/src/libavc/decoder/svc/isvcd_thread_parse_decode.c
Line | Count | Source |
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 | | * @file |
23 | | * isvcd_thread_parse_decode.c |
24 | | * |
25 | | * @brief |
26 | | * Contains routines that for multi-thread decoder |
27 | | * |
28 | | * @author |
29 | | * Kishore |
30 | | * |
31 | | * @remarks |
32 | | * None |
33 | | * |
34 | | ******************************************************************************* |
35 | | */ |
36 | | |
37 | | #include <string.h> |
38 | | #include "ih264d_error_handler.h" |
39 | | #include "ih264d_debug.h" |
40 | | #include "ithread.h" |
41 | | #include "ih264d_defs.h" |
42 | | #include "ih264d_debug.h" |
43 | | #include "ih264d_tables.h" |
44 | | #include "isvcd_structs.h" |
45 | | #include "ih264d_defs.h" |
46 | | #include "ih264d_mb_utils.h" |
47 | | #include "ih264d_thread_parse_decode.h" |
48 | | #include "ih264d_inter_pred.h" |
49 | | #include "ih264d_process_pslice.h" |
50 | | #include "isvcd_process_epslice.h" |
51 | | #include "ih264d_process_intra_mb.h" |
52 | | #include "ih264d_deblocking.h" |
53 | | #include "ih264d_format_conv.h" |
54 | | #include "isvcd_thread_parse_decode.h" |
55 | | |
56 | | /*****************************************************************************/ |
57 | | /* */ |
58 | | /* Function Name : isvcd_decode_recon_tfr_nmb_thread */ |
59 | | /* */ |
60 | | /* Description : */ |
61 | | /* : */ |
62 | | /* : */ |
63 | | /* Inputs : */ |
64 | | /* Processing :Only for Target Layer processing */ |
65 | | /* */ |
66 | | /* Outputs : */ |
67 | | /* Returns : None */ |
68 | | /* */ |
69 | | /* Revision History: */ |
70 | | /* */ |
71 | | /* DD MM YYYY Author(s) Changes (Describe the changes made) */ |
72 | | /* ITTIAM */ |
73 | | /*****************************************************************************/ |
74 | | WORD32 isvcd_decode_recon_tfr_nmb_thread(svc_dec_lyr_struct_t *ps_svc_lyr_dec, UWORD32 u4_num_mbs, |
75 | | UWORD32 u4_num_mbs_next, UWORD32 u4_end_of_row) |
76 | 192k | { |
77 | 192k | WORD32 i, j; |
78 | 192k | dec_mb_info_t *ps_cur_mb_info; |
79 | 192k | dec_svc_mb_info_t *ps_svc_cur_mb_info; |
80 | 192k | dec_struct_t *ps_dec = &ps_svc_lyr_dec->s_dec; |
81 | 192k | const UWORD32 u1_mbaff = ps_dec->ps_cur_slice->u1_mbaff_frame_flag; |
82 | 192k | UWORD32 u1_slice_type, u1_B; |
83 | 192k | WORD32 u1_skip_th; |
84 | 192k | UWORD32 u1_ipcm_th; |
85 | 192k | UWORD32 u4_cond; |
86 | 192k | UWORD16 u2_slice_num, u2_cur_dec_mb_num; |
87 | 192k | UWORD32 u4_mb_num; |
88 | 192k | WORD32 nop_cnt = 8 * 128; |
89 | 192k | UWORD16 *pu2_res_luma_csbp; |
90 | 192k | WORD32 ret; |
91 | 192k | u1_slice_type = ps_dec->ps_decode_cur_slice->slice_type; |
92 | | |
93 | 192k | u1_B = (u1_slice_type == B_SLICE); |
94 | 192k | u1_skip_th = ((u1_slice_type != I_SLICE) ? (u1_B ? B_8x8 : PRED_8x8R0) : -1); |
95 | 192k | u1_ipcm_th = ((u1_slice_type != I_SLICE) ? (u1_B ? 23 : 5) : 0); |
96 | 192k | u2_cur_dec_mb_num = ps_dec->cur_dec_mb_num; |
97 | | |
98 | 256k | while(1) |
99 | 256k | { |
100 | 256k | UWORD32 u4_max_mb = |
101 | 256k | (UWORD32) (ps_dec->i2_dec_thread_mb_y + (1 << u1_mbaff)) * ps_dec->u2_frm_wd_in_mbs - 1; |
102 | 256k | u4_mb_num = u2_cur_dec_mb_num; |
103 | | /*introducing 1 MB delay*/ |
104 | 256k | u4_mb_num = MIN(u4_mb_num + u4_num_mbs + 1, u4_max_mb); |
105 | | |
106 | 256k | CHECK_MB_MAP_BYTE(u4_mb_num, ps_dec->pu1_dec_mb_map, u4_cond); |
107 | 256k | if(u4_cond) |
108 | 191k | { |
109 | 191k | break; |
110 | 191k | } |
111 | 64.9k | else |
112 | 64.9k | { |
113 | 64.9k | if(nop_cnt > 0) |
114 | 55.9k | { |
115 | 55.9k | nop_cnt -= 128; |
116 | 55.9k | NOP(128); |
117 | 55.9k | } |
118 | 9.01k | else |
119 | 9.01k | { |
120 | 9.01k | if(ps_dec->u4_output_present && (2 == ps_dec->u4_num_cores) && |
121 | 7.98k | (ps_dec->u4_fmt_conv_cur_row < ps_dec->s_disp_frame_info.u4_y_ht)) |
122 | 3.83k | { |
123 | 3.83k | ps_dec->u4_fmt_conv_num_rows = |
124 | 3.83k | MIN(FMT_CONV_NUM_ROWS, |
125 | 3.83k | (ps_dec->s_disp_frame_info.u4_y_ht - ps_dec->u4_fmt_conv_cur_row)); |
126 | 3.83k | ih264d_format_convert(ps_dec, &(ps_dec->s_disp_op), ps_dec->u4_fmt_conv_cur_row, |
127 | 3.83k | ps_dec->u4_fmt_conv_num_rows); |
128 | 3.83k | ps_dec->u4_fmt_conv_cur_row += ps_dec->u4_fmt_conv_num_rows; |
129 | 3.83k | } |
130 | 5.17k | else |
131 | 5.17k | { |
132 | 5.17k | nop_cnt = 8 * 128; |
133 | 5.17k | ithread_yield(); |
134 | 5.17k | } |
135 | 9.01k | if(1 == ps_svc_lyr_dec->u1_error_in_cur_frame) |
136 | 1.00k | { |
137 | 1.00k | return NOT_OK; |
138 | 1.00k | } |
139 | 9.01k | } |
140 | 64.9k | } |
141 | 256k | } |
142 | | |
143 | | /* N Mb MC Loop */ |
144 | 812k | for(i = 0; i < u4_num_mbs; i++) |
145 | 627k | { |
146 | 627k | u4_mb_num = u2_cur_dec_mb_num; |
147 | 627k | GET_SLICE_NUM_MAP(ps_dec->pu2_slice_num_map, u2_cur_dec_mb_num, u2_slice_num); |
148 | | |
149 | 627k | if(u2_slice_num != ps_dec->u2_cur_slice_num_dec_thread) |
150 | 7.33k | { |
151 | 7.33k | ps_dec->u4_cur_slice_decode_done = 1; |
152 | 7.33k | break; |
153 | 7.33k | } |
154 | | |
155 | 620k | ps_cur_mb_info = &ps_dec->ps_frm_mb_info[u2_cur_dec_mb_num]; |
156 | | |
157 | 620k | ps_dec->u4_dma_buf_idx = 0; |
158 | 620k | ps_dec->u4_pred_info_idx = 0; |
159 | | |
160 | | /*Pointer assignment for Residual NNZ */ |
161 | 620k | pu2_res_luma_csbp = ps_svc_lyr_dec->pu2_frm_res_luma_csbp + ps_cur_mb_info->u2_mbx; |
162 | 620k | pu2_res_luma_csbp += ps_cur_mb_info->u2_mby * ps_svc_lyr_dec->i4_frm_res_luma_csbp_stride; |
163 | | |
164 | 620k | if(ps_cur_mb_info->u1_mb_type <= u1_skip_th) |
165 | 139k | { |
166 | 139k | WORD32 pred_cnt = 0; |
167 | 139k | pred_info_pkd_t *ps_pred_pkd; |
168 | 139k | UWORD32 u4_pred_info_pkd_idx; |
169 | | |
170 | 139k | u4_pred_info_pkd_idx = ps_cur_mb_info->u4_pred_info_pkd_idx; |
171 | | |
172 | 399k | while(pred_cnt < ps_cur_mb_info->u1_num_pred_parts) |
173 | 260k | { |
174 | 260k | ps_pred_pkd = ps_dec->ps_pred_pkd + u4_pred_info_pkd_idx; |
175 | | |
176 | 260k | ps_dec->p_form_mb_part_info_thread(ps_pred_pkd, ps_dec, ps_cur_mb_info->u2_mbx, |
177 | 260k | ps_cur_mb_info->u2_mby, (i >> u1_mbaff), |
178 | 260k | ps_cur_mb_info); |
179 | | |
180 | 260k | u4_pred_info_pkd_idx++; |
181 | 260k | pred_cnt++; |
182 | 260k | } |
183 | 139k | ps_dec->p_mc_dec_thread(ps_dec, ps_cur_mb_info); |
184 | 139k | } |
185 | 481k | else if(ps_cur_mb_info->u1_mb_type == MB_SKIP) |
186 | 410k | { |
187 | 410k | WORD32 pred_cnt = 0; |
188 | 410k | pred_info_pkd_t *ps_pred_pkd; |
189 | 410k | UWORD32 u4_pred_info_pkd_idx; |
190 | | |
191 | 410k | u4_pred_info_pkd_idx = ps_cur_mb_info->u4_pred_info_pkd_idx; |
192 | | |
193 | 860k | while(pred_cnt < ps_cur_mb_info->u1_num_pred_parts) |
194 | 450k | { |
195 | 450k | ps_pred_pkd = ps_dec->ps_pred_pkd + u4_pred_info_pkd_idx; |
196 | | |
197 | 450k | ps_dec->p_form_mb_part_info_thread(ps_pred_pkd, ps_dec, ps_cur_mb_info->u2_mbx, |
198 | 450k | ps_cur_mb_info->u2_mby, (i >> u1_mbaff), |
199 | 450k | ps_cur_mb_info); |
200 | | |
201 | 450k | u4_pred_info_pkd_idx++; |
202 | 450k | pred_cnt++; |
203 | 450k | } |
204 | | /* Decode MB skip */ |
205 | 410k | ps_dec->p_mc_dec_thread(ps_dec, ps_cur_mb_info); |
206 | | |
207 | 410k | *pu2_res_luma_csbp = 0; |
208 | 410k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb = |
209 | 410k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_frm_start + ps_cur_mb_info->u2_mbx + |
210 | 410k | (ps_svc_lyr_dec->u2_inter_lyr_mb_prms_stride * (ps_cur_mb_info->u2_mby)); |
211 | 410k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb->i1_mb_mode = SVC_INTER_MB; |
212 | 410k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb->i1_tx_size = |
213 | 410k | ps_cur_mb_info->u1_tran_form8x8; |
214 | 410k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb->u2_luma_nnz = 0; |
215 | 410k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb->u1_chroma_nnz = 0; |
216 | 410k | } |
217 | | |
218 | 620k | u2_cur_dec_mb_num++; |
219 | 620k | } |
220 | | |
221 | | /* N Mb IQ IT RECON Loop */ |
222 | 812k | for(j = 0; j < i; j++) |
223 | 620k | { |
224 | 620k | ps_cur_mb_info = &ps_dec->ps_frm_mb_info[ps_dec->cur_dec_mb_num]; |
225 | 620k | ps_svc_cur_mb_info = &ps_svc_lyr_dec->ps_svc_frm_mb_info[ps_dec->cur_dec_mb_num]; |
226 | | |
227 | 620k | if(NULL == ps_cur_mb_info->ps_curmb) |
228 | 0 | { |
229 | 0 | return NOT_OK; |
230 | 0 | } |
231 | | |
232 | | /*Pointer assignment for Residual NNZ */ |
233 | 620k | pu2_res_luma_csbp = ps_svc_lyr_dec->pu2_frm_res_luma_csbp + ps_cur_mb_info->u2_mbx; |
234 | 620k | pu2_res_luma_csbp += ps_cur_mb_info->u2_mby * ps_svc_lyr_dec->i4_frm_res_luma_csbp_stride; |
235 | | |
236 | 620k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb = |
237 | 620k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_frm_start + ps_cur_mb_info->u2_mbx + |
238 | 620k | (ps_svc_lyr_dec->u2_inter_lyr_mb_prms_stride * (ps_cur_mb_info->u2_mby)); |
239 | | |
240 | 620k | ps_svc_lyr_dec->ps_inter_lyr_mb_prms_cur_mb->i1_slice_id = (WORD8) ps_dec->u2_cur_slice_num; |
241 | | |
242 | 620k | if((ps_dec->u4_num_cores == 2) || !ps_dec->i1_recon_in_thread3_flag) |
243 | 620k | { |
244 | 620k | if(ps_cur_mb_info->u1_mb_type <= u1_skip_th) |
245 | 139k | { |
246 | 139k | { |
247 | | /* inter intra pred generation */ |
248 | 139k | if(SVCD_FALSE == ps_svc_lyr_dec->u1_dyadic_flag) |
249 | 100k | { |
250 | 100k | ret = isvcd_process_ii_mb(ps_svc_lyr_dec, ps_cur_mb_info, |
251 | 100k | ps_svc_cur_mb_info, j); |
252 | 100k | if(ret != OK) return ret; |
253 | 100k | } |
254 | 139k | if(0 == ps_svc_cur_mb_info->u1_residual_prediction_flag) |
255 | 29.1k | { |
256 | | /* IT + Recon */ |
257 | 29.1k | ih264d_process_inter_mb(ps_dec, ps_cur_mb_info, j); |
258 | 29.1k | isvcd_update_inter_mb_inter_layer_info(ps_svc_lyr_dec, ps_cur_mb_info, 0); |
259 | 29.1k | *pu2_res_luma_csbp = ps_cur_mb_info->u2_luma_csbp; |
260 | 29.1k | } |
261 | 110k | else |
262 | 110k | { |
263 | | /* IT + Residual + Recon */ |
264 | 110k | ret = isvcd_process_inter_mb_rsd_pred_target_lyr( |
265 | 110k | ps_svc_lyr_dec, ps_cur_mb_info, j, 0, pu2_res_luma_csbp); |
266 | 110k | if(ret != OK) return ret; |
267 | 110k | } |
268 | 139k | } |
269 | 139k | } |
270 | | |
271 | 481k | else if((ps_cur_mb_info->u1_mb_type != MB_SKIP) && |
272 | 70.6k | (ps_cur_mb_info->u1_mb_type != MB_INFER)) |
273 | 12.6k | { |
274 | 12.6k | if((u1_ipcm_th + 25) != ps_cur_mb_info->u1_mb_type) |
275 | 12.4k | { |
276 | 12.4k | ps_cur_mb_info->u1_mb_type -= (u1_skip_th + 1); |
277 | 12.4k | ih264d_process_intra_mb(ps_dec, ps_cur_mb_info, j); |
278 | 12.4k | isvcd_update_intra_mb_inter_layer_info(ps_svc_lyr_dec, ps_cur_mb_info); |
279 | 12.4k | } |
280 | 197 | else |
281 | 197 | { |
282 | 197 | isvcd_update_ipcm_mb_inter_layer_info(ps_svc_lyr_dec, ps_cur_mb_info); |
283 | 197 | } |
284 | 12.6k | *pu2_res_luma_csbp = 0; |
285 | 12.6k | } |
286 | 468k | else if(ps_cur_mb_info->u1_mb_type == MB_INFER) |
287 | 57.9k | { |
288 | | /* inter layer intra prediction : intra upsample, IQ, IT ,deblock */ |
289 | 57.9k | { |
290 | | /* Intra resample for IBL mode */ |
291 | 57.9k | ret = isvcd_process_ibl_mb(ps_svc_lyr_dec, ps_cur_mb_info, j, 0); |
292 | 57.9k | if(ret != OK) return ret; |
293 | | /* Pass intra resample as pred to Recon generation */ |
294 | 57.9k | ih264d_process_inter_mb(ps_dec, ps_cur_mb_info, j); |
295 | 57.9k | isvcd_update_inter_mb_inter_layer_info(ps_svc_lyr_dec, ps_cur_mb_info, 1); |
296 | 57.9k | *pu2_res_luma_csbp = ps_cur_mb_info->u2_luma_csbp; |
297 | 57.9k | } |
298 | 57.9k | ps_dec->pi1_left_pred_mode[0] = DC; |
299 | 57.9k | ps_dec->pi1_left_pred_mode[1] = DC; |
300 | 57.9k | ps_dec->pi1_left_pred_mode[2] = DC; |
301 | 57.9k | ps_dec->pi1_left_pred_mode[3] = DC; |
302 | | |
303 | 57.9k | ps_cur_mb_info->ps_curmb->pi1_intrapredmodes[0] = DC; |
304 | 57.9k | ps_cur_mb_info->ps_curmb->pi1_intrapredmodes[1] = DC; |
305 | 57.9k | ps_cur_mb_info->ps_curmb->pi1_intrapredmodes[2] = DC; |
306 | 57.9k | ps_cur_mb_info->ps_curmb->pi1_intrapredmodes[3] = DC; |
307 | | |
308 | 57.9k | isvcd_update_ibl_mb_inter_layer_info(ps_svc_lyr_dec, ps_cur_mb_info); |
309 | 57.9k | } |
310 | | |
311 | 620k | if(ps_dec->u4_use_intrapred_line_copy == 1) |
312 | 620k | ih264d_copy_intra_pred_line(ps_dec, ps_cur_mb_info, j); |
313 | 620k | } |
314 | | |
315 | 620k | DATA_SYNC(); |
316 | | |
317 | 620k | u4_mb_num = ps_cur_mb_info->u2_mbx + ps_dec->u2_frm_wd_in_mbs * ps_cur_mb_info->u2_mby; |
318 | 620k | UPDATE_MB_MAP_MBNUM_BYTE(ps_dec->pu1_recon_mb_map, u4_mb_num); |
319 | 620k | ps_dec->cur_dec_mb_num++; |
320 | 620k | } |
321 | | |
322 | | /*N MB deblocking*/ |
323 | 191k | if(ps_dec->u4_nmb_deblk == 1) |
324 | 0 | { |
325 | 0 | UWORD32 u4_wd_y, u4_wd_uv; |
326 | 0 | tfr_ctxt_t *ps_tfr_cxt = &(ps_dec->s_tran_addrecon); |
327 | 0 | UWORD8 u1_field_pic_flag = ps_dec->ps_cur_slice->u1_field_pic_flag; |
328 | 0 | const WORD32 i4_cb_qp_idx_ofst = ps_dec->ps_cur_pps->i1_chroma_qp_index_offset; |
329 | 0 | const WORD32 i4_cr_qp_idx_ofst = ps_dec->ps_cur_pps->i1_second_chroma_qp_index_offset; |
330 | |
|
331 | 0 | u4_wd_y = ps_dec->u2_frm_wd_y << u1_field_pic_flag; |
332 | 0 | u4_wd_uv = ps_dec->u2_frm_wd_uv << u1_field_pic_flag; |
333 | |
|
334 | 0 | ps_cur_mb_info = &ps_dec->ps_frm_mb_info[ps_dec->u4_cur_deblk_mb_num]; |
335 | |
|
336 | 0 | ps_dec->u4_deblk_mb_x = ps_cur_mb_info->u2_mbx; |
337 | 0 | ps_dec->u4_deblk_mb_y = ps_cur_mb_info->u2_mby; |
338 | |
|
339 | 0 | for(j = 0; j < i; j++) |
340 | 0 | { |
341 | 0 | ih264d_deblock_mb_nonmbaff(ps_dec, ps_tfr_cxt, i4_cb_qp_idx_ofst, i4_cr_qp_idx_ofst, |
342 | 0 | u4_wd_y, u4_wd_uv); |
343 | 0 | } |
344 | 0 | } |
345 | | |
346 | | /*handle the last mb in picture case*/ |
347 | 191k | if(ps_dec->cur_dec_mb_num > ps_dec->ps_cur_sps->u4_max_mb_addr) |
348 | 19.6k | ps_dec->u4_cur_slice_decode_done = 1; |
349 | | |
350 | 191k | if(i != u4_num_mbs) |
351 | 7.33k | { |
352 | 7.33k | u4_end_of_row = 0; |
353 | | /*Number of MB's left in row*/ |
354 | 7.33k | u4_num_mbs_next = u4_num_mbs_next + ((u4_num_mbs - i) >> u1_mbaff); |
355 | 7.33k | } |
356 | | |
357 | 191k | ih264d_decode_tfr_nmb(ps_dec, (i), u4_num_mbs_next, u4_end_of_row); |
358 | | |
359 | 191k | return OK; |
360 | 191k | } |
361 | | |
362 | | /*****************************************************************************/ |
363 | | /* */ |
364 | | /* Function Name : isvcd_decode_slice_thread */ |
365 | | /* */ |
366 | | /* Description : */ |
367 | | /* : */ |
368 | | /* : */ |
369 | | /* Inputs : */ |
370 | | /* Processing : */ |
371 | | /* */ |
372 | | /* Outputs : */ |
373 | | /* Returns : 0 on success */ |
374 | | /* */ |
375 | | /* Revision History: */ |
376 | | /* */ |
377 | | /* DD MM YYYY Author(s) Changes (Describe the changes made) */ |
378 | | /* ITTIAM */ |
379 | | /*****************************************************************************/ |
380 | | WORD32 isvcd_decode_slice_thread(svc_dec_lyr_struct_t *ps_svc_lyr_dec) |
381 | 28.8k | { |
382 | 28.8k | UWORD32 u4_num_mbs_next, u4_num_mbsleft, u4_end_of_row = 0; |
383 | 28.8k | dec_struct_t *ps_dec = &ps_svc_lyr_dec->s_dec; |
384 | 28.8k | const UWORD32 i2_pic_wdin_mbs = ps_dec->u2_frm_wd_in_mbs; |
385 | 28.8k | UWORD32 u4_mbaff, u4_num_mbs; |
386 | | |
387 | 28.8k | UWORD16 u2_first_mb_in_slice; |
388 | 28.8k | UWORD16 i16_mb_x, i16_mb_y; |
389 | 28.8k | UWORD8 u1_field_pic; |
390 | 28.8k | UWORD32 u4_frame_stride, x_offset, y_offset; |
391 | 28.8k | WORD32 ret; |
392 | 28.8k | tfr_ctxt_t *ps_trns_addr; |
393 | | |
394 | | /*check for mb map of first mb in slice to ensure slice header is parsed*/ |
395 | 172k | while(1) |
396 | 172k | { |
397 | 172k | UWORD32 u4_mb_num = ps_dec->cur_dec_mb_num; |
398 | 172k | UWORD32 u4_cond = 0; |
399 | 172k | WORD32 nop_cnt = 8 * 128; |
400 | 172k | CHECK_MB_MAP_BYTE(u4_mb_num, ps_dec->pu1_dec_mb_map, u4_cond); |
401 | 172k | if(u4_cond) |
402 | 27.9k | { |
403 | 27.9k | break; |
404 | 27.9k | } |
405 | 144k | else |
406 | 144k | { |
407 | 144k | if(nop_cnt > 0) |
408 | 144k | { |
409 | 144k | nop_cnt -= 128; |
410 | 144k | NOP(128); |
411 | 144k | } |
412 | 0 | else if(ps_dec->u4_output_present && (2 == ps_dec->u4_num_cores) && |
413 | 0 | (ps_dec->u4_fmt_conv_cur_row < ps_dec->s_disp_frame_info.u4_y_ht)) |
414 | 0 | { |
415 | 0 | ps_dec->u4_fmt_conv_num_rows = |
416 | 0 | MIN(FMT_CONV_NUM_ROWS, |
417 | 0 | (ps_dec->s_disp_frame_info.u4_y_ht - ps_dec->u4_fmt_conv_cur_row)); |
418 | 0 | ih264d_format_convert(ps_dec, &(ps_dec->s_disp_op), ps_dec->u4_fmt_conv_cur_row, |
419 | 0 | ps_dec->u4_fmt_conv_num_rows); |
420 | 0 | ps_dec->u4_fmt_conv_cur_row += ps_dec->u4_fmt_conv_num_rows; |
421 | 0 | } |
422 | 0 | else |
423 | 0 | { |
424 | 0 | nop_cnt = 8 * 128; |
425 | 0 | ithread_yield(); |
426 | 0 | } |
427 | 144k | if(1 == ps_svc_lyr_dec->u1_error_in_cur_frame) |
428 | 861 | { |
429 | 861 | return NOT_OK; |
430 | 861 | } |
431 | 144k | DEBUG_THREADS_PRINTF( |
432 | 144k | "waiting for mb mapcur_dec_mb_num = %d,ps_dec->u2_cur_mb_addr = " |
433 | 144k | "%d\n", |
434 | 144k | u2_cur_dec_mb_num, ps_dec->u2_cur_mb_addr); |
435 | 144k | } |
436 | 172k | } |
437 | | |
438 | 27.9k | u4_mbaff = ps_dec->ps_cur_slice->u1_mbaff_frame_flag; |
439 | 27.9k | u2_first_mb_in_slice = ps_dec->ps_decode_cur_slice->u4_first_mb_in_slice; |
440 | 27.9k | i16_mb_x = MOD(u2_first_mb_in_slice, i2_pic_wdin_mbs); |
441 | 27.9k | i16_mb_y = DIV(u2_first_mb_in_slice, i2_pic_wdin_mbs); |
442 | 27.9k | i16_mb_y <<= u4_mbaff; |
443 | 27.9k | ps_dec->i2_dec_thread_mb_y = i16_mb_y; |
444 | 27.9k | ps_dec->cur_dec_mb_num = u2_first_mb_in_slice << u4_mbaff; |
445 | | |
446 | 27.9k | if((ps_dec->u4_num_cores == 2) || !ps_dec->i1_recon_in_thread3_flag) |
447 | 27.9k | { |
448 | 27.9k | ps_dec->pv_proc_tu_coeff_data = |
449 | 27.9k | (void *) ps_dec->ps_decode_cur_slice->pv_tu_coeff_data_start; |
450 | 27.9k | } |
451 | | |
452 | | // recalculate recon pointers |
453 | 27.9k | u1_field_pic = ps_dec->ps_cur_slice->u1_field_pic_flag; |
454 | 27.9k | u4_frame_stride = ps_dec->u2_frm_wd_y << u1_field_pic; |
455 | 27.9k | x_offset = i16_mb_x << 4; |
456 | 27.9k | y_offset = (i16_mb_y * u4_frame_stride) << 4; |
457 | | |
458 | 27.9k | ps_trns_addr = &(ps_dec->s_tran_addrecon); |
459 | | |
460 | 27.9k | ps_trns_addr->pu1_dest_y = ps_dec->s_cur_pic.pu1_buf1 + x_offset + y_offset; |
461 | | |
462 | 27.9k | u4_frame_stride = ps_dec->u2_frm_wd_uv << u1_field_pic; |
463 | 27.9k | x_offset >>= 1; |
464 | 27.9k | y_offset = (i16_mb_y * u4_frame_stride) << 3; |
465 | 27.9k | x_offset *= YUV420SP_FACTOR; |
466 | | |
467 | 27.9k | ps_trns_addr->pu1_dest_u = ps_dec->s_cur_pic.pu1_buf2 + x_offset + y_offset; |
468 | 27.9k | ps_trns_addr->pu1_dest_v = ps_dec->s_cur_pic.pu1_buf3 + x_offset + y_offset; |
469 | | |
470 | 27.9k | ps_trns_addr->pu1_mb_y = ps_trns_addr->pu1_dest_y; |
471 | 27.9k | ps_trns_addr->pu1_mb_u = ps_trns_addr->pu1_dest_u; |
472 | 27.9k | ps_trns_addr->pu1_mb_v = ps_trns_addr->pu1_dest_v; |
473 | | |
474 | | /* Initialise MC and formMbPartInfo fn ptrs one time based on profile_idc */ |
475 | 27.9k | { |
476 | 27.9k | ps_dec->p_mc_dec_thread = ih264d_motion_compensate_bp; |
477 | 27.9k | ps_dec->p_form_mb_part_info_thread = ih264d_form_mb_part_info_bp; |
478 | 27.9k | } |
479 | 27.9k | { |
480 | 27.9k | UWORD8 uc_nofield_nombaff; |
481 | 27.9k | uc_nofield_nombaff = ((ps_dec->ps_cur_slice->u1_field_pic_flag == 0) && |
482 | 27.9k | (ps_dec->ps_cur_slice->u1_mbaff_frame_flag == 0) && |
483 | 27.9k | (ps_dec->ps_decode_cur_slice->slice_type != B_SLICE) && |
484 | 24.3k | (ps_dec->ps_cur_pps->u1_wted_pred_flag == 0)); |
485 | | |
486 | 27.9k | if(uc_nofield_nombaff == 0) |
487 | 7.83k | { |
488 | 7.83k | ps_dec->p_mc_dec_thread = ih264d_motion_compensate_mp; |
489 | 7.83k | ps_dec->p_form_mb_part_info_thread = ih264d_form_mb_part_info_mp; |
490 | 7.83k | } |
491 | 27.9k | } |
492 | | |
493 | 27.9k | ps_dec->u4_cur_slice_decode_done = 0; |
494 | | |
495 | 219k | while(ps_dec->u4_cur_slice_decode_done != 1) |
496 | 192k | { |
497 | 192k | u4_num_mbsleft = ((i2_pic_wdin_mbs - i16_mb_x) << u4_mbaff); |
498 | | |
499 | 192k | if(u4_num_mbsleft <= ps_dec->u4_recon_mb_grp) |
500 | 192k | { |
501 | 192k | u4_num_mbs = u4_num_mbsleft; |
502 | | |
503 | | /*Indicate number of mb's left in a row*/ |
504 | 192k | u4_num_mbs_next = 0; |
505 | 192k | u4_end_of_row = 1; |
506 | 192k | i16_mb_x = 0; |
507 | 192k | } |
508 | 0 | else |
509 | 0 | { |
510 | 0 | u4_num_mbs = ps_dec->u4_recon_mb_grp; |
511 | | |
512 | | /*Indicate number of mb's left in a row*/ |
513 | 0 | u4_num_mbs_next = i2_pic_wdin_mbs - i16_mb_x - (ps_dec->u4_recon_mb_grp >> u4_mbaff); |
514 | 0 | i16_mb_x += (u4_num_mbs >> u4_mbaff); |
515 | 0 | u4_end_of_row = 0; |
516 | 0 | } |
517 | 192k | if(ps_svc_lyr_dec->u1_layer_identifier == TARGET_LAYER) |
518 | 192k | { |
519 | 192k | ret = isvcd_decode_recon_tfr_nmb_thread(ps_svc_lyr_dec, u4_num_mbs, u4_num_mbs_next, |
520 | 192k | u4_end_of_row); |
521 | 192k | } |
522 | 192k | if(ret != OK) return ret; |
523 | 192k | } |
524 | 26.9k | return OK; |
525 | 27.9k | } |
526 | | |
527 | | /*****************************************************************************/ |
528 | | /* */ |
529 | | /* Function Name : isvcd_decode_picture_thread */ |
530 | | /* */ |
531 | | /* Description : */ |
532 | | /* : */ |
533 | | /* : */ |
534 | | /* Inputs : */ |
535 | | /* Processing : */ |
536 | | /* */ |
537 | | /* Outputs : */ |
538 | | /* Returns : None */ |
539 | | /* */ |
540 | | /* Revision History: */ |
541 | | /* */ |
542 | | /* DD MM YYYY Author(s) Changes (Describe the changes made) */ |
543 | | /* ITTIAM */ |
544 | | /*****************************************************************************/ |
545 | | void isvcd_decode_picture_thread(svc_dec_lyr_struct_t *ps_svc_lyr_dec) |
546 | 21.5k | { |
547 | 21.5k | dec_struct_t *ps_dec = &ps_svc_lyr_dec->s_dec; |
548 | 21.5k | WORD32 ret; |
549 | 21.5k | ithread_set_name("isvcd_decode_picture_thread"); |
550 | 21.5k | while(1) |
551 | 21.5k | { |
552 | | #ifdef KEEP_THREADS_ACTIVE |
553 | | ret = ithread_mutex_lock(ps_dec->apv_proc_start_mutex[0]); |
554 | | if(OK != ret) break; |
555 | | |
556 | | while(ps_dec->ai4_process_start[0] != PROC_START) |
557 | | { |
558 | | ithread_cond_wait(ps_dec->apv_proc_start_condition[0], ps_dec->apv_proc_start_mutex[0]); |
559 | | } |
560 | | ps_dec->ai4_process_start[0] = PROC_IN_PROGRESS; |
561 | | |
562 | | ret = ithread_mutex_unlock(ps_dec->apv_proc_start_mutex[0]); |
563 | | if(OK != ret || ps_dec->i4_break_threads == 1) break; |
564 | | #endif |
565 | 28.8k | while(1) |
566 | 28.8k | { |
567 | | /*Complete all writes before processing next slice*/ |
568 | | |
569 | 28.8k | DEBUG_THREADS_PRINTF(" Entering decode slice svc ext\n"); |
570 | | |
571 | 28.8k | ret = isvcd_decode_slice_thread(ps_svc_lyr_dec); |
572 | 28.8k | if(OK != ret) break; |
573 | 26.9k | DEBUG_THREADS_PRINTF(" Exit isvcd_decode_slice_thread\n"); |
574 | | |
575 | 26.9k | if(ps_dec->cur_dec_mb_num > ps_dec->ps_cur_sps->u4_max_mb_addr) |
576 | 19.6k | { |
577 | | /*Last slice in frame*/ |
578 | 19.6k | break; |
579 | 19.6k | } |
580 | 7.33k | else |
581 | 7.33k | { |
582 | 7.33k | ps_dec->ps_decode_cur_slice++; |
583 | 7.33k | ps_dec->u2_cur_slice_num_dec_thread++; |
584 | 7.33k | } |
585 | 26.9k | } |
586 | 21.5k | if(ps_dec->u4_output_present && (2 == ps_dec->u4_num_cores) && |
587 | 8.78k | (ps_dec->u4_fmt_conv_cur_row < ps_dec->s_disp_frame_info.u4_y_ht)) |
588 | 8.49k | { |
589 | 8.49k | ps_dec->u4_fmt_conv_num_rows = |
590 | 8.49k | (ps_dec->s_disp_frame_info.u4_y_ht - ps_dec->u4_fmt_conv_cur_row); |
591 | 8.49k | ih264d_format_convert(ps_dec, &(ps_dec->s_disp_op), ps_dec->u4_fmt_conv_cur_row, |
592 | 8.49k | ps_dec->u4_fmt_conv_num_rows); |
593 | 8.49k | ps_dec->u4_fmt_conv_cur_row += ps_dec->u4_fmt_conv_num_rows; |
594 | 8.49k | } |
595 | | #ifdef KEEP_THREADS_ACTIVE |
596 | | ret = ithread_mutex_lock(ps_dec->apv_proc_done_mutex[0]); |
597 | | if(OK != ret) break; |
598 | | |
599 | | ps_dec->ai4_process_done[0] = PROC_DONE; |
600 | | ithread_cond_signal(ps_dec->apv_proc_done_condition[0]); |
601 | | |
602 | | ret = ithread_mutex_unlock(ps_dec->apv_proc_done_mutex[0]); |
603 | | if(OK != ret) break; |
604 | | #else |
605 | 21.5k | break; |
606 | 21.5k | #endif |
607 | 21.5k | } |
608 | 21.5k | } |