ref_by_nearest_poc(ps_dpb_mgr, ps_slice_hdr->i4_abs_pic_order_cnt); |
1027 | 10.9k | if(NULL == ps_pic_buf_ref) |
1028 | 4.37k | { |
1029 | 4.37k | WORD32 size; |
1030 | | |
1031 | 4.37k | WORD32 num_pu; |
1032 | 4.37k | WORD32 num_ctb; |
1033 | 4.37k | WORD32 pic_size; |
1034 | | /* In case current mv buffer itself is being used as reference mv buffer for colocated |
1035 | | * calculations, then memset all the buffers to zero. |
1036 | | */ |
1037 | 4.37k | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
1038 | 4.37k | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
1039 | | |
1040 | 4.37k | num_pu = pic_size / (MIN_PU_SIZE * MIN_PU_SIZE); |
1041 | 4.37k | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
1042 | | |
1043 | 4.37k | memset(ps_mv_buf->ai4_l0_collocated_poc, 0, sizeof(ps_mv_buf->ai4_l0_collocated_poc)); |
1044 | 4.37k | memset(ps_mv_buf->ai1_l0_collocated_poc_lt, 0, sizeof(ps_mv_buf->ai1_l0_collocated_poc_lt)); |
1045 | 4.37k | memset(ps_mv_buf->ai4_l1_collocated_poc, 0, sizeof(ps_mv_buf->ai4_l1_collocated_poc)); |
1046 | 4.37k | memset(ps_mv_buf->ai1_l1_collocated_poc_lt, 0, sizeof(ps_mv_buf->ai1_l1_collocated_poc_lt)); |
1047 | | |
1048 | 4.37k | size = (num_ctb + 1) * sizeof(WORD32); |
1049 | 4.37k | memset(ps_mv_buf->pu4_pic_pu_idx, 0, size); |
1050 | | |
1051 | 4.37k | size = num_pu; |
1052 | 4.37k | memset(ps_mv_buf->pu1_pic_pu_map, 0, size); |
1053 | 4.37k | size = ALIGN4(num_ctb * sizeof(UWORD16)); |
1054 | 4.37k | memset(ps_mv_buf->pu1_pic_slice_map, 0, size); |
1055 | 4.37k | size = num_pu * sizeof(pu_t); |
1056 | 4.37k | memset(ps_mv_buf->ps_pic_pu, 0, size); |
1057 | | |
1058 | 4.37k | ps_pic_buf_ref = ps_cur_pic; |
1059 | 4.37k | ps_mv_buf_ref = ps_mv_buf; |
1060 | 4.37k | } |
1061 | 6.54k | else |
1062 | 6.54k | { |
1063 | 6.54k | ps_mv_buf_ref = ihevcd_mv_mgr_get_poc(ps_mv_buf_mgr, ps_pic_buf_ref->i4_abs_poc); |
1064 | 6.54k | } |
1065 | | |
1066 | 25.9k | for(r_idx = 0; r_idx < ps_slice_hdr->i1_num_ref_idx_l0_active; r_idx++) |
1067 | 15.0k | { |
1068 | 15.0k | if(NULL == ps_slice_hdr->as_ref_pic_list0[r_idx].pv_pic_buf) |
1069 | 1.09k | { |
1070 | 1.09k | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
1071 | 1.09k | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
1072 | 1.09k | } |
1073 | 15.0k | } |
1074 | | |
1075 | 170k | for(r_idx = ps_slice_hdr->i1_num_ref_idx_l0_active; r_idx < MAX_DPB_SIZE; r_idx++) |
1076 | 159k | { |
1077 | 159k | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
1078 | 159k | ps_slice_hdr->as_ref_pic_list0[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
1079 | 159k | } |
1080 | | |
1081 | 23.0k | for(r_idx = 0; r_idx < ps_slice_hdr->i1_num_ref_idx_l1_active; r_idx++) |
1082 | 12.1k | { |
1083 | 12.1k | if(NULL == ps_slice_hdr->as_ref_pic_list1[r_idx].pv_pic_buf) |
1084 | 526 | { |
1085 | 526 | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
1086 | 526 | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
1087 | 526 | } |
1088 | 12.1k | } |
1089 | | |
1090 | 173k | for(r_idx = ps_slice_hdr->i1_num_ref_idx_l1_active; r_idx < MAX_DPB_SIZE; r_idx++) |
1091 | 162k | { |
1092 | 162k | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_pic_buf = (void *)ps_pic_buf_ref; |
1093 | 162k | ps_slice_hdr->as_ref_pic_list1[r_idx].pv_mv_buf = (void *)ps_mv_buf_ref; |
1094 | 162k | } |
1095 | 10.9k | } |
1096 | | |
1097 | | |
1098 | | /* Reset the jobq to start of the jobq buffer */ |
1099 | 10.9k | ihevcd_jobq_reset((jobq_t *)ps_codec->pv_proc_jobq); |
1100 | | |
1101 | 10.9k | if(ps_codec->i4_threads_active) |
1102 | 10.9k | { |
1103 | 10.9k | ps_codec->i4_break_threads = 0; |
1104 | 10.9k | } |
1105 | 10.9k | ps_codec->s_parse.i4_pic_pu_idx = 0; |
1106 | 10.9k | ps_codec->s_parse.i4_pic_tu_idx = 0; |
1107 | | |
1108 | 10.9k | ps_codec->s_parse.pu1_pic_pu_map = ps_mv_buf->pu1_pic_pu_map; |
1109 | 10.9k | ps_codec->s_parse.ps_pic_pu = ps_mv_buf->ps_pic_pu; |
1110 | 10.9k | ps_codec->s_parse.pu4_pic_pu_idx = ps_mv_buf->pu4_pic_pu_idx; |
1111 | 10.9k | ps_codec->s_parse.pu1_slice_idx = (UWORD16 *)ps_mv_buf->pu1_pic_slice_map; |
1112 | 98.3k | for(i = 0; i < MAX_PROCESS_THREADS; i++) |
1113 | 87.3k | { |
1114 | 87.3k | ps_codec->as_process[i].pu1_slice_idx = (UWORD16 *)ps_mv_buf->pu1_pic_slice_map; |
1115 | 87.3k | } |
1116 | 10.9k | ps_codec->s_parse.pu1_pu_map = ps_codec->s_parse.pu1_pic_pu_map; |
1117 | 10.9k | ps_codec->s_parse.ps_pu = ps_codec->s_parse.ps_pic_pu; |
1118 | | |
1119 | 10.9k | { |
1120 | 10.9k | UWORD8 *pu1_buf; |
1121 | 10.9k | WORD32 ctb_luma_min_tu_cnt, ctb_chroma_min_tu_cnt, ctb_min_tu_cnt; |
1122 | 10.9k | WORD32 pic_size; |
1123 | 10.9k | WORD32 num_ctb; |
1124 | | |
1125 | 10.9k | pic_size = ALIGN64(ps_sps->i2_pic_width_in_luma_samples) * |
1126 | 10.9k | ALIGN64(ps_sps->i2_pic_height_in_luma_samples); |
1127 | | |
1128 | 10.9k | ctb_luma_min_tu_cnt = pic_size / (MIN_TU_SIZE * MIN_TU_SIZE); |
1129 | | |
1130 | 10.9k | if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV444) |
1131 | 0 | { |
1132 | 0 | ctb_chroma_min_tu_cnt = ctb_luma_min_tu_cnt << 1; |
1133 | 0 | } |
1134 | 10.9k | else if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV422) |
1135 | 0 | { |
1136 | 0 | ctb_chroma_min_tu_cnt = ctb_luma_min_tu_cnt; |
1137 | 0 | } |
1138 | 10.9k | else if(ps_sps->i1_chroma_format_idc == CHROMA_FMT_IDC_YUV420) |
1139 | 10.9k | { |
1140 | 10.9k | ctb_chroma_min_tu_cnt = ctb_luma_min_tu_cnt >> 1; |
1141 | 10.9k | } |
1142 | 0 | else |
1143 | 0 | { |
1144 | 0 | ctb_chroma_min_tu_cnt = 0; |
1145 | 0 | } |
1146 | | |
1147 | 10.9k | ctb_min_tu_cnt = ctb_luma_min_tu_cnt + ctb_chroma_min_tu_cnt; |
1148 | | |
1149 | 10.9k | num_ctb = pic_size / (MIN_CTB_SIZE * MIN_CTB_SIZE); |
1150 | 10.9k | pu1_buf = (UWORD8 *)ps_codec->pv_tu_data; |
1151 | 10.9k | ps_codec->s_parse.pu4_pic_tu_idx = (UWORD32 *)pu1_buf; |
1152 | 10.9k | pu1_buf += (num_ctb + 1) * sizeof(WORD32); |
1153 | | |
1154 | 10.9k | ps_codec->s_parse.pu1_pic_tu_map = pu1_buf; |
1155 | 10.9k | pu1_buf += ctb_min_tu_cnt; |
1156 | | |
1157 | 10.9k | ps_codec->s_parse.ps_pic_tu = (tu_t *)pu1_buf; |
1158 | 10.9k | pu1_buf += ctb_min_tu_cnt * sizeof(tu_t); |
1159 | | |
1160 | 10.9k | ps_codec->s_parse.pv_pic_tu_coeff_data = pu1_buf; |
1161 | | |
1162 | 10.9k | ps_codec->s_parse.pu1_tu_map = ps_codec->s_parse.pu1_pic_tu_map; |
1163 | 10.9k | ps_codec->s_parse.ps_tu = ps_codec->s_parse.ps_pic_tu; |
1164 | 10.9k | ps_codec->s_parse.pv_tu_coeff_data = ps_codec->s_parse.pv_pic_tu_coeff_data; |
1165 | 10.9k | } |
1166 | | |
1167 | 10.9k | ps_codec->s_parse.s_bs_ctxt.ps_pic_pu = ps_codec->s_parse.ps_pic_pu; |
1168 | 10.9k | ps_codec->s_parse.s_bs_ctxt.pu4_pic_pu_idx = ps_codec->s_parse.pu4_pic_pu_idx; |
1169 | 10.9k | ps_codec->s_parse.s_bs_ctxt.pu4_pic_tu_idx = ps_codec->s_parse.pu4_pic_tu_idx; |
1170 | | |
1171 | | |
1172 | | /* Set number of CTBs to be processed simultaneously */ |
1173 | 10.9k | ps_codec->i4_proc_nctb = ihevcd_nctb_cnt(ps_codec, ps_sps); |
1174 | | |
1175 | | /* Memset Parse Map and process map at the start of frame */ |
1176 | | //TODO: In case of asynchronous API proc_map can not be set to zero here |
1177 | 10.9k | { |
1178 | 10.9k | WORD32 num_ctb; |
1179 | | |
1180 | 10.9k | num_ctb = ps_sps->i4_pic_size_in_ctb; |
1181 | | |
1182 | 10.9k | memset(ps_codec->pu1_parse_map, 0, num_ctb); |
1183 | | |
1184 | 10.9k | memset(ps_codec->pu1_proc_map, 0, num_ctb); |
1185 | 10.9k | } |
1186 | | |
1187 | | |
1188 | | |
1189 | | /* Initialize disp buf id to -1, this will be updated at the end of frame if there is |
1190 | | * buffer to be displayed |
1191 | | */ |
1192 | 10.9k | ps_codec->i4_disp_buf_id = -1; |
1193 | 10.9k | ps_codec->ps_disp_buf = NULL; |
1194 | | |
1195 | 10.9k | ps_codec->i4_disable_deblk_pic = 0; |
1196 | 10.9k | ps_codec->i4_disable_sao_pic = 0; |
1197 | 10.9k | ps_codec->i4_fullpel_inter_pred = 0; |
1198 | 10.9k | ps_codec->i4_mv_frac_mask = 0x7FFFFFFF; |
1199 | | |
1200 | | /* If degrade is enabled, set the degrade flags appropriately */ |
1201 | 10.9k | if(ps_codec->i4_degrade_type && ps_codec->i4_degrade_pics) |
1202 | 0 | { |
1203 | 0 | WORD32 degrade_pic; |
1204 | 0 | ps_codec->i4_degrade_pic_cnt++; |
1205 | 0 | degrade_pic = 0; |
1206 | | |
1207 | | /* If degrade is to be done in all frames, then do not check further */ |
1208 | 0 | switch(ps_codec->i4_degrade_pics) |
1209 | 0 | { |
1210 | 0 | case 4: |
1211 | 0 | { |
1212 | 0 | degrade_pic = 1; |
1213 | 0 | break; |
1214 | 0 | } |
1215 | 0 | case 3: |
1216 | 0 | { |
1217 | 0 | if(ps_slice_hdr->i1_slice_type != ISLICE) |
1218 | 0 | degrade_pic = 1; |
1219 | |
|
1220 | 0 | break; |
1221 | 0 | } |
1222 | 0 | case 2: |
1223 | 0 | { |
1224 | | |
1225 | | /* If pic count hits non-degrade interval or it is an islice, then do not degrade */ |
1226 | 0 | if((ps_slice_hdr->i1_slice_type != ISLICE) && |
1227 | 0 | (ps_codec->i4_degrade_pic_cnt != ps_codec->i4_nondegrade_interval)) |
1228 | 0 | degrade_pic = 1; |
1229 | |
|
1230 | 0 | break; |
1231 | 0 | } |
1232 | 0 | case 1: |
1233 | 0 | { |
1234 | | /* Check if the current picture is non-ref */ |
1235 | 0 | if((ps_slice_hdr->i1_nal_unit_type < NAL_BLA_W_LP) && |
1236 | 0 | (ps_slice_hdr->i1_nal_unit_type % 2 == 0)) |
1237 | 0 | { |
1238 | 0 | degrade_pic = 1; |
1239 | 0 | } |
1240 | 0 | break; |
1241 | 0 | } |
1242 | | |
1243 | |
|
1244 | 0 | } |
1245 | 0 | if(degrade_pic) |
1246 | 0 | { |
1247 | 0 | if(ps_codec->i4_degrade_type & 0x1) |
1248 | 0 | ps_codec->i4_disable_sao_pic = 1; |
1249 | |
|
1250 | 0 | if(ps_codec->i4_degrade_type & 0x2) |
1251 | 0 | ps_codec->i4_disable_deblk_pic = 1; |
1252 | | |
1253 | | /* MC degrading is done only for non-ref pictures */ |
1254 | 0 | if((ps_slice_hdr->i1_nal_unit_type < NAL_BLA_W_LP) && |
1255 | 0 | (ps_slice_hdr->i1_nal_unit_type % 2 == 0)) |
1256 | 0 | { |
1257 | 0 | if(ps_codec->i4_degrade_type & 0x4) |
1258 | 0 | ps_codec->i4_mv_frac_mask = 0; |
1259 | |
|
1260 | 0 | if(ps_codec->i4_degrade_type & 0x8) |
1261 | 0 | ps_codec->i4_mv_frac_mask = 0; |
1262 | 0 | } |
1263 | 0 | } |
1264 | 0 | else |
1265 | 0 | ps_codec->i4_degrade_pic_cnt = 0; |
1266 | 0 | } |
1267 | | |
1268 | | |
1269 | 10.9k | { |
1270 | 10.9k | WORD32 i; |
1271 | 98.3k | for(i = 0; i < MAX_PROCESS_THREADS; i++) |
1272 | 87.3k | { |
1273 | 87.3k | ps_codec->as_process[i].pu4_pic_pu_idx = ps_codec->s_parse.pu4_pic_pu_idx; |
1274 | 87.3k | ps_codec->as_process[i].ps_pic_pu = ps_codec->s_parse.ps_pic_pu; |
1275 | 87.3k | ps_codec->as_process[i].pu1_pic_pu_map = ps_codec->s_parse.pu1_pic_pu_map; |
1276 | 87.3k | ps_codec->as_process[i].pu4_pic_tu_idx = ps_codec->s_parse.pu4_pic_tu_idx; |
1277 | 87.3k | ps_codec->as_process[i].ps_pic_tu = ps_codec->s_parse.ps_pic_tu; |
1278 | 87.3k | ps_codec->as_process[i].pu1_pic_tu_map = ps_codec->s_parse.pu1_pic_tu_map; |
1279 | 87.3k | ps_codec->as_process[i].pv_pic_tu_coeff_data = ps_codec->s_parse.pv_pic_tu_coeff_data; |
1280 | 87.3k | ps_codec->as_process[i].i4_cur_mv_bank_buf_id = cur_mv_bank_buf_id; |
1281 | 87.3k | ps_codec->as_process[i].s_sao_ctxt.pu1_slice_idx = ps_codec->as_process[i].pu1_slice_idx; |
1282 | 87.3k | ps_codec->as_process[i].s_sao_ctxt.pu1_tile_idx = ps_codec->as_process[i].pu1_tile_idx; |
1283 | | |
1284 | | /* TODO: For asynchronous api the following initializations related to picture |
1285 | | * buffer should be moved to processing side |
1286 | | */ |
1287 | 87.3k | ps_codec->as_process[i].pu1_cur_pic_luma = pu1_cur_pic_luma; |
1288 | 87.3k | ps_codec->as_process[i].pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
1289 | 87.3k | ps_codec->as_process[i].ps_cur_pic = ps_cur_pic; |
1290 | 87.3k | ps_codec->as_process[i].i4_cur_pic_buf_id = cur_pic_buf_id; |
1291 | | |
1292 | 87.3k | ps_codec->as_process[i].ps_out_buffer = ps_codec->ps_out_buffer; |
1293 | 87.3k | if(1 < ps_codec->i4_num_cores) |
1294 | 39.6k | { |
1295 | 39.6k | ps_codec->as_process[i].i4_check_parse_status = 1; |
1296 | 39.6k | ps_codec->as_process[i].i4_check_proc_status = 1; |
1297 | 39.6k | } |
1298 | 47.7k | else |
1299 | 47.7k | { |
1300 | 47.7k | ps_codec->as_process[i].i4_check_parse_status = 0; |
1301 | 47.7k | ps_codec->as_process[i].i4_check_proc_status = 0; |
1302 | 47.7k | } |
1303 | 87.3k | ps_codec->as_process[i].pu1_pic_intra_flag = ps_codec->s_parse.pu1_pic_intra_flag; |
1304 | 87.3k | ps_codec->as_process[i].pu1_pic_no_loop_filter_flag = ps_codec->s_parse.pu1_pic_no_loop_filter_flag; |
1305 | 87.3k | ps_codec->as_process[i].i4_init_done = 0; |
1306 | | |
1307 | 87.3k | ps_codec->as_process[i].s_bs_ctxt.pu4_pic_tu_idx = ps_codec->as_process[i].pu4_pic_tu_idx; |
1308 | 87.3k | ps_codec->as_process[i].s_bs_ctxt.pu4_pic_pu_idx = ps_codec->as_process[i].pu4_pic_pu_idx; |
1309 | 87.3k | ps_codec->as_process[i].s_bs_ctxt.ps_pic_pu = ps_codec->as_process[i].ps_pic_pu; |
1310 | 87.3k | ps_codec->as_process[i].s_deblk_ctxt.pu1_pic_no_loop_filter_flag = ps_codec->s_parse.pu1_pic_no_loop_filter_flag; |
1311 | 87.3k | ps_codec->as_process[i].s_deblk_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
1312 | 87.3k | ps_codec->as_process[i].s_deblk_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
1313 | 87.3k | ps_codec->as_process[i].s_sao_ctxt.pu1_pic_no_loop_filter_flag = ps_codec->s_parse.pu1_pic_no_loop_filter_flag; |
1314 | 87.3k | ps_codec->as_process[i].s_sao_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
1315 | 87.3k | ps_codec->as_process[i].s_sao_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
1316 | 87.3k | if(i < (ps_codec->i4_num_cores - 1)) |
1317 | 10.2k | { |
1318 | 10.2k | if (!ps_codec->ai4_process_thread_created[i]) |
1319 | 936 | { |
1320 | 936 | ithread_create(ps_codec->apv_process_thread_handle[i], NULL, |
1321 | 936 | (void *)ihevcd_process_thread, |
1322 | 936 | (void *)&ps_codec->as_process[i]); |
1323 | 936 | ps_codec->ai4_process_thread_created[i] = 1; |
1324 | 936 | } |
1325 | 10.2k | if(ps_codec->i4_threads_active) |
1326 | 10.2k | { |
1327 | 10.2k | ret = ithread_mutex_lock(ps_codec->apv_proc_start_mutex[i]); |
1328 | 10.2k | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
1329 | | |
1330 | 10.2k | ps_codec->ai4_process_start[i] = 1; |
1331 | 10.2k | ithread_cond_signal(ps_codec->apv_proc_start_condition[i]); |
1332 | | |
1333 | 10.2k | ret = ithread_mutex_unlock(ps_codec->apv_proc_start_mutex[i]); |
1334 | 10.2k | RETURN_IF((ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS), ret); |
1335 | 10.2k | } |
1336 | 10.2k | } |
1337 | 77.1k | else |
1338 | 77.1k | { |
1339 | 77.1k | ps_codec->ai4_process_thread_created[i] = 0; |
1340 | 77.1k | } |
1341 | | |
1342 | 87.3k | } |
1343 | 10.9k | if(ps_codec->u1_enable_cu_info) |
1344 | 0 | { |
1345 | 0 | ps_codec->as_buf_id_info_map[cur_pic_buf_id].pu1_qp_map = |
1346 | 0 | ps_codec->pu1_qp_map_base + (cur_pic_buf_id * ps_codec->u4_num_8x8_blks); |
1347 | 0 | ps_codec->as_buf_id_info_map[cur_pic_buf_id].pu1_cu_type_map = |
1348 | 0 | ps_codec->pu1_cu_type_map_base + (cur_pic_buf_id * ps_codec->u4_num_8x8_blks); |
1349 | 0 | memset(ps_codec->as_buf_id_info_map[cur_pic_buf_id].pu1_qp_map, |
1350 | 0 | 0, ps_codec->u4_num_8x8_blks); |
1351 | 0 | memset(ps_codec->as_buf_id_info_map[cur_pic_buf_id].pu1_qp_map, |
1352 | 0 | 0, ps_codec->u4_num_8x8_blks); |
1353 | 0 | } |
1354 | | |
1355 | 10.9k | ps_codec->s_parse.s_deblk_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
1356 | 10.9k | ps_codec->s_parse.s_deblk_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
1357 | | |
1358 | 10.9k | ps_codec->s_parse.s_sao_ctxt.pu1_cur_pic_luma = pu1_cur_pic_luma; |
1359 | 10.9k | ps_codec->s_parse.s_sao_ctxt.pu1_cur_pic_chroma = pu1_cur_pic_chroma; |
1360 | 10.9k | } |
1361 | | /* Since any input bitstream buffer that contains slice data will be sent to output(even in |
1362 | | * case of error, this buffer is added to display queue and next buffer in the display queue |
1363 | | * will be returned as the display buffer. |
1364 | | * Note: If format conversion (or frame copy) is used and is scheduled |
1365 | | * in a different thread then it has to check if the processing for the current row is complete before |
1366 | | * it copies/converts a given row. In case of low delay or in case of B pictures, current frame being decoded has to be |
1367 | | * returned, which requires a status check to ensure that the current row is reconstructed before copying. |
1368 | | */ |
1369 | | /* Add current picture to display manager */ |
1370 | 0 | { |
1371 | 10.9k | WORD32 abs_poc; |
1372 | 10.9k | slice_header_t *ps_slice_hdr; |
1373 | 10.9k | ps_slice_hdr = ps_codec->s_parse.ps_slice_hdr; |
1374 | 10.9k | abs_poc = ps_slice_hdr->i4_abs_pic_order_cnt; |
1375 | 10.9k | ihevc_disp_mgr_add((disp_mgr_t *)ps_codec->pv_disp_buf_mgr, |
1376 | 10.9k | ps_codec->as_process[0].i4_cur_pic_buf_id, |
1377 | 10.9k | abs_poc, |
1378 | 10.9k | ps_codec->as_process[0].ps_cur_pic); |
1379 | 10.9k | } |
1380 | 10.9k | ps_codec->ps_disp_buf = NULL; |
1381 | | /* Get picture to be displayed if number of pictures decoded is more than max allowed reorder */ |
1382 | | /* Since the current will be decoded, check is fore >= instead of > */ |
1383 | 10.9k | 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]) || |
1384 | 1.45k | (ps_codec->e_frm_out_mode == IVD_DECODE_FRAME_OUT)) |
1385 | | |
1386 | 9.46k | { |
1387 | 9.46k | 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); |
1388 | 9.46k | ps_codec->u4_disp_cnt++; |
1389 | 9.46k | } |
1390 | | |
1391 | 10.9k | ps_codec->s_fmt_conv.i4_cur_row = 0; |
1392 | | /* Set number of rows to be processed at a time */ |
1393 | 10.9k | ps_codec->s_fmt_conv.i4_num_rows = 4; |
1394 | | |
1395 | 10.9k | if(ps_codec->u4_enable_fmt_conv_ahead && (ps_codec->i4_num_cores > 1)) |
1396 | 0 | { |
1397 | 0 | process_ctxt_t *ps_proc; |
1398 | | |
1399 | | /* i4_num_cores - 1 contexts are currently being used by other threads */ |
1400 | 0 | ps_proc = &ps_codec->as_process[ps_codec->i4_num_cores - 1]; |
1401 | | |
1402 | | /* If the frame being decoded and displayed are different, schedule format conversion jobs |
1403 | | * this will keep the proc threads busy and lets parse thread decode few CTBs ahead |
1404 | | * If the frame being decoded and displayed are same, then format conversion is scheduled later. |
1405 | | */ |
1406 | 0 | if((ps_codec->ps_disp_buf) && (ps_codec->i4_disp_buf_id != ps_proc->i4_cur_pic_buf_id) && |
1407 | 0 | ((0 == ps_codec->i4_share_disp_buf) || (IV_YUV_420P == ps_codec->e_chroma_fmt))) |
1408 | 0 | { |
1409 | |
|
1410 | 0 | for(i = 0; i < ps_sps->i2_pic_ht_in_ctb; i++) |
1411 | 0 | { |
1412 | 0 | proc_job_t s_job; |
1413 | 0 | IHEVCD_ERROR_T ret; |
1414 | 0 | s_job.i4_cmd = CMD_FMTCONV; |
1415 | 0 | s_job.i2_ctb_cnt = 0; |
1416 | 0 | s_job.i2_ctb_x = 0; |
1417 | 0 | s_job.i2_ctb_y = i; |
1418 | 0 | s_job.i2_slice_idx = 0; |
1419 | 0 | s_job.i4_tu_coeff_data_ofst = 0; |
1420 | 0 | ret = ihevcd_jobq_queue((jobq_t *)ps_codec->s_parse.pv_proc_jobq, |
1421 | 0 | &s_job, sizeof(proc_job_t), 1); |
1422 | 0 | if(ret != (IHEVCD_ERROR_T)IHEVCD_SUCCESS) |
1423 | 0 | return ret; |
1424 | 0 | } |
1425 | 0 | } |
1426 | 0 | } |
1427 | | |
1428 | | /* If parse_pic_init is called, then slice data is present in the input bitstrea stream */ |
1429 | 10.9k | ps_codec->i4_pic_present = 1; |
1430 | | |
1431 | 10.9k | return ret; |
1432 | 10.9k | } |
1433 | | |
1434 | | |