Commit | Line | Data |
---|---|---|
31fe1f25 MT |
1 | /* |
2 | * This file is part of Libav. | |
3 | * | |
4 | * Libav is free software; you can redistribute it and/or | |
5 | * modify it under the terms of the GNU Lesser General Public | |
6 | * License as published by the Free Software Foundation; either | |
7 | * version 2.1 of the License, or (at your option) any later version. | |
8 | * | |
9 | * Libav is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
12 | * Lesser General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU Lesser General Public | |
15 | * License along with Libav; if not, write to the Free Software | |
16 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
17 | */ | |
18 | ||
19 | #include <va/va.h> | |
20 | #include <va/va_enc_hevc.h> | |
21 | ||
22 | #include "libavutil/avassert.h" | |
23 | #include "libavutil/internal.h" | |
24 | #include "libavutil/opt.h" | |
25 | #include "libavutil/pixfmt.h" | |
26 | ||
27 | #include "avcodec.h" | |
28 | #include "hevc.h" | |
29 | #include "internal.h" | |
30 | #include "put_bits.h" | |
31 | #include "vaapi_encode.h" | |
32 | #include "vaapi_encode_h26x.h" | |
33 | ||
34 | ||
35 | #define MAX_ST_REF_PIC_SETS 32 | |
36 | #define MAX_DPB_PICS 16 | |
37 | #define MAX_LAYERS 1 | |
38 | ||
39 | ||
40 | typedef struct VAAPIEncodeH265STRPS { | |
41 | char inter_ref_pic_set_prediction_flag; | |
42 | ||
43 | unsigned int num_negative_pics; | |
44 | unsigned int num_positive_pics; | |
45 | ||
46 | unsigned int delta_poc_s0_minus1[MAX_DPB_PICS]; | |
47 | char used_by_curr_pic_s0_flag[MAX_DPB_PICS]; | |
48 | ||
49 | unsigned int delta_poc_s1_minus1[MAX_DPB_PICS]; | |
50 | char used_by_curr_pic_s1_flag[MAX_DPB_PICS]; | |
51 | } VAAPIEncodeH265STRPS; | |
52 | ||
53 | // This structure contains all possibly-useful per-sequence syntax elements | |
54 | // which are not already contained in the various VAAPI structures. | |
55 | typedef struct VAAPIEncodeH265MiscSequenceParams { | |
56 | ||
57 | // Parameter set IDs. | |
58 | unsigned int video_parameter_set_id; | |
59 | unsigned int seq_parameter_set_id; | |
60 | ||
61 | // Layering. | |
62 | unsigned int vps_max_layers_minus1; | |
63 | unsigned int vps_max_sub_layers_minus1; | |
64 | char vps_temporal_id_nesting_flag; | |
65 | unsigned int vps_max_layer_id; | |
66 | unsigned int vps_num_layer_sets_minus1; | |
67 | unsigned int sps_max_sub_layers_minus1; | |
68 | char sps_temporal_id_nesting_flag; | |
69 | char layer_id_included_flag[MAX_LAYERS][64]; | |
70 | ||
71 | // Profile/tier/level parameters. | |
72 | char general_profile_compatibility_flag[32]; | |
73 | char general_progressive_source_flag; | |
74 | char general_interlaced_source_flag; | |
75 | char general_non_packed_constraint_flag; | |
76 | char general_frame_only_constraint_flag; | |
77 | char general_inbld_flag; | |
78 | ||
79 | // Decode/display ordering parameters. | |
80 | unsigned int log2_max_pic_order_cnt_lsb_minus4; | |
81 | char vps_sub_layer_ordering_info_present_flag; | |
82 | unsigned int vps_max_dec_pic_buffering_minus1[MAX_LAYERS]; | |
83 | unsigned int vps_max_num_reorder_pics[MAX_LAYERS]; | |
84 | unsigned int vps_max_latency_increase_plus1[MAX_LAYERS]; | |
85 | char sps_sub_layer_ordering_info_present_flag; | |
86 | unsigned int sps_max_dec_pic_buffering_minus1[MAX_LAYERS]; | |
87 | unsigned int sps_max_num_reorder_pics[MAX_LAYERS]; | |
88 | unsigned int sps_max_latency_increase_plus1[MAX_LAYERS]; | |
89 | ||
90 | // Timing information. | |
91 | char vps_timing_info_present_flag; | |
92 | unsigned int vps_num_units_in_tick; | |
93 | unsigned int vps_time_scale; | |
94 | char vps_poc_proportional_to_timing_flag; | |
95 | unsigned int vps_num_ticks_poc_diff_minus1; | |
96 | ||
97 | // Cropping information. | |
98 | char conformance_window_flag; | |
99 | unsigned int conf_win_left_offset; | |
100 | unsigned int conf_win_right_offset; | |
101 | unsigned int conf_win_top_offset; | |
102 | unsigned int conf_win_bottom_offset; | |
103 | ||
104 | // Short-term reference picture sets. | |
105 | unsigned int num_short_term_ref_pic_sets; | |
106 | VAAPIEncodeH265STRPS st_ref_pic_set[MAX_ST_REF_PIC_SETS]; | |
107 | ||
108 | // Long-term reference pictures. | |
109 | char long_term_ref_pics_present_flag; | |
110 | unsigned int num_long_term_ref_pics_sps; | |
111 | struct { | |
112 | unsigned int lt_ref_pic_poc_lsb_sps; | |
113 | char used_by_curr_pic_lt_sps_flag; | |
114 | } lt_ref_pic; | |
115 | ||
116 | // Deblocking filter control. | |
117 | char deblocking_filter_control_present_flag; | |
118 | char deblocking_filter_override_enabled_flag; | |
119 | char pps_deblocking_filter_disabled_flag; | |
120 | int pps_beta_offset_div2; | |
121 | int pps_tc_offset_div2; | |
122 | ||
123 | // Video Usability Information. | |
124 | char vui_parameters_present_flag; | |
125 | char aspect_ratio_info_present_flag; | |
126 | unsigned int aspect_ratio_idc; | |
127 | unsigned int sar_width; | |
128 | unsigned int sar_height; | |
129 | char video_signal_type_present_flag; | |
130 | unsigned int video_format; | |
131 | char video_full_range_flag; | |
132 | char colour_description_present_flag; | |
133 | unsigned int colour_primaries; | |
134 | unsigned int transfer_characteristics; | |
135 | unsigned int matrix_coeffs; | |
136 | ||
137 | // Oddments. | |
138 | char uniform_spacing_flag; | |
139 | char output_flag_present_flag; | |
140 | char cabac_init_present_flag; | |
141 | unsigned int num_extra_slice_header_bits; | |
142 | char lists_modification_present_flag; | |
143 | char pps_slice_chroma_qp_offsets_present_flag; | |
144 | char pps_slice_chroma_offset_list_enabled_flag; | |
145 | } VAAPIEncodeH265MiscSequenceParams; | |
146 | ||
147 | // This structure contains all possibly-useful per-slice syntax elements | |
148 | // which are not already contained in the various VAAPI structures. | |
149 | typedef struct VAAPIEncodeH265MiscSliceParams { | |
150 | // Slice segments. | |
151 | char first_slice_segment_in_pic_flag; | |
152 | unsigned int slice_segment_address; | |
153 | ||
154 | // Short-term reference picture sets. | |
155 | char short_term_ref_pic_set_sps_flag; | |
156 | unsigned int short_term_ref_pic_idx; | |
157 | VAAPIEncodeH265STRPS st_ref_pic_set; | |
158 | ||
159 | // Deblocking filter. | |
160 | char deblocking_filter_override_flag; | |
161 | ||
162 | // Oddments. | |
163 | char slice_reserved_flag[8]; | |
164 | char no_output_of_prior_pics_flag; | |
165 | char pic_output_flag; | |
166 | } VAAPIEncodeH265MiscSliceParams; | |
167 | ||
168 | typedef struct VAAPIEncodeH265Slice { | |
169 | VAAPIEncodeH265MiscSliceParams misc_slice_params; | |
170 | ||
171 | int64_t pic_order_cnt; | |
172 | } VAAPIEncodeH265Slice; | |
173 | ||
174 | typedef struct VAAPIEncodeH265Context { | |
175 | VAAPIEncodeH265MiscSequenceParams misc_sequence_params; | |
176 | ||
177 | unsigned int ctu_width; | |
178 | unsigned int ctu_height; | |
179 | ||
180 | int fixed_qp_idr; | |
181 | int fixed_qp_p; | |
182 | int fixed_qp_b; | |
183 | ||
184 | int64_t last_idr_frame; | |
f70e4627 MT |
185 | |
186 | // Rate control configuration. | |
187 | struct { | |
188 | VAEncMiscParameterBuffer misc; | |
189 | VAEncMiscParameterRateControl rc; | |
190 | } rc_params; | |
191 | struct { | |
192 | VAEncMiscParameterBuffer misc; | |
193 | VAEncMiscParameterHRD hrd; | |
194 | } hrd_params; | |
31fe1f25 MT |
195 | } VAAPIEncodeH265Context; |
196 | ||
92fdea37 MT |
197 | typedef struct VAAPIEncodeH265Options { |
198 | int qp; | |
199 | } VAAPIEncodeH265Options; | |
200 | ||
31fe1f25 MT |
201 | |
202 | #define vseq_var(name) vseq->name, name | |
203 | #define vseq_field(name) vseq->seq_fields.bits.name, name | |
204 | #define vpic_var(name) vpic->name, name | |
205 | #define vpic_field(name) vpic->pic_fields.bits.name, name | |
206 | #define vslice_var(name) vslice->name, name | |
207 | #define vslice_field(name) vslice->slice_fields.bits.name, name | |
208 | #define mseq_var(name) mseq->name, name | |
209 | #define mslice_var(name) mslice->name, name | |
210 | #define mstrps_var(name) mstrps->name, name | |
211 | ||
212 | static void vaapi_encode_h265_write_nal_unit_header(PutBitContext *pbc, | |
213 | int nal_unit_type) | |
214 | { | |
215 | u(1, 0, forbidden_zero_bit); | |
216 | u(6, nal_unit_type, nal_unit_type); | |
217 | u(6, 0, nuh_layer_id); | |
218 | u(3, 1, nuh_temporal_id_plus1); | |
219 | } | |
220 | ||
221 | static void vaapi_encode_h265_write_rbsp_trailing_bits(PutBitContext *pbc) | |
222 | { | |
223 | u(1, 1, rbsp_stop_one_bit); | |
224 | while (put_bits_count(pbc) & 7) | |
225 | u(1, 0, rbsp_alignment_zero_bit); | |
226 | } | |
227 | ||
228 | static void vaapi_encode_h265_write_profile_tier_level(PutBitContext *pbc, | |
229 | VAAPIEncodeContext *ctx) | |
230 | { | |
231 | VAEncSequenceParameterBufferHEVC *vseq = ctx->codec_sequence_params; | |
232 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
233 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
234 | int j; | |
235 | ||
236 | if (1) { | |
237 | u(2, 0, general_profile_space); | |
238 | u(1, vseq_var(general_tier_flag)); | |
239 | u(5, vseq_var(general_profile_idc)); | |
240 | ||
241 | for (j = 0; j < 32; j++) { | |
242 | u(1, mseq_var(general_profile_compatibility_flag[j])); | |
243 | } | |
244 | ||
245 | u(1, mseq_var(general_progressive_source_flag)); | |
246 | u(1, mseq_var(general_interlaced_source_flag)); | |
247 | u(1, mseq_var(general_non_packed_constraint_flag)); | |
248 | u(1, mseq_var(general_frame_only_constraint_flag)); | |
249 | ||
250 | if (0) { | |
251 | // Not main profile. | |
252 | // Lots of extra constraint flags. | |
253 | } else { | |
254 | // put_bits only handles up to 31 bits. | |
255 | u(23, 0, general_reserved_zero_43bits); | |
256 | u(20, 0, general_reserved_zero_43bits); | |
257 | } | |
258 | ||
259 | if (vseq->general_profile_idc >= 1 && vseq->general_profile_idc <= 5) { | |
260 | u(1, mseq_var(general_inbld_flag)); | |
261 | } else { | |
262 | u(1, 0, general_reserved_zero_bit); | |
263 | } | |
264 | } | |
265 | ||
266 | u(8, vseq_var(general_level_idc)); | |
267 | ||
268 | // No sublayers. | |
269 | } | |
270 | ||
271 | static void vaapi_encode_h265_write_vps(PutBitContext *pbc, | |
272 | VAAPIEncodeContext *ctx) | |
273 | { | |
274 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
275 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
276 | int i, j; | |
277 | ||
278 | vaapi_encode_h265_write_nal_unit_header(pbc, NAL_VPS); | |
279 | ||
280 | u(4, mseq->video_parameter_set_id, vps_video_parameter_set_id); | |
281 | ||
282 | u(1, 1, vps_base_layer_internal_flag); | |
283 | u(1, 1, vps_base_layer_available_flag); | |
284 | u(6, mseq_var(vps_max_layers_minus1)); | |
285 | u(3, mseq_var(vps_max_sub_layers_minus1)); | |
286 | u(1, mseq_var(vps_temporal_id_nesting_flag)); | |
287 | ||
288 | u(16, 0xffff, vps_reserved_0xffff_16bits); | |
289 | ||
290 | vaapi_encode_h265_write_profile_tier_level(pbc, ctx); | |
291 | ||
292 | u(1, mseq_var(vps_sub_layer_ordering_info_present_flag)); | |
293 | for (i = (mseq->vps_sub_layer_ordering_info_present_flag ? | |
294 | 0 : mseq->vps_max_sub_layers_minus1); | |
295 | i <= mseq->vps_max_sub_layers_minus1; i++) { | |
296 | ue(mseq_var(vps_max_dec_pic_buffering_minus1[i])); | |
297 | ue(mseq_var(vps_max_num_reorder_pics[i])); | |
298 | ue(mseq_var(vps_max_latency_increase_plus1[i])); | |
299 | } | |
300 | ||
301 | u(6, mseq_var(vps_max_layer_id)); | |
302 | ue(mseq_var(vps_num_layer_sets_minus1)); | |
303 | for (i = 1; i <= mseq->vps_num_layer_sets_minus1; i++) { | |
304 | for (j = 0; j < mseq->vps_max_layer_id; j++) | |
305 | u(1, mseq_var(layer_id_included_flag[i][j])); | |
306 | } | |
307 | ||
308 | u(1, mseq_var(vps_timing_info_present_flag)); | |
309 | if (mseq->vps_timing_info_present_flag) { | |
310 | u(1, 0, put_bits_hack_zero_bit); | |
311 | u(31, mseq_var(vps_num_units_in_tick)); | |
312 | u(1, 0, put_bits_hack_zero_bit); | |
313 | u(31, mseq_var(vps_time_scale)); | |
314 | u(1, mseq_var(vps_poc_proportional_to_timing_flag)); | |
315 | if (mseq->vps_poc_proportional_to_timing_flag) { | |
316 | ue(mseq_var(vps_num_ticks_poc_diff_minus1)); | |
317 | } | |
318 | ue(0, vps_num_hrd_parameters); | |
319 | } | |
320 | ||
321 | u(1, 0, vps_extension_flag); | |
322 | ||
323 | vaapi_encode_h265_write_rbsp_trailing_bits(pbc); | |
324 | } | |
325 | ||
326 | static void vaapi_encode_h265_write_st_ref_pic_set(PutBitContext *pbc, | |
327 | int st_rps_idx, | |
328 | VAAPIEncodeH265STRPS *mstrps) | |
329 | { | |
330 | int i; | |
331 | ||
332 | if (st_rps_idx != 0) | |
333 | u(1, mstrps_var(inter_ref_pic_set_prediction_flag)); | |
334 | ||
335 | if (mstrps->inter_ref_pic_set_prediction_flag) { | |
336 | av_assert0(0 && "inter ref pic set prediction not supported"); | |
337 | } else { | |
338 | ue(mstrps_var(num_negative_pics)); | |
339 | ue(mstrps_var(num_positive_pics)); | |
340 | ||
341 | for (i = 0; i < mstrps->num_negative_pics; i++) { | |
342 | ue(mstrps_var(delta_poc_s0_minus1[i])); | |
343 | u(1, mstrps_var(used_by_curr_pic_s0_flag[i])); | |
344 | } | |
345 | for (i = 0; i < mstrps->num_positive_pics; i++) { | |
346 | ue(mstrps_var(delta_poc_s1_minus1[i])); | |
347 | u(1, mstrps_var(used_by_curr_pic_s1_flag[i])); | |
348 | } | |
349 | } | |
350 | } | |
351 | ||
352 | static void vaapi_encode_h265_write_vui_parameters(PutBitContext *pbc, | |
353 | VAAPIEncodeContext *ctx) | |
354 | { | |
355 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
356 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
357 | ||
358 | u(1, mseq_var(aspect_ratio_info_present_flag)); | |
359 | if (mseq->aspect_ratio_info_present_flag) { | |
360 | u(8, mseq_var(aspect_ratio_idc)); | |
361 | if (mseq->aspect_ratio_idc == 255) { | |
362 | u(16, mseq_var(sar_width)); | |
363 | u(16, mseq_var(sar_height)); | |
364 | } | |
365 | } | |
366 | ||
367 | u(1, 0, overscan_info_present_flag); | |
368 | ||
369 | u(1, mseq_var(video_signal_type_present_flag)); | |
370 | if (mseq->video_signal_type_present_flag) { | |
371 | u(3, mseq_var(video_format)); | |
372 | u(1, mseq_var(video_full_range_flag)); | |
373 | u(1, mseq_var(colour_description_present_flag)); | |
374 | if (mseq->colour_description_present_flag) { | |
375 | u(8, mseq_var(colour_primaries)); | |
376 | u(8, mseq_var(transfer_characteristics)); | |
377 | u(8, mseq_var(matrix_coeffs)); | |
378 | } | |
379 | } | |
380 | ||
381 | u(1, 0, chroma_loc_info_present_flag); | |
382 | u(1, 0, neutral_chroma_indication_flag); | |
383 | u(1, 0, field_seq_flag); | |
384 | u(1, 0, frame_field_info_present_flag); | |
385 | u(1, 0, default_display_window_flag); | |
386 | u(1, 0, vui_timing_info_present_flag); | |
387 | u(1, 0, bitstream_restriction_flag_flag); | |
388 | } | |
389 | ||
390 | static void vaapi_encode_h265_write_sps(PutBitContext *pbc, | |
391 | VAAPIEncodeContext *ctx) | |
392 | { | |
393 | VAEncSequenceParameterBufferHEVC *vseq = ctx->codec_sequence_params; | |
394 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
395 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
396 | int i; | |
397 | ||
398 | vaapi_encode_h265_write_nal_unit_header(pbc, NAL_SPS); | |
399 | ||
400 | u(4, mseq->video_parameter_set_id, sps_video_parameter_set_id); | |
401 | ||
402 | u(3, mseq_var(sps_max_sub_layers_minus1)); | |
403 | u(1, mseq_var(sps_temporal_id_nesting_flag)); | |
404 | ||
405 | vaapi_encode_h265_write_profile_tier_level(pbc, ctx); | |
406 | ||
407 | ue(mseq->seq_parameter_set_id, sps_seq_parameter_set_id); | |
408 | ue(vseq_field(chroma_format_idc)); | |
409 | if (vseq->seq_fields.bits.chroma_format_idc == 3) | |
410 | u(1, 0, separate_colour_plane_flag); | |
411 | ||
412 | ue(vseq_var(pic_width_in_luma_samples)); | |
413 | ue(vseq_var(pic_height_in_luma_samples)); | |
414 | ||
415 | u(1, mseq_var(conformance_window_flag)); | |
416 | if (mseq->conformance_window_flag) { | |
417 | ue(mseq_var(conf_win_left_offset)); | |
418 | ue(mseq_var(conf_win_right_offset)); | |
419 | ue(mseq_var(conf_win_top_offset)); | |
420 | ue(mseq_var(conf_win_bottom_offset)); | |
421 | } | |
422 | ||
423 | ue(vseq_field(bit_depth_luma_minus8)); | |
424 | ue(vseq_field(bit_depth_chroma_minus8)); | |
425 | ||
426 | ue(mseq_var(log2_max_pic_order_cnt_lsb_minus4)); | |
427 | ||
428 | u(1, mseq_var(sps_sub_layer_ordering_info_present_flag)); | |
429 | for (i = (mseq->sps_sub_layer_ordering_info_present_flag ? | |
430 | 0 : mseq->sps_max_sub_layers_minus1); | |
431 | i <= mseq->sps_max_sub_layers_minus1; i++) { | |
432 | ue(mseq_var(sps_max_dec_pic_buffering_minus1[i])); | |
433 | ue(mseq_var(sps_max_num_reorder_pics[i])); | |
434 | ue(mseq_var(sps_max_latency_increase_plus1[i])); | |
435 | } | |
436 | ||
437 | ue(vseq_var(log2_min_luma_coding_block_size_minus3)); | |
438 | ue(vseq_var(log2_diff_max_min_luma_coding_block_size)); | |
439 | ue(vseq_var(log2_min_transform_block_size_minus2)); | |
440 | ue(vseq_var(log2_diff_max_min_transform_block_size)); | |
441 | ue(vseq_var(max_transform_hierarchy_depth_inter)); | |
442 | ue(vseq_var(max_transform_hierarchy_depth_intra)); | |
443 | ||
444 | u(1, vseq_field(scaling_list_enabled_flag)); | |
445 | if (vseq->seq_fields.bits.scaling_list_enabled_flag) { | |
446 | u(1, 0, sps_scaling_list_data_present_flag); | |
447 | } | |
448 | ||
449 | u(1, vseq_field(amp_enabled_flag)); | |
450 | u(1, vseq_field(sample_adaptive_offset_enabled_flag)); | |
451 | ||
452 | u(1, vseq_field(pcm_enabled_flag)); | |
453 | if (vseq->seq_fields.bits.pcm_enabled_flag) { | |
454 | u(4, vseq_var(pcm_sample_bit_depth_luma_minus1)); | |
455 | u(4, vseq_var(pcm_sample_bit_depth_chroma_minus1)); | |
456 | ue(vseq_var(log2_min_pcm_luma_coding_block_size_minus3)); | |
457 | ue(vseq->log2_max_pcm_luma_coding_block_size_minus3 - | |
458 | vseq->log2_min_pcm_luma_coding_block_size_minus3, | |
459 | log2_diff_max_min_pcm_luma_coding_block_size); | |
460 | u(1, vseq_field(pcm_loop_filter_disabled_flag)); | |
461 | } | |
462 | ||
463 | ue(mseq_var(num_short_term_ref_pic_sets)); | |
464 | for (i = 0; i < mseq->num_short_term_ref_pic_sets; i++) | |
465 | vaapi_encode_h265_write_st_ref_pic_set(pbc, i, | |
466 | &mseq->st_ref_pic_set[i]); | |
467 | ||
468 | u(1, mseq_var(long_term_ref_pics_present_flag)); | |
469 | if (mseq->long_term_ref_pics_present_flag) { | |
470 | ue(0, num_long_term_ref_pics_sps); | |
471 | } | |
472 | ||
473 | u(1, vseq_field(sps_temporal_mvp_enabled_flag)); | |
474 | u(1, vseq_field(strong_intra_smoothing_enabled_flag)); | |
475 | ||
476 | u(1, mseq_var(vui_parameters_present_flag)); | |
477 | if (mseq->vui_parameters_present_flag) { | |
478 | vaapi_encode_h265_write_vui_parameters(pbc, ctx); | |
479 | } | |
480 | ||
481 | u(1, 0, sps_extension_present_flag); | |
482 | ||
483 | vaapi_encode_h265_write_rbsp_trailing_bits(pbc); | |
484 | } | |
485 | ||
486 | static void vaapi_encode_h265_write_pps(PutBitContext *pbc, | |
487 | VAAPIEncodeContext *ctx) | |
488 | { | |
489 | VAEncPictureParameterBufferHEVC *vpic = ctx->codec_picture_params; | |
490 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
491 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
492 | int i; | |
493 | ||
494 | vaapi_encode_h265_write_nal_unit_header(pbc, NAL_PPS); | |
495 | ||
496 | ue(vpic->slice_pic_parameter_set_id, pps_pic_parameter_set_id); | |
497 | ue(mseq->seq_parameter_set_id, pps_seq_parameter_set_id); | |
498 | ||
499 | u(1, vpic_field(dependent_slice_segments_enabled_flag)); | |
500 | u(1, mseq_var(output_flag_present_flag)); | |
501 | u(3, mseq_var(num_extra_slice_header_bits)); | |
502 | u(1, vpic_field(sign_data_hiding_enabled_flag)); | |
503 | u(1, mseq_var(cabac_init_present_flag)); | |
504 | ||
505 | ue(vpic_var(num_ref_idx_l0_default_active_minus1)); | |
506 | ue(vpic_var(num_ref_idx_l1_default_active_minus1)); | |
507 | ||
508 | se(vpic->pic_init_qp - 26, init_qp_minus26); | |
509 | ||
510 | u(1, vpic_field(constrained_intra_pred_flag)); | |
511 | u(1, vpic_field(transform_skip_enabled_flag)); | |
512 | ||
513 | u(1, vpic_field(cu_qp_delta_enabled_flag)); | |
514 | if (vpic->pic_fields.bits.cu_qp_delta_enabled_flag) | |
515 | ue(vpic_var(diff_cu_qp_delta_depth)); | |
516 | ||
517 | se(vpic_var(pps_cb_qp_offset)); | |
518 | se(vpic_var(pps_cr_qp_offset)); | |
519 | ||
520 | u(1, mseq_var(pps_slice_chroma_qp_offsets_present_flag)); | |
521 | u(1, vpic_field(weighted_pred_flag)); | |
522 | u(1, vpic_field(weighted_bipred_flag)); | |
523 | u(1, vpic_field(transquant_bypass_enabled_flag)); | |
524 | u(1, vpic_field(tiles_enabled_flag)); | |
525 | u(1, vpic_field(entropy_coding_sync_enabled_flag)); | |
526 | ||
527 | if (vpic->pic_fields.bits.tiles_enabled_flag) { | |
528 | ue(vpic_var(num_tile_columns_minus1)); | |
529 | ue(vpic_var(num_tile_rows_minus1)); | |
530 | u(1, mseq_var(uniform_spacing_flag)); | |
531 | if (!mseq->uniform_spacing_flag) { | |
532 | for (i = 0; i < vpic->num_tile_columns_minus1; i++) | |
533 | ue(vpic_var(column_width_minus1[i])); | |
534 | for (i = 0; i < vpic->num_tile_rows_minus1; i++) | |
535 | ue(vpic_var(row_height_minus1[i])); | |
536 | } | |
537 | u(1, vpic_field(loop_filter_across_tiles_enabled_flag)); | |
538 | } | |
539 | ||
540 | u(1, vpic_field(pps_loop_filter_across_slices_enabled_flag)); | |
541 | u(1, mseq_var(deblocking_filter_control_present_flag)); | |
542 | if (mseq->deblocking_filter_control_present_flag) { | |
543 | u(1, mseq_var(deblocking_filter_override_enabled_flag)); | |
544 | u(1, mseq_var(pps_deblocking_filter_disabled_flag)); | |
545 | if (!mseq->pps_deblocking_filter_disabled_flag) { | |
546 | se(mseq_var(pps_beta_offset_div2)); | |
547 | se(mseq_var(pps_tc_offset_div2)); | |
548 | } | |
549 | } | |
550 | ||
551 | u(1, 0, pps_scaling_list_data_present_flag); | |
552 | // No scaling list data. | |
553 | ||
554 | u(1, mseq_var(lists_modification_present_flag)); | |
555 | ue(vpic_var(log2_parallel_merge_level_minus2)); | |
556 | u(1, 0, slice_segment_header_extension_present_flag); | |
557 | u(1, 0, pps_extension_present_flag); | |
558 | ||
559 | vaapi_encode_h265_write_rbsp_trailing_bits(pbc); | |
560 | } | |
561 | ||
562 | static void vaapi_encode_h265_write_slice_header2(PutBitContext *pbc, | |
563 | VAAPIEncodeContext *ctx, | |
564 | VAAPIEncodePicture *pic, | |
565 | VAAPIEncodeSlice *slice) | |
566 | { | |
567 | VAEncSequenceParameterBufferHEVC *vseq = ctx->codec_sequence_params; | |
568 | VAEncPictureParameterBufferHEVC *vpic = pic->codec_picture_params; | |
569 | VAEncSliceParameterBufferHEVC *vslice = slice->codec_slice_params; | |
570 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
571 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
572 | VAAPIEncodeH265Slice *pslice = slice->priv_data; | |
573 | VAAPIEncodeH265MiscSliceParams *mslice = &pslice->misc_slice_params; | |
574 | int i; | |
575 | ||
576 | vaapi_encode_h265_write_nal_unit_header(pbc, vpic->nal_unit_type); | |
577 | ||
578 | u(1, mslice_var(first_slice_segment_in_pic_flag)); | |
579 | if (vpic->nal_unit_type >= NAL_BLA_W_LP && | |
580 | vpic->nal_unit_type <= 23) | |
581 | u(1, mslice_var(no_output_of_prior_pics_flag)); | |
582 | ||
583 | ue(vslice_var(slice_pic_parameter_set_id)); | |
584 | ||
585 | if (!mslice->first_slice_segment_in_pic_flag) { | |
586 | if (vpic->pic_fields.bits.dependent_slice_segments_enabled_flag) | |
587 | u(1, vslice_field(dependent_slice_segment_flag)); | |
588 | u(av_log2((priv->ctu_width * priv->ctu_height) - 1) + 1, | |
589 | mslice_var(slice_segment_address)); | |
590 | } | |
591 | if (!vslice->slice_fields.bits.dependent_slice_segment_flag) { | |
592 | for (i = 0; i < mseq->num_extra_slice_header_bits; i++) | |
593 | u(1, mslice_var(slice_reserved_flag[i])); | |
594 | ||
595 | ue(vslice_var(slice_type)); | |
596 | if (mseq->output_flag_present_flag) | |
597 | u(1, 1, pic_output_flag); | |
598 | if (vseq->seq_fields.bits.separate_colour_plane_flag) | |
599 | u(2, vslice_field(colour_plane_id)); | |
600 | if (vpic->nal_unit_type != NAL_IDR_W_RADL && | |
601 | vpic->nal_unit_type != NAL_IDR_N_LP) { | |
602 | u(4 + mseq->log2_max_pic_order_cnt_lsb_minus4, | |
603 | (pslice->pic_order_cnt & | |
604 | ((1 << (mseq->log2_max_pic_order_cnt_lsb_minus4 + 4)) - 1)), | |
605 | slice_pic_order_cnt_lsb); | |
606 | ||
607 | u(1, mslice_var(short_term_ref_pic_set_sps_flag)); | |
608 | if (!mslice->short_term_ref_pic_set_sps_flag) { | |
609 | vaapi_encode_h265_write_st_ref_pic_set(pbc, mseq->num_short_term_ref_pic_sets, | |
610 | &mslice->st_ref_pic_set); | |
611 | } else if (mseq->num_short_term_ref_pic_sets > 1) { | |
612 | u(av_log2(mseq->num_short_term_ref_pic_sets - 1) + 1, | |
613 | mslice_var(short_term_ref_pic_idx)); | |
614 | } | |
615 | ||
616 | if (mseq->long_term_ref_pics_present_flag) { | |
617 | av_assert0(0); | |
618 | } | |
619 | ||
620 | if (vseq->seq_fields.bits.sps_temporal_mvp_enabled_flag) { | |
621 | u(1, vslice_field(slice_temporal_mvp_enabled_flag)); | |
622 | } | |
623 | ||
624 | if (vseq->seq_fields.bits.sample_adaptive_offset_enabled_flag) { | |
625 | u(1, vslice_field(slice_sao_luma_flag)); | |
626 | if (!vseq->seq_fields.bits.separate_colour_plane_flag && | |
627 | vseq->seq_fields.bits.chroma_format_idc != 0) { | |
628 | u(1, vslice_field(slice_sao_chroma_flag)); | |
629 | } | |
630 | } | |
631 | ||
632 | if (vslice->slice_type == P_SLICE || vslice->slice_type == B_SLICE) { | |
633 | u(1, vslice_field(num_ref_idx_active_override_flag)); | |
634 | if (vslice->slice_fields.bits.num_ref_idx_active_override_flag) { | |
635 | ue(vslice_var(num_ref_idx_l0_active_minus1)); | |
636 | if (vslice->slice_type == B_SLICE) { | |
637 | ue(vslice_var(num_ref_idx_l1_active_minus1)); | |
638 | } | |
639 | } | |
640 | ||
641 | if (mseq->lists_modification_present_flag) { | |
642 | av_assert0(0); | |
643 | // ref_pic_lists_modification() | |
644 | } | |
645 | if (vslice->slice_type == B_SLICE) { | |
646 | u(1, vslice_field(mvd_l1_zero_flag)); | |
647 | } | |
648 | if (mseq->cabac_init_present_flag) { | |
649 | u(1, vslice_field(cabac_init_flag)); | |
650 | } | |
651 | if (vslice->slice_fields.bits.slice_temporal_mvp_enabled_flag) { | |
652 | if (vslice->slice_type == B_SLICE) | |
653 | u(1, vslice_field(collocated_from_l0_flag)); | |
654 | ue(vpic->collocated_ref_pic_index, collocated_ref_idx); | |
655 | } | |
656 | if ((vpic->pic_fields.bits.weighted_pred_flag && | |
657 | vslice->slice_type == P_SLICE) || | |
658 | (vpic->pic_fields.bits.weighted_bipred_flag && | |
659 | vslice->slice_type == B_SLICE)) { | |
660 | av_assert0(0); | |
661 | // pred_weight_table() | |
662 | } | |
663 | ue(5 - vslice->max_num_merge_cand, five_minus_max_num_merge_cand); | |
664 | } | |
665 | ||
666 | se(vslice_var(slice_qp_delta)); | |
667 | if (mseq->pps_slice_chroma_qp_offsets_present_flag) { | |
668 | se(vslice_var(slice_cb_qp_offset)); | |
669 | se(vslice_var(slice_cr_qp_offset)); | |
670 | } | |
671 | if (mseq->pps_slice_chroma_offset_list_enabled_flag) { | |
672 | u(1, 0, cu_chroma_qp_offset_enabled_flag); | |
673 | } | |
674 | if (mseq->deblocking_filter_override_enabled_flag) { | |
675 | u(1, mslice_var(deblocking_filter_override_flag)); | |
676 | } | |
677 | if (mslice->deblocking_filter_override_flag) { | |
678 | u(1, vslice_field(slice_deblocking_filter_disabled_flag)); | |
679 | if (!vslice->slice_fields.bits.slice_deblocking_filter_disabled_flag) { | |
680 | se(vslice_var(slice_beta_offset_div2)); | |
681 | se(vslice_var(slice_tc_offset_div2)); | |
682 | } | |
683 | } | |
684 | if (vpic->pic_fields.bits.pps_loop_filter_across_slices_enabled_flag && | |
685 | (vslice->slice_fields.bits.slice_sao_luma_flag || | |
686 | vslice->slice_fields.bits.slice_sao_chroma_flag || | |
687 | vslice->slice_fields.bits.slice_deblocking_filter_disabled_flag)) { | |
688 | u(1, vslice_field(slice_loop_filter_across_slices_enabled_flag)); | |
689 | } | |
690 | } | |
691 | ||
692 | if (vpic->pic_fields.bits.tiles_enabled_flag || | |
693 | vpic->pic_fields.bits.entropy_coding_sync_enabled_flag) { | |
694 | // num_entry_point_offsets | |
695 | } | |
696 | ||
697 | if (0) { | |
698 | // slice_segment_header_extension_length | |
699 | } | |
700 | } | |
701 | ||
702 | u(1, 1, alignment_bit_equal_to_one); | |
703 | while (put_bits_count(pbc) & 7) | |
704 | u(1, 0, alignment_bit_equal_to_zero); | |
705 | } | |
706 | ||
707 | static int vaapi_encode_h265_write_sequence_header(AVCodecContext *avctx, | |
708 | char *data, size_t *data_len) | |
709 | { | |
710 | VAAPIEncodeContext *ctx = avctx->priv_data; | |
711 | PutBitContext pbc; | |
712 | char tmp[256]; | |
713 | int err; | |
714 | size_t nal_len, bit_len, bit_pos, next_len; | |
715 | ||
716 | bit_len = *data_len; | |
717 | bit_pos = 0; | |
718 | ||
719 | init_put_bits(&pbc, tmp, sizeof(tmp)); | |
720 | vaapi_encode_h265_write_vps(&pbc, ctx); | |
721 | nal_len = put_bits_count(&pbc); | |
722 | flush_put_bits(&pbc); | |
723 | ||
724 | next_len = bit_len - bit_pos; | |
725 | err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8, | |
726 | &next_len, | |
727 | tmp, nal_len); | |
728 | if (err < 0) | |
729 | return err; | |
730 | bit_pos += next_len; | |
731 | ||
732 | init_put_bits(&pbc, tmp, sizeof(tmp)); | |
733 | vaapi_encode_h265_write_sps(&pbc, ctx); | |
734 | nal_len = put_bits_count(&pbc); | |
735 | flush_put_bits(&pbc); | |
736 | ||
737 | next_len = bit_len - bit_pos; | |
738 | err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8, | |
739 | &next_len, | |
740 | tmp, nal_len); | |
741 | if (err < 0) | |
742 | return err; | |
743 | bit_pos += next_len; | |
744 | ||
745 | init_put_bits(&pbc, tmp, sizeof(tmp)); | |
746 | vaapi_encode_h265_write_pps(&pbc, ctx); | |
747 | nal_len = put_bits_count(&pbc); | |
748 | flush_put_bits(&pbc); | |
749 | ||
750 | next_len = bit_len - bit_pos; | |
751 | err = ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data + bit_pos / 8, | |
752 | &next_len, | |
753 | tmp, nal_len); | |
754 | if (err < 0) | |
755 | return err; | |
756 | bit_pos += next_len; | |
757 | ||
758 | *data_len = bit_pos; | |
759 | return 0; | |
760 | } | |
761 | ||
762 | static int vaapi_encode_h265_write_slice_header(AVCodecContext *avctx, | |
763 | VAAPIEncodePicture *pic, | |
764 | VAAPIEncodeSlice *slice, | |
765 | char *data, size_t *data_len) | |
766 | { | |
767 | VAAPIEncodeContext *ctx = avctx->priv_data; | |
768 | PutBitContext pbc; | |
769 | char tmp[256]; | |
770 | size_t header_len; | |
771 | ||
772 | init_put_bits(&pbc, tmp, sizeof(tmp)); | |
773 | vaapi_encode_h265_write_slice_header2(&pbc, ctx, pic, slice); | |
774 | header_len = put_bits_count(&pbc); | |
775 | flush_put_bits(&pbc); | |
776 | ||
777 | return ff_vaapi_encode_h26x_nal_unit_to_byte_stream(data, data_len, | |
778 | tmp, header_len); | |
779 | } | |
780 | ||
781 | static int vaapi_encode_h265_init_sequence_params(AVCodecContext *avctx) | |
782 | { | |
783 | VAAPIEncodeContext *ctx = avctx->priv_data; | |
784 | VAEncSequenceParameterBufferHEVC *vseq = ctx->codec_sequence_params; | |
785 | VAEncPictureParameterBufferHEVC *vpic = ctx->codec_picture_params; | |
786 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
787 | VAAPIEncodeH265MiscSequenceParams *mseq = &priv->misc_sequence_params; | |
788 | int i; | |
789 | ||
790 | { | |
791 | // general_profile_space == 0. | |
792 | vseq->general_profile_idc = 1; // Main profile (ctx->codec_profile?) | |
793 | vseq->general_tier_flag = 0; | |
794 | ||
795 | vseq->general_level_idc = avctx->level * 3; | |
796 | ||
797 | vseq->intra_period = 0; | |
798 | vseq->intra_idr_period = 0; | |
799 | vseq->ip_period = 0; | |
800 | ||
80a5d051 MT |
801 | vseq->pic_width_in_luma_samples = ctx->surface_width; |
802 | vseq->pic_height_in_luma_samples = ctx->surface_height; | |
31fe1f25 MT |
803 | |
804 | vseq->seq_fields.bits.chroma_format_idc = 1; // 4:2:0. | |
805 | vseq->seq_fields.bits.separate_colour_plane_flag = 0; | |
806 | vseq->seq_fields.bits.bit_depth_luma_minus8 = 0; // 8-bit luma. | |
807 | vseq->seq_fields.bits.bit_depth_chroma_minus8 = 0; // 8-bit chroma. | |
808 | // Other misc flags all zero. | |
809 | ||
810 | // These have to come from the capabilities of the encoder. We have | |
811 | // no way to query it, so just hardcode ones which worked for me... | |
812 | // CTB size from 8x8 to 32x32. | |
813 | vseq->log2_min_luma_coding_block_size_minus3 = 0; | |
814 | vseq->log2_diff_max_min_luma_coding_block_size = 2; | |
815 | // Transform size from 4x4 to 32x32. | |
816 | vseq->log2_min_transform_block_size_minus2 = 0; | |
817 | vseq->log2_diff_max_min_transform_block_size = 3; | |
818 | // Full transform hierarchy allowed (2-5). | |
819 | vseq->max_transform_hierarchy_depth_inter = 3; | |
820 | vseq->max_transform_hierarchy_depth_intra = 3; | |
821 | ||
822 | vseq->vui_parameters_present_flag = 0; | |
f70e4627 MT |
823 | |
824 | vseq->bits_per_second = avctx->bit_rate; | |
825 | if (avctx->framerate.num > 0 && avctx->framerate.den > 0) { | |
826 | vseq->vui_num_units_in_tick = avctx->framerate.num; | |
827 | vseq->vui_time_scale = avctx->framerate.den; | |
828 | } else { | |
829 | vseq->vui_num_units_in_tick = avctx->time_base.num; | |
830 | vseq->vui_time_scale = avctx->time_base.den; | |
831 | } | |
832 | ||
833 | vseq->intra_period = ctx->p_per_i * (ctx->b_per_p + 1); | |
834 | vseq->intra_idr_period = vseq->intra_period; | |
835 | vseq->ip_period = ctx->b_per_p + 1; | |
31fe1f25 MT |
836 | } |
837 | ||
838 | { | |
839 | vpic->decoded_curr_pic.picture_id = VA_INVALID_ID; | |
840 | vpic->decoded_curr_pic.flags = VA_PICTURE_HEVC_INVALID; | |
841 | ||
842 | for (i = 0; i < FF_ARRAY_ELEMS(vpic->reference_frames); i++) { | |
843 | vpic->reference_frames[i].picture_id = VA_INVALID_ID; | |
844 | vpic->reference_frames[i].flags = VA_PICTURE_HEVC_INVALID; | |
845 | } | |
846 | ||
847 | vpic->collocated_ref_pic_index = 0xff; | |
848 | ||
849 | vpic->last_picture = 0; | |
850 | ||
851 | vpic->pic_init_qp = priv->fixed_qp_idr; | |
852 | ||
853 | vpic->diff_cu_qp_delta_depth = 0; | |
854 | vpic->pps_cb_qp_offset = 0; | |
855 | vpic->pps_cr_qp_offset = 0; | |
856 | ||
857 | // tiles_enabled_flag == 0, so ignore num_tile_(rows|columns)_minus1. | |
858 | ||
859 | vpic->log2_parallel_merge_level_minus2 = 0; | |
860 | ||
861 | // No limit on size. | |
862 | vpic->ctu_max_bitsize_allowed = 0; | |
863 | ||
864 | vpic->num_ref_idx_l0_default_active_minus1 = 0; | |
865 | vpic->num_ref_idx_l1_default_active_minus1 = 0; | |
866 | ||
867 | vpic->slice_pic_parameter_set_id = 0; | |
868 | ||
869 | vpic->pic_fields.bits.screen_content_flag = 0; | |
870 | vpic->pic_fields.bits.enable_gpu_weighted_prediction = 0; | |
2940e196 MT |
871 | |
872 | // Per-CU QP changes are required for non-constant-QP modes. | |
873 | vpic->pic_fields.bits.cu_qp_delta_enabled_flag = | |
874 | ctx->va_rc_mode != VA_RC_CQP; | |
31fe1f25 MT |
875 | } |
876 | ||
877 | { | |
878 | mseq->video_parameter_set_id = 5; | |
879 | mseq->seq_parameter_set_id = 5; | |
880 | ||
881 | mseq->vps_max_layers_minus1 = 0; | |
882 | mseq->vps_max_sub_layers_minus1 = 0; | |
883 | mseq->vps_temporal_id_nesting_flag = 1; | |
884 | mseq->sps_max_sub_layers_minus1 = 0; | |
885 | mseq->sps_temporal_id_nesting_flag = 1; | |
886 | ||
887 | for (i = 0; i < 32; i++) { | |
888 | mseq->general_profile_compatibility_flag[i] = | |
889 | (i == vseq->general_profile_idc); | |
890 | } | |
891 | ||
892 | mseq->general_progressive_source_flag = 1; | |
893 | mseq->general_interlaced_source_flag = 0; | |
894 | mseq->general_non_packed_constraint_flag = 0; | |
895 | mseq->general_frame_only_constraint_flag = 1; | |
896 | mseq->general_inbld_flag = 0; | |
897 | ||
898 | mseq->log2_max_pic_order_cnt_lsb_minus4 = 8; | |
899 | mseq->vps_sub_layer_ordering_info_present_flag = 0; | |
900 | mseq->vps_max_dec_pic_buffering_minus1[0] = 1; | |
901 | mseq->vps_max_num_reorder_pics[0] = ctx->b_per_p; | |
902 | mseq->vps_max_latency_increase_plus1[0] = 0; | |
903 | mseq->sps_sub_layer_ordering_info_present_flag = 0; | |
904 | mseq->sps_max_dec_pic_buffering_minus1[0] = 1; | |
905 | mseq->sps_max_num_reorder_pics[0] = ctx->b_per_p; | |
906 | mseq->sps_max_latency_increase_plus1[0] = 0; | |
907 | ||
908 | mseq->vps_timing_info_present_flag = 1; | |
909 | mseq->vps_num_units_in_tick = avctx->time_base.num; | |
910 | mseq->vps_time_scale = avctx->time_base.den; | |
911 | mseq->vps_poc_proportional_to_timing_flag = 1; | |
912 | mseq->vps_num_ticks_poc_diff_minus1 = 0; | |
913 | ||
80a5d051 MT |
914 | if (avctx->width != ctx->surface_width || |
915 | avctx->height != ctx->surface_height) { | |
31fe1f25 MT |
916 | mseq->conformance_window_flag = 1; |
917 | mseq->conf_win_left_offset = 0; | |
918 | mseq->conf_win_right_offset = | |
80a5d051 | 919 | (ctx->surface_width - avctx->width) / 2; |
31fe1f25 MT |
920 | mseq->conf_win_top_offset = 0; |
921 | mseq->conf_win_bottom_offset = | |
80a5d051 | 922 | (ctx->surface_height - avctx->height) / 2; |
31fe1f25 MT |
923 | } else { |
924 | mseq->conformance_window_flag = 0; | |
925 | } | |
926 | ||
927 | mseq->num_short_term_ref_pic_sets = 0; | |
928 | // STRPSs should ideally be here rather than repeated in each slice. | |
929 | ||
930 | mseq->vui_parameters_present_flag = 1; | |
931 | if (avctx->sample_aspect_ratio.num != 0) { | |
932 | mseq->aspect_ratio_info_present_flag = 1; | |
933 | if (avctx->sample_aspect_ratio.num == | |
934 | avctx->sample_aspect_ratio.den) { | |
935 | mseq->aspect_ratio_idc = 1; | |
936 | } else { | |
937 | mseq->aspect_ratio_idc = 255; // Extended SAR. | |
938 | mseq->sar_width = avctx->sample_aspect_ratio.num; | |
939 | mseq->sar_height = avctx->sample_aspect_ratio.den; | |
940 | } | |
941 | } | |
942 | if (1) { | |
943 | // Should this be conditional on some of these being set? | |
944 | mseq->video_signal_type_present_flag = 1; | |
945 | mseq->video_format = 5; // Unspecified. | |
946 | mseq->video_full_range_flag = 0; | |
947 | mseq->colour_description_present_flag = 1; | |
948 | mseq->colour_primaries = avctx->color_primaries; | |
949 | mseq->transfer_characteristics = avctx->color_trc; | |
950 | mseq->matrix_coeffs = avctx->colorspace; | |
951 | } | |
952 | } | |
953 | ||
954 | return 0; | |
955 | } | |
956 | ||
957 | static int vaapi_encode_h265_init_picture_params(AVCodecContext *avctx, | |
958 | VAAPIEncodePicture *pic) | |
959 | { | |
960 | VAAPIEncodeContext *ctx = avctx->priv_data; | |
961 | VAEncPictureParameterBufferHEVC *vpic = pic->codec_picture_params; | |
962 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
963 | int i; | |
964 | ||
965 | if (pic->type == PICTURE_TYPE_IDR) { | |
966 | av_assert0(pic->display_order == pic->encode_order); | |
967 | priv->last_idr_frame = pic->display_order; | |
968 | } else { | |
969 | av_assert0(pic->encode_order > priv->last_idr_frame); | |
970 | // Display order need not be if we have RA[SD]L pictures, though. | |
971 | } | |
972 | ||
973 | vpic->decoded_curr_pic.picture_id = pic->recon_surface; | |
974 | vpic->decoded_curr_pic.pic_order_cnt = | |
975 | pic->display_order - priv->last_idr_frame; | |
976 | vpic->decoded_curr_pic.flags = 0; | |
977 | ||
978 | for (i = 0; i < pic->nb_refs; i++) { | |
979 | VAAPIEncodePicture *ref = pic->refs[i]; | |
980 | av_assert0(ref); | |
981 | vpic->reference_frames[i].picture_id = ref->recon_surface; | |
982 | vpic->reference_frames[i].pic_order_cnt = | |
983 | ref->display_order - priv->last_idr_frame; | |
984 | vpic->reference_frames[i].flags = | |
985 | (ref->display_order < pic->display_order ? | |
986 | VA_PICTURE_HEVC_RPS_ST_CURR_BEFORE : 0) | | |
987 | (ref->display_order > pic->display_order ? | |
988 | VA_PICTURE_HEVC_RPS_ST_CURR_AFTER : 0); | |
989 | } | |
990 | for (; i < FF_ARRAY_ELEMS(vpic->reference_frames); i++) { | |
991 | vpic->reference_frames[i].picture_id = VA_INVALID_ID; | |
992 | vpic->reference_frames[i].flags = VA_PICTURE_HEVC_INVALID; | |
993 | } | |
994 | ||
995 | vpic->coded_buf = pic->output_buffer; | |
996 | ||
997 | switch (pic->type) { | |
998 | case PICTURE_TYPE_IDR: | |
999 | vpic->nal_unit_type = NAL_IDR_W_RADL; | |
1000 | vpic->pic_fields.bits.idr_pic_flag = 1; | |
1001 | vpic->pic_fields.bits.coding_type = 1; | |
1002 | vpic->pic_fields.bits.reference_pic_flag = 1; | |
1003 | break; | |
1004 | case PICTURE_TYPE_I: | |
1005 | vpic->nal_unit_type = NAL_TRAIL_R; | |
1006 | vpic->pic_fields.bits.idr_pic_flag = 0; | |
1007 | vpic->pic_fields.bits.coding_type = 1; | |
1008 | vpic->pic_fields.bits.reference_pic_flag = 1; | |
1009 | break; | |
1010 | case PICTURE_TYPE_P: | |
1011 | vpic->nal_unit_type = NAL_TRAIL_R; | |
1012 | vpic->pic_fields.bits.idr_pic_flag = 0; | |
1013 | vpic->pic_fields.bits.coding_type = 2; | |
1014 | vpic->pic_fields.bits.reference_pic_flag = 1; | |
1015 | break; | |
1016 | case PICTURE_TYPE_B: | |
1017 | vpic->nal_unit_type = NAL_TRAIL_R; | |
1018 | vpic->pic_fields.bits.idr_pic_flag = 0; | |
1019 | vpic->pic_fields.bits.coding_type = 3; | |
1020 | vpic->pic_fields.bits.reference_pic_flag = 0; | |
1021 | break; | |
1022 | default: | |
1023 | av_assert0(0 && "invalid picture type"); | |
1024 | } | |
1025 | ||
1026 | pic->nb_slices = 1; | |
1027 | ||
1028 | return 0; | |
1029 | } | |
1030 | ||
1031 | static int vaapi_encode_h265_init_slice_params(AVCodecContext *avctx, | |
1032 | VAAPIEncodePicture *pic, | |
1033 | VAAPIEncodeSlice *slice) | |
1034 | { | |
1035 | VAAPIEncodeContext *ctx = avctx->priv_data; | |
1036 | VAEncPictureParameterBufferHEVC *vpic = pic->codec_picture_params; | |
1037 | VAEncSliceParameterBufferHEVC *vslice = slice->codec_slice_params; | |
1038 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
1039 | VAAPIEncodeH265Slice *pslice; | |
1040 | VAAPIEncodeH265MiscSliceParams *mslice; | |
1041 | int i; | |
1042 | ||
1043 | slice->priv_data = av_mallocz(sizeof(*pslice)); | |
1044 | if (!slice->priv_data) | |
1045 | return AVERROR(ENOMEM); | |
1046 | pslice = slice->priv_data; | |
1047 | mslice = &pslice->misc_slice_params; | |
1048 | ||
1049 | // Currently we only support one slice per frame. | |
1050 | vslice->slice_segment_address = 0; | |
1051 | vslice->num_ctu_in_slice = priv->ctu_width * priv->ctu_height; | |
1052 | ||
1053 | switch (pic->type) { | |
1054 | case PICTURE_TYPE_IDR: | |
1055 | case PICTURE_TYPE_I: | |
1056 | vslice->slice_type = I_SLICE; | |
1057 | break; | |
1058 | case PICTURE_TYPE_P: | |
1059 | vslice->slice_type = P_SLICE; | |
1060 | break; | |
1061 | case PICTURE_TYPE_B: | |
1062 | vslice->slice_type = B_SLICE; | |
1063 | break; | |
1064 | default: | |
1065 | av_assert0(0 && "invalid picture type"); | |
1066 | } | |
1067 | ||
1068 | vslice->slice_pic_parameter_set_id = vpic->slice_pic_parameter_set_id; | |
1069 | ||
1070 | pslice->pic_order_cnt = pic->display_order - priv->last_idr_frame; | |
1071 | ||
1072 | for (i = 0; i < FF_ARRAY_ELEMS(vslice->ref_pic_list0); i++) { | |
1073 | vslice->ref_pic_list0[i].picture_id = VA_INVALID_ID; | |
1074 | vslice->ref_pic_list0[i].flags = VA_PICTURE_HEVC_INVALID; | |
1075 | vslice->ref_pic_list1[i].picture_id = VA_INVALID_ID; | |
1076 | vslice->ref_pic_list1[i].flags = VA_PICTURE_HEVC_INVALID; | |
1077 | } | |
1078 | ||
1079 | av_assert0(pic->nb_refs <= 2); | |
1080 | if (pic->nb_refs >= 1) { | |
41ed7ab4 | 1081 | // Backward reference for P- or B-frame. |
31fe1f25 MT |
1082 | av_assert0(pic->type == PICTURE_TYPE_P || |
1083 | pic->type == PICTURE_TYPE_B); | |
1084 | ||
1085 | vslice->num_ref_idx_l0_active_minus1 = 0; | |
1086 | vslice->ref_pic_list0[0] = vpic->reference_frames[0]; | |
1087 | } | |
1088 | if (pic->nb_refs >= 2) { | |
41ed7ab4 | 1089 | // Forward reference for B-frame. |
31fe1f25 MT |
1090 | av_assert0(pic->type == PICTURE_TYPE_B); |
1091 | ||
1092 | vslice->num_ref_idx_l1_active_minus1 = 0; | |
1093 | vslice->ref_pic_list1[0] = vpic->reference_frames[1]; | |
1094 | } | |
1095 | ||
1096 | vslice->max_num_merge_cand = 5; | |
1097 | ||
1098 | if (pic->type == PICTURE_TYPE_B) | |
1099 | vslice->slice_qp_delta = priv->fixed_qp_b - vpic->pic_init_qp; | |
1100 | else if (pic->type == PICTURE_TYPE_P) | |
1101 | vslice->slice_qp_delta = priv->fixed_qp_p - vpic->pic_init_qp; | |
1102 | else | |
1103 | vslice->slice_qp_delta = priv->fixed_qp_idr - vpic->pic_init_qp; | |
1104 | ||
1105 | vslice->slice_fields.bits.last_slice_of_pic_flag = 1; | |
1106 | ||
1107 | mslice->first_slice_segment_in_pic_flag = 1; | |
1108 | ||
1109 | if (pic->type == PICTURE_TYPE_IDR) { | |
1110 | // No reference pictures. | |
1111 | } else if (0) { | |
1112 | mslice->short_term_ref_pic_set_sps_flag = 1; | |
1113 | mslice->short_term_ref_pic_idx = 0; | |
1114 | } else { | |
1115 | VAAPIEncodePicture *st; | |
1116 | int used; | |
1117 | ||
1118 | mslice->short_term_ref_pic_set_sps_flag = 0; | |
1119 | mslice->st_ref_pic_set.inter_ref_pic_set_prediction_flag = 0; | |
1120 | ||
1121 | for (st = ctx->pic_start; st; st = st->next) { | |
1122 | if (st->encode_order >= pic->encode_order) { | |
1123 | // Not yet in DPB. | |
1124 | continue; | |
1125 | } | |
1126 | used = 0; | |
1127 | for (i = 0; i < pic->nb_refs; i++) { | |
1128 | if (pic->refs[i] == st) | |
1129 | used = 1; | |
1130 | } | |
1131 | if (!used) { | |
1132 | // Currently true, but need not be. | |
1133 | continue; | |
1134 | } | |
1135 | // This only works for one instance of each (delta_poc_sN_minus1 | |
1136 | // is relative to the previous frame in the list, not relative to | |
1137 | // the current frame directly). | |
1138 | if (st->display_order < pic->display_order) { | |
1139 | i = mslice->st_ref_pic_set.num_negative_pics; | |
1140 | mslice->st_ref_pic_set.delta_poc_s0_minus1[i] = | |
1141 | pic->display_order - st->display_order - 1; | |
1142 | mslice->st_ref_pic_set.used_by_curr_pic_s0_flag[i] = used; | |
1143 | ++mslice->st_ref_pic_set.num_negative_pics; | |
1144 | } else { | |
1145 | i = mslice->st_ref_pic_set.num_positive_pics; | |
1146 | mslice->st_ref_pic_set.delta_poc_s1_minus1[i] = | |
1147 | st->display_order - pic->display_order - 1; | |
1148 | mslice->st_ref_pic_set.used_by_curr_pic_s1_flag[i] = used; | |
1149 | ++mslice->st_ref_pic_set.num_positive_pics; | |
1150 | } | |
1151 | } | |
1152 | } | |
1153 | ||
1154 | return 0; | |
1155 | } | |
1156 | ||
80a5d051 | 1157 | static av_cold int vaapi_encode_h265_configure(AVCodecContext *avctx) |
f70e4627 MT |
1158 | { |
1159 | VAAPIEncodeContext *ctx = avctx->priv_data; | |
1160 | VAAPIEncodeH265Context *priv = ctx->priv_data; | |
80a5d051 MT |
1161 | VAAPIEncodeH265Options *opt = ctx->codec_options; |
1162 | ||
1163 | priv->ctu_width = FFALIGN(ctx->surface_width, 32) / 32; | |
1164 | priv->ctu_height = FFALIGN(ctx->surface_height, 32) / 32; | |
1165 | ||
1166 | av_log(avctx, AV_LOG_VERBOSE, "Input %ux%u -> Surface %ux%u -> CTU %ux%u.\n", | |
1167 | avctx->width, avctx->height, ctx->surface_width, | |
1168 | ctx->surface_height, priv->ctu_width, priv->ctu_height); | |
1169 | ||
1170 | if (ctx->va_rc_mode == VA_RC_CQP) { | |
1171 | priv->fixed_qp_p = opt->qp; | |
1172 | if (avctx->i_quant_factor > 0.0) | |
1173 | priv->fixed_qp_idr = (int)((priv->fixed_qp_p * avctx->i_quant_factor + | |
1174 | avctx->i_quant_offset) + 0.5); | |
1175 | else | |
1176 | priv->fixed_qp_idr = priv->fixed_qp_p; | |
1177 | if (avctx->b_quant_factor > 0.0) | |
1178 | priv->fixed_qp_b = (int)((priv->fixed_qp_p * avctx->b_quant_factor + | |
1179 | avctx->b_quant_offset) + 0.5); | |
1180 | else | |
1181 | priv->fixed_qp_b = priv->fixed_qp_p; | |
1182 | ||
1183 | av_log(avctx, AV_LOG_DEBUG, "Using fixed QP = " | |
1184 | "%d / %d / %d for IDR- / P- / B-frames.\n", | |
1185 | priv->fixed_qp_idr, priv->fixed_qp_p, priv->fixed_qp_b); | |
1186 | ||
1187 | } else if (ctx->va_rc_mode == VA_RC_CBR) { | |
1188 | // These still need to be set for pic_init_qp/slice_qp_delta. | |
1189 | priv->fixed_qp_idr = 30; | |
1190 | priv->fixed_qp_p = 30; | |
1191 | priv->fixed_qp_b = 30; | |
1192 | ||
1193 | av_log(avctx, AV_LOG_DEBUG, "Using constant-bitrate = %d bps.\n", | |
1194 | avctx->bit_rate); | |
1195 | ||
1196 | } else { | |
1197 | av_assert0(0 && "Invalid RC mode."); | |
1198 | } | |
f70e4627 | 1199 | |
f70e4627 MT |
1200 | return 0; |
1201 | } | |
1202 | ||
80a5d051 MT |
1203 | static const VAAPIEncodeType vaapi_encode_type_h265 = { |
1204 | .priv_data_size = sizeof(VAAPIEncodeH265Context), | |
f70e4627 | 1205 | |
80a5d051 | 1206 | .configure = &vaapi_encode_h265_configure, |
f70e4627 | 1207 | |
80a5d051 MT |
1208 | .sequence_params_size = sizeof(VAEncSequenceParameterBufferHEVC), |
1209 | .init_sequence_params = &vaapi_encode_h265_init_sequence_params, | |
f70e4627 | 1210 | |
80a5d051 MT |
1211 | .picture_params_size = sizeof(VAEncPictureParameterBufferHEVC), |
1212 | .init_picture_params = &vaapi_encode_h265_init_picture_params, | |
1213 | ||
1214 | .slice_params_size = sizeof(VAEncSliceParameterBufferHEVC), | |
1215 | .init_slice_params = &vaapi_encode_h265_init_slice_params, | |
1216 | ||
1217 | .sequence_header_type = VAEncPackedHeaderSequence, | |
1218 | .write_sequence_header = &vaapi_encode_h265_write_sequence_header, | |
1219 | ||
1220 | .slice_header_type = VAEncPackedHeaderHEVC_Slice, | |
1221 | .write_slice_header = &vaapi_encode_h265_write_slice_header, | |
1222 | }; | |
1223 | ||
1224 | static av_cold int vaapi_encode_h265_init(AVCodecContext *avctx) | |
31fe1f25 | 1225 | { |
80a5d051 | 1226 | VAAPIEncodeContext *ctx = avctx->priv_data; |
f6b85523 | 1227 | |
80a5d051 | 1228 | ctx->codec = &vaapi_encode_type_h265; |
31fe1f25 MT |
1229 | |
1230 | switch (avctx->profile) { | |
1231 | case FF_PROFILE_HEVC_MAIN: | |
1232 | case FF_PROFILE_UNKNOWN: | |
1233 | ctx->va_profile = VAProfileHEVCMain; | |
1234 | break; | |
1235 | case FF_PROFILE_HEVC_MAIN_10: | |
1236 | av_log(avctx, AV_LOG_ERROR, "H.265 main 10-bit profile " | |
1237 | "is not supported.\n"); | |
1238 | return AVERROR_PATCHWELCOME; | |
1239 | default: | |
1240 | av_log(avctx, AV_LOG_ERROR, "Unknown H.265 profile %d.\n", | |
1241 | avctx->profile); | |
1242 | return AVERROR(EINVAL); | |
1243 | } | |
1244 | ctx->va_entrypoint = VAEntrypointEncSlice; | |
1245 | ||
80a5d051 MT |
1246 | // This will be dependent on profile when 10-bit is supported. |
1247 | ctx->va_rt_format = VA_RT_FORMAT_YUV420; | |
31fe1f25 | 1248 | |
80a5d051 | 1249 | if (avctx->bit_rate > 0) |
f70e4627 | 1250 | ctx->va_rc_mode = VA_RC_CBR; |
80a5d051 | 1251 | else |
f70e4627 | 1252 | ctx->va_rc_mode = VA_RC_CQP; |
f6b85523 | 1253 | |
892bbbcd MT |
1254 | ctx->va_packed_headers = |
1255 | VA_ENC_PACKED_HEADER_SEQUENCE | // VPS, SPS and PPS. | |
1256 | VA_ENC_PACKED_HEADER_SLICE; // Slice headers. | |
31fe1f25 | 1257 | |
80a5d051 MT |
1258 | ctx->surface_width = FFALIGN(avctx->width, 16); |
1259 | ctx->surface_height = FFALIGN(avctx->height, 16); | |
31fe1f25 | 1260 | |
80a5d051 | 1261 | return ff_vaapi_encode_init(avctx); |
31fe1f25 MT |
1262 | } |
1263 | ||
92fdea37 MT |
1264 | #define OFFSET(x) (offsetof(VAAPIEncodeContext, codec_options_data) + \ |
1265 | offsetof(VAAPIEncodeH265Options, x)) | |
1266 | #define FLAGS (AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM) | |
1267 | static const AVOption vaapi_encode_h265_options[] = { | |
41ed7ab4 | 1268 | { "qp", "Constant QP (for P-frames; scaled by qfactor/qoffset for I/B)", |
92fdea37 MT |
1269 | OFFSET(qp), AV_OPT_TYPE_INT, { .i64 = 25 }, 0, 52, FLAGS }, |
1270 | { NULL }, | |
1271 | }; | |
1272 | ||
31fe1f25 MT |
1273 | static const AVCodecDefault vaapi_encode_h265_defaults[] = { |
1274 | { "profile", "1" }, | |
1275 | { "level", "51" }, | |
1276 | { "b", "0" }, | |
1277 | { "bf", "2" }, | |
1278 | { "g", "120" }, | |
31fe1f25 MT |
1279 | { "i_qfactor", "1.0" }, |
1280 | { "i_qoffset", "0.0" }, | |
1281 | { "b_qfactor", "1.2" }, | |
1282 | { "b_qoffset", "0.0" }, | |
1283 | { NULL }, | |
1284 | }; | |
1285 | ||
1286 | static const AVClass vaapi_encode_h265_class = { | |
1287 | .class_name = "h265_vaapi", | |
1288 | .item_name = av_default_item_name, | |
92fdea37 | 1289 | .option = vaapi_encode_h265_options, |
31fe1f25 MT |
1290 | .version = LIBAVUTIL_VERSION_INT, |
1291 | }; | |
1292 | ||
1293 | AVCodec ff_hevc_vaapi_encoder = { | |
1294 | .name = "hevc_vaapi", | |
1295 | .long_name = NULL_IF_CONFIG_SMALL("H.265/HEVC (VAAPI)"), | |
1296 | .type = AVMEDIA_TYPE_VIDEO, | |
1297 | .id = AV_CODEC_ID_HEVC, | |
92fdea37 MT |
1298 | .priv_data_size = (sizeof(VAAPIEncodeContext) + |
1299 | sizeof(VAAPIEncodeH265Options)), | |
31fe1f25 MT |
1300 | .init = &vaapi_encode_h265_init, |
1301 | .encode2 = &ff_vaapi_encode2, | |
1302 | .close = &ff_vaapi_encode_close, | |
1303 | .priv_class = &vaapi_encode_h265_class, | |
1304 | .capabilities = AV_CODEC_CAP_DELAY, | |
1305 | .defaults = vaapi_encode_h265_defaults, | |
1306 | .pix_fmts = (const enum AVPixelFormat[]) { | |
1307 | AV_PIX_FMT_VAAPI, | |
1308 | AV_PIX_FMT_NONE, | |
1309 | }, | |
1310 | }; |