/src/libavc/encoder/svc/isvce_utils.c
Line | Count | Source (jump to first uncovered line) |
1 | | /****************************************************************************** |
2 | | * |
3 | | * Copyright (C) 2022 The Android Open Source Project |
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 | | * Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore |
19 | | */ |
20 | | |
21 | | /** |
22 | | ******************************************************************************* |
23 | | * @file |
24 | | * ih264e_svc_utils.c |
25 | | * |
26 | | * @brief |
27 | | * Contains utility functions used for SVC encoding |
28 | | * |
29 | | * @author |
30 | | * ittiam |
31 | | * |
32 | | * @par List of Functions: |
33 | | * - ih264e_svc_ref_list_refresh() |
34 | | * - ih264e_svc_inp_params_validate() |
35 | | * |
36 | | * @remarks |
37 | | * None |
38 | | * |
39 | | ******************************************************************************* |
40 | | */ |
41 | | #include <math.h> |
42 | | #include <limits.h> |
43 | | |
44 | | #include "ih264_typedefs.h" |
45 | | |
46 | | /* Dependencies of ih264_buf_mgr.h */ |
47 | | /* Dependencies of ih264_list.h */ |
48 | | #include "ih264_error.h" |
49 | | |
50 | | #include "ih264_buf_mgr.h" |
51 | | #include "ih264_list.h" |
52 | | #include "ih264_trans_data.h" |
53 | | #include "ih264_size_defs.h" |
54 | | |
55 | | /* Dependencies of ih264_common_tables.h */ |
56 | | #include "ih264_defs.h" |
57 | | #include "ih264_structs.h" |
58 | | |
59 | | #include "ih264_common_tables.h" |
60 | | |
61 | | /* Dependencies of ih264e_bitstream.h */ |
62 | | #include "ih264e_error.h" |
63 | | |
64 | | /* Dependencies of ih264e_cabac_structs.h */ |
65 | | #include "ih264_cabac_tables.h" |
66 | | |
67 | | /* Dependencies of ime_structs.h */ |
68 | | #include "ime_defs.h" |
69 | | #include "ime_distortion_metrics.h" |
70 | | |
71 | | /* Dependencies of ih264e_structs.h */ |
72 | | #include "iv2.h" |
73 | | #include "ive2.h" |
74 | | #include "ih264_defs.h" |
75 | | #include "ih264_deblk_edge_filters.h" |
76 | | #include "ih264_inter_pred_filters.h" |
77 | | #include "ih264_structs.h" |
78 | | #include "ih264_trans_quant_itrans_iquant.h" |
79 | | #include "ih264e_bitstream.h" |
80 | | #include "ih264e_cabac_structs.h" |
81 | | #include "ime_statistics.h" |
82 | | #include "ime_structs.h" |
83 | | /* Dependencies of 'irc_picture_type.h' */ |
84 | | #include "irc_cntrl_param.h" |
85 | | #include "irc_frame_info_collector.h" |
86 | | #include "irc_mem_req_and_acq.h" |
87 | | /* Dependencies of 'irc_rate_control_api_structs' */ |
88 | | #include "irc_picture_type.h" |
89 | | #include "irc_rd_model.h" |
90 | | #include "irc_vbr_storage_vbv.h" |
91 | | #include "irc_est_sad.h" |
92 | | #include "irc_bit_allocation.h" |
93 | | #include "irc_mb_model_based.h" |
94 | | #include "irc_cbr_buffer_control.h" |
95 | | #include "irc_vbr_str_prms.h" |
96 | | #include "irc_common.h" |
97 | | #include "irc_rate_control_api_structs.h" |
98 | | #include "irc_rate_control_api.h" |
99 | | #include "irc_svc_rate_control_api.h" |
100 | | /* Dependencies of 'ih264e_utils.h' */ |
101 | | #include "ih264e_defs.h" |
102 | | #include "ih264e_structs.h" |
103 | | /* Dependencies of 'ih264e_utils.h' */ |
104 | | #include "irc_mem_req_and_acq.h" |
105 | | #include "ih264e_rc_mem_interface.h" |
106 | | #include "ih264e_time_stamp.h" |
107 | | #include "ih264e_utils.h" |
108 | | #include "ime.h" |
109 | | #include "isvc_macros.h" |
110 | | #include "isvce_cabac.h" |
111 | | #include "isvce_core_coding.h" |
112 | | #include "isvce_defs.h" |
113 | | #include "isvce_error.h" |
114 | | #include "isvce_me.h" |
115 | | #include "isvce_utils.h" |
116 | | #include "isvce_downscaler.h" |
117 | | #include "isvce_encode_header.h" |
118 | | #include "isvce_rate_control.h" |
119 | | #include "isvce_sub_pic_rc.h" |
120 | | |
121 | | static const UWORD32 gu4_downscaler_blk_size = 96; |
122 | | |
123 | | static FORCEINLINE UWORD32 isvce_get_downscaler_blk_dims(UWORD32 u4_frame_dim, UWORD32 u4_blk_pos, |
124 | | UWORD32 u4_default_blk_size) |
125 | 2.02M | { |
126 | 2.02M | return ((u4_frame_dim - u4_blk_pos * u4_default_blk_size) < u4_default_blk_size) |
127 | 2.02M | ? (u4_frame_dim - u4_blk_pos * u4_default_blk_size) |
128 | 2.02M | : u4_default_blk_size; |
129 | 2.02M | } |
130 | | |
131 | | /** |
132 | | ******************************************************************************* |
133 | | * |
134 | | * @brief |
135 | | * Reference and MV bank Buffer Manager for SVC |
136 | | * |
137 | | * @par Description: |
138 | | * Here we will |
139 | | * 1) Find the correct ref pics for the current frame |
140 | | * 2) Free the ref pics that are not going to be used anymore |
141 | | * |
142 | | * 1) Finding correct ref pic |
143 | | * All pics needed for future are arranged in a picture list called |
144 | | * ps_codec->as_ref_set. Each picture in this will have a pic buffer and |
145 | | * MV buffer that is marked appropriately as BUF_MGR_REF, BUF_MGR_IO or |
146 | | * BUF_MGR_CODEC. pic_cnt, poc, and temporal_id will also be present. |
147 | | * The strategy is to pick the closest references that belongs to the |
148 | | * same temporal_id or lesser. The closeness is measured via the |
149 | | * smallest absolute difference between ref and cur pocs. |
150 | | * |
151 | | * Note that i4_pic_cnt == -1 is used to filter uninitialised ref pics. |
152 | | * Now since we only have max two ref pics, we will always find max 2 |
153 | | * ref pics. |
154 | | * |
155 | | * 2) Self explanatory |
156 | | * |
157 | | * @param[in] ps_codec |
158 | | * Pointer to codeec context |
159 | | * |
160 | | * @param[in] pps_ref_pics |
161 | | * Array of pointers to refPicBufs |
162 | | * |
163 | | * @param[in] pps_mv_bufs |
164 | | * Array of pointers to refMVBufs |
165 | | * |
166 | | * @param[in] e_pic_type |
167 | | * Picture type |
168 | | * |
169 | | * @returns Nothing |
170 | | * |
171 | | ******************************************************************************* |
172 | | */ |
173 | | static WORD32 isvce_ref_list_refresh(isvce_codec_t *ps_codec, svc_au_buf_t **pps_ref_pics, |
174 | | svc_au_data_t **pps_mv_bufs, WORD32 *pi4_ref_set_id, |
175 | | PIC_TYPE_T e_pic_type) |
176 | 29.5k | { |
177 | 29.5k | typedef struct |
178 | 29.5k | { |
179 | 29.5k | WORD32 i4_buf_id; |
180 | | |
181 | 29.5k | WORD32 i4_abs_poc_diff; |
182 | | |
183 | 29.5k | WORD8 i1_temporal_id; |
184 | 29.5k | } ref_pic_props_t; |
185 | | |
186 | 29.5k | ref_pic_props_t s_ref_pic_props = {0, 0, -1}; |
187 | | |
188 | 29.5k | WORD32 i, buf_status; |
189 | | |
190 | 29.5k | WORD32 i4_cur_pic_poc = ps_codec->i4_poc; |
191 | 29.5k | WORD32 i4_cur_pic_temporal_id = isvce_svc_temporal_id_compute( |
192 | 29.5k | ps_codec->i4_poc, ps_codec->s_cfg.s_svc_params.u1_num_temporal_layers, e_pic_type); |
193 | | |
194 | 29.5k | if(e_pic_type == PIC_B) |
195 | 0 | { |
196 | 0 | return IH264E_FAIL; |
197 | 0 | } |
198 | | |
199 | 29.5k | ASSERT(1 == MAX_LAYER_REFERENCE_PICS); |
200 | | |
201 | | /* Pick a ref_pic for the current picture */ |
202 | 29.5k | if(e_pic_type != PIC_IDR) |
203 | 21.0k | { |
204 | 123k | for(i = 0; i < ps_codec->i4_ref_buf_cnt; i++) |
205 | 102k | { |
206 | 102k | WORD32 i4_abs_poc_diff; |
207 | 102k | WORD8 i1_temporal_id; |
208 | | |
209 | 102k | if(ps_codec->as_ref_set[i].i4_pic_cnt == -1) |
210 | 53.4k | { |
211 | 53.4k | continue; |
212 | 53.4k | } |
213 | | |
214 | 49.1k | buf_status = ih264_buf_mgr_get_status(ps_codec->pv_ref_buf_mgr, |
215 | 49.1k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id); |
216 | | |
217 | 49.1k | if(buf_status & BUF_MGR_REF) |
218 | 46.4k | { |
219 | 46.4k | i4_abs_poc_diff = ABS(ps_codec->as_ref_set[i].i4_poc - i4_cur_pic_poc); |
220 | 46.4k | i1_temporal_id = ps_codec->as_ref_set[i].ps_pic_buf->i1_temporal_id; |
221 | | |
222 | 46.4k | if(s_ref_pic_props.i1_temporal_id > -1) |
223 | 20.1k | { |
224 | 20.1k | if((i1_temporal_id <= i4_cur_pic_temporal_id) && |
225 | 20.1k | (s_ref_pic_props.i4_abs_poc_diff > i4_abs_poc_diff)) |
226 | 6.15k | { |
227 | 6.15k | s_ref_pic_props.i4_abs_poc_diff = i4_abs_poc_diff; |
228 | 6.15k | s_ref_pic_props.i1_temporal_id = i1_temporal_id; |
229 | 6.15k | s_ref_pic_props.i4_buf_id = i; |
230 | 6.15k | } |
231 | 20.1k | } |
232 | 26.2k | else if(i1_temporal_id <= i4_cur_pic_temporal_id) |
233 | 20.8k | { |
234 | 20.8k | s_ref_pic_props.i4_abs_poc_diff = i4_abs_poc_diff; |
235 | 20.8k | s_ref_pic_props.i1_temporal_id = i1_temporal_id; |
236 | 20.8k | s_ref_pic_props.i4_buf_id = i; |
237 | 20.8k | } |
238 | 46.4k | } |
239 | 49.1k | } |
240 | | |
241 | 21.0k | if(s_ref_pic_props.i1_temporal_id < 0) |
242 | 173 | { |
243 | 173 | return IH264E_FAIL; |
244 | 173 | } |
245 | | |
246 | 20.8k | pps_ref_pics[0] = pps_ref_pics[1] = |
247 | 20.8k | ps_codec->as_ref_set[s_ref_pic_props.i4_buf_id].ps_pic_buf; |
248 | 20.8k | pps_mv_bufs[0] = pps_mv_bufs[1] = |
249 | 20.8k | ps_codec->as_ref_set[s_ref_pic_props.i4_buf_id].ps_svc_au_data; |
250 | | |
251 | | /* Pick all ref pic_bufs to be freed. */ |
252 | 122k | for(i = 0; i < ps_codec->i4_ref_buf_cnt; i++) |
253 | 101k | { |
254 | 101k | if(ps_codec->as_ref_set[i].i4_pic_cnt == -1) |
255 | 53.4k | { |
256 | 53.4k | continue; |
257 | 53.4k | } |
258 | | |
259 | 48.4k | buf_status = ih264_buf_mgr_get_status(ps_codec->pv_ref_buf_mgr, |
260 | 48.4k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id); |
261 | | |
262 | 48.4k | if((buf_status & (BUF_MGR_REF | BUF_MGR_CODEC | BUF_MGR_IO)) == 0) |
263 | 229 | { |
264 | 229 | ps_codec->as_ref_set[i].i4_pic_cnt = -1; |
265 | 229 | ps_codec->as_ref_set[i].i4_poc = 32768; |
266 | | |
267 | 229 | continue; |
268 | 229 | } |
269 | | |
270 | 48.1k | if(buf_status & BUF_MGR_REF) |
271 | 46.4k | { |
272 | 46.4k | if((i4_cur_pic_temporal_id <= ps_codec->as_ref_set[i].ps_pic_buf->i1_temporal_id) && |
273 | 46.4k | (pps_ref_pics[0]->i4_frame_num != |
274 | 32.0k | ps_codec->as_ref_set[i].ps_pic_buf->i4_frame_num) && |
275 | 46.4k | (pps_ref_pics[0]->i4_frame_num != |
276 | 15.7k | ps_codec->as_ref_set[i].ps_pic_buf->i4_frame_num)) |
277 | 15.7k | { |
278 | 15.7k | ih264_buf_mgr_release(ps_codec->pv_svc_au_data_store_mgr, |
279 | 15.7k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id, |
280 | 15.7k | BUF_MGR_REF); |
281 | | |
282 | 15.7k | ih264_buf_mgr_release(ps_codec->pv_ref_buf_mgr, |
283 | 15.7k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id, |
284 | 15.7k | BUF_MGR_REF); |
285 | 15.7k | } |
286 | 46.4k | } |
287 | 48.1k | } |
288 | 20.8k | } |
289 | 8.51k | else |
290 | 8.51k | { |
291 | 50.4k | for(i = 0; i < ps_codec->i4_ref_buf_cnt; i++) |
292 | 41.9k | { |
293 | 41.9k | if(ps_codec->as_ref_set[i].i4_pic_cnt == -1) |
294 | 36.2k | { |
295 | 36.2k | continue; |
296 | 36.2k | } |
297 | | |
298 | 5.63k | buf_status = ih264_buf_mgr_get_status(ps_codec->pv_ref_buf_mgr, |
299 | 5.63k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id); |
300 | | |
301 | 5.63k | if((buf_status & (BUF_MGR_REF | BUF_MGR_CODEC | BUF_MGR_IO)) == 0) |
302 | 277 | { |
303 | 277 | ps_codec->as_ref_set[i].i4_pic_cnt = -1; |
304 | 277 | ps_codec->as_ref_set[i].i4_poc = 32768; |
305 | | |
306 | 277 | continue; |
307 | 277 | } |
308 | | |
309 | 5.35k | if(buf_status & BUF_MGR_REF) |
310 | 4.01k | { |
311 | 4.01k | ih264_buf_mgr_release(ps_codec->pv_svc_au_data_store_mgr, |
312 | 4.01k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id, BUF_MGR_REF); |
313 | | |
314 | 4.01k | ih264_buf_mgr_release(ps_codec->pv_ref_buf_mgr, |
315 | 4.01k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id, BUF_MGR_REF); |
316 | 4.01k | } |
317 | 5.35k | } |
318 | 8.51k | } |
319 | | |
320 | | /* |
321 | | * Mark all reference pic with unused buffers to be free |
322 | | * We need this step since each one, ie ref, recon io etc only unset their |
323 | | * respective flags. Hence we need to combine togather and mark the ref set |
324 | | * accordingly |
325 | | */ |
326 | 29.3k | pi4_ref_set_id[0] = -1; |
327 | | |
328 | 173k | for(i = 0; i < ps_codec->i4_ref_buf_cnt; i++) |
329 | 143k | { |
330 | 143k | if(ps_codec->as_ref_set[i].i4_pic_cnt == -1) |
331 | 90.2k | { |
332 | 90.2k | pi4_ref_set_id[0] = i; |
333 | 90.2k | continue; |
334 | 90.2k | } |
335 | | |
336 | 53.5k | buf_status = ih264_buf_mgr_get_status(ps_codec->pv_ref_buf_mgr, |
337 | 53.5k | ps_codec->as_ref_set[i].ps_pic_buf->i4_buf_id); |
338 | | |
339 | 53.5k | if((buf_status & (BUF_MGR_REF | BUF_MGR_CODEC | BUF_MGR_IO)) == 0) |
340 | 18.4k | { |
341 | 18.4k | ps_codec->as_ref_set[i].i4_pic_cnt = -1; |
342 | 18.4k | ps_codec->as_ref_set[i].i4_poc = 32768; |
343 | | |
344 | 18.4k | pi4_ref_set_id[0] = i; |
345 | 18.4k | } |
346 | 53.5k | } |
347 | | |
348 | | /* An asssert failure here means we donot have any free buffs */ |
349 | 29.3k | if(pi4_ref_set_id[0] < 0) |
350 | 1.13k | { |
351 | 1.13k | return IH264E_FAIL; |
352 | 1.13k | } |
353 | | |
354 | 28.2k | return IH264E_SUCCESS; |
355 | 29.3k | } |
356 | | |
357 | | /** |
358 | | ******************************************************************************* |
359 | | * |
360 | | * @brief |
361 | | * Validates SVC AU properties |
362 | | * |
363 | | * @param[in] ps_cfg |
364 | | * Cfg parameters |
365 | | * |
366 | | * @returns error code in conformance with 'IH264E_ERROR_T' |
367 | | * |
368 | | ******************************************************************************* |
369 | | */ |
370 | | WORD32 isvce_svc_au_props_validate(svc_inp_params_t *ps_svc_inp_params, UWORD32 u4_inp_wd, |
371 | | UWORD32 u4_inp_ht, UWORD32 u4_svc_comp_wd, |
372 | | UWORD32 u4_svc_comp_ht) |
373 | 14.8k | { |
374 | 14.8k | typedef struct |
375 | 14.8k | { |
376 | 14.8k | DOUBLE d_spatial_res_ratio; |
377 | | |
378 | 14.8k | UWORD8 u1_max_num_spatial_layers; |
379 | 14.8k | } spatial_layer_props_t; |
380 | | |
381 | 14.8k | UWORD8 i; |
382 | 14.8k | UWORD32 au4_svc_wd[MAX_NUM_SPATIAL_LAYERS]; |
383 | 14.8k | UWORD32 au4_svc_ht[MAX_NUM_SPATIAL_LAYERS]; |
384 | | |
385 | 14.8k | DOUBLE d_scaling_factor = ps_svc_inp_params->d_spatial_res_ratio; |
386 | 14.8k | UWORD8 u1_num_spatial_layers = ps_svc_inp_params->u1_num_spatial_layers; |
387 | 14.8k | const spatial_layer_props_t gas_valid_spatial_layer_props[] = {{1.5, 2}, {2, 3}}; |
388 | 14.8k | UWORD32 u4_error_code = IV_SUCCESS; |
389 | 14.8k | const UWORD8 u1_min_num_temporal_layers = 1; |
390 | 14.8k | const UWORD8 u1_min_num_spatial_layers = 1; |
391 | 14.8k | const UWORD8 u1_max_num_temporal_layers = MAX_NUM_TEMPORAL_LAYERS; |
392 | 14.8k | const UWORD8 u1_max_num_spatial_layers = MAX_NUM_SPATIAL_LAYERS; |
393 | 14.8k | const UWORD8 u1_num_valid_spatial_layer_props = |
394 | 14.8k | sizeof(gas_valid_spatial_layer_props) / sizeof(gas_valid_spatial_layer_props[0]); |
395 | | |
396 | 14.8k | if((ps_svc_inp_params->u1_num_temporal_layers < u1_min_num_temporal_layers) || |
397 | 14.8k | (ps_svc_inp_params->u1_num_temporal_layers > u1_max_num_temporal_layers)) |
398 | 0 | { |
399 | 0 | u4_error_code |= IH264E_INVALID_SVC_PARAMS | IH264E_INVALID_NUM_TEMPORAL_LAYERS; |
400 | 0 | } |
401 | | |
402 | 14.8k | if((ps_svc_inp_params->u1_num_spatial_layers < u1_min_num_spatial_layers) || |
403 | 14.8k | (ps_svc_inp_params->u1_num_spatial_layers > u1_max_num_spatial_layers)) |
404 | 0 | { |
405 | 0 | u4_error_code |= IH264E_INVALID_SVC_PARAMS | IH264E_INVALID_NUM_SPATIAL_LAYERS; |
406 | 0 | } |
407 | | |
408 | 14.8k | { |
409 | 14.8k | UWORD8 u1_is_input_ratio_valid = 0; |
410 | | |
411 | 24.5k | for(i = 0; i < u1_num_valid_spatial_layer_props; i++) |
412 | 24.5k | { |
413 | 24.5k | if(ps_svc_inp_params->d_spatial_res_ratio == |
414 | 24.5k | gas_valid_spatial_layer_props[i].d_spatial_res_ratio) |
415 | 14.8k | { |
416 | 14.8k | u1_is_input_ratio_valid = 1; |
417 | | |
418 | 14.8k | if(ps_svc_inp_params->u1_num_spatial_layers > |
419 | 14.8k | gas_valid_spatial_layer_props[i].u1_max_num_spatial_layers) |
420 | 1 | { |
421 | 1 | u4_error_code |= IH264E_INVALID_SVC_PARAMS | IH264E_INVALID_NUM_SPATIAL_LAYERS; |
422 | 1 | } |
423 | | |
424 | 14.8k | break; |
425 | 14.8k | } |
426 | 24.5k | } |
427 | | |
428 | 14.8k | if(!u1_is_input_ratio_valid) |
429 | 0 | { |
430 | 0 | u4_error_code |= IH264E_INVALID_SVC_PARAMS | IH264E_INVALID_SPATIAL_RES_RATIO; |
431 | 0 | } |
432 | 14.8k | } |
433 | | |
434 | 14.8k | if((u4_svc_comp_wd > SVCE_MAX_INP_DIM) || (u4_svc_comp_ht > SVCE_MAX_INP_DIM) || |
435 | 14.8k | ((u4_svc_comp_wd * u4_svc_comp_ht) > SVCE_MAX_INP_FRAME_SIZE) || |
436 | 14.8k | (u4_svc_comp_wd % 16 != 0) || (u4_svc_comp_ht % 16 != 0)) |
437 | 7 | { |
438 | 7 | u4_error_code |= IH264E_INVALID_SVC_INPUT_DIMENSIONS; |
439 | 7 | } |
440 | | |
441 | | /* Constraint from padding intrinsics */ |
442 | 14.8k | if((u4_svc_comp_wd - u4_inp_wd) % 16) |
443 | 1 | { |
444 | 1 | u4_error_code |= IH264E_INVALID_SVC_INPUT_DIMENSIONS; |
445 | 1 | } |
446 | | |
447 | | /* Constraint from 420p to 420sp conversion */ |
448 | 14.8k | if((u4_svc_comp_ht - u4_inp_ht) % 4) |
449 | 1 | { |
450 | 1 | u4_error_code |= IH264E_INVALID_SVC_INPUT_DIMENSIONS; |
451 | 1 | } |
452 | | |
453 | 14.8k | au4_svc_wd[u1_num_spatial_layers - 1] = u4_svc_comp_wd; |
454 | 14.8k | au4_svc_ht[u1_num_spatial_layers - 1] = u4_svc_comp_ht; |
455 | | |
456 | 34.2k | for(i = (u1_num_spatial_layers - 1); i > 0; i--) |
457 | 19.4k | { |
458 | 19.4k | au4_svc_wd[i - 1] = au4_svc_wd[i] / d_scaling_factor; |
459 | 19.4k | au4_svc_ht[i - 1] = au4_svc_ht[i] / d_scaling_factor; |
460 | | |
461 | 19.4k | if((au4_svc_wd[i - 1] * d_scaling_factor != au4_svc_wd[i]) || |
462 | 19.4k | (au4_svc_ht[i - 1] * d_scaling_factor != au4_svc_ht[i]) || |
463 | 19.4k | (au4_svc_ht[i - 1] % 16 != 0) || (au4_svc_ht[i - 1] % 16 != 0)) |
464 | 0 | { |
465 | 0 | u4_error_code |= IH264E_INVALID_SVC_INPUT_DIMENSIONS; |
466 | 0 | } |
467 | 19.4k | } |
468 | | |
469 | 14.8k | return u4_error_code; |
470 | 14.8k | } |
471 | | |
472 | | /** |
473 | | ******************************************************************************* |
474 | | * |
475 | | * @brief |
476 | | * Validates SVC input params |
477 | | * |
478 | | * @param[in] ps_cfg |
479 | | * Cfg parameters |
480 | | * |
481 | | * @returns error code in conformance with 'IH264E_ERROR_T' |
482 | | * |
483 | | ******************************************************************************* |
484 | | */ |
485 | | WORD32 isvce_svc_inp_params_validate(isvce_init_ip_t *ps_ip, isvce_cfg_params_t *ps_cfg) |
486 | 4.93k | { |
487 | 4.93k | UWORD32 u4_error_code = isvce_svc_au_props_validate(&ps_ip->s_svc_inp_params, ps_ip->u4_wd, |
488 | 4.93k | ps_ip->u4_ht, ps_cfg->u4_wd, ps_cfg->u4_ht); |
489 | | |
490 | 4.93k | if(ps_cfg->u4_enable_alt_ref) |
491 | 0 | { |
492 | 0 | u4_error_code |= IH264E_INVALID_ALT_REF_OPTION; |
493 | 0 | } |
494 | | |
495 | 4.93k | if(ps_cfg->u4_num_bframes) |
496 | 0 | { |
497 | 0 | u4_error_code |= IH264E_BFRAMES_NOT_SUPPORTED; |
498 | 0 | } |
499 | | |
500 | 4.93k | if(ps_cfg->e_slice_mode != IVE_SLICE_MODE_NONE) |
501 | 0 | { |
502 | 0 | u4_error_code |= IH264E_SLICE_TYPE_INPUT_INVALID; |
503 | 0 | } |
504 | | |
505 | 4.93k | if(ps_cfg->e_content_type != IV_PROGRESSIVE) |
506 | 0 | { |
507 | 0 | u4_error_code |= IH264E_CONTENT_TYPE_NOT_SUPPORTED; |
508 | 0 | } |
509 | | |
510 | 4.93k | if(ps_cfg->u4_weighted_prediction) |
511 | 0 | { |
512 | 0 | u4_error_code |= IH264E_WEIGHTED_PRED_NOT_SUPPORTED; |
513 | 0 | } |
514 | | |
515 | 4.93k | return u4_error_code; |
516 | 4.93k | } |
517 | | |
518 | | /** |
519 | | ******************************************************************************* |
520 | | * |
521 | | * @brief |
522 | | * Validates SVC RC params |
523 | | * |
524 | | * @param[in] ps_cfg |
525 | | * Cfg parameters |
526 | | * |
527 | | * @returns error code in conformance with 'IH264E_ERROR_T' |
528 | | * |
529 | | ******************************************************************************* |
530 | | */ |
531 | | static WORD32 isvce_svc_rc_params_validate(isvce_cfg_params_t *ps_cfg) |
532 | 13.1k | { |
533 | 13.1k | WORD32 i; |
534 | | |
535 | | /* RC requires total bits in a second to fit int32_t */ |
536 | 42.7k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
537 | 30.0k | { |
538 | 30.0k | if((((((UWORD64) ps_cfg->au4_target_bitrate[i]) * 1000llu) / ps_cfg->u4_tgt_frame_rate) * |
539 | 30.0k | ps_cfg->u4_idr_frm_interval) > ((UWORD64) INT32_MAX)) |
540 | 433 | { |
541 | 433 | return IH264E_BITRATE_NOT_SUPPORTED; |
542 | 433 | } |
543 | 30.0k | } |
544 | | |
545 | 12.7k | return IH264E_SUCCESS; |
546 | 13.1k | } |
547 | | |
548 | | /** |
549 | | ******************************************************************************* |
550 | | * |
551 | | * @brief |
552 | | * Validates SVC frame-level input params |
553 | | * |
554 | | * @param[in] ps_cfg |
555 | | * Cfg parameters |
556 | | * |
557 | | * @returns error code in conformance with 'IH264E_ERROR_T' |
558 | | * |
559 | | ******************************************************************************* |
560 | | */ |
561 | | WORD32 isvce_svc_frame_params_validate( |
562 | | rate_control_api_t *aps_rate_control_api[MAX_NUM_SPATIAL_LAYERS], UWORD8 u1_num_spatial_layers) |
563 | 43.1k | { |
564 | 43.1k | WORD32 i; |
565 | | |
566 | | /* RC requires total bits in a second to fit int32_t */ |
567 | 133k | for(i = 0; i < u1_num_spatial_layers; i++) |
568 | 91.7k | { |
569 | 91.7k | if((((UWORD64) irc_get_bits_per_frame(aps_rate_control_api[i])) * |
570 | 91.7k | irc_get_intra_frame_interval(aps_rate_control_api[i])) > ((UWORD64) INT32_MAX)) |
571 | 1.92k | { |
572 | 1.92k | return IH264E_BITRATE_NOT_SUPPORTED; |
573 | 1.92k | } |
574 | 91.7k | } |
575 | | |
576 | 41.2k | return IV_SUCCESS; |
577 | 43.1k | } |
578 | | |
579 | | /** |
580 | | ******************************************************************************* |
581 | | * |
582 | | * @brief |
583 | | * Used to get reference picture buffer size for a given level and |
584 | | * and padding used |
585 | | * |
586 | | * @param[in] ps_svc_inp_params |
587 | | * Struct containing SVC specific input params |
588 | | * |
589 | | * @param[in] i4_pic_size |
590 | | * Number of luma samples (Width * Height) |
591 | | * |
592 | | * @param[in] i4_level |
593 | | * Level |
594 | | * |
595 | | * @param[in] i4_horz_pad |
596 | | * Total padding used in horizontal direction |
597 | | * |
598 | | * @param[in] i4_vert_pad |
599 | | * Total padding used in vertical direction |
600 | | * |
601 | | * @param[in] i4_num_ref_frames |
602 | | * Num Reference Frames |
603 | | * |
604 | | * @param[in] i4_num_reorder_frames |
605 | | * Num Reorder Frames |
606 | | * |
607 | | * @returns Total picture buffer size |
608 | | * |
609 | | ******************************************************************************* |
610 | | */ |
611 | | WORD32 isvce_get_total_svc_au_buf_size(svc_inp_params_t *ps_svc_inp_params, WORD32 i4_pic_size, |
612 | | WORD32 i4_level, WORD32 i4_horz_pad, WORD32 i4_vert_pad, |
613 | | WORD32 i4_num_ref_frames, WORD32 i4_num_reorder_frames) |
614 | 9.87k | { |
615 | 9.87k | WORD32 i; |
616 | 9.87k | WORD32 size; |
617 | 9.87k | WORD32 num_luma_samples; |
618 | 9.87k | WORD32 lvl_idx; |
619 | 9.87k | WORD32 max_wd, min_ht; |
620 | 9.87k | WORD32 num_samples; |
621 | 9.87k | WORD32 max_num_bufs; |
622 | | |
623 | 9.87k | WORD32 pad = MAX(i4_horz_pad, i4_vert_pad); |
624 | 9.87k | DOUBLE d_svc_size_multiplier = 1; |
625 | | |
626 | 22.8k | for(i = 1; i < ps_svc_inp_params->u1_num_spatial_layers; i++) |
627 | 12.9k | { |
628 | 12.9k | d_svc_size_multiplier += 1. / pow(ps_svc_inp_params->d_spatial_res_ratio, i); |
629 | 12.9k | } |
630 | | |
631 | | /* |
632 | | * If i4_num_ref_frames and num_reorder_frmaes is specified |
633 | | * Use minimum value |
634 | | */ |
635 | 9.87k | max_num_bufs = (i4_num_ref_frames + i4_num_reorder_frames + MAX_CTXT_SETS + |
636 | 9.87k | ps_svc_inp_params->u1_num_temporal_layers); |
637 | | |
638 | | /* Get i4_level index */ |
639 | 9.87k | lvl_idx = ih264e_get_lvl_idx(i4_level); |
640 | | |
641 | | /* Maximum number of luma samples in a picture at given i4_level */ |
642 | 9.87k | num_luma_samples = gai4_ih264_max_luma_pic_size[lvl_idx]; |
643 | 9.87k | num_luma_samples = MAX(num_luma_samples, i4_pic_size); |
644 | | |
645 | | /* Account for chroma */ |
646 | 9.87k | num_samples = num_luma_samples * 3 / 2; |
647 | | |
648 | | /* Maximum width of luma samples in a picture at given i4_level */ |
649 | 9.87k | max_wd = gai4_ih264_max_wd_ht[lvl_idx]; |
650 | | |
651 | | /* Minimum height of luma samples in a picture at given i4_level */ |
652 | 9.87k | min_ht = gai4_ih264_min_wd_ht[lvl_idx]; |
653 | | |
654 | | /* Allocation is required for |
655 | | * (Wd + i4_horz_pad) * (Ht + i4_vert_pad) * (2 * max_dpb_size + 1) |
656 | | * |
657 | | * Above expanded as |
658 | | * ((Wd * Ht) + (i4_horz_pad * i4_vert_pad) + Wd * i4_vert_pad + Ht * |
659 | | * i4_horz_pad) * (2 * max_dpb_size + 1) (Wd * Ht) * (2 * max_dpb_size + 1) + |
660 | | * ((i4_horz_pad * i4_vert_pad) + Wd * i4_vert_pad + Ht * i4_horz_pad) * (2 * |
661 | | * max_dpb_size + 1) Now max_dpb_size increases with smaller Wd and Ht, but Wd |
662 | | * * ht * max_dpb_size will still be lesser or equal to max_wd * max_ht * |
663 | | * dpb_size |
664 | | * |
665 | | * In the above equation (Wd * Ht) * (2 * max_dpb_size + 1) is accounted by |
666 | | * using num_samples * (2 * max_dpb_size + 1) below |
667 | | * |
668 | | * For the padded area use MAX(i4_horz_pad, i4_vert_pad) as pad |
669 | | * ((pad * pad) + pad * (Wd + Ht)) * (2 * max_dpb_size + 1) has to accounted |
670 | | * from the above for padding |
671 | | * |
672 | | * Since Width and Height can change worst Wd + Ht is when One of the |
673 | | * dimensions is max and other is min So use max_wd and min_ht |
674 | | */ |
675 | | |
676 | | /* Number of bytes in reference pictures */ |
677 | 9.87k | size = num_samples * max_num_bufs; |
678 | | |
679 | | /* Account for Spatial Layers */ |
680 | 9.87k | size = (WORD32) (size * d_svc_size_multiplier + 0.99); |
681 | | |
682 | | /* Account for padding area */ |
683 | 9.87k | size += ((pad * pad) + pad * (max_wd + min_ht)) * 3 / 2 * max_num_bufs * |
684 | 9.87k | ps_svc_inp_params->u1_num_spatial_layers; |
685 | | |
686 | 9.87k | size += ps_svc_inp_params->u1_num_spatial_layers * sizeof(yuv_buf_props_t); |
687 | | |
688 | 9.87k | return size; |
689 | 9.87k | } |
690 | | |
691 | | /** |
692 | | ******************************************************************************* |
693 | | * |
694 | | * @brief |
695 | | * Used to get size of buffers used for storing prediction data |
696 | | * |
697 | | * @param[in] ps_svc_inp_params |
698 | | * Struct containing SVC specific input params |
699 | | * |
700 | | * @param[in] i4_num_luma_samples |
701 | | * Number of luma samples (Width * Height) |
702 | | * |
703 | | * @returns Size of buffers used for storing prediction data |
704 | | * |
705 | | ******************************************************************************* |
706 | | */ |
707 | | UWORD32 isvce_get_total_svc_au_data_size(WORD32 i4_num_luma_samples, UWORD8 u1_num_spatial_layers, |
708 | | DOUBLE d_spatial_res_ratio) |
709 | 9.87k | { |
710 | 9.87k | WORD32 i; |
711 | | |
712 | 9.87k | UWORD32 u4_svc_au_data_size = 0; |
713 | | |
714 | 9.87k | u4_svc_au_data_size += u1_num_spatial_layers * sizeof(svc_layer_data_t); |
715 | | |
716 | 32.7k | for(i = 0; i < u1_num_spatial_layers; i++) |
717 | 22.8k | { |
718 | 22.8k | WORD32 i4_layer_luma_samples = |
719 | 22.8k | ((DOUBLE) i4_num_luma_samples) / pow(pow(d_spatial_res_ratio, i), 2) + 0.99; |
720 | 22.8k | WORD32 i4_num_mbs = i4_layer_luma_samples / (MB_SIZE * MB_SIZE); |
721 | | |
722 | | /* isvce_mb_info_t */ |
723 | 22.8k | u4_svc_au_data_size += i4_num_mbs * sizeof(isvce_mb_info_t); |
724 | | |
725 | | /* pu4_num_pus_in_mb */ |
726 | 22.8k | u4_svc_au_data_size += i4_num_mbs * sizeof(UWORD32); |
727 | 22.8k | } |
728 | | |
729 | 9.87k | return u4_svc_au_data_size; |
730 | 9.87k | } |
731 | | |
732 | | /** |
733 | | ******************************************************************************* |
734 | | * |
735 | | * @brief Function to add buffers to SVC AU Data Store Manager |
736 | | * |
737 | | * @param[in] ps_codec |
738 | | * Pointer to codec context |
739 | | * |
740 | | * @returns error status |
741 | | * |
742 | | ******************************************************************************* |
743 | | */ |
744 | | IH264E_ERROR_T isvce_svc_au_data_mgr_add_bufs(isvce_codec_t *ps_codec) |
745 | 9.77k | { |
746 | 9.77k | IH264_ERROR_T ret; |
747 | | |
748 | 9.77k | WORD32 i, j; |
749 | 9.77k | UWORD8 *pu1_buf; |
750 | | |
751 | 9.77k | svc_au_data_t *ps_svc_au_data = ps_codec->ps_svc_au_data_base; |
752 | | |
753 | 9.77k | WORD32 i4_max_dpb_size = ps_codec->i4_ref_buf_cnt; |
754 | 9.77k | WORD64 i8_alloc_mem_size = ps_codec->i4_svc_au_data_size; |
755 | 9.77k | WORD32 i4_num_luma_samples = ALIGN16(ps_codec->s_cfg.u4_wd) * ALIGN16(ps_codec->s_cfg.u4_ht); |
756 | 9.77k | UWORD8 u1_num_spatial_layers = ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers; |
757 | 9.77k | DOUBLE d_spatial_res_ratio = ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio; |
758 | | |
759 | 9.77k | ps_codec->ps_svc_au_data = ps_svc_au_data; |
760 | 9.77k | pu1_buf = (UWORD8 *) ps_svc_au_data; |
761 | 9.77k | pu1_buf += BUF_MGR_MAX_CNT * sizeof(ps_svc_au_data[0]); |
762 | | |
763 | 9.77k | i8_alloc_mem_size -= (BUF_MGR_MAX_CNT * sizeof(ps_svc_au_data[0])); |
764 | | |
765 | 9.77k | i = 0; |
766 | | |
767 | 57.2k | while(i < i4_max_dpb_size) |
768 | 47.4k | { |
769 | 47.4k | ps_svc_au_data->ps_svc_layer_data = (svc_layer_data_t *) pu1_buf; |
770 | 47.4k | pu1_buf += u1_num_spatial_layers * sizeof(ps_svc_au_data->ps_svc_layer_data[0]); |
771 | 47.4k | i8_alloc_mem_size -= u1_num_spatial_layers * sizeof(ps_svc_au_data->ps_svc_layer_data[0]); |
772 | | |
773 | 158k | for(j = u1_num_spatial_layers - 1; j >= 0; j--) |
774 | 110k | { |
775 | 110k | WORD32 i4_layer_luma_samples = |
776 | 110k | ((DOUBLE) i4_num_luma_samples) / |
777 | 110k | pow(pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - j), 2) + |
778 | 110k | 0.99; |
779 | 110k | WORD32 i4_num_mbs = i4_layer_luma_samples / (MB_SIZE * MB_SIZE); |
780 | | |
781 | 110k | ps_svc_au_data->ps_svc_layer_data[j].pu4_num_pus_in_mb = (UWORD32 *) pu1_buf; |
782 | 110k | pu1_buf += |
783 | 110k | i4_num_mbs * sizeof(ps_svc_au_data->ps_svc_layer_data[j].pu4_num_pus_in_mb[0]); |
784 | 110k | i8_alloc_mem_size -= |
785 | 110k | i4_num_mbs * sizeof(ps_svc_au_data->ps_svc_layer_data[j].pu4_num_pus_in_mb[0]); |
786 | | |
787 | 110k | ps_svc_au_data->ps_svc_layer_data[j].ps_mb_info = (isvce_mb_info_t *) pu1_buf; |
788 | 110k | pu1_buf += i4_num_mbs * sizeof(ps_svc_au_data->ps_svc_layer_data[j].ps_mb_info[0]); |
789 | 110k | i8_alloc_mem_size -= |
790 | 110k | i4_num_mbs * sizeof(ps_svc_au_data->ps_svc_layer_data[j].ps_mb_info[0]); |
791 | | |
792 | 110k | ASSERT(i8_alloc_mem_size >= 0); |
793 | 110k | } |
794 | | |
795 | 47.4k | if(i8_alloc_mem_size < 0) |
796 | 0 | { |
797 | 0 | ps_codec->i4_error_code = IH264E_INSUFFICIENT_MEM_MVBANK; |
798 | |
|
799 | 0 | return IH264E_INSUFFICIENT_MEM_MVBANK; |
800 | 0 | } |
801 | | |
802 | 47.4k | ret = |
803 | 47.4k | ih264_buf_mgr_add((buf_mgr_t *) ps_codec->pv_svc_au_data_store_mgr, ps_svc_au_data, i); |
804 | | |
805 | 47.4k | if(IH264_SUCCESS != ret) |
806 | 0 | { |
807 | 0 | ps_codec->i4_error_code = IH264E_BUF_MGR_ERROR; |
808 | |
|
809 | 0 | return IH264E_BUF_MGR_ERROR; |
810 | 0 | } |
811 | | |
812 | 47.4k | ps_svc_au_data++; |
813 | 47.4k | i++; |
814 | 47.4k | } |
815 | | |
816 | 9.77k | return IH264E_SUCCESS; |
817 | 9.77k | } |
818 | | |
819 | | /** |
820 | | ******************************************************************************* |
821 | | * |
822 | | * @brief |
823 | | * Function to initialize svc_au_buf_t structs add au buffers to |
824 | | * buffer manager in case of non-shared mode |
825 | | * |
826 | | * @param[in] ps_codec |
827 | | * Pointer to codec context |
828 | | * |
829 | | * @returns error status |
830 | | * |
831 | | ******************************************************************************* |
832 | | */ |
833 | | IH264E_ERROR_T isvce_svc_au_buf_mgr_add_bufs(isvce_codec_t *ps_codec) |
834 | 9.77k | { |
835 | 9.77k | WORD32 i, j; |
836 | 9.77k | WORD32 buf_ret; |
837 | | |
838 | 9.77k | svc_au_buf_t *ps_pic_buf = ps_codec->ps_pic_buf; |
839 | | |
840 | 9.77k | IH264E_ERROR_T ret = IH264E_SUCCESS; |
841 | | |
842 | 9.77k | WORD32 i4_max_dpb_size = ps_codec->i4_ref_buf_cnt; |
843 | 9.77k | WORD64 i8_alloc_mem_size = |
844 | 9.77k | ps_codec->i4_total_pic_buf_size - BUF_MGR_MAX_CNT * sizeof(ps_pic_buf[0]); |
845 | 9.77k | UWORD8 *pu1_buf = (UWORD8 *) ps_codec->ps_pic_buf; |
846 | 9.77k | UWORD8 u1_num_spatial_layers = ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers; |
847 | 9.77k | DOUBLE d_spatial_res_ratio = ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio; |
848 | | |
849 | 9.77k | pu1_buf += BUF_MGR_MAX_CNT * sizeof(svc_au_buf_t); |
850 | | |
851 | 57.2k | for(i = 0; i < i4_max_dpb_size; i++) |
852 | 47.4k | { |
853 | 47.4k | WORD32 i4_total_fpel_mem_size = 0; |
854 | | |
855 | 47.4k | ps_pic_buf->ps_layer_yuv_buf_props = (yuv_buf_props_t *) pu1_buf; |
856 | 47.4k | pu1_buf += u1_num_spatial_layers * sizeof(ps_pic_buf->ps_layer_yuv_buf_props[0]); |
857 | 47.4k | i8_alloc_mem_size -= u1_num_spatial_layers * sizeof(ps_pic_buf->ps_layer_yuv_buf_props[0]); |
858 | | |
859 | 47.4k | if(i8_alloc_mem_size < 0) |
860 | 0 | { |
861 | 0 | ps_codec->i4_error_code = IH264E_INSUFFICIENT_MEM_PICBUF; |
862 | 0 | return IH264E_INSUFFICIENT_MEM_PICBUF; |
863 | 0 | } |
864 | | |
865 | 158k | for(j = u1_num_spatial_layers - 1; j >= 0; j--) |
866 | 110k | { |
867 | 110k | WORD32 i4_layer_luma_wd = ((DOUBLE) ps_codec->s_cfg.u4_wd / |
868 | 110k | pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - j)) + |
869 | 110k | 0.99; |
870 | 110k | WORD32 i4_layer_luma_ht = ((DOUBLE) ps_codec->s_cfg.u4_ht / |
871 | 110k | pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - j)) + |
872 | 110k | 0.99; |
873 | 110k | WORD32 i4_layer_luma_samples = |
874 | 110k | (ALIGN16(i4_layer_luma_wd) + PAD_WD) * (i4_layer_luma_ht + PAD_HT); |
875 | 110k | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
876 | 110k | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
877 | 110k | WORD32 i4_layer_uv_samples = |
878 | 110k | (ALIGN16(i4_layer_uv_wd) + PAD_WD) * (i4_layer_uv_ht + PAD_HT); |
879 | | |
880 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].as_component_bufs[0].i4_data_stride = |
881 | 110k | ALIGN16(i4_layer_luma_wd) + PAD_WD; |
882 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].as_component_bufs[0].pv_data = |
883 | 110k | pu1_buf + |
884 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].as_component_bufs[0].i4_data_stride * |
885 | 110k | PAD_TOP + |
886 | 110k | PAD_LEFT; |
887 | | |
888 | 110k | pu1_buf += i4_layer_luma_samples; |
889 | | |
890 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].as_component_bufs[1].i4_data_stride = |
891 | 110k | ALIGN16(i4_layer_uv_wd) + PAD_WD; |
892 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].as_component_bufs[1].pv_data = |
893 | 110k | pu1_buf + |
894 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].as_component_bufs[1].i4_data_stride * |
895 | 110k | (PAD_TOP / 2) + |
896 | 110k | PAD_LEFT; |
897 | | |
898 | 110k | pu1_buf += i4_layer_uv_samples; |
899 | | |
900 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].u4_width = i4_layer_luma_wd; |
901 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].u4_height = i4_layer_luma_ht; |
902 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].u1_bit_depth = 8; |
903 | 110k | ps_pic_buf->ps_layer_yuv_buf_props[j].e_color_format = IV_YUV_420SP_UV; |
904 | | |
905 | 110k | i8_alloc_mem_size -= i4_layer_luma_samples + i4_layer_uv_samples; |
906 | 110k | i4_total_fpel_mem_size += i4_layer_luma_samples + i4_layer_uv_samples; |
907 | | |
908 | 110k | if(i8_alloc_mem_size < 0) |
909 | 23 | { |
910 | 23 | ps_codec->i4_error_code = IH264E_INSUFFICIENT_MEM_PICBUF; |
911 | 23 | return IH264E_INSUFFICIENT_MEM_PICBUF; |
912 | 23 | } |
913 | 110k | } |
914 | | |
915 | 47.4k | buf_ret = ih264_buf_mgr_add((buf_mgr_t *) ps_codec->pv_ref_buf_mgr, ps_pic_buf, i); |
916 | | |
917 | 47.4k | if(0 != buf_ret) |
918 | 0 | { |
919 | 0 | ps_codec->i4_error_code = IH264E_BUF_MGR_ERROR; |
920 | 0 | return IH264E_BUF_MGR_ERROR; |
921 | 0 | } |
922 | | |
923 | 47.4k | pu1_buf += (HPEL_PLANES_CNT - 1) * i4_total_fpel_mem_size; |
924 | 47.4k | ps_pic_buf++; |
925 | 47.4k | } |
926 | | |
927 | 9.75k | return ret; |
928 | 9.77k | } |
929 | | |
930 | | /** |
931 | | ******************************************************************************* |
932 | | * |
933 | | * @brief |
934 | | * Returns size of buffers for storing SVC input data |
935 | | * |
936 | | * @param[in] u1_num_spatial_layers |
937 | | * Num Spatial Layers |
938 | | * |
939 | | * @param[in] d_spatial_res_ratio |
940 | | * Resolution Ratio b/w spatial layers |
941 | | * |
942 | | * @param[in] u4_wd |
943 | | * Input Width |
944 | | * |
945 | | * @param[in] u4_ht |
946 | | * Input Height |
947 | | * |
948 | | * @returns Size of buffers |
949 | | * |
950 | | ******************************************************************************* |
951 | | */ |
952 | | UWORD32 isvce_get_svc_inp_buf_size(UWORD8 u1_num_spatial_layers, DOUBLE d_spatial_res_ratio, |
953 | | UWORD32 u4_wd, UWORD32 u4_ht) |
954 | 14.8k | { |
955 | 14.8k | padding_dims_t s_pad_dims; |
956 | | |
957 | 14.8k | UWORD32 i; |
958 | 14.8k | UWORD8 u1_filter_padding_size_x, u1_filter_padding_size_y; |
959 | | |
960 | 14.8k | UWORD32 u4_size = 0; |
961 | | |
962 | 14.8k | isvce_get_downscaler_padding_dims(&s_pad_dims); |
963 | | |
964 | 14.8k | u1_filter_padding_size_x = s_pad_dims.u1_left_pad_size + s_pad_dims.u1_right_pad_size; |
965 | | |
966 | 14.8k | u1_filter_padding_size_y = s_pad_dims.u1_top_pad_size + s_pad_dims.u1_bottom_pad_size; |
967 | | |
968 | 49.0k | for(i = 0; i < u1_num_spatial_layers; i++) |
969 | 34.2k | { |
970 | 34.2k | WORD32 i4_layer_luma_wd = ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, i)) + 0.99; |
971 | 34.2k | WORD32 i4_layer_luma_ht = ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, i)) + 0.99; |
972 | 34.2k | WORD32 i4_layer_luma_samples = |
973 | 34.2k | (ALIGN16(i4_layer_luma_wd) + PAD_WD + u1_filter_padding_size_x) * |
974 | 34.2k | (i4_layer_luma_ht + PAD_HT + u1_filter_padding_size_y); |
975 | 34.2k | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
976 | 34.2k | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
977 | | /* u1_filter_padding_size_x * 2 because U and V |
978 | | both need same amount of padding */ |
979 | 34.2k | WORD32 i4_layer_uv_samples = |
980 | 34.2k | (ALIGN16(i4_layer_uv_wd) + PAD_WD + u1_filter_padding_size_x * 2) * |
981 | 34.2k | (i4_layer_uv_ht + PAD_HT + u1_filter_padding_size_y); |
982 | | |
983 | 34.2k | u4_size += (i4_layer_luma_samples + i4_layer_uv_samples) * sizeof(UWORD8); |
984 | 34.2k | } |
985 | | |
986 | 14.8k | return SVC_MAX_NUM_INP_FRAMES * u4_size; |
987 | 14.8k | } |
988 | | |
989 | | /** |
990 | | ******************************************************************************* |
991 | | * |
992 | | * @brief |
993 | | * Function to initialize svc input buffers |
994 | | * |
995 | | * @param[in] ps_codec |
996 | | * Pointer to codec context |
997 | | * |
998 | | * @param[in] ps_mem_rec |
999 | | * Pointer to memory allocated for input buffers |
1000 | | * |
1001 | | ******************************************************************************* |
1002 | | */ |
1003 | | void isvce_svc_inp_buf_init(isvce_codec_t *ps_codec, iv_mem_rec_t *ps_mem_rec) |
1004 | 4.93k | { |
1005 | 4.93k | padding_dims_t s_pad_dims; |
1006 | | |
1007 | 4.93k | WORD32 i, j; |
1008 | 4.93k | UWORD8 u1_filter_padding_size_x, u1_filter_padding_size_y; |
1009 | | |
1010 | 4.93k | DOUBLE d_spatial_res_ratio = ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio; |
1011 | 4.93k | UWORD8 u1_num_spatial_layers = ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers; |
1012 | 4.93k | UWORD32 u4_wd = ps_codec->s_cfg.u4_wd; |
1013 | 4.93k | UWORD32 u4_ht = ps_codec->s_cfg.u4_ht; |
1014 | 4.93k | UWORD8 *pu1_buf = ps_mem_rec->pv_base; |
1015 | 4.93k | WORD64 i8_alloc_mem_size = |
1016 | 4.93k | isvce_get_svc_inp_buf_size(u1_num_spatial_layers, d_spatial_res_ratio, u4_wd, u4_ht); |
1017 | | |
1018 | 4.93k | isvce_get_downscaler_padding_dims(&s_pad_dims); |
1019 | | |
1020 | 4.93k | u1_filter_padding_size_x = s_pad_dims.u1_left_pad_size + s_pad_dims.u1_right_pad_size; |
1021 | | |
1022 | 4.93k | u1_filter_padding_size_y = s_pad_dims.u1_top_pad_size + s_pad_dims.u1_bottom_pad_size; |
1023 | | |
1024 | 14.8k | for(i = 0; i < SVC_MAX_NUM_INP_FRAMES; i++) |
1025 | 9.86k | { |
1026 | 9.86k | ps_codec->as_inp_list[i].s_svc_params = ps_codec->s_cfg.s_svc_params; |
1027 | | |
1028 | 32.6k | for(j = u1_num_spatial_layers - 1; j >= 0; j--) |
1029 | 22.8k | { |
1030 | 22.8k | WORD32 i4_layer_luma_wd = |
1031 | 22.8k | ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - j)) + 0.99; |
1032 | 22.8k | WORD32 i4_layer_luma_ht = |
1033 | 22.8k | ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - j)) + 0.99; |
1034 | 22.8k | WORD32 i4_layer_luma_samples = |
1035 | 22.8k | (ALIGN16(i4_layer_luma_wd) + PAD_WD + u1_filter_padding_size_x) * |
1036 | 22.8k | (i4_layer_luma_ht + PAD_HT + u1_filter_padding_size_y); |
1037 | 22.8k | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
1038 | 22.8k | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
1039 | | /* u1_filter_padding_size_x * 2 because U and V |
1040 | | both need same amount of padding */ |
1041 | 22.8k | WORD32 i4_layer_uv_samples = |
1042 | 22.8k | (ALIGN16(i4_layer_uv_wd) + PAD_WD + u1_filter_padding_size_x * 2) * |
1043 | 22.8k | (i4_layer_uv_ht + PAD_HT + u1_filter_padding_size_y); |
1044 | | |
1045 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].as_component_bufs[Y].i4_data_stride = |
1046 | 22.8k | ALIGN16(i4_layer_luma_wd) + PAD_WD + u1_filter_padding_size_x; |
1047 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].as_component_bufs[Y].pv_data = |
1048 | 22.8k | pu1_buf + |
1049 | 22.8k | ps_codec->as_inp_list[i] |
1050 | 22.8k | .as_layer_yuv_buf_props[j] |
1051 | 22.8k | .as_component_bufs[Y] |
1052 | 22.8k | .i4_data_stride * |
1053 | 22.8k | (PAD_TOP + s_pad_dims.u1_top_pad_size) + |
1054 | 22.8k | (PAD_LEFT + s_pad_dims.u1_left_pad_size); |
1055 | 22.8k | pu1_buf += i4_layer_luma_samples * sizeof(UWORD8); |
1056 | 22.8k | i8_alloc_mem_size -= i4_layer_luma_samples * sizeof(UWORD8); |
1057 | | |
1058 | 22.8k | ps_codec->as_inp_list[i] |
1059 | 22.8k | .as_layer_yuv_buf_props[j] |
1060 | 22.8k | .as_component_bufs[UV] |
1061 | 22.8k | .i4_data_stride = ALIGN16(i4_layer_uv_wd) + PAD_WD + u1_filter_padding_size_x * 2; |
1062 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].as_component_bufs[UV].pv_data = |
1063 | 22.8k | pu1_buf + |
1064 | 22.8k | ps_codec->as_inp_list[i] |
1065 | 22.8k | .as_layer_yuv_buf_props[j] |
1066 | 22.8k | .as_component_bufs[UV] |
1067 | 22.8k | .i4_data_stride * |
1068 | 22.8k | (PAD_TOP + s_pad_dims.u1_top_pad_size) + |
1069 | 22.8k | (PAD_LEFT + s_pad_dims.u1_left_pad_size * 2); |
1070 | 22.8k | pu1_buf += i4_layer_uv_samples * sizeof(UWORD8); |
1071 | 22.8k | i8_alloc_mem_size -= i4_layer_uv_samples * sizeof(UWORD8); |
1072 | | |
1073 | | /* Chroma is always stored interleaved */ |
1074 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].as_component_bufs[V].pv_data = NULL; |
1075 | | |
1076 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].u1_bit_depth = 8; |
1077 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].e_color_format = IV_YUV_420SP_UV; |
1078 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].u4_width = i4_layer_luma_wd; |
1079 | 22.8k | ps_codec->as_inp_list[i].as_layer_yuv_buf_props[j].u4_height = i4_layer_luma_ht; |
1080 | | |
1081 | 22.8k | ASSERT(i8_alloc_mem_size >= 0); |
1082 | 22.8k | } |
1083 | 9.86k | } |
1084 | 4.93k | } |
1085 | | |
1086 | | void isvce_init_svc_dimension(isvce_inp_buf_t *ps_inp) |
1087 | 4.93k | { |
1088 | 4.93k | WORD32 i; |
1089 | | |
1090 | 4.93k | UWORD8 u1_num_spatial_layers = ps_inp->s_svc_params.u1_num_spatial_layers; |
1091 | 4.93k | DOUBLE d_spatial_res_ratio = ps_inp->s_svc_params.d_spatial_res_ratio; |
1092 | 4.93k | UWORD32 u4_wd = ps_inp->s_inp_props.s_raw_buf.au4_wd[Y]; |
1093 | 4.93k | UWORD32 u4_ht = ps_inp->s_inp_props.s_raw_buf.au4_ht[Y]; |
1094 | | |
1095 | 16.3k | for(i = 0; i < u1_num_spatial_layers; i++) |
1096 | 11.4k | { |
1097 | 11.4k | ps_inp->as_layer_yuv_buf_props[i].u4_width = |
1098 | 11.4k | ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
1099 | 11.4k | ps_inp->as_layer_yuv_buf_props[i].u4_height = |
1100 | 11.4k | ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
1101 | 11.4k | } |
1102 | 4.93k | } |
1103 | | |
1104 | | /** |
1105 | | ******************************************************************************* |
1106 | | * |
1107 | | * @brief |
1108 | | * Pads input buf as assumed by the downscaler filter |
1109 | | * |
1110 | | * @param[in] ps_codec |
1111 | | * Pointer to codec ctxt |
1112 | | * |
1113 | | * @param[in] ps_inp |
1114 | | * Pointer to svc input buffer |
1115 | | * |
1116 | | * @param[in] u1_svc_layer_index |
1117 | | * SVC layer index of the buffer |
1118 | | * |
1119 | | ******************************************************************************* |
1120 | | */ |
1121 | | |
1122 | | static void isvce_pad_buf_for_filtering(isvce_codec_t *ps_codec, isvce_inp_buf_t *ps_inp, |
1123 | | UWORD8 u1_svc_layer_index) |
1124 | 95.3k | { |
1125 | 95.3k | padding_dims_t s_pad_dims; |
1126 | | |
1127 | 95.3k | UWORD8 *pu1_buf; |
1128 | 95.3k | UWORD32 u4_buf_width, u4_buf_height; |
1129 | | |
1130 | 95.3k | UWORD8 u1_pad_left_size; |
1131 | 95.3k | UWORD8 u1_pad_right_size; |
1132 | 95.3k | UWORD8 u1_pad_top_size; |
1133 | 95.3k | UWORD8 u1_pad_bottom_size; |
1134 | 95.3k | UWORD8 u1_filter_padding_size_x; |
1135 | 95.3k | UWORD8 u1_filter_padding_size_chroma_x; |
1136 | | |
1137 | 95.3k | ASSERT(ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].e_color_format == IV_YUV_420SP_UV); |
1138 | | |
1139 | 95.3k | isvce_get_downscaler_padding_dims(&s_pad_dims); |
1140 | | |
1141 | 95.3k | u1_pad_left_size = s_pad_dims.u1_left_pad_size; |
1142 | 95.3k | u1_pad_right_size = s_pad_dims.u1_right_pad_size; |
1143 | 95.3k | u1_pad_top_size = s_pad_dims.u1_top_pad_size; |
1144 | 95.3k | u1_pad_bottom_size = s_pad_dims.u1_bottom_pad_size; |
1145 | 95.3k | u1_filter_padding_size_x = u1_pad_left_size + u1_pad_right_size; |
1146 | 95.3k | u1_filter_padding_size_chroma_x = u1_filter_padding_size_x * 2; |
1147 | | |
1148 | 95.3k | u4_buf_width = ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].u4_width; |
1149 | | |
1150 | 95.3k | u4_buf_height = ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].u4_height; |
1151 | | |
1152 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1153 | 95.3k | .as_component_bufs[0] |
1154 | 95.3k | .pv_data); |
1155 | | |
1156 | 95.3k | ps_codec->pf_pad_left_luma( |
1157 | 95.3k | pu1_buf, |
1158 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[0].i4_data_stride, |
1159 | 95.3k | u4_buf_height, u1_pad_left_size); |
1160 | | |
1161 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1162 | 95.3k | .as_component_bufs[0] |
1163 | 95.3k | .pv_data); |
1164 | | |
1165 | 95.3k | pu1_buf += u4_buf_width; |
1166 | | |
1167 | 95.3k | ps_codec->pf_pad_right_luma( |
1168 | 95.3k | pu1_buf, |
1169 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[0].i4_data_stride, |
1170 | 95.3k | u4_buf_height, u1_pad_right_size); |
1171 | | |
1172 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1173 | 95.3k | .as_component_bufs[1] |
1174 | 95.3k | .pv_data); |
1175 | | |
1176 | 95.3k | ps_codec->pf_pad_left_chroma( |
1177 | 95.3k | pu1_buf, |
1178 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[1].i4_data_stride, |
1179 | 95.3k | u4_buf_height / 2, u1_pad_left_size * 2); |
1180 | | |
1181 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1182 | 95.3k | .as_component_bufs[1] |
1183 | 95.3k | .pv_data); |
1184 | | |
1185 | 95.3k | pu1_buf += u4_buf_width; |
1186 | | |
1187 | 95.3k | ps_codec->pf_pad_right_chroma( |
1188 | 95.3k | pu1_buf, |
1189 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[1].i4_data_stride, |
1190 | 95.3k | u4_buf_height / 2, u1_pad_right_size * 2); |
1191 | | |
1192 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1193 | 95.3k | .as_component_bufs[0] |
1194 | 95.3k | .pv_data) - |
1195 | 95.3k | u1_pad_left_size; |
1196 | | |
1197 | 95.3k | ps_codec->pf_pad_top( |
1198 | 95.3k | pu1_buf, |
1199 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[0].i4_data_stride, |
1200 | 95.3k | (u4_buf_width + u1_filter_padding_size_x), u1_pad_top_size); |
1201 | | |
1202 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1203 | 95.3k | .as_component_bufs[0] |
1204 | 95.3k | .pv_data) - |
1205 | 95.3k | u1_pad_left_size; |
1206 | | |
1207 | 95.3k | pu1_buf += |
1208 | 95.3k | (u4_buf_height * |
1209 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[0].i4_data_stride); |
1210 | | |
1211 | 95.3k | ps_codec->pf_pad_bottom( |
1212 | 95.3k | pu1_buf, |
1213 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[0].i4_data_stride, |
1214 | 95.3k | (u4_buf_width + u1_filter_padding_size_x), u1_pad_bottom_size); |
1215 | | |
1216 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1217 | 95.3k | .as_component_bufs[1] |
1218 | 95.3k | .pv_data) - |
1219 | 95.3k | u1_pad_left_size * 2; |
1220 | | |
1221 | 95.3k | ps_codec->pf_pad_top( |
1222 | 95.3k | pu1_buf, |
1223 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[1].i4_data_stride, |
1224 | 95.3k | (u4_buf_width + u1_filter_padding_size_chroma_x), u1_pad_top_size); |
1225 | | |
1226 | 95.3k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index] |
1227 | 95.3k | .as_component_bufs[1] |
1228 | 95.3k | .pv_data) - |
1229 | 95.3k | u1_pad_left_size * 2; |
1230 | | |
1231 | 95.3k | pu1_buf += |
1232 | 95.3k | ((u4_buf_height / 2) * |
1233 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[1].i4_data_stride); |
1234 | | |
1235 | 95.3k | ps_codec->pf_pad_bottom( |
1236 | 95.3k | pu1_buf, |
1237 | 95.3k | ps_inp->as_layer_yuv_buf_props[u1_svc_layer_index].as_component_bufs[1].i4_data_stride, |
1238 | 95.3k | (u4_buf_width + u1_filter_padding_size_chroma_x), u1_pad_bottom_size); |
1239 | 95.3k | } |
1240 | | |
1241 | | /** |
1242 | | ******************************************************************************* |
1243 | | * |
1244 | | * @brief |
1245 | | * Pads raw input to satisfy SVC compliant input dimensions |
1246 | | * |
1247 | | * @param[in] ps_codec |
1248 | | * Pointer to codec ctxt |
1249 | | * |
1250 | | * @param[in] ps_inp |
1251 | | * Pointer to svc input buffer |
1252 | | * |
1253 | | ******************************************************************************* |
1254 | | */ |
1255 | | |
1256 | | static void isvce_pad_input_to_svc_compliant_dims(isvce_codec_t *ps_codec, isvce_inp_buf_t *ps_inp) |
1257 | 65.7k | { |
1258 | 65.7k | UWORD8 *pu1_buf; |
1259 | 65.7k | UWORD32 u4_raw_input_wd, u4_raw_input_ht, u4_padded_width, u4_padded_height, u4_width_delta, |
1260 | 65.7k | u4_height_delta; |
1261 | 65.7k | UWORD8 u1_num_layers = ps_inp->s_svc_params.u1_num_spatial_layers; |
1262 | | |
1263 | 65.7k | ASSERT(ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].e_color_format == IV_YUV_420SP_UV); |
1264 | | |
1265 | 65.7k | u4_padded_width = ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].u4_width; |
1266 | 65.7k | u4_padded_height = ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].u4_height; |
1267 | 65.7k | u4_raw_input_wd = ps_inp->s_inp_props.s_raw_buf.au4_wd[0]; |
1268 | 65.7k | u4_raw_input_ht = ps_inp->s_inp_props.s_raw_buf.au4_ht[0]; |
1269 | 65.7k | u4_width_delta = u4_padded_width - u4_raw_input_wd; |
1270 | 65.7k | u4_height_delta = u4_padded_height - u4_raw_input_ht; |
1271 | | |
1272 | 65.7k | ASSERT(!(u4_width_delta & 1)); |
1273 | 65.7k | ASSERT(!(u4_height_delta & 1)); |
1274 | | |
1275 | 65.7k | if(u4_width_delta) |
1276 | 53.1k | { |
1277 | 53.1k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1278 | 53.1k | .as_component_bufs[0] |
1279 | 53.1k | .pv_data); |
1280 | | |
1281 | 53.1k | pu1_buf += ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1282 | 53.1k | .as_component_bufs[0] |
1283 | 53.1k | .i4_data_stride) * |
1284 | 53.1k | (u4_height_delta / 2)); |
1285 | | |
1286 | 53.1k | ps_codec->pf_pad_left_luma( |
1287 | 53.1k | pu1_buf, |
1288 | 53.1k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[0].i4_data_stride, |
1289 | 53.1k | u4_padded_height, u4_width_delta / 2); |
1290 | | |
1291 | 53.1k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1292 | 53.1k | .as_component_bufs[0] |
1293 | 53.1k | .pv_data); |
1294 | | |
1295 | 53.1k | pu1_buf += ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1296 | 53.1k | .as_component_bufs[0] |
1297 | 53.1k | .i4_data_stride) * |
1298 | 53.1k | (u4_height_delta / 2)); |
1299 | | |
1300 | 53.1k | pu1_buf += u4_raw_input_wd; |
1301 | | |
1302 | 53.1k | ps_codec->pf_pad_right_luma( |
1303 | 53.1k | pu1_buf, |
1304 | 53.1k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[0].i4_data_stride, |
1305 | 53.1k | u4_padded_height, u4_width_delta / 2); |
1306 | | |
1307 | 53.1k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1308 | 53.1k | .as_component_bufs[1] |
1309 | 53.1k | .pv_data); |
1310 | | |
1311 | 53.1k | pu1_buf += ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1312 | 53.1k | .as_component_bufs[1] |
1313 | 53.1k | .i4_data_stride) * |
1314 | 53.1k | (u4_height_delta / 4)); |
1315 | | |
1316 | 53.1k | ps_codec->pf_pad_left_chroma( |
1317 | 53.1k | pu1_buf, |
1318 | 53.1k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[1].i4_data_stride, |
1319 | 53.1k | u4_padded_height / 2, u4_width_delta / 2); |
1320 | | |
1321 | 53.1k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1322 | 53.1k | .as_component_bufs[1] |
1323 | 53.1k | .pv_data); |
1324 | | |
1325 | 53.1k | pu1_buf += ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1326 | 53.1k | .as_component_bufs[1] |
1327 | 53.1k | .i4_data_stride) * |
1328 | 53.1k | (u4_height_delta / 4)); |
1329 | | |
1330 | 53.1k | pu1_buf += u4_raw_input_wd; |
1331 | | |
1332 | 53.1k | ps_codec->pf_pad_right_chroma( |
1333 | 53.1k | pu1_buf, |
1334 | 53.1k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[1].i4_data_stride, |
1335 | 53.1k | u4_padded_height / 2, u4_width_delta / 2); |
1336 | 53.1k | } |
1337 | | |
1338 | 65.7k | if(u4_height_delta) |
1339 | 61.0k | { |
1340 | 61.0k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1341 | 61.0k | .as_component_bufs[0] |
1342 | 61.0k | .pv_data); |
1343 | | |
1344 | 61.0k | pu1_buf += ((ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1345 | 61.0k | .as_component_bufs[0] |
1346 | 61.0k | .i4_data_stride) * |
1347 | 61.0k | (u4_height_delta / 2)); |
1348 | | |
1349 | 61.0k | ps_codec->pf_pad_top( |
1350 | 61.0k | pu1_buf, |
1351 | 61.0k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[0].i4_data_stride, |
1352 | 61.0k | u4_padded_width, u4_height_delta / 2); |
1353 | | |
1354 | 61.0k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1355 | 61.0k | .as_component_bufs[0] |
1356 | 61.0k | .pv_data); |
1357 | | |
1358 | 61.0k | pu1_buf += ((ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1359 | 61.0k | .as_component_bufs[0] |
1360 | 61.0k | .i4_data_stride) * |
1361 | 61.0k | (u4_height_delta / 2)); |
1362 | | |
1363 | 61.0k | pu1_buf += |
1364 | 61.0k | (u4_raw_input_ht * |
1365 | 61.0k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[0].i4_data_stride); |
1366 | | |
1367 | 61.0k | ps_codec->pf_pad_bottom( |
1368 | 61.0k | pu1_buf, |
1369 | 61.0k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[0].i4_data_stride, |
1370 | 61.0k | u4_padded_width, u4_height_delta / 2); |
1371 | | |
1372 | 61.0k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1373 | 61.0k | .as_component_bufs[1] |
1374 | 61.0k | .pv_data); |
1375 | | |
1376 | 61.0k | pu1_buf += ((ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1377 | 61.0k | .as_component_bufs[1] |
1378 | 61.0k | .i4_data_stride) * |
1379 | 61.0k | (u4_height_delta / 4)); |
1380 | | |
1381 | 61.0k | ps_codec->pf_pad_top( |
1382 | 61.0k | pu1_buf, |
1383 | 61.0k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[1].i4_data_stride, |
1384 | 61.0k | u4_padded_width, u4_height_delta / 4); |
1385 | | |
1386 | 61.0k | pu1_buf = (UWORD8 *) (ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1387 | 61.0k | .as_component_bufs[1] |
1388 | 61.0k | .pv_data); |
1389 | | |
1390 | 61.0k | pu1_buf += ((ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1] |
1391 | 61.0k | .as_component_bufs[1] |
1392 | 61.0k | .i4_data_stride) * |
1393 | 61.0k | (u4_height_delta / 4)); |
1394 | | |
1395 | 61.0k | pu1_buf += |
1396 | 61.0k | ((u4_raw_input_ht / 2) * |
1397 | 61.0k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[1].i4_data_stride); |
1398 | | |
1399 | 61.0k | ps_codec->pf_pad_bottom( |
1400 | 61.0k | pu1_buf, |
1401 | 61.0k | ps_inp->as_layer_yuv_buf_props[u1_num_layers - 1].as_component_bufs[1].i4_data_stride, |
1402 | 61.0k | u4_padded_width, u4_height_delta / 4); |
1403 | 61.0k | } |
1404 | 65.7k | } |
1405 | | |
1406 | | /** |
1407 | | ******************************************************************************* |
1408 | | * |
1409 | | * @brief |
1410 | | * Format conversion and downsampling for deriving spatial layer inputs |
1411 | | * |
1412 | | * @param[in] ps_inp |
1413 | | * Pointer to input buffer |
1414 | | * |
1415 | | ******************************************************************************* |
1416 | | */ |
1417 | | void isvce_svc_inp_buf_populate(isvce_codec_t *ps_codec, isvce_inp_buf_t *ps_inp) |
1418 | 65.7k | { |
1419 | 65.7k | yuv_buf_props_t s_src_buf_props, s_dst_buf_props; |
1420 | | |
1421 | 65.7k | UWORD32 i; |
1422 | 65.7k | UWORD32 u4_blk_x, u4_blk_y; |
1423 | 65.7k | UWORD8 *pu1_planar_y, *pu1_planar_u, *pu1_planar_v, *pu1_semi_planar_y, *pu1_semi_planar_uv; |
1424 | 65.7k | UWORD8 *pu1_src_luma, *pu1_src_chroma, *pu1_dst_luma, *pu1_dst_chroma; |
1425 | 65.7k | UWORD32 u4_num_blocks_x, u4_num_blocks_y; |
1426 | 65.7k | UWORD32 u4_scaled_block_wd, u4_scaled_block_ht; |
1427 | 65.7k | UWORD32 u4_blk_wd_luma, u4_blk_ht_luma; |
1428 | | |
1429 | 65.7k | downscaler_ctxt_t *ps_scaler = &ps_codec->s_scaler; |
1430 | 65.7k | isa_dependent_fxns_t *ps_isa_dependent_fxns = &ps_codec->s_isa_dependent_fxns; |
1431 | 65.7k | mem_fxns_t *ps_mem_fxns = &ps_isa_dependent_fxns->s_mem_fxns; |
1432 | | |
1433 | 65.7k | const UWORD8 u1_num_yuv_components_420sp = NUM_SP_COMPONENTS; |
1434 | 65.7k | UWORD8 u1_num_spatial_layers = ps_inp->s_svc_params.u1_num_spatial_layers; |
1435 | 65.7k | UWORD32 u4_padded_width = ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1].u4_width; |
1436 | 65.7k | UWORD32 u4_padded_height = ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1].u4_height; |
1437 | 65.7k | UWORD32 u4_raw_input_wd = ps_inp->s_inp_props.s_raw_buf.au4_wd[0]; |
1438 | 65.7k | UWORD32 u4_raw_input_ht = ps_inp->s_inp_props.s_raw_buf.au4_ht[0]; |
1439 | 65.7k | UWORD32 u4_width_delta = u4_padded_width - u4_raw_input_wd; |
1440 | 65.7k | UWORD32 u4_height_delta = u4_padded_height - u4_raw_input_ht; |
1441 | | |
1442 | 65.7k | ASSERT(!(u4_width_delta & 1)); |
1443 | 65.7k | ASSERT(!(u4_height_delta & 1)); |
1444 | | |
1445 | 65.7k | ASSERT((ps_inp->s_inp_props.s_raw_buf.e_color_fmt == IV_YUV_420P) || |
1446 | 65.7k | (ps_inp->s_inp_props.s_raw_buf.e_color_fmt == IV_YUV_420SP_UV)); |
1447 | | |
1448 | | /* Check is input is valid */ |
1449 | 65.7k | if(!(ps_inp->s_inp_props.s_raw_buf.apv_bufs[0])) |
1450 | 0 | { |
1451 | 0 | ASSERT(0); |
1452 | | |
1453 | 0 | return; |
1454 | 0 | } |
1455 | | |
1456 | | /* Convert the input into semi-planar in case of other formats */ |
1457 | 65.7k | if(ps_inp->s_inp_props.s_raw_buf.e_color_fmt == IV_YUV_420P) |
1458 | 55.7k | { |
1459 | 55.7k | pu1_planar_y = (UWORD8 *) ps_inp->s_inp_props.s_raw_buf.apv_bufs[0]; |
1460 | 55.7k | pu1_planar_u = (UWORD8 *) ps_inp->s_inp_props.s_raw_buf.apv_bufs[1]; |
1461 | 55.7k | pu1_planar_v = (UWORD8 *) ps_inp->s_inp_props.s_raw_buf.apv_bufs[2]; |
1462 | | |
1463 | 55.7k | pu1_semi_planar_y = (UWORD8 *) ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1464 | 55.7k | .as_component_bufs[0] |
1465 | 55.7k | .pv_data; |
1466 | | |
1467 | 55.7k | pu1_semi_planar_uv = (UWORD8 *) ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1468 | 55.7k | .as_component_bufs[1] |
1469 | 55.7k | .pv_data; |
1470 | | |
1471 | 55.7k | pu1_semi_planar_y += |
1472 | 55.7k | ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1473 | 55.7k | .as_component_bufs[0] |
1474 | 55.7k | .i4_data_stride) * |
1475 | 55.7k | (u4_height_delta / 2)); |
1476 | | |
1477 | 55.7k | pu1_semi_planar_uv += |
1478 | 55.7k | ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1479 | 55.7k | .as_component_bufs[1] |
1480 | 55.7k | .i4_data_stride) * |
1481 | 55.7k | (u4_height_delta / 4)); |
1482 | | |
1483 | 55.7k | ps_codec->pf_ih264e_conv_420p_to_420sp( |
1484 | 55.7k | pu1_planar_y, pu1_planar_u, pu1_planar_v, pu1_semi_planar_y, pu1_semi_planar_uv, |
1485 | 55.7k | ps_inp->s_inp_props.s_raw_buf.au4_ht[0], ps_inp->s_inp_props.s_raw_buf.au4_wd[0], |
1486 | 55.7k | ps_inp->s_inp_props.s_raw_buf.au4_strd[0], ps_inp->s_inp_props.s_raw_buf.au4_strd[1], |
1487 | 55.7k | ps_inp->s_inp_props.s_raw_buf.au4_strd[2], |
1488 | 55.7k | ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1489 | 55.7k | .as_component_bufs[0] |
1490 | 55.7k | .i4_data_stride, |
1491 | 55.7k | ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1492 | 55.7k | .as_component_bufs[1] |
1493 | 55.7k | .i4_data_stride, |
1494 | 55.7k | 0); |
1495 | 55.7k | } |
1496 | 9.98k | else |
1497 | 9.98k | { |
1498 | 9.98k | UWORD32 u4_wd, u4_ht; |
1499 | 9.98k | UWORD8 u1_comp; |
1500 | 9.98k | UWORD32 au4_arr_dims[4]; |
1501 | 9.98k | UWORD8 *pu1_src, *pu1_dst; |
1502 | | |
1503 | 9.98k | au4_arr_dims[0] = ps_inp->s_inp_props.s_raw_buf.au4_wd[0]; |
1504 | 9.98k | au4_arr_dims[1] = ps_inp->s_inp_props.s_raw_buf.au4_ht[0]; |
1505 | 9.98k | au4_arr_dims[2] = ps_inp->s_inp_props.s_raw_buf.au4_wd[1]; |
1506 | 9.98k | au4_arr_dims[3] = ps_inp->s_inp_props.s_raw_buf.au4_ht[1]; |
1507 | | |
1508 | 29.9k | for(u1_comp = 0; u1_comp < u1_num_yuv_components_420sp; u1_comp++) |
1509 | 19.9k | { |
1510 | 19.9k | u4_wd = au4_arr_dims[u1_comp * 2]; |
1511 | 19.9k | u4_ht = au4_arr_dims[(u1_comp * 2) + 1]; |
1512 | | |
1513 | 19.9k | pu1_dst = (UWORD8 *) ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1514 | 19.9k | .as_component_bufs[u1_comp] |
1515 | 19.9k | .pv_data; |
1516 | | |
1517 | 19.9k | pu1_dst += |
1518 | 19.9k | ((u4_width_delta / 2) + (ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1519 | 19.9k | .as_component_bufs[u1_comp] |
1520 | 19.9k | .i4_data_stride) * |
1521 | 19.9k | ((u4_height_delta / 2) / (u1_comp + 1))); |
1522 | | |
1523 | 19.9k | pu1_src = ps_inp->s_inp_props.s_raw_buf.apv_bufs[u1_comp]; |
1524 | | |
1525 | 19.9k | ps_mem_fxns->pf_copy_2d(pu1_dst, |
1526 | 19.9k | ps_inp->as_layer_yuv_buf_props[u1_num_spatial_layers - 1] |
1527 | 19.9k | .as_component_bufs[u1_comp] |
1528 | 19.9k | .i4_data_stride, |
1529 | 19.9k | pu1_src, ps_inp->s_inp_props.s_raw_buf.au4_strd[u1_comp], u4_wd, |
1530 | 19.9k | u4_ht); |
1531 | 19.9k | } |
1532 | 9.98k | } |
1533 | | |
1534 | | /* Padding input to satisfy SVC constraints */ |
1535 | 65.7k | isvce_pad_input_to_svc_compliant_dims(ps_codec, ps_inp); |
1536 | | |
1537 | | /* Downscaling */ |
1538 | 161k | for(i = u1_num_spatial_layers - 1; i > 0; i--) |
1539 | 95.3k | { |
1540 | 95.3k | const UWORD32 u4_default_scaled_blk_wd = |
1541 | 95.3k | gu4_downscaler_blk_size / ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio + 0.5; |
1542 | 95.3k | const UWORD32 u4_default_scaled_blk_ht = |
1543 | 95.3k | gu4_downscaler_blk_size / ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio + 0.5; |
1544 | | |
1545 | 95.3k | isvce_pad_buf_for_filtering(ps_codec, ps_inp, i); |
1546 | | |
1547 | 95.3k | s_src_buf_props = ps_inp->as_layer_yuv_buf_props[i]; |
1548 | 95.3k | s_dst_buf_props = ps_inp->as_layer_yuv_buf_props[i - 1]; |
1549 | | |
1550 | 95.3k | u4_num_blocks_x = |
1551 | 95.3k | (s_src_buf_props.u4_width + (gu4_downscaler_blk_size - 1)) / gu4_downscaler_blk_size; |
1552 | | |
1553 | 95.3k | u4_num_blocks_y = |
1554 | 95.3k | (s_src_buf_props.u4_height + (gu4_downscaler_blk_size - 1)) / gu4_downscaler_blk_size; |
1555 | | |
1556 | 95.3k | pu1_src_luma = (UWORD8 *) s_src_buf_props.as_component_bufs[Y].pv_data; |
1557 | 95.3k | pu1_src_chroma = (UWORD8 *) s_src_buf_props.as_component_bufs[U].pv_data; |
1558 | 95.3k | pu1_dst_luma = (UWORD8 *) s_dst_buf_props.as_component_bufs[Y].pv_data; |
1559 | 95.3k | pu1_dst_chroma = (UWORD8 *) s_dst_buf_props.as_component_bufs[U].pv_data; |
1560 | | |
1561 | 284k | for(u4_blk_x = 0; u4_blk_x < u4_num_blocks_x; u4_blk_x++) |
1562 | 188k | { |
1563 | 695k | for(u4_blk_y = 0; u4_blk_y < u4_num_blocks_y; u4_blk_y++) |
1564 | 507k | { |
1565 | 507k | u4_blk_wd_luma = isvce_get_downscaler_blk_dims(s_src_buf_props.u4_width, u4_blk_x, |
1566 | 507k | gu4_downscaler_blk_size); |
1567 | | |
1568 | 507k | u4_blk_ht_luma = isvce_get_downscaler_blk_dims(s_src_buf_props.u4_height, u4_blk_y, |
1569 | 507k | gu4_downscaler_blk_size); |
1570 | | |
1571 | 507k | u4_scaled_block_wd = isvce_get_downscaler_blk_dims( |
1572 | 507k | s_dst_buf_props.u4_width, u4_blk_x, u4_default_scaled_blk_wd); |
1573 | | |
1574 | 507k | u4_scaled_block_ht = isvce_get_downscaler_blk_dims( |
1575 | 507k | s_dst_buf_props.u4_height, u4_blk_y, u4_default_scaled_blk_ht); |
1576 | | |
1577 | 507k | s_src_buf_props.as_component_bufs[Y].pv_data = |
1578 | 507k | pu1_src_luma + (u4_blk_x * gu4_downscaler_blk_size + |
1579 | 507k | u4_blk_y * gu4_downscaler_blk_size * |
1580 | 507k | s_src_buf_props.as_component_bufs[Y].i4_data_stride); |
1581 | | |
1582 | 507k | s_src_buf_props.as_component_bufs[U].pv_data = |
1583 | 507k | pu1_src_chroma + (u4_blk_x * gu4_downscaler_blk_size + |
1584 | 507k | u4_blk_y * (gu4_downscaler_blk_size / 2) * |
1585 | 507k | s_src_buf_props.as_component_bufs[U].i4_data_stride); |
1586 | | |
1587 | 507k | s_dst_buf_props.as_component_bufs[Y].pv_data = |
1588 | 507k | pu1_dst_luma + (u4_blk_x * u4_default_scaled_blk_wd + |
1589 | 507k | u4_blk_y * u4_default_scaled_blk_ht * |
1590 | 507k | s_dst_buf_props.as_component_bufs[Y].i4_data_stride); |
1591 | | |
1592 | 507k | s_dst_buf_props.as_component_bufs[U].pv_data = |
1593 | 507k | pu1_dst_chroma + (u4_blk_x * u4_default_scaled_blk_wd + |
1594 | 507k | u4_blk_y * (u4_default_scaled_blk_ht / 2) * |
1595 | 507k | s_dst_buf_props.as_component_bufs[U].i4_data_stride); |
1596 | | |
1597 | 507k | ASSERT(!(u4_scaled_block_wd & 1)); |
1598 | 507k | ASSERT(!(u4_scaled_block_ht & 1)); |
1599 | | |
1600 | 507k | isvce_process_downscaler(ps_scaler, &s_src_buf_props, &s_dst_buf_props, |
1601 | 507k | u4_blk_wd_luma, u4_blk_ht_luma); |
1602 | 507k | } |
1603 | 188k | } |
1604 | 95.3k | } |
1605 | | |
1606 | 65.7k | UNUSED(u4_scaled_block_wd); |
1607 | 65.7k | UNUSED(u4_scaled_block_ht); |
1608 | 65.7k | } |
1609 | | |
1610 | | /** |
1611 | | ******************************************************************************* |
1612 | | * |
1613 | | * @brief |
1614 | | * calculates the greatest common divisor between the two parameters. |
1615 | | * |
1616 | | ******************************************************************************* |
1617 | | */ |
1618 | | |
1619 | | static DOUBLE isvce_get_GCD(DOUBLE a, DOUBLE b) |
1620 | 22.0k | { |
1621 | 22.0k | if(b == 0) |
1622 | 5.51k | { |
1623 | 5.51k | return a; |
1624 | 5.51k | } |
1625 | | |
1626 | 16.5k | return isvce_get_GCD(b, fmod(a, b)); |
1627 | 22.0k | } |
1628 | | |
1629 | | /** |
1630 | | ******************************************************************************* |
1631 | | * |
1632 | | * @brief |
1633 | | * calculates the least common multiple between the two parameters |
1634 | | * |
1635 | | ******************************************************************************* |
1636 | | */ |
1637 | | |
1638 | 5.51k | static DOUBLE isvce_get_LCM(DOUBLE a, DOUBLE b) { return (a / isvce_get_GCD(a, b)) * b; } |
1639 | | |
1640 | | /** |
1641 | | ******************************************************************************* |
1642 | | * |
1643 | | * @brief |
1644 | | * sets the width and height in config structure to SVC compliant width and |
1645 | | * height |
1646 | | * |
1647 | | * @param[in] ps_cfg |
1648 | | * Pointer to config struct |
1649 | | * |
1650 | | * @param[in] u4_app_wd |
1651 | | * width of the YUV as read by the app |
1652 | | * |
1653 | | * @param[in] u4_app_ht |
1654 | | * height of the YUV as read by the app |
1655 | | * |
1656 | | ******************************************************************************* |
1657 | | */ |
1658 | | |
1659 | | void isvce_get_svc_compliant_dimensions(UWORD8 u1_num_spatial_layers, DOUBLE d_scaling_factor, |
1660 | | UWORD32 u4_wd, UWORD32 u4_ht, UWORD32 *pu4_svc_comp_wd, |
1661 | | UWORD32 *pu4_svc_comp_ht) |
1662 | 22.7k | { |
1663 | 22.7k | DOUBLE d_scaling_factor_power_num_layers_minus1 = 0; |
1664 | 22.7k | UWORD32 u4_constraint_offset = 0; |
1665 | | |
1666 | 22.7k | d_scaling_factor_power_num_layers_minus1 = pow(d_scaling_factor, u1_num_spatial_layers - 1); |
1667 | | |
1668 | 22.7k | if(fmod(16, d_scaling_factor_power_num_layers_minus1)) |
1669 | 5.51k | { |
1670 | 5.51k | u4_constraint_offset = |
1671 | 5.51k | (UWORD32) isvce_get_LCM(16, d_scaling_factor_power_num_layers_minus1); |
1672 | 5.51k | } |
1673 | 17.2k | else |
1674 | 17.2k | { |
1675 | 17.2k | u4_constraint_offset = (UWORD32) (16 * d_scaling_factor_power_num_layers_minus1); |
1676 | 17.2k | } |
1677 | | |
1678 | 22.7k | if(u4_wd % u4_constraint_offset) |
1679 | 10.1k | { |
1680 | 10.1k | *pu4_svc_comp_wd = u4_wd - ((u4_wd) % u4_constraint_offset) + u4_constraint_offset; |
1681 | 10.1k | } |
1682 | 12.6k | else |
1683 | 12.6k | { |
1684 | 12.6k | *pu4_svc_comp_wd = u4_wd; |
1685 | 12.6k | } |
1686 | | |
1687 | 22.7k | if(u4_ht % u4_constraint_offset) |
1688 | 11.8k | { |
1689 | 11.8k | *pu4_svc_comp_ht = u4_ht - ((u4_ht) % u4_constraint_offset) + u4_constraint_offset; |
1690 | 11.8k | } |
1691 | 10.9k | else |
1692 | 10.9k | { |
1693 | 10.9k | *pu4_svc_comp_ht = u4_ht; |
1694 | 10.9k | } |
1695 | 22.7k | } |
1696 | | |
1697 | | /** |
1698 | | ******************************************************************************* |
1699 | | * |
1700 | | * @brief |
1701 | | * Returns size of buffers for storing SVC layer nbr info |
1702 | | * |
1703 | | * @param[in] u1_num_spatial_layers |
1704 | | * Num Spatial Layers |
1705 | | * |
1706 | | * @param[in] d_spatial_res_ratio |
1707 | | * Resolution Ratio b/w spatial layers |
1708 | | * |
1709 | | * @param[in] u4_wd |
1710 | | * Input Width |
1711 | | * |
1712 | | * @returns Size of buffers |
1713 | | * |
1714 | | ******************************************************************************* |
1715 | | */ |
1716 | | UWORD32 isvce_get_svc_nbr_info_buf_size(UWORD8 u1_num_spatial_layers, DOUBLE d_spatial_res_ratio, |
1717 | | UWORD32 u4_wd, UWORD32 u4_ht) |
1718 | 14.8k | { |
1719 | 14.8k | UWORD32 i; |
1720 | | |
1721 | 14.8k | UWORD32 u4_size = 0; |
1722 | | |
1723 | 14.8k | ASSERT(1 == MAX_CTXT_SETS); |
1724 | | |
1725 | 14.8k | u4_size += MAX_PROCESS_CTXT * u1_num_spatial_layers * sizeof(nbr_info_t); |
1726 | | |
1727 | 49.0k | for(i = 0; i < u1_num_spatial_layers; i++) |
1728 | 34.2k | { |
1729 | 34.2k | WORD32 i4_layer_luma_wd = ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, i)) + 0.99; |
1730 | 34.2k | WORD32 i4_layer_luma_ht = ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, i)) + 0.99; |
1731 | 34.2k | WORD32 i4_num_mbs_in_row = i4_layer_luma_wd / MB_SIZE; |
1732 | 34.2k | WORD32 i4_num_mbs_in_col = i4_layer_luma_ht / MB_SIZE; |
1733 | | |
1734 | | /* ps_top_row_mb_info */ |
1735 | 34.2k | u4_size += (i4_num_mbs_in_row + 1) * i4_num_mbs_in_col * sizeof(isvce_mb_info_t); |
1736 | | |
1737 | | /* ps_left_mb_info */ |
1738 | 34.2k | u4_size += MAX_PROCESS_CTXT * sizeof(isvce_mb_info_t); |
1739 | | |
1740 | | /* ps_top_mb_intra_modes */ |
1741 | 34.2k | u4_size += (i4_num_mbs_in_row + 1) * i4_num_mbs_in_col * sizeof(mb_intra_modes_t); |
1742 | | |
1743 | | /* ps_left_mb_intra_modes */ |
1744 | 34.2k | u4_size += MAX_PROCESS_CTXT * sizeof(mb_intra_modes_t); |
1745 | 34.2k | } |
1746 | | |
1747 | 14.8k | return u4_size; |
1748 | 14.8k | } |
1749 | | |
1750 | | /** |
1751 | | ******************************************************************************* |
1752 | | * |
1753 | | * @brief |
1754 | | * Function to initialize svc nbr info buffers |
1755 | | * |
1756 | | * @param[in] ps_codec |
1757 | | * Pointer to codec context |
1758 | | * |
1759 | | * @param[in] ps_mem_rec |
1760 | | * Pointer to memory allocated for input buffers |
1761 | | * |
1762 | | ******************************************************************************* |
1763 | | */ |
1764 | | void isvce_svc_nbr_info_buf_init(isvce_codec_t *ps_codec, iv_mem_rec_t *ps_mem_rec) |
1765 | 4.93k | { |
1766 | 4.93k | WORD32 i, j; |
1767 | | |
1768 | 4.93k | DOUBLE d_spatial_res_ratio = ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio; |
1769 | 4.93k | UWORD8 u1_num_spatial_layers = ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers; |
1770 | 4.93k | UWORD32 u4_wd = ps_codec->s_cfg.u4_wd; |
1771 | 4.93k | UWORD32 u4_ht = ps_codec->s_cfg.u4_ht; |
1772 | | |
1773 | 4.93k | UWORD8 *pu1_buf = ps_mem_rec->pv_base; |
1774 | 4.93k | WORD64 i8_alloc_mem_size = |
1775 | 4.93k | isvce_get_svc_nbr_info_buf_size(u1_num_spatial_layers, d_spatial_res_ratio, u4_wd, u4_ht); |
1776 | | |
1777 | 4.93k | ASSERT(1 == MAX_CTXT_SETS); |
1778 | | |
1779 | 44.4k | for(i = 0; i < MAX_PROCESS_CTXT; i++) |
1780 | 39.4k | { |
1781 | 39.4k | ps_codec->as_process[i].s_nbr_info_base.ps_layer_nbr_info = (nbr_info_t *) pu1_buf; |
1782 | 39.4k | pu1_buf += u1_num_spatial_layers * |
1783 | 39.4k | sizeof(ps_codec->as_process[i].s_nbr_info_base.ps_layer_nbr_info[0]); |
1784 | 39.4k | i8_alloc_mem_size -= u1_num_spatial_layers * |
1785 | 39.4k | sizeof(ps_codec->as_process[i].s_nbr_info_base.ps_layer_nbr_info[0]); |
1786 | | |
1787 | 130k | for(j = u1_num_spatial_layers - 1; j >= 0; j--) |
1788 | 91.2k | { |
1789 | 91.2k | ps_codec->as_process[i].s_nbr_info_base.ps_layer_nbr_info[j].ps_left_mb_info = |
1790 | 91.2k | (isvce_mb_info_t *) pu1_buf; |
1791 | 91.2k | ps_codec->as_process[i].s_nbr_info.ps_left_mb_info = (isvce_mb_info_t *) pu1_buf; |
1792 | 91.2k | pu1_buf += sizeof(ps_codec->as_process[i].s_nbr_info.ps_left_mb_info[0]); |
1793 | 91.2k | i8_alloc_mem_size -= sizeof(ps_codec->as_process[i].s_nbr_info.ps_left_mb_info[0]); |
1794 | | |
1795 | 91.2k | ps_codec->as_process[i].s_nbr_info_base.ps_layer_nbr_info[j].ps_left_mb_intra_modes = |
1796 | 91.2k | (mb_intra_modes_t *) pu1_buf; |
1797 | 91.2k | ps_codec->as_process[i].s_nbr_info.ps_left_mb_intra_modes = |
1798 | 91.2k | (mb_intra_modes_t *) pu1_buf; |
1799 | 91.2k | pu1_buf += sizeof(ps_codec->as_process[i].s_nbr_info.ps_left_mb_intra_modes[0]); |
1800 | 91.2k | i8_alloc_mem_size -= |
1801 | 91.2k | sizeof(ps_codec->as_process[i].s_nbr_info.ps_left_mb_intra_modes[0]); |
1802 | 91.2k | } |
1803 | | |
1804 | 39.4k | ASSERT(i8_alloc_mem_size >= 0); |
1805 | 39.4k | } |
1806 | | |
1807 | 16.3k | for(i = u1_num_spatial_layers - 1; i >= 0; i--) |
1808 | 11.4k | { |
1809 | 11.4k | isvce_mb_info_t *ps_top_mb_info; |
1810 | 11.4k | mb_intra_modes_t *ps_top_intra_modes; |
1811 | | |
1812 | 11.4k | WORD32 i4_layer_luma_wd = |
1813 | 11.4k | ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
1814 | 11.4k | WORD32 i4_layer_luma_ht = |
1815 | 11.4k | ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
1816 | 11.4k | WORD32 i4_num_mbs_in_row = i4_layer_luma_wd / MB_SIZE; |
1817 | 11.4k | WORD32 i4_num_mbs_in_col = i4_layer_luma_ht / MB_SIZE; |
1818 | | |
1819 | 11.4k | ps_top_mb_info = (isvce_mb_info_t *) pu1_buf; |
1820 | 11.4k | pu1_buf += (i4_num_mbs_in_row + 1) * i4_num_mbs_in_col * sizeof(ps_top_mb_info[0]); |
1821 | 11.4k | i8_alloc_mem_size -= |
1822 | 11.4k | (i4_num_mbs_in_row + 1) * i4_num_mbs_in_col * sizeof(ps_top_mb_info[0]); |
1823 | | |
1824 | 11.4k | ps_top_intra_modes = (mb_intra_modes_t *) pu1_buf; |
1825 | 11.4k | pu1_buf += (i4_num_mbs_in_row + 1) * i4_num_mbs_in_col * sizeof(ps_top_intra_modes[0]); |
1826 | 11.4k | i8_alloc_mem_size -= |
1827 | 11.4k | (i4_num_mbs_in_row + 1) * i4_num_mbs_in_col * sizeof(ps_top_intra_modes[0]); |
1828 | | |
1829 | 102k | for(j = 0; j < MAX_PROCESS_CTXT; j++) |
1830 | 91.2k | { |
1831 | 91.2k | ps_codec->as_process[j].s_nbr_info_base.ps_layer_nbr_info[i].ps_top_row_mb_info = |
1832 | 91.2k | ps_top_mb_info; |
1833 | 91.2k | ps_codec->as_process[j].s_nbr_info.ps_top_row_mb_info = NULL; |
1834 | | |
1835 | 91.2k | ps_codec->as_process[j].s_nbr_info_base.ps_layer_nbr_info[i].ps_top_mb_intra_modes = |
1836 | 91.2k | ps_top_intra_modes; |
1837 | 91.2k | ps_codec->as_process[j].s_nbr_info.ps_top_mb_intra_modes = NULL; |
1838 | 91.2k | } |
1839 | | |
1840 | 11.4k | ASSERT(i8_alloc_mem_size >= 0); |
1841 | 11.4k | } |
1842 | 4.93k | } |
1843 | | |
1844 | | /** |
1845 | | ******************************************************************************* |
1846 | | * |
1847 | | * @brief |
1848 | | * isvce_codec_t and proc_t initialisations for an Access Unit |
1849 | | * |
1850 | | * @par Description: |
1851 | | * Before beginning to encode the frame, the current function initializes all |
1852 | | * the ctxts (proc, entropy, me, ...) basing on the input configured params. |
1853 | | * It locates space for storing recon in the encoder picture buffer set, fetches |
1854 | | * reference frame from encoder picture buffer set. Calls RC pre-enc to get |
1855 | | * qp and pic type for the current frame. Queues proc jobs so that |
1856 | | * the other threads can begin encoding. In brief, this function sets up the |
1857 | | * tone for the entire encoder. |
1858 | | * |
1859 | | * @param[in] ps_codec |
1860 | | * Pointer to codec context |
1861 | | * |
1862 | | * @param[in] ps_inp_buf |
1863 | | * Pointer to input buffer context |
1864 | | * |
1865 | | * @returns error_status |
1866 | | * |
1867 | | * @remarks |
1868 | | * |
1869 | | * |
1870 | | ******************************************************************************* |
1871 | | */ |
1872 | | IH264E_ERROR_T isvce_svc_au_init(isvce_codec_t *ps_codec, isvce_inp_buf_t *ps_inp_buf) |
1873 | 29.5k | { |
1874 | 29.5k | svc_au_buf_t *ps_cur_pic; |
1875 | | |
1876 | 29.5k | WORD32 cur_mv_bank_buf_id; |
1877 | 29.5k | WORD32 cur_pic_buf_id; |
1878 | 29.5k | WORD32 ref_set_id; |
1879 | 29.5k | WORD32 i, j; |
1880 | | |
1881 | 29.5k | svc_au_data_t *ps_mv_buf = NULL; |
1882 | 29.5k | svc_au_buf_t *aps_ref_pic[MAX_REF_PIC_CNT] = {NULL, NULL}; |
1883 | 29.5k | svc_au_data_t *aps_mv_buf[MAX_REF_PIC_CNT] = {NULL, NULL}; |
1884 | | |
1885 | 29.5k | IH264E_ERROR_T error_status = IH264E_SUCCESS; |
1886 | 29.5k | PIC_TYPE_T *pic_type = &ps_codec->pic_type; |
1887 | | |
1888 | 29.5k | UWORD32 u4_timestamp_high = ps_inp_buf->s_inp_props.u4_timestamp_high; |
1889 | 29.5k | UWORD32 u4_timestamp_low = ps_inp_buf->s_inp_props.u4_timestamp_low; |
1890 | 29.5k | WORD32 ctxt_sel = ps_codec->i4_encode_api_call_cnt % MAX_CTXT_SETS; |
1891 | | /* Diamond search Iteration Max Cnt */ |
1892 | 29.5k | UWORD32 u4_num_layers = |
1893 | 29.5k | (ps_codec->s_cfg.u4_enc_speed_preset == IVE_FASTEST) ? (NUM_LAYERS >> 2) : NUM_LAYERS; |
1894 | 29.5k | UWORD32 u4_enable_fast_sad = ps_codec->s_cfg.u4_enable_fast_sad; |
1895 | | |
1896 | 29.5k | if((PIC_I == *pic_type) || (PIC_IDR == *pic_type)) |
1897 | 10.4k | { |
1898 | 10.4k | ps_codec->i4_slice_type = ISLICE; |
1899 | 10.4k | } |
1900 | 19.1k | else if(PIC_P == *pic_type) |
1901 | 19.1k | { |
1902 | 19.1k | ps_codec->i4_slice_type = PSLICE; |
1903 | 19.1k | } |
1904 | 0 | else if(PIC_B == *pic_type) |
1905 | 0 | { |
1906 | 0 | ps_codec->i4_slice_type = BSLICE; |
1907 | 0 | } |
1908 | | |
1909 | 29.5k | ps_codec->u4_is_curr_frm_ref = 0; |
1910 | 29.5k | ps_codec->u4_is_curr_frm_ref = (*pic_type != PIC_B); |
1911 | | |
1912 | 29.5k | if(ps_codec->s_cfg.u4_enable_alt_ref && (*pic_type == PIC_P) && |
1913 | 29.5k | (ps_codec->i4_pic_cnt % (ps_codec->s_cfg.u4_enable_alt_ref + 1))) |
1914 | 0 | { |
1915 | 0 | ps_codec->u4_is_curr_frm_ref = 0; |
1916 | 0 | } |
1917 | | |
1918 | 29.5k | ps_codec->u4_is_idr = 0; |
1919 | | |
1920 | 29.5k | if(PIC_IDR == *pic_type) |
1921 | 8.51k | { |
1922 | 8.51k | ps_codec->u4_is_idr = 1; |
1923 | | |
1924 | 8.51k | ps_codec->i4_frame_num = 0; |
1925 | | |
1926 | 8.51k | ps_codec->i4_idr_pic_id++; |
1927 | 8.51k | } |
1928 | | |
1929 | 29.5k | ps_codec->u4_disable_deblock_level = 1; |
1930 | | |
1931 | 29.5k | if(ps_codec->s_cfg.u4_disable_deblock_level == DISABLE_DEBLK_LEVEL_0) |
1932 | 15.9k | { |
1933 | 15.9k | ps_codec->u4_disable_deblock_level = 0; |
1934 | 15.9k | } |
1935 | 13.5k | else if(ps_codec->s_cfg.u4_disable_deblock_level == DISABLE_DEBLK_LEVEL_2) |
1936 | 5.37k | { |
1937 | 5.37k | if(ps_codec->u4_disable_deblock_level_cnt == DISABLE_DEBLOCK_INTERVAL || |
1938 | 5.37k | ps_codec->i4_slice_type == ISLICE) |
1939 | 2.08k | { |
1940 | 2.08k | ps_codec->u4_disable_deblock_level = 0; |
1941 | 2.08k | } |
1942 | 5.37k | } |
1943 | 8.22k | else if(ps_codec->s_cfg.u4_disable_deblock_level == DISABLE_DEBLK_LEVEL_3) |
1944 | 2.90k | { |
1945 | 2.90k | if(ps_codec->i4_slice_type == ISLICE) |
1946 | 1.30k | { |
1947 | 1.30k | ps_codec->u4_disable_deblock_level = 0; |
1948 | 1.30k | } |
1949 | 2.90k | } |
1950 | | |
1951 | 29.5k | if(ps_codec->u4_disable_deblock_level) |
1952 | 10.2k | { |
1953 | 10.2k | ps_codec->u4_disable_deblock_level_cnt++; |
1954 | 10.2k | } |
1955 | 19.3k | else |
1956 | 19.3k | { |
1957 | 19.3k | ps_codec->u4_disable_deblock_level_cnt = 0; |
1958 | 19.3k | } |
1959 | | |
1960 | 29.5k | if(ps_codec->u4_disable_deblock_level == 0) |
1961 | 19.3k | { |
1962 | 19.3k | if(ps_codec->s_cfg.e_slice_mode != IVE_SLICE_MODE_NONE) |
1963 | 0 | { |
1964 | 0 | ps_codec->i4_error_code = IH264E_SLICE_TYPE_INPUT_INVALID; |
1965 | |
|
1966 | 0 | return IH264E_SLICE_TYPE_INPUT_INVALID; |
1967 | 0 | } |
1968 | 19.3k | } |
1969 | | |
1970 | 29.5k | ps_codec->i4_error_code = IH264E_SUCCESS; |
1971 | | |
1972 | 29.5k | if(ps_codec->i4_gen_header) |
1973 | 0 | { |
1974 | 0 | sps_t *ps_sps = NULL; |
1975 | 0 | pps_t *ps_pps = NULL; |
1976 | 0 | subset_sps_t *ps_subset_sps = NULL; |
1977 | 0 | UWORD8 u1_profile_idc = IH264_PROFILE_BASELINE; |
1978 | |
|
1979 | 0 | if(ps_codec->as_process[ctxt_sel * MAX_PROCESS_THREADS].u1_spatial_layer_id > 0) |
1980 | 0 | { |
1981 | 0 | u1_profile_idc = IH264_SCALABLE_BASELINE; |
1982 | 0 | } |
1983 | |
|
1984 | 0 | ps_sps = ps_codec->ps_sps_base; |
1985 | 0 | isvce_populate_sps(ps_codec, ps_sps, 0, u1_profile_idc, ps_inp_buf, 0); |
1986 | |
|
1987 | 0 | ps_pps = ps_codec->ps_pps_base; |
1988 | 0 | isvce_populate_pps(ps_codec, ps_pps, 0, 0, 0); |
1989 | |
|
1990 | 0 | for(i = 1; i < ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers; i++) |
1991 | 0 | { |
1992 | 0 | ps_subset_sps = ps_codec->ps_subset_sps_base + i; |
1993 | 0 | isvce_populate_subset_sps(ps_codec, ps_subset_sps, i, ps_inp_buf, i); |
1994 | | |
1995 | | /* populate pps header */ |
1996 | 0 | ps_pps = ps_codec->ps_pps_base + i; |
1997 | 0 | isvce_populate_pps(ps_codec, ps_pps, i, i, i); |
1998 | 0 | } |
1999 | 0 | } |
2000 | | |
2001 | 29.5k | if(IH264E_SUCCESS != |
2002 | 29.5k | isvce_ref_list_refresh(ps_codec, aps_ref_pic, aps_mv_buf, &ref_set_id, pic_type[0])) |
2003 | 1.30k | { |
2004 | 1.30k | ps_codec->i4_error_code = IH264E_NO_FREE_PICBUF; |
2005 | | |
2006 | 1.30k | return IH264E_NO_FREE_PICBUF; |
2007 | 1.30k | } |
2008 | | |
2009 | 28.2k | { |
2010 | 28.2k | ps_mv_buf = (svc_au_data_t *) ih264_buf_mgr_get_next_free( |
2011 | 28.2k | (buf_mgr_t *) ps_codec->pv_svc_au_data_store_mgr, &cur_mv_bank_buf_id); |
2012 | | |
2013 | 28.2k | if(NULL == ps_mv_buf) |
2014 | 99 | { |
2015 | 99 | ps_codec->i4_error_code = IH264E_NO_FREE_MVBANK; |
2016 | | |
2017 | 99 | return IH264E_NO_FREE_MVBANK; |
2018 | 99 | } |
2019 | | |
2020 | 28.1k | if(ps_codec->u4_is_curr_frm_ref) |
2021 | 28.1k | { |
2022 | 28.1k | ih264_buf_mgr_set_status(ps_codec->pv_svc_au_data_store_mgr, cur_mv_bank_buf_id, |
2023 | 28.1k | BUF_MGR_REF); |
2024 | 28.1k | } |
2025 | | |
2026 | 28.1k | ps_mv_buf->i4_abs_poc = ps_codec->i4_abs_pic_order_cnt; |
2027 | 28.1k | ps_mv_buf->i4_buf_id = cur_mv_bank_buf_id; |
2028 | 28.1k | } |
2029 | | |
2030 | 0 | { |
2031 | 28.1k | ps_cur_pic = (svc_au_buf_t *) ih264_buf_mgr_get_next_free( |
2032 | 28.1k | (buf_mgr_t *) ps_codec->pv_ref_buf_mgr, &cur_pic_buf_id); |
2033 | | |
2034 | 28.1k | if(NULL == ps_cur_pic) |
2035 | 1 | { |
2036 | 1 | ps_codec->i4_error_code = IH264E_NO_FREE_PICBUF; |
2037 | | |
2038 | 1 | return IH264E_NO_FREE_PICBUF; |
2039 | 1 | } |
2040 | | |
2041 | 28.1k | if(ps_codec->u4_is_curr_frm_ref) |
2042 | 28.1k | { |
2043 | 28.1k | ih264_buf_mgr_set_status(ps_codec->pv_ref_buf_mgr, cur_pic_buf_id, BUF_MGR_REF); |
2044 | 28.1k | } |
2045 | | |
2046 | 28.1k | if(1 == ps_codec->s_cfg.u4_enable_recon) |
2047 | 5.03k | { |
2048 | 5.03k | ih264_buf_mgr_set_status(ps_codec->pv_ref_buf_mgr, cur_pic_buf_id, BUF_MGR_IO); |
2049 | 5.03k | } |
2050 | | |
2051 | 28.1k | ps_cur_pic->u4_timestamp_high = ps_inp_buf->s_inp_props.u4_timestamp_high; |
2052 | 28.1k | ps_cur_pic->u4_timestamp_low = ps_inp_buf->s_inp_props.u4_timestamp_low; |
2053 | | |
2054 | 28.1k | ps_cur_pic->i4_abs_poc = ps_codec->i4_poc; |
2055 | 28.1k | ps_cur_pic->i4_poc_lsb = ps_codec->i4_pic_order_cnt_lsb; |
2056 | 28.1k | ps_cur_pic->i4_frame_num = ps_codec->i4_frame_num; |
2057 | | |
2058 | 28.1k | ps_cur_pic->i4_buf_id = cur_pic_buf_id; |
2059 | | |
2060 | 28.1k | ps_cur_pic->i1_temporal_id = isvce_svc_temporal_id_compute( |
2061 | 28.1k | ps_codec->i4_poc, ps_codec->s_cfg.s_svc_params.u1_num_temporal_layers, pic_type[0]); |
2062 | 28.1k | } |
2063 | | |
2064 | | /* |
2065 | | * Add the current picture to ref list independent of the fact that it is used |
2066 | | * as reference or not. This is because, now recon is not in sync with output |
2067 | | * hence we may need the current recon after some delay. By adding it to ref |
2068 | | * list we can retrieve the recon any time we want. The information that it is |
2069 | | * used for ref can still be found by checking the buffer status of pic buf. |
2070 | | */ |
2071 | 0 | ps_codec->as_ref_set[ref_set_id].i4_pic_cnt = ps_codec->i4_pic_cnt; |
2072 | 28.1k | ps_codec->as_ref_set[ref_set_id].i4_poc = ps_codec->i4_poc; |
2073 | 28.1k | ps_codec->as_ref_set[ref_set_id].ps_svc_au_data = ps_mv_buf; |
2074 | 28.1k | ps_codec->as_ref_set[ref_set_id].ps_pic_buf = ps_cur_pic; |
2075 | | |
2076 | 28.1k | ps_codec->s_svc_ilp_data.ps_svc_au_data = ps_mv_buf; |
2077 | | |
2078 | 28.1k | { |
2079 | 28.1k | isvce_process_ctxt_t *ps_proc = NULL; |
2080 | | |
2081 | 28.1k | j = ctxt_sel * MAX_PROCESS_THREADS; |
2082 | | |
2083 | 253k | for(i = j; i < (j + MAX_PROCESS_THREADS); i++) |
2084 | 225k | { |
2085 | 225k | ps_proc = &ps_codec->as_process[i]; |
2086 | | |
2087 | 225k | ps_proc->s_svc_params = ps_codec->s_cfg.s_svc_params; |
2088 | | |
2089 | 225k | ps_proc->i4_frame_num = ps_codec->i4_frame_num; |
2090 | 225k | ps_proc->u4_is_idr = ps_codec->u4_is_idr; |
2091 | 225k | ps_proc->u4_idr_pic_id = ps_codec->i4_idr_pic_id; |
2092 | 225k | ps_proc->i4_slice_type = ps_codec->i4_slice_type; |
2093 | | |
2094 | 225k | ps_proc->u4_half_x_offset = 0; |
2095 | 225k | ps_proc->u4_half_y_offset = 0; |
2096 | 225k | ps_proc->u4_half_xy_offset = 0; |
2097 | | |
2098 | 225k | ps_proc->u4_disable_deblock_level = ps_codec->u4_disable_deblock_level; |
2099 | | |
2100 | 225k | ps_proc->i4_cur_mv_bank_buf_id = cur_mv_bank_buf_id; |
2101 | 225k | ps_proc->ps_cur_pic = ps_cur_pic; |
2102 | 225k | ps_proc->ps_cur_mv_buf = ps_mv_buf; |
2103 | | |
2104 | | /* |
2105 | | * pointer to ref picture |
2106 | | * 0 : Temporal back reference |
2107 | | * 1 : Temporal forward reference |
2108 | | */ |
2109 | 225k | ps_proc->aps_ref_pic[L0] = aps_ref_pic[L0]; |
2110 | 225k | ps_proc->aps_ref_pic[L1] = aps_ref_pic[L1]; |
2111 | 225k | if(ps_codec->pic_type == PIC_B) |
2112 | 0 | { |
2113 | 0 | ps_proc->aps_mv_buf[L0] = aps_mv_buf[L0]; |
2114 | 0 | ps_proc->aps_mv_buf[L1] = aps_mv_buf[L1]; |
2115 | 0 | } |
2116 | 225k | else |
2117 | 225k | { |
2118 | | /* |
2119 | | * Else is dummy since for non B pic we does not need this |
2120 | | * But an assignment here will help in not having a segfault |
2121 | | * when we calcualte colpic in P slices |
2122 | | */ |
2123 | 225k | ps_proc->aps_mv_buf[L0] = ps_mv_buf; |
2124 | 225k | ps_proc->aps_mv_buf[L1] = ps_mv_buf; |
2125 | 225k | } |
2126 | | |
2127 | 225k | ps_proc->s_inp_buf = ps_inp_buf[0]; |
2128 | | |
2129 | 225k | ps_proc->i4_encode_api_call_cnt = ps_codec->i4_encode_api_call_cnt; |
2130 | | |
2131 | 225k | ps_proc->i4_pic_cnt = ps_codec->i4_pic_cnt; |
2132 | | |
2133 | 225k | ps_proc->i4_error_code = 0; |
2134 | | |
2135 | 225k | { |
2136 | 225k | isvce_entropy_ctxt_t *ps_entropy = &ps_proc->s_entropy; |
2137 | | |
2138 | 225k | ps_entropy->i4_sof = 0; |
2139 | 225k | ps_entropy->i4_eof = 0; |
2140 | 225k | ps_entropy->ps_sps_base = ps_codec->ps_sps_base; |
2141 | 225k | ps_entropy->ps_pps_base = ps_codec->ps_pps_base; |
2142 | 225k | ps_entropy->pu1_slice_idx = ps_proc->pu1_slice_idx; |
2143 | 225k | ps_entropy->ps_svc_nalu_ext_base = ps_proc->ps_svc_nalu_ext_base; |
2144 | 225k | ps_entropy->ps_subset_sps_base = ps_proc->ps_subset_sps_base; |
2145 | 225k | ps_entropy->ps_slice_hdr_base = ps_proc->ps_slice_hdr_base; |
2146 | 225k | ps_entropy->ps_svc_slice_hdr_base = ps_proc->ps_svc_slice_hdr_base; |
2147 | 225k | ps_entropy->i4_abs_pic_order_cnt = ps_codec->i4_poc; |
2148 | | |
2149 | 225k | ps_entropy->i1_transform_8x8_mode_flag = 0; |
2150 | | |
2151 | 225k | ps_entropy->i4_error_code = IH264E_SUCCESS; |
2152 | 225k | ps_proc->s_entropy.u4_is_last = ps_inp_buf->s_inp_props.u4_is_last; |
2153 | 225k | ps_proc->s_entropy.i4_pic_cnt = ps_codec->i4_pic_cnt; |
2154 | | |
2155 | 225k | ps_entropy->u4_timestamp_low = u4_timestamp_low; |
2156 | 225k | ps_entropy->u4_timestamp_high = u4_timestamp_high; |
2157 | 225k | } |
2158 | | |
2159 | 225k | { |
2160 | 225k | isvce_me_ctxt_t *ps_me_ctxt = &ps_proc->s_me_ctxt; |
2161 | | |
2162 | 225k | ps_me_ctxt->ai2_srch_boundaries[0] = ps_codec->s_cfg.u4_srch_rng_x; |
2163 | 225k | ps_me_ctxt->ai2_srch_boundaries[1] = ps_codec->s_cfg.u4_srch_rng_y; |
2164 | | |
2165 | 225k | ps_me_ctxt->u4_half_x_offset = ps_proc->u4_half_x_offset; |
2166 | 225k | ps_me_ctxt->u4_half_y_offset = ps_proc->u4_half_y_offset; |
2167 | 225k | ps_me_ctxt->u4_half_xy_offset = ps_proc->u4_half_xy_offset; |
2168 | | |
2169 | 225k | ps_me_ctxt->u4_enable_fast_sad = u4_enable_fast_sad; |
2170 | 225k | ps_me_ctxt->u4_enable_hpel = ps_codec->s_cfg.u4_enable_hpel; |
2171 | 225k | ps_me_ctxt->u4_num_layers = u4_num_layers; |
2172 | 225k | ps_me_ctxt->u4_me_speed_preset = ps_codec->s_cfg.u4_me_speed_preset; |
2173 | | |
2174 | 225k | if((i == j) && (0 == ps_codec->i4_poc)) |
2175 | 4.95k | { |
2176 | 4.95k | isvce_init_mv_bits(ps_me_ctxt); |
2177 | 4.95k | } |
2178 | 225k | } |
2179 | | |
2180 | 225k | ps_proc->ps_ngbr_avbl = &(ps_proc->s_ngbr_avbl); |
2181 | 225k | } |
2182 | 28.1k | } |
2183 | | |
2184 | 28.1k | return error_status; |
2185 | 28.1k | } |
2186 | | |
2187 | | void isvce_init_quant_params(isvce_process_ctxt_t *ps_proc, WORD32 qp) |
2188 | 853k | { |
2189 | 853k | isvce_codec_t *ps_codec = ps_proc->ps_codec; |
2190 | | /* quant params */ |
2191 | 853k | quant_params_t *ps_qp_params; |
2192 | | |
2193 | | /* ptr to forward quant threshold matrix */ |
2194 | 853k | const UWORD16 *pu2_thres_mat = NULL; |
2195 | | |
2196 | | /* ptr to forward scale matrix */ |
2197 | 853k | const UWORD16 *pu2_scale_mat = gu2_quant_scale_matrix_4x4; |
2198 | | |
2199 | | /* ptr to inverse scale matrix */ |
2200 | 853k | const UWORD16 *pu2_iscale_mat = gau2_ih264_iquant_scale_matrix_4x4; |
2201 | | |
2202 | | /* temp var */ |
2203 | 853k | UWORD32 u4_qp[3], u4_qp_div6, u4_qp_mod6; |
2204 | 853k | COMPONENT_TYPE plane; |
2205 | 853k | WORD32 i; |
2206 | 853k | UWORD32 u4_satdq_t; |
2207 | 853k | const UWORD16 *pu2_smat; |
2208 | | |
2209 | | /********************************************************************/ |
2210 | | /* init quant params for all planes Y, U and V */ |
2211 | | /********************************************************************/ |
2212 | | /* luma qp */ |
2213 | 853k | u4_qp[Y] = qp; |
2214 | | |
2215 | | /* chroma qp |
2216 | | * TODO_LATER : just in case if the chroma planes use different qp's this |
2217 | | * needs to be corrected accordingly. |
2218 | | */ |
2219 | 853k | u4_qp[U] = gu1_qpc_fqpi[qp]; |
2220 | 853k | u4_qp[V] = gu1_qpc_fqpi[qp]; |
2221 | | |
2222 | 853k | plane = Y; |
2223 | 3.40M | while(plane <= V) |
2224 | 2.55M | { |
2225 | 2.55M | u4_qp_div6 = (u4_qp[plane] / 6); |
2226 | 2.55M | u4_qp_mod6 = (u4_qp[plane] % 6); |
2227 | | |
2228 | 2.55M | ps_qp_params = ps_proc->ps_qp_params[plane]; |
2229 | | |
2230 | | /* mb qp */ |
2231 | 2.55M | ps_qp_params->u1_mb_qp = u4_qp[plane]; |
2232 | | |
2233 | | /* mb qp / 6 */ |
2234 | 2.55M | ps_qp_params->u1_qp_div = u4_qp_div6; |
2235 | | |
2236 | | /* mb qp % 6 */ |
2237 | 2.55M | ps_qp_params->u1_qp_rem = u4_qp_mod6; |
2238 | | |
2239 | | /* QP bits */ |
2240 | 2.55M | ps_qp_params->u1_qbits = QP_BITS_h264_4x4 + u4_qp_div6; |
2241 | | |
2242 | | /* forward scale matrix */ |
2243 | 2.55M | ps_qp_params->pu2_scale_mat = pu2_scale_mat + (u4_qp_mod6 * 16); |
2244 | | |
2245 | | /* threshold matrix & weight for quantization */ |
2246 | 2.55M | pu2_thres_mat = gu2_forward_quant_threshold_4x4 + (u4_qp_mod6 * 16); |
2247 | 43.3M | for(i = 0; i < 16; i++) |
2248 | 40.7M | { |
2249 | 40.7M | ps_qp_params->pu2_thres_mat[i] = pu2_thres_mat[i] >> (8 - u4_qp_div6); |
2250 | 40.7M | ps_qp_params->pu2_weigh_mat[i] = 16; |
2251 | 40.7M | } |
2252 | | |
2253 | | /* qp dependent rounding constant */ |
2254 | 2.55M | ps_qp_params->u4_dead_zone = gu4_forward_quant_round_factor_4x4[u4_qp_div6]; |
2255 | | |
2256 | | /* slice dependent rounding constant */ |
2257 | 2.55M | if(ps_proc->i4_slice_type != ISLICE && ps_proc->i4_slice_type != SISLICE) |
2258 | 1.19M | { |
2259 | 1.19M | ps_qp_params->u4_dead_zone >>= 1; |
2260 | 1.19M | } |
2261 | | |
2262 | | /* SATQD threshold for zero block prediction */ |
2263 | 2.55M | if(ps_codec->s_cfg.u4_enable_satqd) |
2264 | 2.55M | { |
2265 | 2.55M | pu2_smat = ps_qp_params->pu2_scale_mat; |
2266 | | |
2267 | 2.55M | u4_satdq_t = ((1 << (ps_qp_params->u1_qbits)) - ps_qp_params->u4_dead_zone); |
2268 | | |
2269 | 2.55M | ps_qp_params->pu2_sad_thrsh[0] = u4_satdq_t / MAX(pu2_smat[3], pu2_smat[11]); |
2270 | 2.55M | ps_qp_params->pu2_sad_thrsh[1] = u4_satdq_t / MAX(pu2_smat[1], pu2_smat[9]); |
2271 | 2.55M | ps_qp_params->pu2_sad_thrsh[2] = u4_satdq_t / pu2_smat[15]; |
2272 | 2.55M | ps_qp_params->pu2_sad_thrsh[3] = u4_satdq_t / pu2_smat[7]; |
2273 | 2.55M | ps_qp_params->pu2_sad_thrsh[4] = u4_satdq_t / MAX(pu2_smat[12], pu2_smat[14]); |
2274 | 2.55M | ps_qp_params->pu2_sad_thrsh[5] = u4_satdq_t / MAX(pu2_smat[4], pu2_smat[6]); |
2275 | 2.55M | ps_qp_params->pu2_sad_thrsh[6] = u4_satdq_t / pu2_smat[13]; |
2276 | 2.55M | ps_qp_params->pu2_sad_thrsh[7] = u4_satdq_t / pu2_smat[5]; |
2277 | 2.55M | ps_qp_params->pu2_sad_thrsh[8] = |
2278 | 2.55M | u4_satdq_t / MAX(MAX3(pu2_smat[0], pu2_smat[2], pu2_smat[8]), pu2_smat[10]); |
2279 | 2.55M | } |
2280 | | |
2281 | | /* inverse scale matrix */ |
2282 | 2.55M | ps_qp_params->pu2_iscale_mat = pu2_iscale_mat + (u4_qp_mod6 * 16); |
2283 | | |
2284 | 2.55M | plane += 1; |
2285 | 2.55M | } |
2286 | 853k | } |
2287 | | |
2288 | | /** |
2289 | | ******************************************************************************* |
2290 | | * |
2291 | | * @brief |
2292 | | * isvce_codec_t and proc_t initialisations for an Access Unit |
2293 | | * |
2294 | | * @par Description: |
2295 | | * Before beginning to encode the frame, the current function initializes all |
2296 | | * the ctxts (proc, entropy, me, ...) basing on the input configured params. |
2297 | | * It locates space for storing recon in the encoder picture buffer set, fetches |
2298 | | * reference frame from encoder picture buffer set. Calls RC pre-enc to get |
2299 | | * qp and pic type for the current frame. Queues proc jobs so that |
2300 | | * the other threads can begin encoding. In brief, this function sets up the |
2301 | | * tone for the entire encoder. |
2302 | | * |
2303 | | * @param[in] ps_codec |
2304 | | * Pointer to codec context |
2305 | | * |
2306 | | * @param[in] ps_inp_buf |
2307 | | * Pointer to input buffer context |
2308 | | * |
2309 | | * @param[in] u1_spatial_layer_id |
2310 | | * Spatial Layer IDl 0 => Base layer |
2311 | | * |
2312 | | * @returns error_status |
2313 | | * |
2314 | | * @remarks |
2315 | | * |
2316 | | * |
2317 | | ******************************************************************************* |
2318 | | */ |
2319 | | IH264E_ERROR_T isvce_svc_layer_pic_init(isvce_codec_t *ps_codec, isvce_inp_buf_t *ps_inp_buf, |
2320 | | UWORD8 u1_spatial_layer_id) |
2321 | 61.1k | { |
2322 | 61.1k | WORD32 i; |
2323 | | |
2324 | 61.1k | IH264E_ERROR_T error_status = IH264E_SUCCESS; |
2325 | 61.1k | IH264_ERROR_T ret = IH264_SUCCESS; |
2326 | 61.1k | PIC_TYPE_T e_pic_type = ps_codec->pic_type; |
2327 | | |
2328 | 61.1k | ASSERT(MAX_CTXT_SETS == 1); |
2329 | | |
2330 | 550k | for(i = 0; i < MAX_PROCESS_THREADS; i++) |
2331 | 489k | { |
2332 | 489k | isvce_process_ctxt_t *ps_proc = &ps_codec->as_process[i]; |
2333 | 489k | isvce_entropy_ctxt_t *ps_entropy = &ps_proc->s_entropy; |
2334 | 489k | isvce_deblk_ctxt_t *ps_deblk = &ps_proc->s_deblk_ctxt; |
2335 | 489k | isvce_me_ctxt_t *ps_me_ctxt = &ps_proc->s_me_ctxt; |
2336 | 489k | svc_au_buf_t *ps_cur_pic = ps_proc->ps_cur_pic; |
2337 | 489k | svc_au_buf_t *aps_ref_pic[MAX_REF_PIC_CNT] = {ps_proc->aps_ref_pic[L0], |
2338 | 489k | ps_proc->aps_ref_pic[L1]}; |
2339 | | |
2340 | 489k | ps_proc->u1_spatial_layer_id = u1_spatial_layer_id; |
2341 | | |
2342 | 489k | ps_proc->s_src_pic_buf_props = ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id]; |
2343 | | |
2344 | 489k | ps_proc->s_rec_pic_buf_props = ps_cur_pic->ps_layer_yuv_buf_props[u1_spatial_layer_id]; |
2345 | | |
2346 | 489k | ASSERT(0 == (ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id].u4_width % MB_SIZE)); |
2347 | 489k | ASSERT(0 == (ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id].u4_height % MB_SIZE)); |
2348 | | |
2349 | 489k | ps_proc->i4_wd_mbs = |
2350 | 489k | ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id].u4_width / MB_SIZE; |
2351 | 489k | ps_proc->i4_ht_mbs = |
2352 | 489k | ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id].u4_height / MB_SIZE; |
2353 | | |
2354 | 489k | ps_proc->u1_frame_qp = ps_codec->au4_frame_qp[u1_spatial_layer_id]; |
2355 | | |
2356 | 489k | ps_proc->u1_mb_qp = ps_proc->u1_frame_qp; |
2357 | 489k | ps_entropy->ps_mb_qp_ctxt->u1_cur_mb_qp = ps_proc->u1_frame_qp; |
2358 | | |
2359 | 489k | isvce_init_quant_params(ps_proc, ps_proc->u1_frame_qp); |
2360 | | |
2361 | 489k | memset(&ps_proc->s_frame_info, 0, sizeof(frame_info_t)); |
2362 | | |
2363 | | /* row '-1' */ |
2364 | 489k | memset(ps_proc->pu1_proc_map - ps_proc->i4_wd_mbs, 1, |
2365 | 489k | ps_proc->i4_wd_mbs * sizeof(ps_proc->pu1_proc_map[0])); |
2366 | | |
2367 | | /* row 0 to ht in mbs */ |
2368 | 489k | memset(ps_proc->pu1_proc_map, 0, |
2369 | 489k | ps_proc->i4_wd_mbs * ps_proc->i4_ht_mbs * sizeof(ps_proc->pu1_proc_map[0])); |
2370 | | |
2371 | | /* row '-1' */ |
2372 | 489k | memset(ps_proc->pu1_deblk_map - ps_proc->i4_wd_mbs, 1, |
2373 | 489k | ps_proc->i4_wd_mbs * sizeof(ps_proc->pu1_deblk_map[0])); |
2374 | | |
2375 | | /* row 0 to ht in mbs */ |
2376 | 489k | memset(ps_proc->pu1_deblk_map, 0, |
2377 | 489k | ps_proc->i4_wd_mbs * ps_proc->i4_ht_mbs * sizeof(ps_proc->pu1_deblk_map[0])); |
2378 | | |
2379 | | /* row '-1' */ |
2380 | 489k | memset(ps_proc->pu1_me_map - ps_proc->i4_wd_mbs, 1, |
2381 | 489k | ps_proc->i4_wd_mbs * sizeof(ps_proc->pu1_me_map[0])); |
2382 | | |
2383 | | /* row 0 to ht in mbs */ |
2384 | 489k | memset(ps_proc->pu1_me_map, 0, |
2385 | 489k | ps_proc->i4_wd_mbs * ps_proc->i4_ht_mbs * sizeof(ps_proc->pu1_me_map[0])); |
2386 | | |
2387 | 489k | if(IVE_AIR_MODE_NONE != ps_codec->s_cfg.e_air_mode) |
2388 | 0 | { |
2389 | 0 | ps_codec->i4_air_pic_cnt = |
2390 | 0 | (ps_codec->i4_air_pic_cnt + 1) % ps_codec->s_cfg.u4_air_refresh_period; |
2391 | |
|
2392 | 0 | if(!ps_codec->i4_air_pic_cnt) |
2393 | 0 | { |
2394 | 0 | memset(ps_proc->pu1_is_intra_coded, 0, |
2395 | 0 | ps_proc->i4_wd_mbs * ps_proc->i4_ht_mbs * |
2396 | 0 | sizeof(ps_proc->pu1_is_intra_coded[0])); |
2397 | 0 | } |
2398 | 0 | } |
2399 | | |
2400 | 489k | if(ps_codec->s_cfg.e_slice_mode == IVE_SLICE_MODE_NONE) |
2401 | 489k | { |
2402 | 489k | memset(ps_proc->pu1_slice_idx, 0, |
2403 | 489k | ps_proc->i4_wd_mbs * ps_proc->i4_ht_mbs * sizeof(ps_proc->pu1_slice_idx[0])); |
2404 | 489k | } |
2405 | 0 | else if(ps_codec->s_cfg.e_slice_mode == IVE_SLICE_MODE_BLOCKS) |
2406 | 0 | { |
2407 | 0 | UWORD8 *pu1_slice_idx = ps_proc->pu1_slice_idx; |
2408 | 0 | WORD32 i4_mb_y = 0, slice_idx = 0, cnt; |
2409 | |
|
2410 | 0 | while(i4_mb_y < ps_proc->i4_ht_mbs) |
2411 | 0 | { |
2412 | 0 | if(i4_mb_y + (WORD32) ps_codec->s_cfg.u4_slice_param < ps_proc->i4_ht_mbs) |
2413 | 0 | { |
2414 | 0 | cnt = ps_codec->s_cfg.u4_slice_param * ps_proc->i4_wd_mbs; |
2415 | 0 | i4_mb_y += ps_codec->s_cfg.u4_slice_param; |
2416 | 0 | } |
2417 | 0 | else |
2418 | 0 | { |
2419 | 0 | cnt = (ps_proc->i4_ht_mbs - i4_mb_y) * ps_proc->i4_wd_mbs; |
2420 | 0 | i4_mb_y += (ps_proc->i4_ht_mbs - i4_mb_y); |
2421 | 0 | } |
2422 | |
|
2423 | 0 | memset(pu1_slice_idx, slice_idx, cnt * sizeof(pu1_slice_idx[0])); |
2424 | |
|
2425 | 0 | slice_idx++; |
2426 | 0 | pu1_slice_idx += cnt; |
2427 | 0 | } |
2428 | 0 | } |
2429 | | |
2430 | 489k | if((e_pic_type != PIC_IDR) && (e_pic_type != PIC_I)) |
2431 | 322k | { |
2432 | 322k | ps_proc->as_ref_pic_buf_props[L0] = |
2433 | 322k | aps_ref_pic[L0]->ps_layer_yuv_buf_props[u1_spatial_layer_id]; |
2434 | 322k | ps_proc->as_ref_pic_buf_props[L1] = |
2435 | 322k | aps_ref_pic[L1]->ps_layer_yuv_buf_props[u1_spatial_layer_id]; |
2436 | 322k | } |
2437 | | |
2438 | 489k | ps_entropy->i4_gen_header = ps_codec->i4_gen_header && (0 == u1_spatial_layer_id); |
2439 | 489k | ps_entropy->i4_gen_subset_sps = |
2440 | 489k | (ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers > 1) && ps_codec->i4_gen_header; |
2441 | | |
2442 | | /* row '-1' */ |
2443 | 489k | memset(ps_entropy->pu1_entropy_map - ps_proc->i4_wd_mbs, 1, |
2444 | 489k | ps_proc->i4_wd_mbs * sizeof(ps_entropy->pu1_entropy_map[0])); |
2445 | | |
2446 | | /* row 0 to ht in mbs */ |
2447 | 489k | memset(ps_entropy->pu1_entropy_map, 0, |
2448 | 489k | ps_proc->i4_wd_mbs * ps_proc->i4_ht_mbs * sizeof(ps_entropy->pu1_entropy_map[0])); |
2449 | | |
2450 | 489k | isvce_init_cabac_table(ps_entropy); |
2451 | | |
2452 | 489k | ps_entropy->i4_wd_mbs = ps_proc->i4_wd_mbs; |
2453 | 489k | ps_entropy->i4_ht_mbs = ps_proc->i4_ht_mbs; |
2454 | | |
2455 | 489k | ps_entropy->u1_entropy_coding_mode_flag = |
2456 | 489k | ((ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers > 1) && (0 == u1_spatial_layer_id)) |
2457 | 489k | ? CAVLC |
2458 | 489k | : ps_codec->s_cfg.u4_entropy_coding_mode; |
2459 | | |
2460 | 489k | ps_proc->s_entropy.pi4_mb_skip_run[0] = 0; |
2461 | | |
2462 | 489k | ps_entropy->u4_header_bits[MB_TYPE_INTRA] = 0; |
2463 | 489k | ps_entropy->u4_header_bits[MB_TYPE_INTER] = 0; |
2464 | 489k | ps_entropy->u4_residue_bits[MB_TYPE_INTRA] = 0; |
2465 | 489k | ps_entropy->u4_residue_bits[MB_TYPE_INTER] = 0; |
2466 | | |
2467 | 489k | ps_entropy->u1_spatial_layer_id = ps_proc->u1_spatial_layer_id; |
2468 | | |
2469 | 489k | ps_deblk->pu1_slice_idx = ps_proc->pu1_slice_idx; |
2470 | | |
2471 | 489k | ps_me_ctxt->u1_mb_qp = ps_codec->au4_frame_qp[u1_spatial_layer_id]; |
2472 | | |
2473 | 489k | { |
2474 | 489k | UWORD8 u1_min_qp; |
2475 | 489k | UWORD8 u1_max_qp; |
2476 | | |
2477 | 489k | svc_sub_pic_rc_ctxt_t *ps_sub_pic_rc_ctxt = ps_proc->ps_sub_pic_rc_ctxt; |
2478 | 489k | svc_sub_pic_rc_layer_variables_t *ps_layer_variables = |
2479 | 489k | &ps_sub_pic_rc_ctxt->s_sub_pic_rc_variables.s_layer_variables; |
2480 | | |
2481 | 489k | switch(ps_proc->i4_slice_type) |
2482 | 489k | { |
2483 | 166k | case ISLICE: |
2484 | 166k | { |
2485 | 166k | u1_min_qp = ps_codec->s_cfg.au4_i_qp_min[u1_spatial_layer_id]; |
2486 | 166k | u1_max_qp = ps_codec->s_cfg.au4_i_qp_max[u1_spatial_layer_id]; |
2487 | | |
2488 | 166k | break; |
2489 | 0 | } |
2490 | 322k | case PSLICE: |
2491 | 322k | { |
2492 | 322k | u1_min_qp = ps_codec->s_cfg.au4_p_qp_min[u1_spatial_layer_id]; |
2493 | 322k | u1_max_qp = ps_codec->s_cfg.au4_p_qp_max[u1_spatial_layer_id]; |
2494 | | |
2495 | 322k | break; |
2496 | 0 | } |
2497 | 0 | default: |
2498 | 0 | { |
2499 | 0 | u1_min_qp = ps_codec->s_cfg.au4_b_qp_min[u1_spatial_layer_id]; |
2500 | 0 | u1_max_qp = ps_codec->s_cfg.au4_b_qp_max[u1_spatial_layer_id]; |
2501 | |
|
2502 | 0 | break; |
2503 | 0 | } |
2504 | 489k | } |
2505 | | |
2506 | 489k | ps_layer_variables->i4_max_num_reference_frames = ps_codec->i4_max_num_reference_frames; |
2507 | 489k | ps_layer_variables->i4_slice_type = ps_proc->i4_slice_type; |
2508 | 489k | ps_layer_variables->i4_frame_num = ps_proc->i4_frame_num; |
2509 | 489k | ps_layer_variables->u1_frame_qp = ps_proc->u1_frame_qp; |
2510 | 489k | ps_layer_variables->u1_spatial_layer_id = u1_spatial_layer_id; |
2511 | 489k | ps_layer_variables->u1_min_qp = u1_min_qp; |
2512 | 489k | ps_layer_variables->u1_max_qp = u1_max_qp; |
2513 | | |
2514 | 489k | isvce_sub_pic_rc_ctxt_layer_init(ps_proc->ps_sub_pic_rc_ctxt); |
2515 | 489k | } |
2516 | 489k | } |
2517 | | |
2518 | 61.1k | { |
2519 | 61.1k | job_t s_job; |
2520 | | |
2521 | 61.1k | s_job.i4_cmd = CMD_PROCESS; |
2522 | 61.1k | s_job.i2_mb_cnt = |
2523 | 61.1k | ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id].u4_width / MB_SIZE; |
2524 | 61.1k | s_job.i2_mb_x = 0; |
2525 | | |
2526 | 343k | for(i = 0; i < (WORD32) (ps_inp_buf->as_layer_yuv_buf_props[u1_spatial_layer_id].u4_height / |
2527 | 343k | MB_SIZE); |
2528 | 282k | i++) |
2529 | 282k | { |
2530 | 282k | s_job.i2_mb_y = i; |
2531 | | |
2532 | 282k | ret = ih264_list_queue(ps_codec->pv_proc_jobq, &s_job, 1); |
2533 | | |
2534 | 282k | if(ret != IH264_SUCCESS) |
2535 | 0 | { |
2536 | 0 | ps_codec->i4_error_code = ret; |
2537 | |
|
2538 | 0 | return IH264E_FAIL; |
2539 | 0 | } |
2540 | 282k | } |
2541 | | |
2542 | | /* Once all the jobs are queued, terminate the queue */ |
2543 | | /* Since the threads are created and deleted in each call, terminating |
2544 | | here is not an issue */ |
2545 | 61.1k | ih264_list_terminate(ps_codec->pv_proc_jobq); |
2546 | 61.1k | } |
2547 | | |
2548 | 0 | ps_codec->i4_gen_header = 0; |
2549 | | |
2550 | 61.1k | return error_status; |
2551 | 61.1k | } |
2552 | | |
2553 | | /** |
2554 | | ******************************************************************************* |
2555 | | * |
2556 | | * @brief initialize process context. |
2557 | | * |
2558 | | * @par Description: |
2559 | | * Before dispatching the current job to process thread, the process context |
2560 | | * associated with the job is initialized. Usually every job aims to encode one |
2561 | | * row of mb's. Basing on the row indices provided by the job, the process |
2562 | | * context's buffer ptrs, slice indices and other elements that are necessary |
2563 | | * during core-coding are initialized. |
2564 | | * |
2565 | | * @param[in] ps_proc |
2566 | | * Pointer to the current process context |
2567 | | * |
2568 | | * @returns error status |
2569 | | * |
2570 | | * @remarks none |
2571 | | * |
2572 | | ******************************************************************************* |
2573 | | */ |
2574 | | IH264E_ERROR_T isvce_init_layer_proc_ctxt(isvce_process_ctxt_t *ps_proc) |
2575 | 277k | { |
2576 | 277k | WORD32 i4_mb_x, i4_mb_y; |
2577 | | |
2578 | 277k | isvce_codec_t *ps_codec = ps_proc->ps_codec; |
2579 | 277k | n_mb_process_ctxt_t *ps_n_mb_ctxt = &ps_proc->s_n_mb_ctxt; |
2580 | 277k | quant_params_t *ps_qp_params = ps_proc->ps_qp_params[0]; |
2581 | 277k | isvce_deblk_ctxt_t *ps_deblk = &ps_proc->s_deblk_ctxt; |
2582 | 277k | isvce_bs_ctxt_t *ps_bs = &(ps_deblk->s_bs_ctxt); |
2583 | 277k | svc_au_data_t *ps_cur_mv_buf = ps_proc->ps_cur_mv_buf; |
2584 | | |
2585 | 277k | i4_mb_x = ps_proc->i4_mb_x; |
2586 | 277k | i4_mb_y = ps_proc->i4_mb_y; |
2587 | | |
2588 | 277k | ASSERT((ps_codec->s_cfg.u4_wd - ps_codec->s_cfg.u4_disp_wd) == 0); |
2589 | 277k | ASSERT((ps_codec->s_cfg.u4_ht - ps_codec->s_cfg.u4_disp_ht) == 0); |
2590 | | |
2591 | 277k | ps_proc->i4_nmb_ntrpy = ps_proc->i4_wd_mbs; |
2592 | 277k | ps_proc->u4_nmb_me = 1; |
2593 | | |
2594 | 277k | ps_proc->s_src_buf_props = ps_proc->s_src_pic_buf_props; |
2595 | 277k | ps_proc->s_rec_buf_props = ps_proc->s_rec_pic_buf_props; |
2596 | 277k | ps_proc->as_ref_buf_props[0] = ps_proc->as_ref_pic_buf_props[0]; |
2597 | 277k | ps_proc->as_ref_buf_props[1] = ps_proc->as_ref_pic_buf_props[1]; |
2598 | | |
2599 | 277k | ps_proc->s_src_buf_props.as_component_bufs[0].pv_data = |
2600 | 277k | ((UWORD8 *) ps_proc->s_src_buf_props.as_component_bufs[0].pv_data) + (i4_mb_x * MB_SIZE) + |
2601 | 277k | ps_proc->s_src_buf_props.as_component_bufs[0].i4_data_stride * (i4_mb_y * MB_SIZE); |
2602 | 277k | ps_proc->s_src_buf_props.as_component_bufs[1].pv_data = |
2603 | 277k | ((UWORD8 *) ps_proc->s_src_pic_buf_props.as_component_bufs[1].pv_data) + |
2604 | 277k | (i4_mb_x * MB_SIZE) + |
2605 | 277k | ps_proc->s_src_buf_props.as_component_bufs[1].i4_data_stride * (i4_mb_y * BLK8x8SIZE); |
2606 | | |
2607 | 277k | ps_proc->s_rec_buf_props.as_component_bufs[0].pv_data = |
2608 | 277k | ((UWORD8 *) ps_proc->s_rec_buf_props.as_component_bufs[0].pv_data) + (i4_mb_x * MB_SIZE) + |
2609 | 277k | ps_proc->s_rec_buf_props.as_component_bufs[0].i4_data_stride * (i4_mb_y * MB_SIZE); |
2610 | 277k | ps_proc->s_rec_buf_props.as_component_bufs[1].pv_data = |
2611 | 277k | ((UWORD8 *) ps_proc->s_rec_buf_props.as_component_bufs[1].pv_data) + (i4_mb_x * MB_SIZE) + |
2612 | 277k | ps_proc->s_rec_buf_props.as_component_bufs[1].i4_data_stride * (i4_mb_y * BLK8x8SIZE); |
2613 | | |
2614 | 277k | ps_proc->as_ref_buf_props[0].as_component_bufs[0].pv_data = |
2615 | 277k | ((UWORD8 *) ps_proc->as_ref_buf_props[0].as_component_bufs[0].pv_data) + |
2616 | 277k | (i4_mb_x * MB_SIZE) + |
2617 | 277k | ps_proc->as_ref_buf_props[0].as_component_bufs[0].i4_data_stride * (i4_mb_y * MB_SIZE); |
2618 | 277k | ps_proc->as_ref_buf_props[0].as_component_bufs[1].pv_data = |
2619 | 277k | ((UWORD8 *) ps_proc->as_ref_buf_props[0].as_component_bufs[1].pv_data) + |
2620 | 277k | (i4_mb_x * MB_SIZE) + |
2621 | 277k | ps_proc->as_ref_buf_props[0].as_component_bufs[1].i4_data_stride * (i4_mb_y * BLK8x8SIZE); |
2622 | | |
2623 | 277k | ps_proc->as_ref_buf_props[1].as_component_bufs[0].pv_data = |
2624 | 277k | ((UWORD8 *) ps_proc->as_ref_buf_props[1].as_component_bufs[0].pv_data) + |
2625 | 277k | (i4_mb_x * MB_SIZE) + |
2626 | 277k | ps_proc->as_ref_buf_props[1].as_component_bufs[0].i4_data_stride * (i4_mb_y * MB_SIZE); |
2627 | 277k | ps_proc->as_ref_buf_props[1].as_component_bufs[1].pv_data = |
2628 | 277k | ((UWORD8 *) ps_proc->as_ref_buf_props[1].as_component_bufs[1].pv_data) + |
2629 | 277k | (i4_mb_x * MB_SIZE) + |
2630 | 277k | ps_proc->as_ref_buf_props[1].as_component_bufs[1].i4_data_stride * (i4_mb_y * BLK8x8SIZE); |
2631 | | |
2632 | 277k | ps_proc->pv_mb_coeff_data = |
2633 | 277k | ((UWORD8 *) ps_proc->pv_pic_mb_coeff_data) + i4_mb_y * ps_codec->u4_size_coeff_data; |
2634 | | |
2635 | 277k | ps_proc->pv_mb_header_data = |
2636 | 277k | ((UWORD8 *) ps_proc->pv_pic_mb_header_data) + i4_mb_y * ps_codec->u4_size_header_data; |
2637 | | |
2638 | 277k | ps_proc->i4_cur_slice_idx = ps_proc->pu1_slice_idx[i4_mb_y * ps_proc->i4_wd_mbs + i4_mb_x]; |
2639 | | |
2640 | 277k | ps_proc->ps_mb_info = |
2641 | 277k | ps_cur_mv_buf->ps_svc_layer_data[ps_proc->u1_spatial_layer_id].ps_mb_info + |
2642 | 277k | i4_mb_y * ps_proc->i4_wd_mbs; |
2643 | | |
2644 | 277k | ps_proc->ps_col_mb = |
2645 | 277k | ps_proc->aps_mv_buf[1]->ps_svc_layer_data[ps_proc->u1_spatial_layer_id].ps_mb_info + |
2646 | 277k | i4_mb_y * ps_proc->i4_wd_mbs; |
2647 | | |
2648 | 277k | { |
2649 | 277k | ps_proc->s_nbr_info.ps_top_row_mb_info = |
2650 | 277k | ps_proc->s_nbr_info_base.ps_layer_nbr_info[ps_proc->u1_spatial_layer_id] |
2651 | 277k | .ps_top_row_mb_info + |
2652 | 277k | (i4_mb_x + (i4_mb_y - 1) * ps_proc->i4_wd_mbs); |
2653 | | |
2654 | 277k | ps_proc->s_nbr_info.ps_top_mb_intra_modes = |
2655 | 277k | ps_proc->s_nbr_info_base.ps_layer_nbr_info[ps_proc->u1_spatial_layer_id] |
2656 | 277k | .ps_top_mb_intra_modes + |
2657 | 277k | (i4_mb_x + (i4_mb_y - 1) * ps_proc->i4_wd_mbs); |
2658 | 277k | } |
2659 | | |
2660 | 277k | ps_proc->pu4_mb_pu_cnt = |
2661 | 277k | ps_cur_mv_buf->ps_svc_layer_data[ps_proc->u1_spatial_layer_id].pu4_num_pus_in_mb + |
2662 | 277k | (i4_mb_y * ps_proc->i4_wd_mbs); |
2663 | | |
2664 | 277k | ps_proc->ps_mb_info->u2_mb_type = I16x16; |
2665 | | |
2666 | 277k | ps_proc->u4_lambda = gu1_qp0[ps_qp_params->u1_mb_qp]; |
2667 | | |
2668 | 277k | ps_proc->i4_mb_distortion = SHRT_MAX; |
2669 | | |
2670 | 277k | if(i4_mb_x == 0) |
2671 | 276k | { |
2672 | 276k | ps_proc->s_nbr_info.ps_left_mb_info[0].i4_mb_distortion = 0; |
2673 | 276k | } |
2674 | | |
2675 | 277k | ps_proc->i4_mb_cost = INT_MAX; |
2676 | | |
2677 | 277k | ps_deblk->i4_mb_x = ps_proc->i4_mb_x; |
2678 | | /* deblk lags the current mb proc by 1 row */ |
2679 | | /* NOTE: Intra prediction has to happen with non deblocked samples used as |
2680 | | * reference */ |
2681 | | /* Hence to deblk MB 0 of row 0, you have wait till MB 0 of row 1 is encoded. |
2682 | | */ |
2683 | | /* For simplicity, we chose to lag deblking by 1 Row wrt to proc */ |
2684 | 277k | ps_deblk->i4_mb_y = ps_proc->i4_mb_y - 1; |
2685 | | |
2686 | 277k | ps_deblk->s_rec_pic_buf_props = ps_proc->s_rec_pic_buf_props; |
2687 | | |
2688 | 277k | ps_bs->i4_mb_x = ps_proc->i4_mb_x; |
2689 | 277k | ps_bs->i4_mb_y = ps_proc->i4_mb_y; |
2690 | | |
2691 | 277k | ps_n_mb_ctxt->i4_mb_x = 0; |
2692 | 277k | ps_n_mb_ctxt->i4_mb_y = ps_deblk->i4_mb_y; |
2693 | 277k | ps_n_mb_ctxt->i4_n_mbs = ps_proc->i4_nmb_ntrpy; |
2694 | | |
2695 | 277k | return IH264E_SUCCESS; |
2696 | 277k | } |
2697 | | |
2698 | | /** |
2699 | | ******************************************************************************* |
2700 | | * |
2701 | | * @brief |
2702 | | * Returns size of buffers for storing SVC ILP data |
2703 | | * |
2704 | | * @param[in] u1_num_spatial_layers |
2705 | | * Num Spatial Layers |
2706 | | * |
2707 | | * @param[in] d_spatial_res_ratio |
2708 | | * Resolution Ratio b/w spatial layers |
2709 | | * |
2710 | | * @param[in] u4_wd |
2711 | | * Input Width |
2712 | | * |
2713 | | * @param[in] u4_ht |
2714 | | * Input Height |
2715 | | * |
2716 | | * @returns Size of buffers |
2717 | | * |
2718 | | ******************************************************************************* |
2719 | | */ |
2720 | | UWORD32 isvce_get_svc_ilp_buf_size(UWORD8 u1_num_spatial_layers, DOUBLE d_spatial_res_ratio, |
2721 | | UWORD32 u4_wd, UWORD32 u4_ht) |
2722 | 14.8k | { |
2723 | 14.8k | WORD32 i; |
2724 | | |
2725 | 14.8k | UWORD32 u4_size = 0; |
2726 | | |
2727 | 14.8k | if(u1_num_spatial_layers > 1) |
2728 | 12.1k | { |
2729 | | /* ps_intra_recon_bufs */ |
2730 | 12.1k | u4_size += u1_num_spatial_layers * sizeof(yuv_buf_props_t); |
2731 | | |
2732 | | /* ps_residual_bufs */ |
2733 | 12.1k | u4_size += u1_num_spatial_layers * sizeof(yuv_buf_props_t); |
2734 | | |
2735 | | /* aps_layer_resampler_props[Y] */ |
2736 | 12.1k | u4_size += u1_num_spatial_layers * sizeof(layer_resampler_props_t); |
2737 | | |
2738 | | /* aps_layer_resampler_props[UV] */ |
2739 | 12.1k | u4_size += u1_num_spatial_layers * sizeof(layer_resampler_props_t); |
2740 | | |
2741 | 43.6k | for(i = u1_num_spatial_layers - 1; i >= 0; i--) |
2742 | 31.5k | { |
2743 | 31.5k | WORD32 i4_layer_luma_wd = |
2744 | 31.5k | ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
2745 | 31.5k | WORD32 i4_layer_luma_ht = |
2746 | 31.5k | ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
2747 | 31.5k | WORD32 i4_layer_luma_samples = |
2748 | 31.5k | (ALIGN16(i4_layer_luma_wd) + PAD_WD) * (i4_layer_luma_ht + PAD_HT); |
2749 | 31.5k | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
2750 | 31.5k | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
2751 | 31.5k | WORD32 i4_layer_uv_samples = |
2752 | 31.5k | (ALIGN16(i4_layer_uv_wd) + PAD_WD) * (i4_layer_uv_ht + PAD_HT); |
2753 | | |
2754 | | /* ps_intra_recon_bufs */ |
2755 | 31.5k | u4_size += (i4_layer_luma_samples + i4_layer_uv_samples) * sizeof(UWORD8); |
2756 | | |
2757 | | /* ps_residual_bufs */ |
2758 | 31.5k | u4_size += (i4_layer_luma_samples + i4_layer_uv_samples) * sizeof(WORD16); |
2759 | 31.5k | } |
2760 | 12.1k | } |
2761 | 2.67k | else |
2762 | 2.67k | { |
2763 | 2.67k | WORD32 i4_layer_luma_wd = u4_wd; |
2764 | 2.67k | WORD32 i4_layer_luma_ht = u4_ht; |
2765 | 2.67k | WORD32 i4_layer_luma_samples = |
2766 | 2.67k | (ALIGN16(i4_layer_luma_wd) + PAD_WD) * (i4_layer_luma_ht + PAD_HT); |
2767 | 2.67k | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
2768 | 2.67k | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
2769 | 2.67k | WORD32 i4_layer_uv_samples = (ALIGN16(i4_layer_uv_wd) + PAD_WD) * (i4_layer_uv_ht + PAD_HT); |
2770 | | |
2771 | | /* ps_residual_bufs */ |
2772 | 2.67k | u4_size += sizeof(yuv_buf_props_t); |
2773 | | |
2774 | | /* ps_residual_bufs */ |
2775 | 2.67k | u4_size += (i4_layer_luma_samples + i4_layer_uv_samples) * sizeof(WORD16); |
2776 | 2.67k | } |
2777 | | |
2778 | 14.8k | return u4_size; |
2779 | 14.8k | } |
2780 | | |
2781 | | static void isvce_layer_resampler_props_init(layer_resampler_props_t *ps_layer_props, |
2782 | | DOUBLE d_spatial_res_ratio, UWORD32 u4_wd, |
2783 | | UWORD32 u4_ht, UWORD8 u1_level_idc, |
2784 | | UWORD8 u1_is_chroma) |
2785 | 12.9k | { |
2786 | 12.9k | const UWORD8 u1_ref_layer_field_pic_flag = 0; |
2787 | 12.9k | const UWORD8 u1_field_pic_flag = 0; |
2788 | 12.9k | const UWORD8 u1_frame_mbs_only_flag = 1; |
2789 | 12.9k | const UWORD8 u1_ref_layer_frame_mbs_only_flag = 1; |
2790 | 12.9k | const UWORD8 u1_bot_field_flag = 0; |
2791 | 12.9k | const WORD32 i4_scaled_ref_layer_left_offset = 0; |
2792 | 12.9k | const WORD32 i4_scaled_ref_layer_top_offset = 0; |
2793 | 12.9k | const WORD32 i4_ref_layer_chroma_phase_x_plus1 = 1; |
2794 | 12.9k | const WORD32 i4_ref_layer_chroma_phase_y_plus1 = 1; |
2795 | 12.9k | const WORD32 i4_chroma_phase_x_plus1 = 1; |
2796 | 12.9k | const WORD32 i4_chroma_phase_y_plus1 = 1; |
2797 | 12.9k | const WORD32 i4_sub_wd_chroma = 2; |
2798 | 12.9k | const WORD32 i4_sub_ht_chroma = 2; |
2799 | 12.9k | UWORD32 u4_ref_wd = (u4_wd / d_spatial_res_ratio); |
2800 | 12.9k | UWORD32 u4_ref_ht = (u4_ht / d_spatial_res_ratio) * (1 + u1_ref_layer_field_pic_flag); |
2801 | 12.9k | UWORD32 u4_scaled_wd = u4_wd; |
2802 | 12.9k | UWORD32 u4_scaled_ht = u4_ht * (1 + u1_field_pic_flag); |
2803 | | |
2804 | 12.9k | u4_ref_wd = u4_ref_wd >> u1_is_chroma; |
2805 | 12.9k | u4_ref_ht = u4_ref_ht >> u1_is_chroma; |
2806 | 12.9k | u4_scaled_wd = u4_scaled_wd >> u1_is_chroma; |
2807 | 12.9k | u4_scaled_ht = u4_scaled_ht >> u1_is_chroma; |
2808 | | |
2809 | 12.9k | if(u1_is_chroma) |
2810 | 6.47k | { |
2811 | 6.47k | ps_layer_props->i4_refphase_x = i4_ref_layer_chroma_phase_x_plus1 - 1; |
2812 | 6.47k | ps_layer_props->i4_refphase_y = i4_ref_layer_chroma_phase_y_plus1 - 1; |
2813 | 6.47k | ps_layer_props->i4_phase_x = i4_chroma_phase_x_plus1 - 1; |
2814 | 6.47k | ps_layer_props->i4_phase_y = i4_chroma_phase_y_plus1 - 1; |
2815 | 6.47k | ps_layer_props->u4_sub_wd = i4_sub_wd_chroma; |
2816 | 6.47k | ps_layer_props->u4_sub_ht = i4_sub_ht_chroma; |
2817 | 6.47k | ps_layer_props->u4_mb_wd = MB_SIZE >> 1; |
2818 | 6.47k | ps_layer_props->u4_mb_ht = MB_SIZE >> 1; |
2819 | 6.47k | } |
2820 | 6.47k | else |
2821 | 6.47k | { |
2822 | 6.47k | ps_layer_props->i4_refphase_x = 0; |
2823 | 6.47k | ps_layer_props->i4_refphase_y = 0; |
2824 | 6.47k | ps_layer_props->i4_phase_x = 0; |
2825 | 6.47k | ps_layer_props->i4_phase_y = 0; |
2826 | 6.47k | ps_layer_props->u4_sub_wd = 1; |
2827 | 6.47k | ps_layer_props->u4_sub_ht = 1; |
2828 | 6.47k | ps_layer_props->u4_mb_wd = MB_SIZE; |
2829 | 6.47k | ps_layer_props->u4_mb_ht = MB_SIZE; |
2830 | 6.47k | } |
2831 | | |
2832 | 12.9k | if(u1_level_idc <= 30) |
2833 | 7.71k | { |
2834 | 7.71k | ps_layer_props->u4_shift_x = 16; |
2835 | 7.71k | ps_layer_props->u4_shift_y = 16; |
2836 | 7.71k | } |
2837 | 5.24k | else |
2838 | 5.24k | { |
2839 | 5.24k | ps_layer_props->u4_shift_x = 31 - isvcd_get_ceil_log2(u4_ref_wd); |
2840 | 5.24k | ps_layer_props->u4_shift_y = 31 - isvcd_get_ceil_log2(u4_ref_ht); |
2841 | 5.24k | } |
2842 | | |
2843 | 12.9k | if((0 == u1_frame_mbs_only_flag) || (0 == u1_ref_layer_frame_mbs_only_flag)) |
2844 | 0 | { |
2845 | 0 | ps_layer_props->i4_phase_y = ps_layer_props->i4_phase_y + 4 * u1_bot_field_flag; |
2846 | |
|
2847 | 0 | if(1 == u1_ref_layer_frame_mbs_only_flag) |
2848 | 0 | { |
2849 | 0 | ps_layer_props->i4_refphase_y = (2 * ps_layer_props->i4_refphase_y) + 2; |
2850 | 0 | } |
2851 | 0 | else |
2852 | 0 | { |
2853 | 0 | ps_layer_props->i4_refphase_y = ps_layer_props->i4_refphase_y + (4 * u1_bot_field_flag); |
2854 | 0 | } |
2855 | 0 | } |
2856 | | |
2857 | 12.9k | ps_layer_props->u4_scale_x = |
2858 | 12.9k | ((u4_ref_wd << ps_layer_props->u4_shift_x) + (u4_scaled_wd >> 1)) / (u4_scaled_wd); |
2859 | 12.9k | ps_layer_props->u4_scale_y = |
2860 | 12.9k | ((u4_ref_ht << ps_layer_props->u4_shift_y) + (u4_scaled_ht >> 1)) / (u4_scaled_ht); |
2861 | | |
2862 | 12.9k | ps_layer_props->i4_offset_x = i4_scaled_ref_layer_left_offset / ps_layer_props->u4_sub_wd; |
2863 | 12.9k | ps_layer_props->i4_add_x = |
2864 | 12.9k | (((u4_ref_wd * (2 + ps_layer_props->i4_phase_x)) << (ps_layer_props->u4_shift_x - 2)) + |
2865 | 12.9k | (u4_scaled_wd >> 1)) / |
2866 | 12.9k | u4_scaled_wd + |
2867 | 12.9k | (1 << (ps_layer_props->u4_shift_x - 5)); |
2868 | 12.9k | ps_layer_props->i4_delta_x = 4 * (2 + ps_layer_props->i4_refphase_x); |
2869 | | |
2870 | 12.9k | if((1 == u1_frame_mbs_only_flag) && (1 == u1_ref_layer_frame_mbs_only_flag)) |
2871 | 12.9k | { |
2872 | 12.9k | ps_layer_props->i4_offset_y = i4_scaled_ref_layer_top_offset / ps_layer_props->u4_sub_ht; |
2873 | 12.9k | ps_layer_props->i4_add_y = |
2874 | 12.9k | (((u4_ref_ht * (2 + ps_layer_props->i4_phase_y)) << (ps_layer_props->u4_shift_y - 2)) + |
2875 | 12.9k | (u4_scaled_ht >> 1)) / |
2876 | 12.9k | u4_scaled_ht + |
2877 | 12.9k | (1 << (ps_layer_props->u4_shift_y - 5)); |
2878 | 12.9k | ps_layer_props->i4_delta_y = 4 * (2 + ps_layer_props->i4_refphase_y); |
2879 | 12.9k | } |
2880 | 0 | else |
2881 | 0 | { |
2882 | 0 | ps_layer_props->i4_offset_y = |
2883 | 0 | i4_scaled_ref_layer_top_offset / (2 * ps_layer_props->u4_sub_ht); |
2884 | 0 | ps_layer_props->i4_add_y = |
2885 | 0 | (((u4_ref_ht * (2 + ps_layer_props->i4_phase_y)) << (ps_layer_props->u4_shift_y - 3)) + |
2886 | 0 | (u4_scaled_ht >> 1)) / |
2887 | 0 | u4_scaled_ht + |
2888 | 0 | (1 << (ps_layer_props->u4_shift_y - 5)); |
2889 | 0 | ps_layer_props->i4_delta_y = 2 * (2 + ps_layer_props->i4_refphase_y); |
2890 | 0 | } |
2891 | 12.9k | } |
2892 | | |
2893 | | /** |
2894 | | ******************************************************************************* |
2895 | | * |
2896 | | * @brief |
2897 | | * Function to initialize svc ilp buffers |
2898 | | * |
2899 | | * @param[in] ps_codec |
2900 | | * Pointer to codec context |
2901 | | * |
2902 | | * @param[in] ps_mem_rec |
2903 | | * Pointer to memory allocated for input buffers |
2904 | | * |
2905 | | ******************************************************************************* |
2906 | | */ |
2907 | | void isvce_svc_ilp_buf_init(isvce_codec_t *ps_codec, iv_mem_rec_t *ps_mem_rec) |
2908 | 4.93k | { |
2909 | 4.93k | UWORD8 u1_num_spatial_layers = ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers; |
2910 | 4.93k | DOUBLE d_spatial_res_ratio = ps_codec->s_cfg.s_svc_params.d_spatial_res_ratio; |
2911 | 4.93k | UWORD32 u4_wd = ps_codec->s_cfg.u4_wd; |
2912 | 4.93k | UWORD32 u4_ht = ps_codec->s_cfg.u4_ht; |
2913 | 4.93k | UWORD8 *pu1_buf = ps_mem_rec->pv_base; |
2914 | 4.93k | WORD64 i8_alloc_mem_size = |
2915 | 4.93k | isvce_get_svc_ilp_buf_size(u1_num_spatial_layers, d_spatial_res_ratio, u4_wd, u4_ht); |
2916 | | |
2917 | 4.93k | if(u1_num_spatial_layers > 1) |
2918 | 4.04k | { |
2919 | 4.04k | WORD32 i, j; |
2920 | | |
2921 | 4.04k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs = (yuv_buf_props_t *) pu1_buf; |
2922 | 4.04k | pu1_buf += u1_num_spatial_layers * sizeof(ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[0]); |
2923 | 4.04k | i8_alloc_mem_size -= |
2924 | 4.04k | u1_num_spatial_layers * sizeof(ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[0]); |
2925 | | |
2926 | 4.04k | ps_codec->s_svc_ilp_data.ps_residual_bufs = (yuv_buf_props_t *) pu1_buf; |
2927 | 4.04k | pu1_buf += u1_num_spatial_layers * sizeof(ps_codec->s_svc_ilp_data.ps_residual_bufs[0]); |
2928 | 4.04k | i8_alloc_mem_size -= |
2929 | 4.04k | u1_num_spatial_layers * sizeof(ps_codec->s_svc_ilp_data.ps_residual_bufs[0]); |
2930 | | |
2931 | 12.1k | for(i = 0; i < NUM_SP_COMPONENTS; i++) |
2932 | 8.08k | { |
2933 | 8.08k | ps_codec->s_svc_ilp_data.aps_layer_resampler_props[i] = |
2934 | 8.08k | (layer_resampler_props_t *) pu1_buf; |
2935 | 8.08k | pu1_buf += u1_num_spatial_layers * |
2936 | 8.08k | sizeof(ps_codec->s_svc_ilp_data.aps_layer_resampler_props[i][0]); |
2937 | 8.08k | i8_alloc_mem_size -= u1_num_spatial_layers * |
2938 | 8.08k | sizeof(ps_codec->s_svc_ilp_data.aps_layer_resampler_props[i][0]); |
2939 | 8.08k | } |
2940 | | |
2941 | 4.04k | ASSERT(i8_alloc_mem_size >= 0); |
2942 | | |
2943 | 14.5k | for(i = u1_num_spatial_layers - 1; i >= 0; i--) |
2944 | 10.5k | { |
2945 | 10.5k | WORD32 i4_stride; |
2946 | | |
2947 | 10.5k | WORD32 i4_layer_luma_wd = |
2948 | 10.5k | ((DOUBLE) u4_wd / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
2949 | 10.5k | WORD32 i4_layer_luma_ht = |
2950 | 10.5k | ((DOUBLE) u4_ht / pow(d_spatial_res_ratio, u1_num_spatial_layers - 1 - i)) + 0.99; |
2951 | 10.5k | WORD32 i4_layer_luma_samples = |
2952 | 10.5k | (ALIGN16(i4_layer_luma_wd) + PAD_WD) * (i4_layer_luma_ht + PAD_HT); |
2953 | 10.5k | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
2954 | 10.5k | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
2955 | 10.5k | WORD32 i4_layer_uv_samples = |
2956 | 10.5k | (ALIGN16(i4_layer_uv_wd) + PAD_WD) * (i4_layer_uv_ht + PAD_HT); |
2957 | | |
2958 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].u4_width = i4_layer_luma_wd; |
2959 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].u4_height = i4_layer_luma_ht; |
2960 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].e_color_format = IV_YUV_420SP_UV; |
2961 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].u1_bit_depth = 8; |
2962 | | |
2963 | 10.5k | i4_stride = ALIGN16(i4_layer_luma_wd) + PAD_WD; |
2964 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].as_component_bufs[Y].pv_data = |
2965 | 10.5k | pu1_buf + PAD_LEFT + PAD_TOP * i4_stride; |
2966 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].as_component_bufs[Y].i4_data_stride = |
2967 | 10.5k | ALIGN16(i4_layer_luma_wd) + PAD_WD; |
2968 | 10.5k | pu1_buf += i4_layer_luma_samples * sizeof(UWORD8); |
2969 | 10.5k | i8_alloc_mem_size -= i4_layer_luma_samples * sizeof(UWORD8); |
2970 | | |
2971 | 10.5k | i4_stride = ALIGN16(i4_layer_uv_wd) + PAD_WD; |
2972 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].as_component_bufs[UV].pv_data = |
2973 | 10.5k | pu1_buf + PAD_LEFT + PAD_TOP * i4_stride; |
2974 | 10.5k | ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[i].as_component_bufs[UV].i4_data_stride = |
2975 | 10.5k | ALIGN16(i4_layer_uv_wd) + PAD_WD; |
2976 | 10.5k | pu1_buf += i4_layer_uv_samples * sizeof(UWORD8); |
2977 | 10.5k | i8_alloc_mem_size -= i4_layer_uv_samples * sizeof(UWORD8); |
2978 | | |
2979 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].u4_width = i4_layer_luma_wd; |
2980 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].u4_height = i4_layer_luma_ht; |
2981 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].e_color_format = IV_YUV_420SP_UV; |
2982 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].u1_bit_depth = 10; |
2983 | | |
2984 | 10.5k | i4_stride = ALIGN16(i4_layer_luma_wd) + PAD_WD; |
2985 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].as_component_bufs[Y].pv_data = |
2986 | 10.5k | pu1_buf + (PAD_LEFT + PAD_TOP * i4_stride) * (sizeof(WORD16) / sizeof(pu1_buf[0])); |
2987 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].as_component_bufs[Y].i4_data_stride = |
2988 | 10.5k | i4_stride; |
2989 | 10.5k | pu1_buf += i4_layer_luma_samples * sizeof(WORD16); |
2990 | 10.5k | i8_alloc_mem_size -= i4_layer_luma_samples * sizeof(WORD16); |
2991 | | |
2992 | 10.5k | i4_stride = ALIGN16(i4_layer_uv_wd) + PAD_WD; |
2993 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].as_component_bufs[UV].pv_data = |
2994 | 10.5k | pu1_buf + (PAD_LEFT + PAD_TOP * i4_stride) * (sizeof(WORD16) / sizeof(pu1_buf[0])); |
2995 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].as_component_bufs[UV].i4_data_stride = |
2996 | 10.5k | i4_stride; |
2997 | 10.5k | pu1_buf += i4_layer_uv_samples * sizeof(WORD16); |
2998 | 10.5k | i8_alloc_mem_size -= i4_layer_uv_samples * sizeof(WORD16); |
2999 | | |
3000 | 10.5k | ps_codec->s_svc_ilp_data.ps_residual_bufs[i].as_component_bufs[V].pv_data = NULL; |
3001 | | |
3002 | 10.5k | ASSERT(i8_alloc_mem_size >= 0); |
3003 | | |
3004 | 10.5k | if(i >= 1) |
3005 | 6.47k | { |
3006 | 19.4k | for(j = 0; j < NUM_SP_COMPONENTS; j++) |
3007 | 12.9k | { |
3008 | 12.9k | isvce_layer_resampler_props_init( |
3009 | 12.9k | &ps_codec->s_svc_ilp_data.aps_layer_resampler_props[j][i], |
3010 | 12.9k | d_spatial_res_ratio, i4_layer_luma_wd, i4_layer_luma_ht, |
3011 | 12.9k | ps_codec->s_cfg.u4_max_level, ((COMPONENT_TYPE) j) == UV); |
3012 | 12.9k | } |
3013 | 6.47k | } |
3014 | 10.5k | } |
3015 | 4.04k | } |
3016 | 891 | else |
3017 | 891 | { |
3018 | 891 | WORD32 i4_stride; |
3019 | | |
3020 | 891 | WORD32 i4_layer_luma_wd = u4_wd; |
3021 | 891 | WORD32 i4_layer_luma_ht = u4_ht; |
3022 | 891 | WORD32 i4_layer_luma_samples = |
3023 | 891 | (ALIGN16(i4_layer_luma_wd) + PAD_WD) * (i4_layer_luma_ht + PAD_HT); |
3024 | 891 | WORD32 i4_layer_uv_wd = i4_layer_luma_wd; |
3025 | 891 | WORD32 i4_layer_uv_ht = i4_layer_luma_ht / 2.0 + 0.99; |
3026 | 891 | WORD32 i4_layer_uv_samples = (ALIGN16(i4_layer_uv_wd) + PAD_WD) * (i4_layer_uv_ht + PAD_HT); |
3027 | | |
3028 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs = (yuv_buf_props_t *) pu1_buf; |
3029 | 891 | pu1_buf += sizeof(ps_codec->s_svc_ilp_data.ps_residual_bufs[0]); |
3030 | 891 | i8_alloc_mem_size -= sizeof(ps_codec->s_svc_ilp_data.ps_residual_bufs[0]); |
3031 | | |
3032 | 891 | ASSERT(i8_alloc_mem_size >= 0); |
3033 | | |
3034 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].u4_width = i4_layer_luma_wd; |
3035 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].u4_height = i4_layer_luma_ht; |
3036 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].e_color_format = IV_YUV_420SP_UV; |
3037 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].u1_bit_depth = 10; |
3038 | | |
3039 | 891 | i4_stride = ALIGN16(i4_layer_luma_wd) + PAD_WD; |
3040 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].as_component_bufs[Y].pv_data = |
3041 | 891 | pu1_buf + (PAD_LEFT + PAD_TOP * i4_stride) * (sizeof(WORD16) / sizeof(pu1_buf[0])); |
3042 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].as_component_bufs[Y].i4_data_stride = |
3043 | 891 | i4_stride; |
3044 | 891 | pu1_buf += i4_layer_luma_samples * sizeof(WORD16); |
3045 | 891 | i8_alloc_mem_size -= i4_layer_luma_samples * sizeof(WORD16); |
3046 | | |
3047 | 891 | i4_stride = ALIGN16(i4_layer_uv_wd) + PAD_WD; |
3048 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].as_component_bufs[UV].pv_data = |
3049 | 891 | pu1_buf + (PAD_LEFT + PAD_TOP * i4_stride) * (sizeof(WORD16) / sizeof(pu1_buf[0])); |
3050 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].as_component_bufs[UV].i4_data_stride = |
3051 | 891 | i4_stride; |
3052 | 891 | pu1_buf += i4_layer_uv_samples * sizeof(WORD16); |
3053 | 891 | i8_alloc_mem_size -= i4_layer_uv_samples * sizeof(WORD16); |
3054 | | |
3055 | 891 | ps_codec->s_svc_ilp_data.ps_residual_bufs[0].as_component_bufs[V].pv_data = NULL; |
3056 | | |
3057 | 891 | ASSERT(i8_alloc_mem_size >= 0); |
3058 | 891 | } |
3059 | 4.93k | } |
3060 | | |
3061 | | static FORCEINLINE UWORD32 isvce_get_residual_csbf(mem_fxns_t *ps_mem_fxns, |
3062 | | buffer_container_t *ps_comp_buf) |
3063 | 247k | { |
3064 | 247k | WORD32 i; |
3065 | | |
3066 | 247k | UWORD32 u4_csbf = 0; |
3067 | | |
3068 | 4.18M | for(i = 0; i < MAX_TU_IN_MB; i++) |
3069 | 3.93M | { |
3070 | 3.93M | UWORD8 u1_zscan_idx = gau1_raster_to_zscan_map[i]; |
3071 | 3.93M | UWORD8 u1_offset_x = (i % MAX_TU_IN_MB_ROW) * MIN_TU_SIZE; |
3072 | 3.93M | UWORD8 u1_offset_y = (i / MAX_TU_IN_MB_ROW) * MIN_TU_SIZE; |
3073 | 3.93M | WORD16 *pi2_res = ((WORD16 *) ps_comp_buf->pv_data) + u1_offset_x + |
3074 | 3.93M | u1_offset_y * ps_comp_buf->i4_data_stride; |
3075 | 3.93M | UWORD8 u1_cbf = ps_mem_fxns->pf_nonzero_checker( |
3076 | 3.93M | (UWORD8 *) pi2_res, ps_comp_buf->i4_data_stride * (sizeof(WORD16) / sizeof(UWORD8)), |
3077 | 3.93M | MIN_TU_SIZE * (sizeof(WORD16) / sizeof(UWORD8)), MIN_TU_SIZE); |
3078 | | |
3079 | 3.93M | u4_csbf |= (u1_cbf << u1_zscan_idx); |
3080 | 3.93M | } |
3081 | | |
3082 | 247k | return u4_csbf; |
3083 | 247k | } |
3084 | | |
3085 | | /** |
3086 | | ******************************************************************************* |
3087 | | * |
3088 | | * @brief |
3089 | | * Function to update svc ilp buffers after every MB |
3090 | | * |
3091 | | * @param[in] ps_proc |
3092 | | * Pointer to process context |
3093 | | * |
3094 | | ******************************************************************************* |
3095 | | */ |
3096 | | void isvce_svc_ilp_buf_update(isvce_process_ctxt_t *ps_proc) |
3097 | 7.19M | { |
3098 | 7.19M | isvce_codec_t *ps_codec = ps_proc->ps_codec; |
3099 | 7.19M | svc_params_t *ps_svc_params = &ps_codec->s_cfg.s_svc_params; |
3100 | | |
3101 | 7.19M | UWORD8 u1_spatial_layer_id = ps_proc->u1_spatial_layer_id; |
3102 | | |
3103 | 7.19M | if(ps_svc_params->u1_num_spatial_layers > 1) |
3104 | 6.25M | { |
3105 | 6.25M | buffer_container_t s_src; |
3106 | 6.25M | buffer_container_t s_dst; |
3107 | | |
3108 | 6.25M | WORD32 i; |
3109 | | |
3110 | 6.25M | svc_ilp_data_t *ps_svc_ilp_data = &ps_codec->s_svc_ilp_data; |
3111 | 6.25M | isa_dependent_fxns_t *ps_isa_dependent_fxns = &ps_codec->s_isa_dependent_fxns; |
3112 | 6.25M | mem_fxns_t *ps_mem_fxns = &ps_isa_dependent_fxns->s_mem_fxns; |
3113 | 6.25M | yuv_buf_props_t *ps_residual_buf = |
3114 | 6.25M | &ps_codec->s_svc_ilp_data.ps_residual_bufs[u1_spatial_layer_id]; |
3115 | | |
3116 | 6.25M | WORD32 i4_mb_x = ps_proc->i4_mb_x; |
3117 | 6.25M | WORD32 i4_mb_y = ps_proc->i4_mb_y; |
3118 | | |
3119 | 6.25M | ASSERT(ps_proc->s_rec_buf_props.e_color_format == IV_YUV_420SP_UV); |
3120 | | |
3121 | 6.25M | if(u1_spatial_layer_id < (ps_svc_params->u1_num_spatial_layers - 1)) |
3122 | 1.57M | { |
3123 | 1.57M | if(ps_proc->ps_mb_info->u1_is_intra) |
3124 | 1.37M | { |
3125 | 4.10M | for(i = 0; i < NUM_SP_COMPONENTS; i++) |
3126 | 2.72M | { |
3127 | 2.72M | UWORD8 u1_is_chroma = (Y != ((COMPONENT_TYPE) i)); |
3128 | | |
3129 | 2.72M | s_src = ps_proc->s_rec_buf_props.as_component_bufs[i]; |
3130 | | |
3131 | 2.72M | s_dst.i4_data_stride = ps_svc_ilp_data->ps_intra_recon_bufs[u1_spatial_layer_id] |
3132 | 2.72M | .as_component_bufs[i] |
3133 | 2.72M | .i4_data_stride; |
3134 | 2.72M | s_dst.pv_data = |
3135 | 2.72M | ((UWORD8 *) ps_svc_ilp_data->ps_intra_recon_bufs[u1_spatial_layer_id] |
3136 | 2.72M | .as_component_bufs[i] |
3137 | 2.72M | .pv_data) + |
3138 | 2.72M | i4_mb_x * MB_SIZE + |
3139 | 2.72M | i4_mb_y * (MB_SIZE >> u1_is_chroma) * s_dst.i4_data_stride; |
3140 | | |
3141 | 2.72M | ps_mem_fxns->pf_copy_2d((UWORD8 *) s_dst.pv_data, s_dst.i4_data_stride, |
3142 | 2.72M | (UWORD8 *) s_src.pv_data, s_src.i4_data_stride, MB_SIZE, |
3143 | 2.72M | (MB_SIZE >> u1_is_chroma)); |
3144 | 2.72M | } |
3145 | 1.37M | } |
3146 | 204k | else |
3147 | 204k | { |
3148 | 612k | for(i = 0; i < NUM_SP_COMPONENTS; i++) |
3149 | 408k | { |
3150 | 408k | UWORD8 u1_is_chroma = (Y != ((COMPONENT_TYPE) i)); |
3151 | | |
3152 | 408k | s_dst.i4_data_stride = ps_svc_ilp_data->ps_intra_recon_bufs[u1_spatial_layer_id] |
3153 | 408k | .as_component_bufs[i] |
3154 | 408k | .i4_data_stride; |
3155 | 408k | s_dst.pv_data = |
3156 | 408k | ((UWORD8 *) ps_svc_ilp_data->ps_intra_recon_bufs[u1_spatial_layer_id] |
3157 | 408k | .as_component_bufs[i] |
3158 | 408k | .pv_data) + |
3159 | 408k | i4_mb_x * MB_SIZE + |
3160 | 408k | i4_mb_y * (MB_SIZE >> u1_is_chroma) * s_dst.i4_data_stride; |
3161 | | |
3162 | 408k | ps_mem_fxns->pf_memset_2d((UWORD8 *) s_dst.pv_data, s_dst.i4_data_stride, 0, |
3163 | 408k | MB_SIZE, (MB_SIZE >> u1_is_chroma)); |
3164 | 408k | } |
3165 | 204k | } |
3166 | 1.57M | } |
3167 | | |
3168 | 6.25M | if(ENABLE_RESIDUAL_PREDICTION && (ps_proc->i4_slice_type != ISLICE) && |
3169 | 6.25M | (u1_spatial_layer_id < (ps_svc_params->u1_num_spatial_layers - 1))) |
3170 | 243k | { |
3171 | 243k | if(ps_proc->ps_mb_info->u1_is_intra || (ps_proc->ps_mb_info->u2_mb_type == PSKIP) || |
3172 | 243k | (ps_proc->ps_mb_info->u2_mb_type == BSKIP)) |
3173 | 90.9k | { |
3174 | 272k | for(i = 0; i < NUM_SP_COMPONENTS; i++) |
3175 | 181k | { |
3176 | 181k | buffer_container_t *ps_comp_buf; |
3177 | | |
3178 | 181k | WORD16 *pi2_res; |
3179 | | |
3180 | 181k | UWORD8 u1_is_chroma = (Y != ((COMPONENT_TYPE) i)); |
3181 | | |
3182 | 181k | ps_comp_buf = &ps_residual_buf->as_component_bufs[u1_is_chroma ? UV : Y]; |
3183 | 181k | pi2_res = |
3184 | 181k | ((WORD16 *) ps_comp_buf->pv_data) + ps_proc->i4_mb_x * MB_SIZE + |
3185 | 181k | ps_proc->i4_mb_y * (MB_SIZE >> u1_is_chroma) * ps_comp_buf->i4_data_stride; |
3186 | | |
3187 | 181k | ps_mem_fxns->pf_memset_2d( |
3188 | 181k | (UWORD8 *) pi2_res, |
3189 | 181k | ps_comp_buf->i4_data_stride * (sizeof(WORD16) / sizeof(UWORD8)), 0, |
3190 | 181k | MB_SIZE * (sizeof(WORD16) / sizeof(UWORD8)), MB_SIZE >> u1_is_chroma); |
3191 | 181k | } |
3192 | 90.9k | } |
3193 | 243k | } |
3194 | | |
3195 | 6.25M | if(ENABLE_RESIDUAL_PREDICTION && (u1_spatial_layer_id > 0) && |
3196 | 6.25M | !(ps_proc->ps_mb_info->u1_is_intra || (ps_proc->ps_mb_info->u2_mb_type == PSKIP) || |
3197 | 5.29M | (ps_proc->ps_mb_info->u2_mb_type == BSKIP))) |
3198 | 247k | { |
3199 | 247k | s_src = ps_residual_buf->as_component_bufs[Y]; |
3200 | 247k | s_src.pv_data = ((WORD16 *) s_src.pv_data) + ps_proc->i4_mb_x * MB_SIZE + |
3201 | 247k | ps_proc->i4_mb_y * MB_SIZE * s_src.i4_data_stride; |
3202 | | |
3203 | 247k | ps_proc->ps_mb_info->u4_res_csbp = isvce_get_residual_csbf(ps_mem_fxns, &s_src); |
3204 | 247k | } |
3205 | 6.01M | else |
3206 | 6.01M | { |
3207 | 6.01M | ps_proc->ps_mb_info->u4_res_csbp = 0; |
3208 | 6.01M | } |
3209 | 6.25M | } |
3210 | 936k | else |
3211 | 936k | { |
3212 | 936k | ps_proc->ps_mb_info->u4_res_csbp = 0; |
3213 | 936k | } |
3214 | 7.19M | } |
3215 | | |
3216 | | /* |
3217 | | * Padding has a one MB row dependency on deblock which |
3218 | | * in turn has a one MB row dependency on encode |
3219 | | */ |
3220 | | static IH264E_ERROR_T isvce_pad_frame(isvce_process_ctxt_t *ps_proc, yuv_buf_props_t *ps_pad_buf) |
3221 | 163k | { |
3222 | | /* codec context */ |
3223 | 163k | isvce_codec_t *ps_codec = ps_proc->ps_codec; |
3224 | | |
3225 | 163k | WORD32 i4_element_size = (ps_pad_buf->u1_bit_depth > 8) ? 2 : 1; |
3226 | | |
3227 | | /* src buffers luma */ |
3228 | 163k | WORD32 i4_luma_stride = ps_pad_buf->as_component_bufs[0].i4_data_stride * i4_element_size; |
3229 | 163k | UWORD8 *pu1_curr_pic_luma = (UWORD8 *) (ps_pad_buf->as_component_bufs[0].pv_data); |
3230 | | |
3231 | | /* src buffers chroma */ |
3232 | 163k | WORD32 i4_chroma_stride = ps_pad_buf->as_component_bufs[1].i4_data_stride * i4_element_size; |
3233 | 163k | UWORD8 *pu1_curr_pic_chroma = (UWORD8 *) (ps_pad_buf->as_component_bufs[1].pv_data); |
3234 | | |
3235 | 163k | WORD32 i4_bottom_offset_luma = ps_pad_buf->u4_height * i4_luma_stride; |
3236 | 163k | WORD32 i4_bottom_offset_chroma = (ps_pad_buf->u4_height >> 1) * i4_chroma_stride; |
3237 | | |
3238 | | /* Pad left */ |
3239 | 163k | ps_codec->pf_pad_left_luma(pu1_curr_pic_luma, i4_luma_stride, ps_pad_buf->u4_height, |
3240 | 163k | PAD_LEFT * i4_element_size); |
3241 | 163k | ps_codec->pf_pad_left_chroma(pu1_curr_pic_chroma, i4_chroma_stride, ps_pad_buf->u4_height >> 1, |
3242 | 163k | PAD_LEFT * i4_element_size); |
3243 | | |
3244 | | /* Pad right */ |
3245 | 163k | ps_codec->pf_pad_right_luma(pu1_curr_pic_luma + ps_pad_buf->u4_width * i4_element_size, |
3246 | 163k | i4_luma_stride, ps_pad_buf->u4_height, PAD_RIGHT * i4_element_size); |
3247 | 163k | ps_codec->pf_pad_right_chroma(pu1_curr_pic_chroma + ps_pad_buf->u4_width * i4_element_size, |
3248 | 163k | i4_chroma_stride, ps_pad_buf->u4_height >> 1, |
3249 | 163k | PAD_RIGHT * i4_element_size); |
3250 | | |
3251 | | /* Pad top */ |
3252 | 163k | ps_codec->pf_pad_top(pu1_curr_pic_luma - (PAD_LEFT * i4_element_size), i4_luma_stride, |
3253 | 163k | (ps_pad_buf->u4_width + PAD_WD) * i4_element_size, PAD_TOP); |
3254 | 163k | ps_codec->pf_pad_top(pu1_curr_pic_chroma - (PAD_LEFT * i4_element_size), i4_chroma_stride, |
3255 | 163k | (ps_pad_buf->u4_width + PAD_WD) * i4_element_size, PAD_TOP >> 1); |
3256 | | |
3257 | | /* Pad bottom */ |
3258 | 163k | ps_codec->pf_pad_bottom( |
3259 | 163k | pu1_curr_pic_luma + i4_bottom_offset_luma - (PAD_LEFT * i4_element_size), i4_luma_stride, |
3260 | 163k | (ps_pad_buf->u4_width + PAD_WD) * i4_element_size, PAD_BOT); |
3261 | 163k | ps_codec->pf_pad_bottom( |
3262 | 163k | pu1_curr_pic_chroma + i4_bottom_offset_chroma - (PAD_LEFT * i4_element_size), |
3263 | 163k | i4_chroma_stride, (ps_pad_buf->u4_width + PAD_WD) * i4_element_size, PAD_BOT >> 1); |
3264 | | |
3265 | 163k | return IH264E_SUCCESS; |
3266 | 163k | } |
3267 | | |
3268 | | void isvce_svc_pad_frame(isvce_process_ctxt_t *ps_proc) |
3269 | 58.7k | { |
3270 | 58.7k | isvce_codec_t *ps_codec = ps_proc->ps_codec; |
3271 | | |
3272 | 58.7k | isvce_pad_frame(ps_proc, &(ps_proc->s_rec_pic_buf_props)); |
3273 | | |
3274 | 58.7k | if(ps_proc->s_svc_params.u1_num_spatial_layers > 1) |
3275 | 52.3k | { |
3276 | 52.3k | isvce_pad_frame( |
3277 | 52.3k | ps_proc, &(ps_codec->s_svc_ilp_data.ps_intra_recon_bufs[ps_proc->u1_spatial_layer_id])); |
3278 | 52.3k | isvce_pad_frame(ps_proc, |
3279 | 52.3k | &(ps_codec->s_svc_ilp_data.ps_residual_bufs[ps_proc->u1_spatial_layer_id])); |
3280 | 52.3k | } |
3281 | 58.7k | } |
3282 | | |
3283 | | /** |
3284 | | ******************************************************************************* |
3285 | | * |
3286 | | * @brief |
3287 | | * Initialize AIR mb frame Map |
3288 | | * |
3289 | | * @par Description: |
3290 | | * Initialize AIR mb frame map |
3291 | | * MB frame map indicates which frame an Mb should be coded as intra according |
3292 | | *to AIR |
3293 | | * |
3294 | | * @param[in] ps_codec |
3295 | | * Pointer to codec context |
3296 | | * |
3297 | | * @returns error_status |
3298 | | * |
3299 | | * @remarks |
3300 | | * |
3301 | | * |
3302 | | ******************************************************************************* |
3303 | | */ |
3304 | | IH264E_ERROR_T isvce_init_air_map(isvce_codec_t *ps_codec) |
3305 | 9.84k | { |
3306 | | /* intra refresh map */ |
3307 | 9.84k | UWORD16 *pu2_intr_rfrsh_map = ps_codec->pu2_intr_rfrsh_map; |
3308 | | |
3309 | | /* air mode */ |
3310 | 9.84k | IVE_AIR_MODE_T air_mode = ps_codec->s_cfg.e_air_mode; |
3311 | | |
3312 | | /* refresh period */ |
3313 | 9.84k | UWORD32 air_period = ps_codec->s_cfg.u4_air_refresh_period; |
3314 | | |
3315 | | /* mb cnt */ |
3316 | 9.84k | UWORD32 u4_mb_cnt = ps_codec->s_cfg.i4_wd_mbs * ps_codec->s_cfg.i4_ht_mbs; |
3317 | | |
3318 | | /* temp var */ |
3319 | 9.84k | UWORD32 curr_mb, seed_rand = 1; |
3320 | | |
3321 | 9.84k | switch(air_mode) |
3322 | 9.84k | { |
3323 | 0 | case IVE_AIR_MODE_CYCLIC: |
3324 | |
|
3325 | 0 | for(curr_mb = 0; curr_mb < u4_mb_cnt; curr_mb++) |
3326 | 0 | { |
3327 | 0 | pu2_intr_rfrsh_map[curr_mb] = curr_mb % air_period; |
3328 | 0 | } |
3329 | 0 | break; |
3330 | | |
3331 | 0 | case IVE_AIR_MODE_RANDOM: |
3332 | |
|
3333 | 0 | for(curr_mb = 0; curr_mb < u4_mb_cnt; curr_mb++) |
3334 | 0 | { |
3335 | 0 | seed_rand = (seed_rand * 32719 + 3) % 32749; |
3336 | 0 | pu2_intr_rfrsh_map[curr_mb] = seed_rand % air_period; |
3337 | 0 | } |
3338 | 0 | break; |
3339 | | |
3340 | 9.84k | default: |
3341 | | |
3342 | 9.84k | break; |
3343 | 9.84k | } |
3344 | | |
3345 | 9.84k | return IH264E_SUCCESS; |
3346 | 9.84k | } |
3347 | | |
3348 | | /** |
3349 | | ****************************************************************************** |
3350 | | * |
3351 | | * @brief |
3352 | | * derivation process for macroblock availability |
3353 | | * |
3354 | | * @par Description |
3355 | | * Calculates the availability of the left, top, topright and topleft macroblocks. |
3356 | | * |
3357 | | * @param[in] ps_proc_ctxt |
3358 | | * pointer to proc context (handle) |
3359 | | * |
3360 | | * @remarks Based on section 6.4.5 in H264 spec |
3361 | | * |
3362 | | * @return none |
3363 | | * |
3364 | | ****************************************************************************** |
3365 | | */ |
3366 | | void isvce_derive_nghbr_avbl_of_mbs(isvce_process_ctxt_t *ps_proc) |
3367 | 8.07M | { |
3368 | 8.07M | UWORD8 *pu1_slice_idx_curr = ps_proc->pu1_slice_idx; |
3369 | 8.07M | UWORD8 *pu1_slice_idx_b; |
3370 | 8.07M | UWORD8 *pu1_slice_idx_a; |
3371 | 8.07M | UWORD8 *pu1_slice_idx_c; |
3372 | 8.07M | UWORD8 *pu1_slice_idx_d; |
3373 | 8.07M | block_neighbors_t *ps_ngbr_avbl; |
3374 | 8.07M | WORD32 i4_mb_x, i4_mb_y; |
3375 | 8.07M | WORD32 i4_wd_mbs; |
3376 | | |
3377 | 8.07M | i4_mb_x = ps_proc->i4_mb_x; |
3378 | 8.07M | i4_mb_y = ps_proc->i4_mb_y; |
3379 | | |
3380 | 8.07M | i4_wd_mbs = ps_proc->i4_wd_mbs; |
3381 | | |
3382 | 8.07M | pu1_slice_idx_curr += (i4_mb_y * i4_wd_mbs) + i4_mb_x; |
3383 | 8.07M | pu1_slice_idx_a = pu1_slice_idx_curr - 1; |
3384 | 8.07M | pu1_slice_idx_b = pu1_slice_idx_curr - i4_wd_mbs; |
3385 | 8.07M | pu1_slice_idx_c = pu1_slice_idx_b + 1; |
3386 | 8.07M | pu1_slice_idx_d = pu1_slice_idx_b - 1; |
3387 | 8.07M | ps_ngbr_avbl = ps_proc->ps_ngbr_avbl; |
3388 | | |
3389 | | /**********************************************************************/ |
3390 | | /* The macroblock is marked as available, unless one of the following */ |
3391 | | /* conditions is true in which case the macroblock shall be marked as */ |
3392 | | /* not available. */ |
3393 | | /* 1. mbAddr < 0 */ |
3394 | | /* 2 mbAddr > CurrMbAddr */ |
3395 | | /* 3. the macroblock with address mbAddr belongs to a different slice */ |
3396 | | /* than the macroblock with address CurrMbAddr */ |
3397 | | /**********************************************************************/ |
3398 | | |
3399 | | /* left macroblock availability */ |
3400 | 8.07M | if(i4_mb_x == 0) |
3401 | 389k | { /* macroblocks along first column */ |
3402 | 389k | ps_ngbr_avbl->u1_mb_a = 0; |
3403 | 389k | } |
3404 | 7.68M | else |
3405 | 7.68M | { /* macroblocks belong to same slice? */ |
3406 | 7.68M | if(*pu1_slice_idx_a != *pu1_slice_idx_curr) |
3407 | 0 | ps_ngbr_avbl->u1_mb_a = 0; |
3408 | 7.68M | else |
3409 | 7.68M | ps_ngbr_avbl->u1_mb_a = 1; |
3410 | 7.68M | } |
3411 | | |
3412 | | /* top macroblock availability */ |
3413 | 8.07M | if(i4_mb_y == 0) |
3414 | 783k | { /* macroblocks along first row */ |
3415 | 783k | ps_ngbr_avbl->u1_mb_b = 0; |
3416 | 783k | } |
3417 | 7.29M | else |
3418 | 7.29M | { /* macroblocks belong to same slice? */ |
3419 | 7.29M | if(*pu1_slice_idx_b != *pu1_slice_idx_curr) |
3420 | 0 | ps_ngbr_avbl->u1_mb_b = 0; |
3421 | 7.29M | else |
3422 | 7.29M | ps_ngbr_avbl->u1_mb_b = 1; |
3423 | 7.29M | } |
3424 | | |
3425 | | /* top right macroblock availability */ |
3426 | 8.07M | if(i4_mb_x == i4_wd_mbs - 1 || i4_mb_y == 0) |
3427 | 1.07M | { /* macroblocks along last column */ |
3428 | 1.07M | ps_ngbr_avbl->u1_mb_c = 0; |
3429 | 1.07M | } |
3430 | 7.00M | else |
3431 | 7.00M | { /* macroblocks belong to same slice? */ |
3432 | 7.00M | if(*pu1_slice_idx_c != *pu1_slice_idx_curr) |
3433 | 0 | ps_ngbr_avbl->u1_mb_c = 0; |
3434 | 7.00M | else |
3435 | 7.00M | ps_ngbr_avbl->u1_mb_c = 1; |
3436 | 7.00M | } |
3437 | | |
3438 | | /* top left macroblock availability */ |
3439 | 8.07M | if(i4_mb_x == 0 || i4_mb_y == 0) |
3440 | 1.07M | { /* macroblocks along first column */ |
3441 | 1.07M | ps_ngbr_avbl->u1_mb_d = 0; |
3442 | 1.07M | } |
3443 | 7.00M | else |
3444 | 7.00M | { /* macroblocks belong to same slice? */ |
3445 | 7.00M | if(*pu1_slice_idx_d != *pu1_slice_idx_curr) |
3446 | 0 | ps_ngbr_avbl->u1_mb_d = 0; |
3447 | 7.00M | else |
3448 | 7.00M | ps_ngbr_avbl->u1_mb_d = 1; |
3449 | 7.00M | } |
3450 | 8.07M | } |
3451 | | |
3452 | | static IH264E_ERROR_T isvce_rc_init_wrapper(isvce_codec_t *ps_codec) |
3453 | 13.1k | { |
3454 | 13.1k | WORD32 i; |
3455 | | |
3456 | 13.1k | isvce_cfg_params_t *ps_cfg = &ps_codec->s_cfg; |
3457 | | |
3458 | 13.1k | IH264E_ERROR_T err = isvce_svc_rc_params_validate(ps_cfg); |
3459 | | |
3460 | 13.1k | if(IH264E_SUCCESS != err) |
3461 | 433 | { |
3462 | 433 | return err; |
3463 | 433 | } |
3464 | | |
3465 | 25.4k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
3466 | 12.7k | { |
3467 | 12.7k | UWORD8 au1_init_qp[MAX_PIC_TYPE]; |
3468 | 12.7k | UWORD8 au1_min_max_qp[2 * MAX_PIC_TYPE]; |
3469 | | |
3470 | 12.7k | au1_init_qp[0] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_i_qp[i]]; |
3471 | 12.7k | au1_init_qp[1] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_p_qp[i]]; |
3472 | 12.7k | au1_init_qp[2] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_b_qp[i]]; |
3473 | | |
3474 | 12.7k | au1_min_max_qp[2 * I_PIC] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_i_qp_min[i]]; |
3475 | 12.7k | au1_min_max_qp[2 * I_PIC + 1] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_i_qp_max[i]]; |
3476 | | |
3477 | 12.7k | au1_min_max_qp[2 * P_PIC] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_p_qp_min[i]]; |
3478 | 12.7k | au1_min_max_qp[2 * P_PIC + 1] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_p_qp_max[i]]; |
3479 | | |
3480 | 12.7k | au1_min_max_qp[2 * B_PIC] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_b_qp_min[i]]; |
3481 | 12.7k | au1_min_max_qp[2 * B_PIC + 1] = gau1_h264_to_mpeg2_qmap[ps_cfg->au4_b_qp_max[i]]; |
3482 | | |
3483 | 12.7k | switch(ps_cfg->e_rc_mode) |
3484 | 12.7k | { |
3485 | 3.33k | case IVE_RC_STORAGE: |
3486 | 3.33k | { |
3487 | 3.33k | ps_codec->s_rate_control.e_rc_type = VBR_STORAGE; |
3488 | 3.33k | break; |
3489 | 0 | } |
3490 | 6.18k | case IVE_RC_CBR_NON_LOW_DELAY: |
3491 | 6.18k | { |
3492 | 6.18k | ps_codec->s_rate_control.e_rc_type = CBR_NLDRC; |
3493 | 6.18k | break; |
3494 | 0 | } |
3495 | 0 | case IVE_RC_CBR_LOW_DELAY: |
3496 | 0 | { |
3497 | 0 | ps_codec->s_rate_control.e_rc_type = CBR_LDRC; |
3498 | 0 | break; |
3499 | 0 | } |
3500 | 3.20k | case IVE_RC_NONE: |
3501 | 3.20k | { |
3502 | 3.20k | ps_codec->s_rate_control.e_rc_type = CONST_QP; |
3503 | 3.20k | break; |
3504 | 0 | } |
3505 | 0 | default: |
3506 | 0 | { |
3507 | 0 | break; |
3508 | 0 | } |
3509 | 12.7k | } |
3510 | | |
3511 | 50.9k | for(i = 0; i < MAX_NUM_SPATIAL_LAYERS; i++) |
3512 | 38.1k | { |
3513 | 38.1k | isvce_rc_init(ps_codec->s_rate_control.apps_rate_control_api[i], |
3514 | 38.1k | ps_codec->s_rate_control.pps_frame_time, |
3515 | 38.1k | ps_codec->s_rate_control.pps_time_stamp, |
3516 | 38.1k | ps_codec->s_rate_control.pps_pd_frm_rate, ps_cfg->u4_max_framerate, |
3517 | 38.1k | ps_cfg->u4_src_frame_rate, ps_cfg->u4_tgt_frame_rate, |
3518 | 38.1k | ps_codec->s_rate_control.e_rc_type, ps_cfg->au4_target_bitrate[i], |
3519 | 38.1k | ps_cfg->au4_max_bitrate[i], ps_cfg->au4_vbv_buffer_delay[i], |
3520 | 38.1k | ps_cfg->u4_i_frm_interval, ps_cfg->u4_num_bframes + 1, au1_init_qp, |
3521 | 38.1k | ps_cfg->u4_num_bframes + 2, au1_min_max_qp, ps_cfg->u4_max_level); |
3522 | 38.1k | } |
3523 | 12.7k | } |
3524 | | |
3525 | 12.7k | return IH264E_SUCCESS; |
3526 | 12.7k | } |
3527 | | |
3528 | | /** |
3529 | | ******************************************************************************* |
3530 | | * |
3531 | | * @brief |
3532 | | * Codec level initializations |
3533 | | * |
3534 | | * @par Description: |
3535 | | * Initializes the codec with parameters that needs to be set before encoding |
3536 | | * first frame |
3537 | | * |
3538 | | * @param[in] ps_codec |
3539 | | * Pointer to codec context |
3540 | | * |
3541 | | * @param[in] ps_inp_buf |
3542 | | * Pointer to input buffer context |
3543 | | * |
3544 | | * @returns error_status |
3545 | | * |
3546 | | * @remarks |
3547 | | * |
3548 | | * |
3549 | | ******************************************************************************* |
3550 | | */ |
3551 | | IH264E_ERROR_T isvce_codec_init(isvce_codec_t *ps_codec) |
3552 | 9.84k | { |
3553 | 9.84k | isa_dependent_fxns_t *ps_isa_dependent_fxns = &ps_codec->s_isa_dependent_fxns; |
3554 | 9.84k | enc_loop_fxns_t *ps_enc_loop_fxns = &ps_isa_dependent_fxns->s_enc_loop_fxns; |
3555 | 9.84k | WORD8 i; |
3556 | | |
3557 | | /******************************************************************** |
3558 | | * INITIALIZE CODEC CONTEXT * |
3559 | | ********************************************************************/ |
3560 | | /* encoder presets */ |
3561 | 9.84k | if(ps_codec->s_cfg.u4_enc_speed_preset != IVE_CONFIG) |
3562 | 8.11k | { |
3563 | 8.11k | if(ps_codec->s_cfg.u4_enc_speed_preset == IVE_SLOWEST) |
3564 | 809 | { /* high quality */ |
3565 | | /* enable diamond search */ |
3566 | 809 | ps_codec->s_cfg.u4_me_speed_preset = DMND_SRCH; |
3567 | 809 | ps_codec->s_cfg.u4_enable_fast_sad = 0; |
3568 | | |
3569 | | /* disable intra 4x4 */ |
3570 | 809 | ps_codec->s_cfg.u4_enable_intra_4x4 = 1; |
3571 | 809 | if(!FORCE_FAST_INTRA4X4) |
3572 | 809 | { |
3573 | 809 | ps_enc_loop_fxns->apf_luma_energy_compaction[1] = |
3574 | 809 | isvce_code_luma_intra_macroblock_4x4_rdopt_on; |
3575 | 809 | } |
3576 | | |
3577 | | /* sub pel off */ |
3578 | 809 | ps_codec->s_cfg.u4_enable_hpel = 1; |
3579 | | |
3580 | | /* disabled intra inter gating in Inter slices */ |
3581 | 809 | ps_codec->u4_inter_gate = 0; |
3582 | 809 | } |
3583 | 7.30k | else if(ps_codec->s_cfg.u4_enc_speed_preset == IVE_NORMAL) |
3584 | 924 | { /* normal */ |
3585 | | /* enable diamond search */ |
3586 | 924 | ps_codec->s_cfg.u4_me_speed_preset = DMND_SRCH; |
3587 | 924 | ps_codec->s_cfg.u4_enable_fast_sad = 0; |
3588 | | |
3589 | | /* disable intra 4x4 */ |
3590 | 924 | ps_codec->s_cfg.u4_enable_intra_4x4 = 1; |
3591 | | |
3592 | | /* sub pel off */ |
3593 | 924 | ps_codec->s_cfg.u4_enable_hpel = 1; |
3594 | | |
3595 | | /* disabled intra inter gating in Inter slices */ |
3596 | 924 | ps_codec->u4_inter_gate = 0; |
3597 | 924 | } |
3598 | 6.38k | else if(ps_codec->s_cfg.u4_enc_speed_preset == IVE_FAST) |
3599 | 692 | { /* normal */ |
3600 | | /* enable diamond search */ |
3601 | 692 | ps_codec->s_cfg.u4_me_speed_preset = DMND_SRCH; |
3602 | 692 | ps_codec->s_cfg.u4_enable_fast_sad = 0; |
3603 | | |
3604 | | /* disable intra 4x4 */ |
3605 | 692 | ps_codec->s_cfg.u4_enable_intra_4x4 = 0; |
3606 | | |
3607 | | /* sub pel off */ |
3608 | 692 | ps_codec->s_cfg.u4_enable_hpel = 1; |
3609 | | |
3610 | | /* disabled intra inter gating in Inter slices */ |
3611 | 692 | ps_codec->u4_inter_gate = 1; |
3612 | 692 | } |
3613 | 5.69k | else if(ps_codec->s_cfg.u4_enc_speed_preset == IVE_HIGH_SPEED) |
3614 | 364 | { /* fast */ |
3615 | | /* enable diamond search */ |
3616 | 364 | ps_codec->s_cfg.u4_me_speed_preset = DMND_SRCH; |
3617 | 364 | ps_codec->s_cfg.u4_enable_fast_sad = 0; |
3618 | | |
3619 | | /* disable intra 4x4 */ |
3620 | 364 | ps_codec->s_cfg.u4_enable_intra_4x4 = 0; |
3621 | | |
3622 | | /* sub pel off */ |
3623 | 364 | ps_codec->s_cfg.u4_enable_hpel = 0; |
3624 | | |
3625 | | /* disabled intra inter gating in Inter slices */ |
3626 | 364 | ps_codec->u4_inter_gate = 0; |
3627 | 364 | } |
3628 | 5.32k | else if(ps_codec->s_cfg.u4_enc_speed_preset == IVE_FASTEST) |
3629 | 5.32k | { /* fastest */ |
3630 | | /* enable diamond search */ |
3631 | 5.32k | ps_codec->s_cfg.u4_me_speed_preset = DMND_SRCH; |
3632 | | |
3633 | | /* disable intra 4x4 */ |
3634 | 5.32k | ps_codec->s_cfg.u4_enable_intra_4x4 = 0; |
3635 | | |
3636 | | /* sub pel off */ |
3637 | 5.32k | ps_codec->s_cfg.u4_enable_hpel = 0; |
3638 | | |
3639 | | /* disabled intra inter gating in Inter slices */ |
3640 | 5.32k | ps_codec->u4_inter_gate = 1; |
3641 | 5.32k | } |
3642 | 8.11k | } |
3643 | | |
3644 | | /***************************************************************** |
3645 | | * Initialize AIR inside codec |
3646 | | *****************************************************************/ |
3647 | 9.84k | if(IVE_AIR_MODE_NONE != ps_codec->s_cfg.e_air_mode) |
3648 | 0 | { |
3649 | 0 | isvce_init_air_map(ps_codec); |
3650 | |
|
3651 | 0 | ps_codec->i4_air_pic_cnt = -1; |
3652 | 0 | } |
3653 | | |
3654 | 9.84k | { |
3655 | 9.84k | WORD32 i4_err_code = isvce_rc_init_wrapper(ps_codec); |
3656 | | |
3657 | 9.84k | if(IH264E_SUCCESS != i4_err_code) |
3658 | 63 | { |
3659 | 63 | return i4_err_code; |
3660 | 63 | } |
3661 | 9.84k | } |
3662 | | |
3663 | | /* recon stride */ |
3664 | 9.78k | ps_codec->i4_rec_strd = ALIGN16(ps_codec->s_cfg.u4_max_wd) + PAD_WD; |
3665 | | |
3666 | | /* max ref and reorder cnt */ |
3667 | 9.78k | ps_codec->i4_ref_buf_cnt = ps_codec->s_cfg.u4_max_ref_cnt + ps_codec->s_cfg.u4_max_reorder_cnt; |
3668 | 9.78k | ps_codec->i4_ref_buf_cnt += MAX_CTXT_SETS; |
3669 | 9.78k | ps_codec->i4_ref_buf_cnt += ps_codec->s_cfg.s_svc_params.u1_num_temporal_layers; |
3670 | | |
3671 | 9.78k | DEBUG_HISTOGRAM_INIT(); |
3672 | | |
3673 | | /* Init dependecy vars */ |
3674 | 9.78k | ps_codec->i4_last_inp_buff_received = 0; |
3675 | | |
3676 | | /* At codec start no IDR is pending */ |
3677 | 9.78k | ps_codec->i4_pending_idr_flag = 0; |
3678 | | |
3679 | 22.6k | for(i = 0; i < ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers - 1; i++) |
3680 | 12.8k | { |
3681 | 12.8k | ps_codec->au4_constrained_intra_pred[i] = 1; |
3682 | 12.8k | } |
3683 | | |
3684 | 9.78k | ps_codec->au4_constrained_intra_pred[ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers - 1] = |
3685 | 9.78k | 0; |
3686 | | |
3687 | 9.78k | return IH264E_SUCCESS; |
3688 | 9.84k | } |
3689 | | |
3690 | | /** |
3691 | | ******************************************************************************* |
3692 | | * |
3693 | | * @brief update encoder configuration parameters |
3694 | | * |
3695 | | * @par Description: |
3696 | | * updates encoder configuration parameters from the given config set. |
3697 | | * Initialize/reinitialize codec parameters according to new configurations. |
3698 | | * |
3699 | | * @param[in] ps_codec |
3700 | | * Pointer to codec context |
3701 | | * |
3702 | | * @param[in] ps_cfg |
3703 | | * Pointer to config param set |
3704 | | * |
3705 | | * @remarks none |
3706 | | * |
3707 | | ******************************************************************************* |
3708 | | */ |
3709 | | IH264E_ERROR_T isvce_codec_update_config(isvce_codec_t *ps_codec, isvce_cfg_params_t *ps_cfg) |
3710 | 85.4k | { |
3711 | | /* config params */ |
3712 | 85.4k | isvce_cfg_params_t *ps_curr_cfg = &ps_codec->s_cfg; |
3713 | | |
3714 | | /* error status */ |
3715 | 85.4k | IH264E_ERROR_T err = IH264E_SUCCESS; |
3716 | | |
3717 | | /* temp var */ |
3718 | 85.4k | UWORD32 u4_init_rc = 0; |
3719 | | |
3720 | 85.4k | WORD8 i; |
3721 | | |
3722 | | /***********************/ |
3723 | | /* UPDATE CODEC CONFIG */ |
3724 | | /***********************/ |
3725 | 85.4k | if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_DIMENSIONS) |
3726 | 1.52k | { |
3727 | 1.52k | UWORD32 wd_aln = ALIGN16(ps_cfg->u4_wd); |
3728 | 1.52k | UWORD32 ht_aln = ALIGN16(ps_cfg->u4_ht); |
3729 | | |
3730 | 1.52k | if(ps_curr_cfg->u4_wd != wd_aln || ps_curr_cfg->u4_ht != ht_aln || |
3731 | 1.52k | ps_curr_cfg->u4_disp_wd != ps_cfg->u4_disp_wd || |
3732 | 1.52k | ps_curr_cfg->u4_disp_ht != ps_cfg->u4_disp_ht) |
3733 | 0 | { |
3734 | 0 | ps_curr_cfg->u4_wd = wd_aln; |
3735 | 0 | ps_curr_cfg->u4_ht = ht_aln; |
3736 | |
|
3737 | 0 | ps_curr_cfg->u4_disp_wd = ps_cfg->u4_disp_wd; |
3738 | 0 | ps_curr_cfg->u4_disp_ht = ps_cfg->u4_disp_ht; |
3739 | |
|
3740 | 0 | ps_curr_cfg->i4_wd_mbs = ps_curr_cfg->u4_wd >> 4; |
3741 | 0 | ps_curr_cfg->i4_ht_mbs = ps_curr_cfg->u4_ht >> 4; |
3742 | |
|
3743 | 0 | ps_codec->i4_rec_strd = ALIGN16(ps_cfg->u4_wd) + PAD_WD; |
3744 | | |
3745 | | /* If number of MBs in a frame changes the air map also changes. |
3746 | | * Hence recompute air map also reset air pic cnt */ |
3747 | 0 | if(ps_codec->s_cfg.e_air_mode != IVE_AIR_MODE_NONE) |
3748 | 0 | { |
3749 | | /* re-init the air map */ |
3750 | 0 | isvce_init_air_map(ps_codec); |
3751 | | |
3752 | | /* reset air counter */ |
3753 | 0 | ps_codec->i4_air_pic_cnt = -1; |
3754 | 0 | } |
3755 | | |
3756 | | /* initialize mv bank buffer manager */ |
3757 | 0 | err = isvce_svc_au_data_mgr_add_bufs(ps_codec); |
3758 | 0 | if(err != IH264E_SUCCESS) return err; |
3759 | | |
3760 | | /* initialize ref bank buffer manager */ |
3761 | 0 | err = isvce_svc_au_buf_mgr_add_bufs(ps_codec); |
3762 | 0 | if(err != IH264E_SUCCESS) return err; |
3763 | | |
3764 | | /* since dimension changed, start new sequence by forcing IDR */ |
3765 | 0 | ps_codec->force_curr_frame_type = IV_IDR_FRAME; |
3766 | | |
3767 | | /* in case dimension changes, we need to reinitialize RC as the |
3768 | | * old model shall not fit further */ |
3769 | 0 | u4_init_rc = 1; |
3770 | | |
3771 | | /* when the dimension changes, the header needs to be regenerated */ |
3772 | 0 | ps_codec->i4_gen_header = 1; |
3773 | 0 | } |
3774 | 1.52k | } |
3775 | 83.8k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_FRAMERATE) |
3776 | 4.29k | { |
3777 | 4.29k | UWORD32 u4_src_ticks, u4_tgt_ticks; |
3778 | | |
3779 | 4.29k | u4_src_ticks = ih264e_frame_time_get_src_ticks(ps_codec->s_rate_control.pps_frame_time); |
3780 | | |
3781 | 4.29k | u4_tgt_ticks = ih264e_frame_time_get_tgt_ticks(ps_codec->s_rate_control.pps_frame_time); |
3782 | | |
3783 | | /* Change frame rate */ |
3784 | 4.29k | if(ps_codec->s_cfg.u4_src_frame_rate != ps_cfg->u4_src_frame_rate * 1000) |
3785 | 4.27k | { |
3786 | 4.27k | ps_codec->s_cfg.u4_src_frame_rate = ps_cfg->u4_src_frame_rate * 1000; |
3787 | | |
3788 | 4.27k | ih264e_frame_time_update_src_frame_rate(ps_codec->s_rate_control.pps_frame_time, |
3789 | 4.27k | ps_codec->s_cfg.u4_src_frame_rate); |
3790 | | |
3791 | 4.27k | ih264_time_stamp_update_frame_rate(ps_codec->s_rate_control.pps_time_stamp, |
3792 | 4.27k | ps_codec->s_cfg.u4_src_frame_rate); |
3793 | | |
3794 | 14.1k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
3795 | 9.88k | { |
3796 | 9.88k | irc_change_frame_rate(ps_codec->s_rate_control.apps_rate_control_api[i], |
3797 | 9.88k | ps_codec->s_cfg.u4_src_frame_rate, u4_src_ticks, |
3798 | 9.88k | u4_tgt_ticks); |
3799 | 9.88k | } |
3800 | 4.27k | } |
3801 | | |
3802 | 4.29k | if(ps_codec->s_cfg.u4_tgt_frame_rate != ps_cfg->u4_tgt_frame_rate * 1000) |
3803 | 4.27k | { |
3804 | 4.27k | ps_codec->s_cfg.u4_tgt_frame_rate = ps_cfg->u4_tgt_frame_rate * 1000; |
3805 | | |
3806 | 4.27k | ih264e_frame_time_update_tgt_frame_rate(ps_codec->s_rate_control.pps_frame_time, |
3807 | 4.27k | ps_codec->s_cfg.u4_tgt_frame_rate); |
3808 | | |
3809 | 14.1k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
3810 | 9.88k | { |
3811 | 9.88k | irc_change_frame_rate(ps_codec->s_rate_control.apps_rate_control_api[i], |
3812 | 9.88k | ps_codec->s_cfg.u4_src_frame_rate, u4_src_ticks, |
3813 | 9.88k | u4_tgt_ticks); |
3814 | | |
3815 | 9.88k | irc_change_frm_rate_for_bit_alloc(ps_codec->s_rate_control.apps_rate_control_api[i], |
3816 | 9.88k | ps_codec->s_cfg.u4_tgt_frame_rate); |
3817 | 9.88k | } |
3818 | 4.27k | } |
3819 | 4.29k | } |
3820 | 79.5k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_BITRATE) |
3821 | 4.83k | { |
3822 | 19.3k | for(i = 0; i < MAX_NUM_SPATIAL_LAYERS; i++) |
3823 | 14.4k | { |
3824 | 14.4k | if(ps_curr_cfg->au4_target_bitrate[i] != ps_cfg->au4_target_bitrate[i]) |
3825 | 11.1k | { |
3826 | 11.1k | if(IVE_RC_NONE != ps_curr_cfg->e_rc_mode) |
3827 | 7.68k | irc_change_avg_bit_rate(ps_codec->s_rate_control.apps_rate_control_api[i], |
3828 | 7.68k | ps_cfg->au4_target_bitrate[i]); |
3829 | | |
3830 | 11.1k | ps_curr_cfg->au4_target_bitrate[i] = ps_cfg->au4_target_bitrate[i]; |
3831 | 11.1k | } |
3832 | 14.4k | } |
3833 | 4.83k | } |
3834 | 74.7k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_FRAMETYPE) |
3835 | 918 | { |
3836 | 918 | switch(ps_cfg->e_frame_type) |
3837 | 918 | { |
3838 | 0 | case IV_I_FRAME: |
3839 | 0 | ps_codec->force_curr_frame_type = IV_I_FRAME; |
3840 | 0 | break; |
3841 | | |
3842 | 918 | case IV_IDR_FRAME: |
3843 | 918 | ps_codec->force_curr_frame_type = IV_IDR_FRAME; |
3844 | 918 | break; |
3845 | | |
3846 | 0 | case IV_P_FRAME: |
3847 | 0 | default: |
3848 | 0 | break; |
3849 | 918 | } |
3850 | 918 | } |
3851 | 73.8k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_ME_PARAMS) |
3852 | 2.83k | { |
3853 | 2.83k | if(ps_curr_cfg->u4_enc_speed_preset == IVE_CONFIG) |
3854 | 1.13k | { |
3855 | 1.13k | ps_codec->s_cfg.u4_enable_hpel = ps_cfg->u4_enable_hpel; |
3856 | 1.13k | ps_codec->s_cfg.u4_enable_fast_sad = ps_cfg->u4_enable_fast_sad; |
3857 | 1.13k | ps_codec->s_cfg.u4_me_speed_preset = ps_cfg->u4_me_speed_preset; |
3858 | 1.13k | ps_codec->s_cfg.u4_enable_qpel = ps_cfg->u4_enable_qpel; |
3859 | 1.13k | } |
3860 | 1.69k | else if(ps_curr_cfg->u4_enc_speed_preset == IVE_FASTEST) |
3861 | 118 | { |
3862 | 118 | ps_codec->s_cfg.u4_enable_fast_sad = ps_cfg->u4_enable_fast_sad; |
3863 | 118 | } |
3864 | 2.83k | ps_codec->s_cfg.u4_srch_rng_x = ps_cfg->u4_srch_rng_x; |
3865 | 2.83k | ps_codec->s_cfg.u4_srch_rng_y = ps_cfg->u4_srch_rng_y; |
3866 | | |
3867 | 2.83k | if(ps_codec->s_cfg.u4_enable_alt_ref != ps_cfg->u4_enable_alt_ref) |
3868 | 0 | { |
3869 | 0 | ps_codec->s_cfg.u4_enable_alt_ref = ps_cfg->u4_enable_alt_ref; |
3870 | 0 | ps_codec->u4_is_curr_frm_ref = 1; |
3871 | 0 | } |
3872 | 2.83k | } |
3873 | 71.0k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_IPE_PARAMS) |
3874 | 4.93k | { |
3875 | 4.93k | ps_curr_cfg->u4_enc_speed_preset = ps_cfg->u4_enc_speed_preset; |
3876 | | |
3877 | 4.93k | if(ps_curr_cfg->u4_enc_speed_preset == IVE_SLOWEST) |
3878 | 810 | { |
3879 | 810 | isa_dependent_fxns_t *ps_isa_dependent_fxns = &ps_codec->s_isa_dependent_fxns; |
3880 | 810 | enc_loop_fxns_t *ps_enc_loop_fxns = &ps_isa_dependent_fxns->s_enc_loop_fxns; |
3881 | | |
3882 | | /* enable diamond search */ |
3883 | 810 | ps_curr_cfg->u4_me_speed_preset = DMND_SRCH; |
3884 | 810 | ps_curr_cfg->u4_enable_fast_sad = 0; |
3885 | | |
3886 | | /* disable intra 4x4 */ |
3887 | 810 | ps_curr_cfg->u4_enable_intra_4x4 = 1; |
3888 | 810 | ps_enc_loop_fxns->apf_luma_energy_compaction[1] = |
3889 | 810 | isvce_code_luma_intra_macroblock_4x4_rdopt_on; |
3890 | | |
3891 | | /* sub pel off */ |
3892 | 810 | ps_curr_cfg->u4_enable_hpel = 1; |
3893 | | |
3894 | | /* disabled intra inter gating in Inter slices */ |
3895 | 810 | ps_codec->u4_inter_gate = 0; |
3896 | 810 | } |
3897 | 4.12k | else if(ps_curr_cfg->u4_enc_speed_preset == IVE_NORMAL) |
3898 | 924 | { /* normal */ |
3899 | | /* enable diamond search */ |
3900 | 924 | ps_curr_cfg->u4_me_speed_preset = DMND_SRCH; |
3901 | 924 | ps_curr_cfg->u4_enable_fast_sad = 0; |
3902 | | |
3903 | | /* disable intra 4x4 */ |
3904 | 924 | ps_curr_cfg->u4_enable_intra_4x4 = 1; |
3905 | | |
3906 | | /* sub pel off */ |
3907 | 924 | ps_curr_cfg->u4_enable_hpel = 1; |
3908 | | |
3909 | | /* disabled intra inter gating in Inter slices */ |
3910 | 924 | ps_codec->u4_inter_gate = 0; |
3911 | 924 | } |
3912 | 3.20k | else if(ps_curr_cfg->u4_enc_speed_preset == IVE_FAST) |
3913 | 693 | { /* normal */ |
3914 | | /* enable diamond search */ |
3915 | 693 | ps_curr_cfg->u4_me_speed_preset = DMND_SRCH; |
3916 | 693 | ps_curr_cfg->u4_enable_fast_sad = 0; |
3917 | | |
3918 | | /* disable intra 4x4 */ |
3919 | 693 | ps_curr_cfg->u4_enable_intra_4x4 = 0; |
3920 | | |
3921 | | /* sub pel off */ |
3922 | 693 | ps_curr_cfg->u4_enable_hpel = 1; |
3923 | | |
3924 | | /* disabled intra inter gating in Inter slices */ |
3925 | 693 | ps_codec->u4_inter_gate = 1; |
3926 | 693 | } |
3927 | 2.50k | else if(ps_curr_cfg->u4_enc_speed_preset == IVE_HIGH_SPEED) |
3928 | 366 | { /* fast */ |
3929 | | /* enable diamond search */ |
3930 | 366 | ps_curr_cfg->u4_me_speed_preset = DMND_SRCH; |
3931 | 366 | ps_curr_cfg->u4_enable_fast_sad = 0; |
3932 | | |
3933 | | /* disable intra 4x4 */ |
3934 | 366 | ps_curr_cfg->u4_enable_intra_4x4 = 0; |
3935 | | |
3936 | | /* sub pel off */ |
3937 | 366 | ps_curr_cfg->u4_enable_hpel = 0; |
3938 | | |
3939 | | /* disabled intra inter gating in Inter slices */ |
3940 | 366 | ps_codec->u4_inter_gate = 0; |
3941 | 366 | } |
3942 | 2.14k | else if(ps_curr_cfg->u4_enc_speed_preset == IVE_FASTEST) |
3943 | 397 | { /* fastest */ |
3944 | | /* enable diamond search */ |
3945 | 397 | ps_curr_cfg->u4_me_speed_preset = DMND_SRCH; |
3946 | | // u4_num_layers = 4; |
3947 | | |
3948 | | /* disable intra 4x4 */ |
3949 | 397 | ps_curr_cfg->u4_enable_intra_4x4 = 0; |
3950 | | |
3951 | | /* sub pel off */ |
3952 | 397 | ps_curr_cfg->u4_enable_hpel = 0; |
3953 | | |
3954 | | /* disabled intra inter gating in Inter slices */ |
3955 | 397 | ps_codec->u4_inter_gate = 1; |
3956 | 397 | } |
3957 | 1.74k | else if(ps_curr_cfg->u4_enc_speed_preset == IVE_CONFIG) |
3958 | 1.74k | { |
3959 | 1.74k | ps_curr_cfg->u4_enable_intra_4x4 = ps_cfg->u4_enable_intra_4x4; |
3960 | 1.74k | } |
3961 | 4.93k | } |
3962 | 66.0k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_GOP_PARAMS) |
3963 | 4.93k | { |
3964 | 4.93k | if(ps_curr_cfg->u4_i_frm_interval != ps_cfg->u4_i_frm_interval) |
3965 | 4.92k | { |
3966 | 4.92k | ps_curr_cfg->u4_i_frm_interval = ps_cfg->u4_i_frm_interval; |
3967 | | |
3968 | | /* reset air counter */ |
3969 | 4.92k | ps_codec->i4_air_pic_cnt = -1; |
3970 | | |
3971 | | /* re-init air map */ |
3972 | 4.92k | isvce_init_air_map(ps_codec); |
3973 | | |
3974 | | /*Effect intra frame interval change*/ |
3975 | 16.3k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
3976 | 11.3k | { |
3977 | 11.3k | irc_change_intra_frm_int_call(ps_codec->s_rate_control.apps_rate_control_api[i], |
3978 | 11.3k | ps_curr_cfg->u4_i_frm_interval); |
3979 | 11.3k | } |
3980 | 4.92k | } |
3981 | | |
3982 | 4.93k | ps_curr_cfg->u4_idr_frm_interval = ps_cfg->u4_idr_frm_interval; |
3983 | 4.93k | } |
3984 | 61.1k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_DEBLOCK_PARAMS) |
3985 | 4.93k | { |
3986 | 4.93k | ps_curr_cfg->u4_disable_deblock_level = ps_cfg->u4_disable_deblock_level; |
3987 | 4.93k | } |
3988 | 56.2k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_QP) |
3989 | 1.92k | { |
3990 | 6.54k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
3991 | 4.61k | { |
3992 | 4.61k | UWORD8 au1_init_qp[MAX_PIC_TYPE]; |
3993 | 4.61k | UWORD8 au1_min_max_qp[2 * MAX_PIC_TYPE]; |
3994 | 4.61k | UWORD8 au1_min_max_avc_qp[2 * MAX_PIC_TYPE]; |
3995 | | |
3996 | 4.61k | ps_codec->s_cfg.au4_i_qp_max[i] = ps_cfg->au4_i_qp_max[i]; |
3997 | 4.61k | ps_codec->s_cfg.au4_i_qp_min[i] = ps_cfg->au4_i_qp_min[i]; |
3998 | 4.61k | ps_codec->s_cfg.au4_i_qp[i] = ps_cfg->au4_i_qp[i]; |
3999 | | |
4000 | 4.61k | ps_codec->s_cfg.au4_p_qp_max[i] = ps_cfg->au4_p_qp_max[i]; |
4001 | 4.61k | ps_codec->s_cfg.au4_p_qp_min[i] = ps_cfg->au4_p_qp_min[i]; |
4002 | 4.61k | ps_codec->s_cfg.au4_p_qp[i] = ps_cfg->au4_p_qp[i]; |
4003 | | |
4004 | 4.61k | ps_codec->s_cfg.au4_b_qp_max[i] = ps_cfg->au4_b_qp_max[i]; |
4005 | 4.61k | ps_codec->s_cfg.au4_b_qp_min[i] = ps_cfg->au4_b_qp_min[i]; |
4006 | 4.61k | ps_codec->s_cfg.au4_b_qp[i] = ps_cfg->au4_b_qp[i]; |
4007 | | |
4008 | | /* update rc lib with modified qp */ |
4009 | 4.61k | au1_init_qp[0] = gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_i_qp[i]]; |
4010 | 4.61k | au1_init_qp[1] = gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_p_qp[i]]; |
4011 | 4.61k | au1_init_qp[2] = gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_b_qp[i]]; |
4012 | | |
4013 | 4.61k | irc_change_init_qp(ps_codec->s_rate_control.apps_rate_control_api[i], au1_init_qp); |
4014 | | |
4015 | 4.61k | au1_min_max_qp[2 * I_PIC] = gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_i_qp_min[i]]; |
4016 | 4.61k | au1_min_max_qp[2 * I_PIC + 1] = |
4017 | 4.61k | gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_i_qp_max[i]]; |
4018 | | |
4019 | 4.61k | au1_min_max_qp[2 * P_PIC] = gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_p_qp_min[i]]; |
4020 | 4.61k | au1_min_max_qp[2 * P_PIC + 1] = |
4021 | 4.61k | gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_p_qp_max[i]]; |
4022 | | |
4023 | 4.61k | au1_min_max_qp[2 * B_PIC] = gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_b_qp_min[i]]; |
4024 | 4.61k | au1_min_max_qp[2 * B_PIC + 1] = |
4025 | 4.61k | gau1_h264_to_mpeg2_qmap[ps_codec->s_cfg.au4_b_qp_max[i]]; |
4026 | | |
4027 | 4.61k | au1_min_max_avc_qp[2 * I_PIC] = ps_codec->s_cfg.au4_i_qp_min[i]; |
4028 | 4.61k | au1_min_max_avc_qp[2 * I_PIC + 1] = ps_codec->s_cfg.au4_i_qp_max[i]; |
4029 | | |
4030 | 4.61k | au1_min_max_avc_qp[2 * P_PIC] = ps_codec->s_cfg.au4_p_qp_min[i]; |
4031 | 4.61k | au1_min_max_avc_qp[2 * P_PIC + 1] = ps_codec->s_cfg.au4_p_qp_max[i]; |
4032 | | |
4033 | 4.61k | au1_min_max_avc_qp[2 * B_PIC] = ps_codec->s_cfg.au4_b_qp_min[i]; |
4034 | 4.61k | au1_min_max_avc_qp[2 * B_PIC + 1] = ps_codec->s_cfg.au4_b_qp_max[i]; |
4035 | | |
4036 | 4.61k | irc_change_qp_constraints(ps_codec->s_rate_control.apps_rate_control_api[i], |
4037 | 4.61k | au1_min_max_qp, au1_min_max_avc_qp); |
4038 | 4.61k | } |
4039 | 1.92k | } |
4040 | 54.2k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_ENC_MODE) |
4041 | 4.93k | { |
4042 | 4.93k | ps_codec->s_cfg.e_enc_mode = ps_cfg->e_enc_mode; |
4043 | | |
4044 | 4.93k | if(ps_codec->s_cfg.e_enc_mode == IVE_ENC_MODE_HEADER) |
4045 | 4.93k | { |
4046 | 4.93k | ps_codec->i4_header_mode = 1; |
4047 | 4.93k | ps_codec->s_cfg.e_enc_mode = IVE_ENC_MODE_PICTURE; |
4048 | 4.93k | } |
4049 | 0 | else |
4050 | 0 | { |
4051 | 0 | ps_codec->i4_header_mode = 0; |
4052 | 0 | } |
4053 | 4.93k | } |
4054 | 49.3k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_VBV_PARAMS && |
4055 | 49.3k | IVE_RC_NONE != ps_codec->s_cfg.e_rc_mode) |
4056 | 3.31k | { |
4057 | 11.0k | for(i = 0; i < ps_cfg->s_svc_params.u1_num_spatial_layers; i++) |
4058 | 7.75k | { |
4059 | 7.75k | ps_codec->s_cfg.au4_vbv_buffer_delay[i] = ps_cfg->au4_vbv_buffer_delay[i]; |
4060 | 7.75k | } |
4061 | | // irc_change_buffer_delay(ps_codec->s_rate_control.pps_rate_control_api, |
4062 | | // ps_codec->s_cfg.u4_vbv_buffer_delay); |
4063 | | |
4064 | | // TODO: remove this when the support for changing buffer dynamically |
4065 | | // is yet to be added. |
4066 | 3.31k | u4_init_rc = 1; |
4067 | 3.31k | } |
4068 | 46.0k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_AIR_PARAMS) |
4069 | 4.93k | { |
4070 | 4.93k | if(ps_curr_cfg->e_air_mode != ps_cfg->e_air_mode || |
4071 | 4.93k | ps_curr_cfg->u4_air_refresh_period != ps_cfg->u4_air_refresh_period) |
4072 | 4.91k | { |
4073 | 4.91k | ps_curr_cfg->e_air_mode = ps_cfg->e_air_mode; |
4074 | 4.91k | ps_curr_cfg->u4_air_refresh_period = ps_cfg->u4_air_refresh_period; |
4075 | | |
4076 | 4.91k | isvce_init_air_map(ps_codec); |
4077 | | |
4078 | | /* reset air counter */ |
4079 | 4.91k | ps_codec->i4_air_pic_cnt = -1; |
4080 | 4.91k | } |
4081 | 4.93k | } |
4082 | 41.0k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_PROFILE_PARAMS) |
4083 | 4.93k | { |
4084 | 4.93k | ps_codec->s_cfg.e_profile = ps_cfg->e_profile; |
4085 | 4.93k | ps_codec->s_cfg.u4_entropy_coding_mode = ps_cfg->u4_entropy_coding_mode; |
4086 | 4.93k | } |
4087 | 36.1k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_NUM_CORES) |
4088 | 4.93k | { |
4089 | 4.93k | ps_codec->s_cfg.u4_num_cores = ps_cfg->u4_num_cores; |
4090 | 4.93k | } |
4091 | 31.2k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_VUI_PARAMS) |
4092 | 4.93k | { |
4093 | 4.93k | ps_codec->s_cfg.s_vui = ps_cfg->s_vui; |
4094 | 4.93k | } |
4095 | | |
4096 | 26.2k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_SEI_MDCV_PARAMS) |
4097 | 4.93k | { |
4098 | 4.93k | ps_codec->s_cfg.s_sei.u1_sei_mdcv_params_present_flag = |
4099 | 4.93k | ps_cfg->s_sei.u1_sei_mdcv_params_present_flag; |
4100 | 4.93k | ps_codec->s_cfg.s_sei.s_sei_mdcv_params = ps_cfg->s_sei.s_sei_mdcv_params; |
4101 | 4.93k | } |
4102 | 21.3k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_SEI_CLL_PARAMS) |
4103 | 4.93k | { |
4104 | 4.93k | ps_codec->s_cfg.s_sei.u1_sei_cll_params_present_flag = |
4105 | 4.93k | ps_cfg->s_sei.u1_sei_cll_params_present_flag; |
4106 | 4.93k | ps_codec->s_cfg.s_sei.s_sei_cll_params = ps_cfg->s_sei.s_sei_cll_params; |
4107 | 4.93k | } |
4108 | 16.4k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_SEI_AVE_PARAMS) |
4109 | 4.93k | { |
4110 | 4.93k | ps_codec->s_cfg.s_sei.u1_sei_ave_params_present_flag = |
4111 | 4.93k | ps_cfg->s_sei.u1_sei_ave_params_present_flag; |
4112 | 4.93k | ps_codec->s_cfg.s_sei.s_sei_ave_params = ps_cfg->s_sei.s_sei_ave_params; |
4113 | 4.93k | } |
4114 | 11.4k | else if(ps_cfg->e_cmd == ISVCE_CMD_CTL_SET_SEI_CCV_PARAMS) |
4115 | 4.93k | { |
4116 | 4.93k | ps_codec->s_cfg.s_sei.u1_sei_ccv_params_present_flag = |
4117 | 4.93k | ps_cfg->s_sei.u1_sei_ccv_params_present_flag; |
4118 | 4.93k | ps_codec->s_cfg.s_sei.s_sei_ccv_params = ps_cfg->s_sei.s_sei_ccv_params; |
4119 | 4.93k | } |
4120 | | |
4121 | | /* reset RC model */ |
4122 | 85.4k | if(u4_init_rc) |
4123 | 3.31k | { |
4124 | 3.31k | err = isvce_rc_init_wrapper(ps_codec); |
4125 | 3.31k | } |
4126 | | |
4127 | 85.4k | return err; |
4128 | 85.4k | } |
4129 | | |
4130 | | static FORCEINLINE void isvce_change_rc_init_qp(void *pv_rate_control_api, UWORD8 u1_qp) |
4131 | 0 | { |
4132 | 0 | UWORD8 au1_pic_qps[MAX_PIC_TYPE]; |
4133 | 0 | WORD32 i; |
4134 | |
|
4135 | 0 | for(i = 0; i < MAX_PIC_TYPE; i++) |
4136 | 0 | { |
4137 | 0 | au1_pic_qps[i] = gau1_h264_to_mpeg2_qmap[CLIP3(MIN_H264_QP, MAX_H264_QP, u1_qp + i)]; |
4138 | 0 | } |
4139 | |
|
4140 | 0 | irc_change_init_qp(pv_rate_control_api, au1_pic_qps); |
4141 | 0 | } |
4142 | | |
4143 | | /** |
4144 | | ******************************************************************************* |
4145 | | * |
4146 | | * @brief |
4147 | | * Queues the current buffer, gets back a another buffer for encoding with |
4148 | | *corrent picture type |
4149 | | * |
4150 | | * @par Description: |
4151 | | * This function performs 3 distinct but related functions. |
4152 | | * 1) Maintains an input queue [Note the the term queue donot imply a |
4153 | | * first-in first-out logic here] that queues input and dequeues them so |
4154 | | * that input frames can be encoded at any predetermined encoding order |
4155 | | * 2) Uses RC library to decide which frame must be encoded in current pass |
4156 | | * and which picture type it must be encoded to. |
4157 | | * 3) Uses RC library to decide the QP at which current frame has to be |
4158 | | * encoded |
4159 | | * 4) Determines if the current picture must be encoded or not based on |
4160 | | * PRE-ENC skip |
4161 | | * |
4162 | | * Input queue is used for storing input buffers till they are used for |
4163 | | * encoding. This queue is maintained at ps_codec->as_inp_list. Whenever a |
4164 | | * valid input comes, it is added to the end of queue. This same input is |
4165 | | * added to RC queue using the identifier as ps_codec->i4_pic_cnt. Hence any |
4166 | | * pic from RC can be located in the input queue easily. |
4167 | | * |
4168 | | * The dequeue operation does not start till we have |
4169 | | *ps_codec->s_cfg.u4_max_num_bframes frames in the queue. THis is done in order |
4170 | | *to ensure that once output starts we will have a constant stream of output |
4171 | | *with no gaps. |
4172 | | * |
4173 | | * THe output frame order is governed by RC library. When ever we dequeue a |
4174 | | * buffer from RC library, it ensures that we will get them in encoding |
4175 | | *order With the output of RC library, we can use the picture id to dequeue the |
4176 | | * corresponding buffer from input queue and encode it. |
4177 | | * |
4178 | | * Condition at the end of stream. |
4179 | | * ------------------------------- |
4180 | | * At the last valid buffer from the app, we will get ps_ive_ip->u4_is_last |
4181 | | * to be set. This will the given to lib when appropriate input buffer is |
4182 | | * given to encoding. |
4183 | | * |
4184 | | * Since we have to output is not in sync with input, we will have frames |
4185 | | *to encode even after we recive the last vaild input buffer. Hence we have to |
4186 | | * make sure that we donot queue any new buffers once we get the flag [It |
4187 | | *may mess up GOP ?]. This is acheived by setting |
4188 | | *ps_codec->i4_last_inp_buff_received to act as a permenent marker for last |
4189 | | *frame recived [This may not be needed, because in our current app, all buffers |
4190 | | *after the last are marked as last. But can we rely on that?] . Hence after |
4191 | | *this flgag is set no new buffers are queued. |
4192 | | * |
4193 | | * @param[in] ps_codec |
4194 | | * Pointer to codec descriptor |
4195 | | * |
4196 | | * @param[in] ps_ive_ip |
4197 | | * Current input buffer to the encoder |
4198 | | * |
4199 | | * @param[out] ps_inp |
4200 | | * Buffer to be encoded in the current pass |
4201 | | * |
4202 | | * @returns |
4203 | | * Flag indicating if we have a pre-enc skip or not |
4204 | | * |
4205 | | * @remarks |
4206 | | * TODO (bpic) |
4207 | | * The check for null ans is last is redudent. |
4208 | | * Need to see if we can remove it |
4209 | | * |
4210 | | ******************************************************************************* |
4211 | | */ |
4212 | | WORD32 isvce_input_queue_update(isvce_codec_t *ps_codec, ive_video_encode_ip_t *ps_ive_ip, |
4213 | | isvce_inp_buf_t *ps_enc_buff, WORD8 i1_layer_id) |
4214 | 78.3k | { |
4215 | 78.3k | isvce_inp_buf_t *ps_inp_buf; |
4216 | 78.3k | picture_type_e e_pictype; |
4217 | 78.3k | WORD32 i4_skip; |
4218 | 78.3k | UWORD32 ctxt_sel, u4_pic_id, u4_pic_disp_id; |
4219 | 78.3k | UWORD8 u1_frame_qp = MAX_H264_QP; |
4220 | 78.3k | UWORD32 max_frame_bits = 0x7FFFFFFF; |
4221 | | |
4222 | 78.3k | WORD32 i; |
4223 | | |
4224 | | /* Mark that the last input frame has been received */ |
4225 | 78.3k | if(ps_ive_ip->u4_is_last == 1) |
4226 | 12.4k | { |
4227 | 12.4k | ps_codec->i4_last_inp_buff_received = 1; |
4228 | 12.4k | } |
4229 | | |
4230 | 78.3k | if(ps_ive_ip->s_inp_buf.apv_bufs[0] == NULL && !ps_codec->i4_last_inp_buff_received) |
4231 | 0 | { |
4232 | 0 | ps_enc_buff->s_inp_props.s_raw_buf.apv_bufs[0] = NULL; |
4233 | 0 | ps_enc_buff->s_inp_props.u4_is_last = ps_ive_ip->u4_is_last; |
4234 | 0 | return 0; |
4235 | 0 | } |
4236 | | |
4237 | | /*************************************************************************** |
4238 | | * Check for pre enc skip |
4239 | | * When src and target frame rates donot match, we skip some frames to |
4240 | | * maintain the relation ship between them |
4241 | | **************************************************************************/ |
4242 | 78.3k | { |
4243 | 78.3k | WORD32 skip_src; |
4244 | | |
4245 | 78.3k | skip_src = isvce_update_rc_framerates( |
4246 | 78.3k | ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id], |
4247 | 78.3k | ps_codec->s_rate_control.pps_pd_frm_rate, ps_codec->s_rate_control.pps_time_stamp, |
4248 | 78.3k | ps_codec->s_rate_control.pps_frame_time); |
4249 | | |
4250 | 78.3k | if(skip_src) |
4251 | 141 | { |
4252 | 141 | ps_enc_buff->s_inp_props.u4_is_last = ps_ive_ip->u4_is_last; |
4253 | 141 | return 1; |
4254 | 141 | } |
4255 | 78.3k | } |
4256 | | |
4257 | | /*************************************************************************** |
4258 | | *Queue the input to the queue |
4259 | | **************************************************************************/ |
4260 | 78.2k | ps_inp_buf = &(ps_codec->as_inp_list[ps_codec->i4_pic_cnt % SVC_MAX_NUM_INP_FRAMES]); |
4261 | | |
4262 | | /* copy input info. to internal structure */ |
4263 | 78.2k | ps_inp_buf->s_inp_props.s_raw_buf = ps_ive_ip->s_inp_buf; |
4264 | 78.2k | ps_inp_buf->s_inp_props.u4_timestamp_low = ps_ive_ip->u4_timestamp_low; |
4265 | 78.2k | ps_inp_buf->s_inp_props.u4_timestamp_high = ps_ive_ip->u4_timestamp_high; |
4266 | 78.2k | ps_inp_buf->s_inp_props.u4_is_last = ps_ive_ip->u4_is_last; |
4267 | 78.2k | ps_inp_buf->s_inp_props.pv_mb_info = ps_ive_ip->pv_mb_info; |
4268 | 78.2k | ps_inp_buf->s_inp_props.u4_mb_info_type = ps_ive_ip->u4_mb_info_type; |
4269 | 78.2k | ps_inp_buf->s_inp_props.pv_pic_info = ps_ive_ip->pv_pic_info; |
4270 | 78.2k | ps_inp_buf->s_inp_props.u4_pic_info_type = ps_ive_ip->u4_pic_info_type; |
4271 | | |
4272 | 78.2k | ps_inp_buf->s_inp_props.u1_sei_ccv_params_present_flag = |
4273 | 78.2k | ps_codec->s_cfg.s_sei.u1_sei_ccv_params_present_flag; |
4274 | 78.2k | ps_inp_buf->s_inp_props.s_sei_ccv = ps_codec->s_cfg.s_sei.s_sei_ccv_params; |
4275 | | |
4276 | 78.2k | if(ps_inp_buf->s_inp_props.s_raw_buf.apv_bufs[0]) |
4277 | 65.7k | isvce_svc_inp_buf_populate(ps_codec, ps_inp_buf); |
4278 | | |
4279 | | /*************************************************************************** |
4280 | | * Now we should add the picture to RC stack here |
4281 | | **************************************************************************/ |
4282 | | /* |
4283 | | * If an I frame has been requested, ask RC to force it |
4284 | | * For IDR requests, we have to ask RC to force I and set IDR by our selves |
4285 | | * since RC Donot know about IDR. For forcing an IDR at dequeue stage we |
4286 | | * should record that an IDR has been requested some where. Hence we will |
4287 | | * store it in the u4_idr_inp_list at a position same as that of input frame |
4288 | | */ |
4289 | 78.2k | { |
4290 | 78.2k | WORD32 i4_force_idr, i4_force_i; |
4291 | | |
4292 | 78.2k | i4_force_idr = (ps_codec->force_curr_frame_type == IV_IDR_FRAME); |
4293 | 78.2k | i4_force_idr |= !(ps_codec->i4_pic_cnt % ps_codec->s_cfg.u4_idr_frm_interval); |
4294 | | |
4295 | 78.2k | i4_force_i = (ps_codec->force_curr_frame_type == IV_I_FRAME); |
4296 | | |
4297 | 78.2k | ps_codec->i4_pending_idr_flag |= i4_force_idr; |
4298 | | |
4299 | 78.2k | if((ps_codec->i4_pic_cnt > 0) && (i4_force_idr || i4_force_i)) |
4300 | 9.51k | { |
4301 | 9.51k | irc_force_I_frame(ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id]); |
4302 | 9.51k | } |
4303 | | |
4304 | 78.2k | if(i1_layer_id == (ps_codec->s_cfg.s_svc_params.u1_num_spatial_layers - 1)) |
4305 | 36.1k | { |
4306 | 36.1k | ps_codec->force_curr_frame_type = IV_NA_FRAME; |
4307 | 36.1k | } |
4308 | 78.2k | } |
4309 | | |
4310 | 78.2k | irc_add_picture_to_stack(ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id], |
4311 | 78.2k | ps_codec->i4_pic_cnt); |
4312 | | |
4313 | | /* Delay */ |
4314 | 78.2k | if(ps_codec->i4_encode_api_call_cnt < (WORD32) (ps_codec->s_cfg.u4_num_bframes)) |
4315 | 0 | { |
4316 | 0 | ps_enc_buff->s_inp_props.s_raw_buf.apv_bufs[0] = NULL; |
4317 | 0 | ps_enc_buff->s_inp_props.u4_is_last = 0; |
4318 | 0 | return 0; |
4319 | 0 | } |
4320 | | |
4321 | | /*************************************************************************** |
4322 | | * Get a new pic to encode |
4323 | | **************************************************************************/ |
4324 | | /* Query the picture_type */ |
4325 | 78.2k | e_pictype = |
4326 | 78.2k | isvce_rc_get_picture_details(ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id], |
4327 | 78.2k | (WORD32 *) (&u4_pic_id), (WORD32 *) (&u4_pic_disp_id)); |
4328 | | |
4329 | 78.2k | switch(e_pictype) |
4330 | 78.2k | { |
4331 | 26.3k | case I_PIC: |
4332 | 26.3k | ps_codec->pic_type = PIC_I; |
4333 | 26.3k | break; |
4334 | 51.8k | case P_PIC: |
4335 | 51.8k | ps_codec->pic_type = PIC_P; |
4336 | 51.8k | break; |
4337 | 0 | case B_PIC: |
4338 | 0 | ps_codec->pic_type = PIC_B; |
4339 | 0 | break; |
4340 | 0 | default: |
4341 | 0 | ps_codec->pic_type = PIC_NA; |
4342 | 0 | ps_enc_buff->s_inp_props.s_raw_buf.apv_bufs[0] = NULL; |
4343 | 0 | return 0; |
4344 | 78.2k | } |
4345 | | |
4346 | | /* Set IDR if it has been requested */ |
4347 | 78.2k | if(ps_codec->pic_type == PIC_I) |
4348 | 26.3k | { |
4349 | 26.3k | ps_codec->pic_type = ps_codec->i4_pending_idr_flag ? PIC_IDR : ps_codec->pic_type; |
4350 | 26.3k | ps_codec->i4_pending_idr_flag = 0; |
4351 | 26.3k | } |
4352 | | |
4353 | 78.2k | if(ps_codec->s_rate_control.e_rc_type != CONST_QP && ps_codec->u1_enable_init_qp && |
4354 | 78.2k | (u4_pic_id == 0 || |
4355 | 0 | irc_is_scenecut(ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id]))) |
4356 | 0 | { |
4357 | 0 | DOUBLE d_bpp; |
4358 | |
|
4359 | 0 | svc_rc_utils_ctxt_t *ps_svc_rc_utils = &ps_codec->s_rate_control.s_rc_utils; |
4360 | |
|
4361 | 0 | UWORD32 u4_src_fps = ps_codec->s_cfg.u4_src_frame_rate / 1000; |
4362 | 0 | UWORD32 u4_wd = ps_inp_buf->as_layer_yuv_buf_props[i1_layer_id].u4_width; |
4363 | 0 | UWORD32 u4_ht = ps_inp_buf->as_layer_yuv_buf_props[i1_layer_id].u4_height; |
4364 | 0 | DOUBLE d_gpp = |
4365 | 0 | isvce_compute_gpp(ps_svc_rc_utils, &ps_inp_buf->as_layer_yuv_buf_props[i1_layer_id]); |
4366 | |
|
4367 | 0 | d_bpp = ((DOUBLE) irc_get_vbv_buf_size( |
4368 | 0 | ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id]) / |
4369 | 0 | 10.) / |
4370 | 0 | ((DOUBLE) (u4_src_fps * u4_wd * u4_ht)); |
4371 | |
|
4372 | 0 | u1_frame_qp = (UWORD8) irc_get_frame_level_init_qp( |
4373 | 0 | ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id], |
4374 | 0 | ps_codec->s_rate_control.e_rc_type, e_pictype, d_bpp, d_gpp); |
4375 | |
|
4376 | 0 | isvce_change_rc_init_qp(ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id], |
4377 | 0 | u1_frame_qp); |
4378 | |
|
4379 | 0 | ps_codec->au4_frame_qp[i1_layer_id] = u1_frame_qp; |
4380 | 0 | } |
4381 | 78.2k | else |
4382 | 78.2k | { |
4383 | | /* Get current frame Qp */ |
4384 | 78.2k | u1_frame_qp = (UWORD8) irc_get_frame_level_qp( |
4385 | 78.2k | ps_codec->s_rate_control.apps_rate_control_api[i1_layer_id], e_pictype, max_frame_bits); |
4386 | 78.2k | ps_codec->au4_frame_qp[i1_layer_id] = gau1_mpeg2_to_h264_qmap[u1_frame_qp]; |
4387 | 78.2k | } |
4388 | | |
4389 | | /* |
4390 | | * copy the pic id to poc because the display order is assumed to be same |
4391 | | * as input order |
4392 | | */ |
4393 | 78.2k | ps_codec->i4_poc = u4_pic_id; |
4394 | | |
4395 | | /*************************************************************************** |
4396 | | * Now retrieve the correct picture from the queue |
4397 | | **************************************************************************/ |
4398 | | |
4399 | | /* Mark the skip flag */ |
4400 | 78.2k | i4_skip = 0; |
4401 | 78.2k | ctxt_sel = ps_codec->i4_encode_api_call_cnt % MAX_CTXT_SETS; |
4402 | 78.2k | ps_codec->s_rate_control.pre_encode_skip[ctxt_sel] = i4_skip; |
4403 | | |
4404 | | /* Get a buffer to encode */ |
4405 | 78.2k | ps_inp_buf = &(ps_codec->as_inp_list[u4_pic_id % SVC_MAX_NUM_INP_FRAMES]); |
4406 | | |
4407 | | /* copy dequeued input to output */ |
4408 | 78.2k | ps_enc_buff[0] = ps_inp_buf[0]; |
4409 | | |
4410 | | /* Special case for encoding trailing B frames |
4411 | | * |
4412 | | * In encoding streams with B frames it may happen that we have a B frame |
4413 | | * at the end without a P/I frame after it. Hence when we are dequeing from |
4414 | | * the RC, it will return the P frame [next in display order but before in |
4415 | | * encoding order] first. Since the dequeue happens for an invalid frame we |
4416 | | * will get a frame with null buff and set u4_is_last. Hence lib with return |
4417 | | * last frame flag at this point and will stop encoding. |
4418 | | * |
4419 | | * Since for the last B frame, we does not have the forward ref frame |
4420 | | * it makes sense to force it into P. |
4421 | | * |
4422 | | * To solve this, in case the current frame is P and if the last frame flag |
4423 | | * is set, we need to see if there is and pending B frames. If there are any, |
4424 | | * we should just encode that picture as the current P frame and set |
4425 | | * that B frame as the last frame. Hence the encoder will terminate naturally |
4426 | | * once that B-frame is encoded after all the in between frames. |
4427 | | * |
4428 | | * Since we cannot touch RC stack directly, the option of actually swapping |
4429 | | * frames in RC is ruled out. We have to modify the as_inp_list to simulate |
4430 | | * such a behavior by RC. We can do that by |
4431 | | * 1) Search through as_inp_list to locate the largest u4_timestamp_low less |
4432 | | * than current u4_timestamp_low. This will give us the last B frame |
4433 | | * before the current P frame. Note that this will handle pre encode skip too |
4434 | | * since queue happens after pre enc skip. 2) Swap the position in |
4435 | | * as_inp_list. Hence now the last B frame is encoded as P frame. And the new |
4436 | | * last B frame will have u4_is_last set so that encoder will end naturally |
4437 | | * once we reached that B frame or any subsequent frame. Also the current GOP |
4438 | | * will have 1 less B frame Since we are swapping, the poc will also be |
4439 | | * in-order. 3) In case we have an IPP stream, the result of our search will |
4440 | | * be an I/P frame which is already encoded. Thus swap and encode will result |
4441 | | * in encoding of duplicate frames. Hence to avoid this we will only |
4442 | | * have this work around in case of u4_num_bframes > 0. |
4443 | | * |
4444 | | * In case we have forced an I/IDR frame In between this P frame and |
4445 | | * the last B frame -> This cannot happen as the current P frame is |
4446 | | * supposed to have u4_is_last set. Thus forcing an I/ IDR after this |
4447 | | * is illogical. |
4448 | | * |
4449 | | * In cae if we have forced an I such that the frame just before last |
4450 | | * frame in is I/P -> This case will never arise. Since we have a closed GOP |
4451 | | * now, once we force an I, the gop gets reset, hence there will be a B |
4452 | | * between I/P and I/P. |
4453 | | */ |
4454 | 78.2k | if(ps_enc_buff->s_inp_props.u4_is_last && (ps_codec->pic_type == PIC_P) && |
4455 | 78.2k | ps_codec->s_cfg.u4_num_bframes) |
4456 | 0 | { |
4457 | 0 | WORD32 cntr; |
4458 | 0 | WORD32 lst_bframe = -1; |
4459 | 0 | UWORD32 u4_timestamp_low = 0; |
4460 | 0 | UWORD32 u4_timestamp_high = 0; |
4461 | 0 | isvce_inp_buf_t *ps_swap_buff, *ps_inp_list; |
4462 | |
|
4463 | 0 | ps_inp_list = &ps_codec->as_inp_list[0]; |
4464 | | |
4465 | | /* Now search the inp list for highest timestamp */ |
4466 | 0 | for(cntr = 0; cntr < SVC_MAX_NUM_INP_FRAMES; cntr++) |
4467 | 0 | { |
4468 | 0 | if(ps_inp_list[cntr].s_inp_props.s_raw_buf.apv_bufs[0] != NULL) |
4469 | 0 | { |
4470 | 0 | if((ps_inp_list[cntr].s_inp_props.u4_timestamp_high > u4_timestamp_high) || |
4471 | 0 | (ps_inp_list[cntr].s_inp_props.u4_timestamp_high == u4_timestamp_high && |
4472 | 0 | ps_inp_list[cntr].s_inp_props.u4_timestamp_low > u4_timestamp_low)) |
4473 | 0 | { |
4474 | 0 | u4_timestamp_low = ps_inp_list[cntr].s_inp_props.u4_timestamp_low; |
4475 | 0 | u4_timestamp_high = ps_inp_list[cntr].s_inp_props.u4_timestamp_high; |
4476 | 0 | lst_bframe = cntr; |
4477 | 0 | } |
4478 | 0 | } |
4479 | 0 | } |
4480 | |
|
4481 | 0 | if(lst_bframe != -1) |
4482 | 0 | { |
4483 | 0 | ps_swap_buff = &(ps_codec->as_inp_list[lst_bframe]); |
4484 | | |
4485 | | /* copy the last B buffer to output */ |
4486 | 0 | *ps_enc_buff = *ps_swap_buff; |
4487 | | |
4488 | | /* Store the current buf into the queue in place of last B buf */ |
4489 | 0 | *ps_swap_buff = *ps_inp_buf; |
4490 | 0 | } |
4491 | 0 | } |
4492 | | |
4493 | 78.2k | if(ps_enc_buff->s_inp_props.u4_is_last) |
4494 | 11.7k | { |
4495 | 11.7k | ps_codec->pic_type = PIC_NA; |
4496 | 11.7k | } |
4497 | | |
4498 | | /* The buffer in the queue is set to NULL to specify that encoding is done for |
4499 | | * that frame */ |
4500 | 312k | for(i = 0; i < 3; i++) |
4501 | 234k | { |
4502 | 234k | ps_inp_buf->s_inp_props.s_raw_buf.apv_bufs[i] = NULL; |
4503 | 234k | } |
4504 | | |
4505 | | /* Return the buffer status */ |
4506 | 78.2k | return (0); |
4507 | 78.2k | } |
4508 | | |
4509 | | /** |
4510 | | ****************************************************************************** |
4511 | | * |
4512 | | * @brief |
4513 | | * This function joins all the spawned threads after successful completion of |
4514 | | * their tasks |
4515 | | * |
4516 | | * @par Description |
4517 | | * |
4518 | | * @param[in] ps_codec |
4519 | | * pointer to codec context |
4520 | | * |
4521 | | * @returns none |
4522 | | * |
4523 | | ****************************************************************************** |
4524 | | */ |
4525 | | void isvce_join_threads(isvce_codec_t *ps_codec) |
4526 | 61.1k | { |
4527 | 61.1k | WORD32 i = 0; |
4528 | 61.1k | WORD32 ret = 0; |
4529 | | |
4530 | | /* join spawned threads */ |
4531 | 138k | while(i < ps_codec->i4_proc_thread_cnt) |
4532 | 77.6k | { |
4533 | 77.6k | if(ps_codec->ai4_process_thread_created[i]) |
4534 | 77.6k | { |
4535 | 77.6k | ret = ithread_join(ps_codec->apv_proc_thread_handle[i], NULL); |
4536 | | |
4537 | 77.6k | if(ret != 0) |
4538 | 0 | { |
4539 | 0 | ASSERT(0); |
4540 | 0 | } |
4541 | | |
4542 | 77.6k | ps_codec->ai4_process_thread_created[i] = 0; |
4543 | 77.6k | i++; |
4544 | 77.6k | } |
4545 | 77.6k | } |
4546 | | |
4547 | 61.1k | ps_codec->i4_proc_thread_cnt = 0; |
4548 | 61.1k | } |
4549 | | |
4550 | | UWORD32 isvce_get_min_outbuf_size(UWORD32 u4_wd, UWORD32 u4_ht, UWORD8 u1_num_spatial_layers) |
4551 | 44.7k | { |
4552 | 44.7k | return MAX((u4_wd * u4_ht * 3), MIN_STREAM_SIZE) * u1_num_spatial_layers; |
4553 | 44.7k | } |