1 /**************************************************************************
2 *
3 * Copyright 2017 Advanced Micro Devices, Inc.
4 *
5 * SPDX-License-Identifier: MIT
6 *
7 **************************************************************************/
8
9 #include "radeon_vcn_dec.h"
10
11 #include "pipe/p_video_codec.h"
12 #include "radeonsi/si_pipe.h"
13 #include "util/u_memory.h"
14 #include "util/u_video.h"
15 #include "vl/vl_mpeg12_decoder.h"
16 #include "vl/vl_probs_table.h"
17 #include "pspdecryptionparam.h"
18
19 #include <assert.h>
20 #include <stdio.h>
21
22 #include "ac_vcn_av1_default.h"
23 #include "ac_debug.h"
24
25 #define FB_BUFFER_OFFSET 0x2000
26 #define FB_BUFFER_SIZE 2048
27 #define IT_SCALING_TABLE_SIZE 992
28 #define VP9_PROBS_TABLE_SIZE (RDECODE_VP9_PROBS_DATA_SIZE + 256)
29
30 #define NUM_MPEG2_REFS 6
31 #define NUM_H264_REFS 17
32 #define NUM_VC1_REFS 5
33 #define NUM_VP9_REFS 8
34 #define NUM_AV1_REFS 8
35 #define NUM_AV1_REFS_PER_FRAME 7
36
37 static enum pipe_format get_buffer_format(struct radeon_decoder *dec);
38 static unsigned calc_dpb_size(struct radeon_decoder *dec);
39 static unsigned calc_ctx_size_h264_perf(struct radeon_decoder *dec);
40 static unsigned calc_ctx_size_h265_main(struct radeon_decoder *dec);
41 static unsigned calc_ctx_size_h265_main10(struct radeon_decoder *dec,
42 struct pipe_h265_picture_desc *pic);
43
radeon_dec_destroy_associated_data(void * data)44 static void radeon_dec_destroy_associated_data(void *data)
45 {
46 /* NOOP, since we only use an intptr */
47 }
48
get_current_pic_index(struct radeon_decoder * dec,struct pipe_video_buffer * target,unsigned char * curr_pic_idx)49 static void get_current_pic_index(struct radeon_decoder *dec,
50 struct pipe_video_buffer *target,
51 unsigned char *curr_pic_idx)
52 {
53 for (int i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
54 if (dec->render_pic_list[i] && dec->render_pic_list[i] == target) {
55 if (target->codec != NULL)
56 *curr_pic_idx = (uintptr_t)vl_video_buffer_get_associated_data(target, &dec->base);
57 else {
58 *curr_pic_idx = i;
59 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
60 &radeon_dec_destroy_associated_data);
61 }
62 break;
63 } else if (!dec->render_pic_list[i]) {
64 dec->render_pic_list[i] = target;
65 *curr_pic_idx = i;
66 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
67 &radeon_dec_destroy_associated_data);
68 break;
69 }
70 }
71 }
72
get_h264_msg(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_h264_picture_desc * pic)73 static rvcn_dec_message_avc_t get_h264_msg(struct radeon_decoder *dec,
74 struct pipe_video_buffer *target,
75 struct pipe_h264_picture_desc *pic)
76 {
77 rvcn_dec_message_avc_t result;
78 unsigned i, j, k;
79
80 memset(&result, 0, sizeof(result));
81 switch (pic->base.profile) {
82 case PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE:
83 case PIPE_VIDEO_PROFILE_MPEG4_AVC_CONSTRAINED_BASELINE:
84 result.profile = RDECODE_H264_PROFILE_BASELINE;
85 break;
86
87 case PIPE_VIDEO_PROFILE_MPEG4_AVC_MAIN:
88 result.profile = RDECODE_H264_PROFILE_MAIN;
89 break;
90
91 case PIPE_VIDEO_PROFILE_MPEG4_AVC_HIGH:
92 result.profile = RDECODE_H264_PROFILE_HIGH;
93 break;
94
95 default:
96 assert(0);
97 break;
98 }
99
100 result.level = dec->base.level;
101
102 result.sps_info_flags = 0;
103 result.sps_info_flags |= pic->pps->sps->direct_8x8_inference_flag << 0;
104 result.sps_info_flags |= pic->pps->sps->mb_adaptive_frame_field_flag << 1;
105 result.sps_info_flags |= pic->pps->sps->frame_mbs_only_flag << 2;
106 result.sps_info_flags |= pic->pps->sps->delta_pic_order_always_zero_flag << 3;
107 result.sps_info_flags |= ((dec->dpb_type == DPB_DYNAMIC_TIER_2) ? 0 : 1)
108 << RDECODE_SPS_INFO_H264_EXTENSION_SUPPORT_FLAG_SHIFT;
109
110 result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
111 result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
112 result.log2_max_frame_num_minus4 = pic->pps->sps->log2_max_frame_num_minus4;
113 result.pic_order_cnt_type = pic->pps->sps->pic_order_cnt_type;
114 result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
115
116 switch (dec->base.chroma_format) {
117 case PIPE_VIDEO_CHROMA_FORMAT_NONE:
118 break;
119 case PIPE_VIDEO_CHROMA_FORMAT_400:
120 result.chroma_format = 0;
121 break;
122 case PIPE_VIDEO_CHROMA_FORMAT_420:
123 result.chroma_format = 1;
124 break;
125 case PIPE_VIDEO_CHROMA_FORMAT_422:
126 result.chroma_format = 2;
127 break;
128 case PIPE_VIDEO_CHROMA_FORMAT_444:
129 result.chroma_format = 3;
130 break;
131 case PIPE_VIDEO_CHROMA_FORMAT_440:
132 result.chroma_format = 4;
133 break;
134 }
135
136 result.pps_info_flags = 0;
137 result.pps_info_flags |= pic->pps->transform_8x8_mode_flag << 0;
138 result.pps_info_flags |= pic->pps->redundant_pic_cnt_present_flag << 1;
139 result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 2;
140 result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag << 3;
141 result.pps_info_flags |= pic->pps->weighted_bipred_idc << 4;
142 result.pps_info_flags |= pic->pps->weighted_pred_flag << 6;
143 result.pps_info_flags |= pic->pps->bottom_field_pic_order_in_frame_present_flag << 7;
144 result.pps_info_flags |= pic->pps->entropy_coding_mode_flag << 8;
145
146 result.num_slice_groups_minus1 = pic->pps->num_slice_groups_minus1;
147 result.slice_group_map_type = pic->pps->slice_group_map_type;
148 result.slice_group_change_rate_minus1 = pic->pps->slice_group_change_rate_minus1;
149 result.pic_init_qp_minus26 = pic->pps->pic_init_qp_minus26;
150 result.chroma_qp_index_offset = pic->pps->chroma_qp_index_offset;
151 result.second_chroma_qp_index_offset = pic->pps->second_chroma_qp_index_offset;
152
153 memcpy(result.scaling_list_4x4, pic->pps->ScalingList4x4, 6 * 16);
154 memcpy(result.scaling_list_8x8, pic->pps->ScalingList8x8, 2 * 64);
155
156 memcpy(dec->it, result.scaling_list_4x4, 6 * 16);
157 memcpy((dec->it + 96), result.scaling_list_8x8, 2 * 64);
158
159 result.num_ref_frames = pic->num_ref_frames;
160
161 result.num_ref_idx_l0_active_minus1 = pic->num_ref_idx_l0_active_minus1;
162 result.num_ref_idx_l1_active_minus1 = pic->num_ref_idx_l1_active_minus1;
163
164 result.frame_num = pic->frame_num;
165 memcpy(result.frame_num_list, pic->frame_num_list, 4 * 16);
166 result.curr_field_order_cnt_list[0] = pic->field_order_cnt[0];
167 result.curr_field_order_cnt_list[1] = pic->field_order_cnt[1];
168 memcpy(result.field_order_cnt_list, pic->field_order_cnt_list, 4 * 16 * 2);
169 result.non_existing_frame_flags = 0;
170 result.used_for_reference_flags = 0;
171
172 if (dec->dpb_type != DPB_DYNAMIC_TIER_2) {
173 result.decoded_pic_idx = pic->frame_num;
174 goto end;
175 }
176
177 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); i++) {
178 for (j = 0; (j < ARRAY_SIZE(pic->ref) && (pic->ref[j] != NULL)); j++) {
179 if (dec->render_pic_list[i] == pic->ref[j])
180 break;
181 if (j == ARRAY_SIZE(pic->ref) - 1)
182 dec->render_pic_list[i] = NULL;
183 else if (pic->ref[j + 1] == NULL)
184 dec->render_pic_list[i] = NULL;
185 }
186 }
187 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
188 if (dec->render_pic_list[i] && dec->render_pic_list[i] == target) {
189 if (target->codec != NULL){
190 result.decoded_pic_idx =
191 (uintptr_t)vl_video_buffer_get_associated_data(target, &dec->base);
192 } else {
193 result.decoded_pic_idx = i;
194 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
195 &radeon_dec_destroy_associated_data);
196 }
197 break;
198 }
199 }
200 if (i == ARRAY_SIZE(dec->render_pic_list)) {
201 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
202 if (!dec->render_pic_list[i]) {
203 dec->render_pic_list[i] = target;
204 result.decoded_pic_idx = i;
205 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
206 &radeon_dec_destroy_associated_data);
207 break;
208 }
209 }
210 }
211 for (i = 0; i < ARRAY_SIZE(result.ref_frame_list); i++) {
212 result.ref_frame_list[i] = pic->ref[i] ?
213 (uintptr_t)vl_video_buffer_get_associated_data(pic->ref[i], &dec->base) : 0xff;
214
215 if (result.ref_frame_list[i] != 0xff) {
216 if (pic->top_is_reference[i])
217 result.used_for_reference_flags |= (1 << (2 * i));
218 if (pic->bottom_is_reference[i])
219 result.used_for_reference_flags |= (1 << (2 * i + 1));
220
221 if (pic->is_long_term[i])
222 result.ref_frame_list[i] |= 0x80;
223
224 result.curr_pic_ref_frame_num++;
225
226 for (j = 0; j < ARRAY_SIZE(dec->h264_valid_ref_num); j++) {
227 if ((dec->h264_valid_ref_num[j] != (unsigned)-1)
228 && (dec->h264_valid_ref_num[j] == result.frame_num_list[i]))
229 break;
230 }
231
232 for (k = 0; k < ARRAY_SIZE(dec->h264_valid_poc_num); k++) {
233 if ((dec->h264_valid_poc_num[k] != (unsigned)-1)
234 && ((dec->h264_valid_poc_num[k] == result.field_order_cnt_list[i][0])
235 || dec->h264_valid_poc_num[k] == result.field_order_cnt_list[i][1]))
236 break;
237 }
238 if ((j == ARRAY_SIZE(dec->h264_valid_ref_num)) && (k == ARRAY_SIZE(dec->h264_valid_poc_num))) {
239 result.non_existing_frame_flags |= 1 << i;
240 result.curr_pic_ref_frame_num--;
241 result.ref_frame_list[i] = 0xff;
242 }
243 }
244 }
245
246 /* need at least one reference for P/B frames */
247 if (result.curr_pic_ref_frame_num == 0 && pic->slice_parameter.slice_info_present) {
248 for (i = 0; i < pic->slice_count; i++) {
249 if (pic->slice_parameter.slice_type[i] % 5 != 2) {
250 result.curr_pic_ref_frame_num++;
251 result.ref_frame_list[0] = 0;
252 result.non_existing_frame_flags &= ~1;
253 break;
254 }
255 }
256 }
257
258 for (i = 0; i < ARRAY_SIZE(result.ref_frame_list); i++) {
259 if (result.ref_frame_list[i] != 0xff) {
260 dec->h264_valid_ref_num[i] = result.frame_num_list[i];
261 dec->h264_valid_poc_num[2 * i] = pic->top_is_reference[i] ?
262 result.field_order_cnt_list[i][0] : (unsigned) -1;
263 dec->h264_valid_poc_num[2 * i + 1] = pic->bottom_is_reference[i] ?
264 result.field_order_cnt_list[i][1] : (unsigned) -1;
265 } else {
266 dec->h264_valid_ref_num[i] =
267 dec->h264_valid_poc_num[2 * i] =
268 dec->h264_valid_poc_num[2 * i + 1] = (unsigned) -1;
269 }
270 }
271
272 dec->h264_valid_ref_num[ARRAY_SIZE(dec->h264_valid_ref_num) - 1] = result.frame_num;
273 dec->h264_valid_poc_num[ARRAY_SIZE(dec->h264_valid_poc_num) - 2] =
274 pic->field_pic_flag && pic->bottom_field_flag ?
275 (unsigned) -1 : result.curr_field_order_cnt_list[0];
276 dec->h264_valid_poc_num[ARRAY_SIZE(dec->h264_valid_poc_num) - 1] =
277 pic->field_pic_flag && !pic->bottom_field_flag ?
278 (unsigned) -1 : result.curr_field_order_cnt_list[1];
279
280 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
281 dec->ref_codec.bts = CODEC_8_BITS;
282 dec->ref_codec.index = result.decoded_pic_idx;
283 dec->ref_codec.ref_size = 16;
284 dec->ref_codec.num_refs = result.curr_pic_ref_frame_num;
285 STATIC_ASSERT(sizeof(dec->ref_codec.ref_list) == sizeof(result.ref_frame_list));
286 memcpy(dec->ref_codec.ref_list, result.ref_frame_list, sizeof(result.ref_frame_list));
287 }
288
289 end:
290 return result;
291 }
292
get_h265_msg(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_h265_picture_desc * pic)293 static rvcn_dec_message_hevc_t get_h265_msg(struct radeon_decoder *dec,
294 struct pipe_video_buffer *target,
295 struct pipe_h265_picture_desc *pic)
296 {
297 rvcn_dec_message_hevc_t result;
298 unsigned i, j, num_refs = 0;
299
300 memset(&result, 0, sizeof(result));
301 result.sps_info_flags = 0;
302 result.sps_info_flags |= pic->pps->sps->scaling_list_enabled_flag << 0;
303 result.sps_info_flags |= pic->pps->sps->amp_enabled_flag << 1;
304 result.sps_info_flags |= pic->pps->sps->sample_adaptive_offset_enabled_flag << 2;
305 result.sps_info_flags |= pic->pps->sps->pcm_enabled_flag << 3;
306 result.sps_info_flags |= pic->pps->sps->pcm_loop_filter_disabled_flag << 4;
307 result.sps_info_flags |= pic->pps->sps->long_term_ref_pics_present_flag << 5;
308 result.sps_info_flags |= pic->pps->sps->sps_temporal_mvp_enabled_flag << 6;
309 result.sps_info_flags |= pic->pps->sps->strong_intra_smoothing_enabled_flag << 7;
310 result.sps_info_flags |= pic->pps->sps->separate_colour_plane_flag << 8;
311 if (((struct si_screen *)dec->screen)->info.family == CHIP_CARRIZO)
312 result.sps_info_flags |= 1 << 9;
313 if (pic->UseRefPicList == true) {
314 result.sps_info_flags |= 1 << 10;
315 result.sps_info_flags |= 1 << 12;
316 }
317 if (pic->UseStRpsBits == true && pic->pps->st_rps_bits != 0) {
318 result.sps_info_flags |= 1 << 11;
319 result.st_rps_bits = pic->pps->st_rps_bits;
320 }
321
322 result.chroma_format = pic->pps->sps->chroma_format_idc;
323 result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
324 result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
325 result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
326 result.sps_max_dec_pic_buffering_minus1 = pic->pps->sps->sps_max_dec_pic_buffering_minus1;
327 result.log2_min_luma_coding_block_size_minus3 =
328 pic->pps->sps->log2_min_luma_coding_block_size_minus3;
329 result.log2_diff_max_min_luma_coding_block_size =
330 pic->pps->sps->log2_diff_max_min_luma_coding_block_size;
331 result.log2_min_transform_block_size_minus2 =
332 pic->pps->sps->log2_min_transform_block_size_minus2;
333 result.log2_diff_max_min_transform_block_size =
334 pic->pps->sps->log2_diff_max_min_transform_block_size;
335 result.max_transform_hierarchy_depth_inter = pic->pps->sps->max_transform_hierarchy_depth_inter;
336 result.max_transform_hierarchy_depth_intra = pic->pps->sps->max_transform_hierarchy_depth_intra;
337 result.pcm_sample_bit_depth_luma_minus1 = pic->pps->sps->pcm_sample_bit_depth_luma_minus1;
338 result.pcm_sample_bit_depth_chroma_minus1 = pic->pps->sps->pcm_sample_bit_depth_chroma_minus1;
339 result.log2_min_pcm_luma_coding_block_size_minus3 =
340 pic->pps->sps->log2_min_pcm_luma_coding_block_size_minus3;
341 result.log2_diff_max_min_pcm_luma_coding_block_size =
342 pic->pps->sps->log2_diff_max_min_pcm_luma_coding_block_size;
343 result.num_short_term_ref_pic_sets = pic->pps->sps->num_short_term_ref_pic_sets;
344
345 result.pps_info_flags = 0;
346 result.pps_info_flags |= pic->pps->dependent_slice_segments_enabled_flag << 0;
347 result.pps_info_flags |= pic->pps->output_flag_present_flag << 1;
348 result.pps_info_flags |= pic->pps->sign_data_hiding_enabled_flag << 2;
349 result.pps_info_flags |= pic->pps->cabac_init_present_flag << 3;
350 result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 4;
351 result.pps_info_flags |= pic->pps->transform_skip_enabled_flag << 5;
352 result.pps_info_flags |= pic->pps->cu_qp_delta_enabled_flag << 6;
353 result.pps_info_flags |= pic->pps->pps_slice_chroma_qp_offsets_present_flag << 7;
354 result.pps_info_flags |= pic->pps->weighted_pred_flag << 8;
355 result.pps_info_flags |= pic->pps->weighted_bipred_flag << 9;
356 result.pps_info_flags |= pic->pps->transquant_bypass_enabled_flag << 10;
357 result.pps_info_flags |= pic->pps->tiles_enabled_flag << 11;
358 result.pps_info_flags |= pic->pps->entropy_coding_sync_enabled_flag << 12;
359 result.pps_info_flags |= pic->pps->uniform_spacing_flag << 13;
360 result.pps_info_flags |= pic->pps->loop_filter_across_tiles_enabled_flag << 14;
361 result.pps_info_flags |= pic->pps->pps_loop_filter_across_slices_enabled_flag << 15;
362 result.pps_info_flags |= pic->pps->deblocking_filter_override_enabled_flag << 16;
363 result.pps_info_flags |= pic->pps->pps_deblocking_filter_disabled_flag << 17;
364 result.pps_info_flags |= pic->pps->lists_modification_present_flag << 18;
365 result.pps_info_flags |= pic->pps->slice_segment_header_extension_present_flag << 19;
366
367 result.num_extra_slice_header_bits = pic->pps->num_extra_slice_header_bits;
368 result.num_long_term_ref_pic_sps = pic->pps->sps->num_long_term_ref_pics_sps;
369 result.num_ref_idx_l0_default_active_minus1 = pic->pps->num_ref_idx_l0_default_active_minus1;
370 result.num_ref_idx_l1_default_active_minus1 = pic->pps->num_ref_idx_l1_default_active_minus1;
371 result.pps_cb_qp_offset = pic->pps->pps_cb_qp_offset;
372 result.pps_cr_qp_offset = pic->pps->pps_cr_qp_offset;
373 result.pps_beta_offset_div2 = pic->pps->pps_beta_offset_div2;
374 result.pps_tc_offset_div2 = pic->pps->pps_tc_offset_div2;
375 result.diff_cu_qp_delta_depth = pic->pps->diff_cu_qp_delta_depth;
376 result.num_tile_columns_minus1 = pic->pps->num_tile_columns_minus1;
377 result.num_tile_rows_minus1 = pic->pps->num_tile_rows_minus1;
378 result.log2_parallel_merge_level_minus2 = pic->pps->log2_parallel_merge_level_minus2;
379 result.init_qp_minus26 = pic->pps->init_qp_minus26;
380
381 for (i = 0; i < 19; ++i)
382 result.column_width_minus1[i] = pic->pps->column_width_minus1[i];
383
384 for (i = 0; i < 21; ++i)
385 result.row_height_minus1[i] = pic->pps->row_height_minus1[i];
386
387 result.num_delta_pocs_ref_rps_idx = pic->NumDeltaPocsOfRefRpsIdx;
388 result.curr_poc = pic->CurrPicOrderCntVal;
389
390 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); i++) {
391 for (j = 0;
392 (j < ARRAY_SIZE(pic->ref) && (pic->ref[j] != NULL));
393 j++) {
394 if (dec->render_pic_list[i] == pic->ref[j])
395 break;
396 if (j == ARRAY_SIZE(pic->ref) - 1)
397 dec->render_pic_list[i] = NULL;
398 else if (pic->ref[j + 1] == NULL)
399 dec->render_pic_list[i] = NULL;
400 }
401 }
402 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); i++) {
403 if (dec->render_pic_list[i] == NULL) {
404 dec->render_pic_list[i] = target;
405 result.curr_idx = i;
406 break;
407 }
408 }
409
410 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)result.curr_idx,
411 &radeon_dec_destroy_associated_data);
412
413 for (i = 0; i < 16; ++i) {
414 struct pipe_video_buffer *ref = pic->ref[i];
415 uintptr_t ref_pic = 0;
416
417 result.poc_list[i] = pic->PicOrderCntVal[i];
418
419 if (ref) {
420 ref_pic = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
421 num_refs++;
422 } else
423 ref_pic = 0x7F;
424 result.ref_pic_list[i] = ref_pic;
425 }
426
427 for (i = 0; i < 8; ++i) {
428 result.ref_pic_set_st_curr_before[i] = 0xFF;
429 result.ref_pic_set_st_curr_after[i] = 0xFF;
430 result.ref_pic_set_lt_curr[i] = 0xFF;
431 }
432
433 for (i = 0; i < pic->NumPocStCurrBefore; ++i)
434 result.ref_pic_set_st_curr_before[i] = pic->RefPicSetStCurrBefore[i];
435
436 for (i = 0; i < pic->NumPocStCurrAfter; ++i)
437 result.ref_pic_set_st_curr_after[i] = pic->RefPicSetStCurrAfter[i];
438
439 for (i = 0; i < pic->NumPocLtCurr; ++i)
440 result.ref_pic_set_lt_curr[i] = pic->RefPicSetLtCurr[i];
441
442 for (i = 0; i < 6; ++i)
443 result.ucScalingListDCCoefSizeID2[i] = pic->pps->sps->ScalingListDCCoeff16x16[i];
444
445 for (i = 0; i < 2; ++i)
446 result.ucScalingListDCCoefSizeID3[i] = pic->pps->sps->ScalingListDCCoeff32x32[i];
447
448 memcpy(dec->it, pic->pps->sps->ScalingList4x4, 6 * 16);
449 memcpy(dec->it + 96, pic->pps->sps->ScalingList8x8, 6 * 64);
450 memcpy(dec->it + 480, pic->pps->sps->ScalingList16x16, 6 * 64);
451 memcpy(dec->it + 864, pic->pps->sps->ScalingList32x32, 2 * 64);
452
453 for (i = 0; i < 2; i++) {
454 for (j = 0; j < 15; j++)
455 result.direct_reflist[i][j] = pic->RefPicList[0][i][j];
456 }
457
458 if (pic->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10) {
459 if (target->buffer_format == PIPE_FORMAT_P010 || target->buffer_format == PIPE_FORMAT_P016) {
460 result.p010_mode = 1;
461 result.msb_mode = 1;
462 } else {
463 result.p010_mode = 0;
464 result.luma_10to8 = 5;
465 result.chroma_10to8 = 5;
466 result.hevc_reserved[0] = 4; /* sclr_luma10to8 */
467 result.hevc_reserved[1] = 4; /* sclr_chroma10to8 */
468 }
469 }
470
471 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
472 dec->ref_codec.bts = (pic->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10) ?
473 CODEC_10_BITS : CODEC_8_BITS;
474 dec->ref_codec.index = result.curr_idx;
475 dec->ref_codec.ref_size = 15;
476 dec->ref_codec.num_refs = num_refs;
477 STATIC_ASSERT(sizeof(dec->ref_codec.ref_list) == sizeof(result.ref_pic_list));
478 memcpy(dec->ref_codec.ref_list, result.ref_pic_list, sizeof(result.ref_pic_list));
479 }
480 return result;
481 }
482
fill_probs_table(void * ptr)483 static void fill_probs_table(void *ptr)
484 {
485 rvcn_dec_vp9_probs_t *probs = (rvcn_dec_vp9_probs_t *)ptr;
486
487 memcpy(&probs->coef_probs[0], default_coef_probs_4x4, sizeof(default_coef_probs_4x4));
488 memcpy(&probs->coef_probs[1], default_coef_probs_8x8, sizeof(default_coef_probs_8x8));
489 memcpy(&probs->coef_probs[2], default_coef_probs_16x16, sizeof(default_coef_probs_16x16));
490 memcpy(&probs->coef_probs[3], default_coef_probs_32x32, sizeof(default_coef_probs_32x32));
491 memcpy(probs->y_mode_prob, default_if_y_probs, sizeof(default_if_y_probs));
492 memcpy(probs->uv_mode_prob, default_if_uv_probs, sizeof(default_if_uv_probs));
493 memcpy(probs->single_ref_prob, default_single_ref_p, sizeof(default_single_ref_p));
494 memcpy(probs->switchable_interp_prob, default_switchable_interp_prob,
495 sizeof(default_switchable_interp_prob));
496 memcpy(probs->partition_prob, default_partition_probs, sizeof(default_partition_probs));
497 memcpy(probs->inter_mode_probs, default_inter_mode_probs, sizeof(default_inter_mode_probs));
498 memcpy(probs->mbskip_probs, default_skip_probs, sizeof(default_skip_probs));
499 memcpy(probs->intra_inter_prob, default_intra_inter_p, sizeof(default_intra_inter_p));
500 memcpy(probs->comp_inter_prob, default_comp_inter_p, sizeof(default_comp_inter_p));
501 memcpy(probs->comp_ref_prob, default_comp_ref_p, sizeof(default_comp_ref_p));
502 memcpy(probs->tx_probs_32x32, default_tx_probs_32x32, sizeof(default_tx_probs_32x32));
503 memcpy(probs->tx_probs_16x16, default_tx_probs_16x16, sizeof(default_tx_probs_16x16));
504 memcpy(probs->tx_probs_8x8, default_tx_probs_8x8, sizeof(default_tx_probs_8x8));
505 memcpy(probs->mv_joints, default_nmv_joints, sizeof(default_nmv_joints));
506 memcpy(&probs->mv_comps[0], default_nmv_components, sizeof(default_nmv_components));
507 memset(&probs->nmvc_mask, 0, sizeof(rvcn_dec_vp9_nmv_ctx_mask_t));
508 }
509
get_vp9_msg(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_vp9_picture_desc * pic)510 static rvcn_dec_message_vp9_t get_vp9_msg(struct radeon_decoder *dec,
511 struct pipe_video_buffer *target,
512 struct pipe_vp9_picture_desc *pic)
513 {
514 rvcn_dec_message_vp9_t result;
515 unsigned i, j, num_refs = 0;
516
517 memset(&result, 0, sizeof(result));
518
519 /* segment table */
520 rvcn_dec_vp9_probs_segment_t *prbs = (rvcn_dec_vp9_probs_segment_t *)(dec->probs);
521
522 if (pic->picture_parameter.pic_fields.segmentation_enabled) {
523 for (i = 0; i < 8; ++i) {
524 prbs->seg.feature_data[i] =
525 (pic->slice_parameter.seg_param[i].alt_quant & 0xffff) |
526 ((pic->slice_parameter.seg_param[i].alt_lf & 0xff) << 16) |
527 ((pic->slice_parameter.seg_param[i].segment_flags.segment_reference & 0xf) << 24);
528 prbs->seg.feature_mask[i] =
529 (pic->slice_parameter.seg_param[i].alt_quant_enabled << 0) |
530 (pic->slice_parameter.seg_param[i].alt_lf_enabled << 1) |
531 (pic->slice_parameter.seg_param[i].segment_flags.segment_reference_enabled << 2) |
532 (pic->slice_parameter.seg_param[i].segment_flags.segment_reference_skipped << 3);
533 }
534
535 for (i = 0; i < 7; ++i)
536 prbs->seg.tree_probs[i] = pic->picture_parameter.mb_segment_tree_probs[i];
537
538 for (i = 0; i < 3; ++i)
539 prbs->seg.pred_probs[i] = pic->picture_parameter.segment_pred_probs[i];
540
541 prbs->seg.abs_delta = pic->picture_parameter.abs_delta;
542 } else
543 memset(prbs->segment_data, 0, sizeof(prbs->segment_data));
544
545 result.frame_header_flags = (pic->picture_parameter.pic_fields.frame_type
546 << RDECODE_FRAME_HDR_INFO_VP9_FRAME_TYPE_SHIFT) &
547 RDECODE_FRAME_HDR_INFO_VP9_FRAME_TYPE_MASK;
548
549 result.frame_header_flags |= (pic->picture_parameter.pic_fields.error_resilient_mode
550 << RDECODE_FRAME_HDR_INFO_VP9_ERROR_RESILIENT_MODE_SHIFT) &
551 RDECODE_FRAME_HDR_INFO_VP9_ERROR_RESILIENT_MODE_MASK;
552
553 result.frame_header_flags |= (pic->picture_parameter.pic_fields.intra_only
554 << RDECODE_FRAME_HDR_INFO_VP9_INTRA_ONLY_SHIFT) &
555 RDECODE_FRAME_HDR_INFO_VP9_INTRA_ONLY_MASK;
556
557 result.frame_header_flags |= (pic->picture_parameter.pic_fields.allow_high_precision_mv
558 << RDECODE_FRAME_HDR_INFO_VP9_ALLOW_HIGH_PRECISION_MV_SHIFT) &
559 RDECODE_FRAME_HDR_INFO_VP9_ALLOW_HIGH_PRECISION_MV_MASK;
560
561 result.frame_header_flags |= (pic->picture_parameter.pic_fields.frame_parallel_decoding_mode
562 << RDECODE_FRAME_HDR_INFO_VP9_FRAME_PARALLEL_DECODING_MODE_SHIFT) &
563 RDECODE_FRAME_HDR_INFO_VP9_FRAME_PARALLEL_DECODING_MODE_MASK;
564
565 result.frame_header_flags |= (pic->picture_parameter.pic_fields.refresh_frame_context
566 << RDECODE_FRAME_HDR_INFO_VP9_REFRESH_FRAME_CONTEXT_SHIFT) &
567 RDECODE_FRAME_HDR_INFO_VP9_REFRESH_FRAME_CONTEXT_MASK;
568
569 result.frame_header_flags |= (pic->picture_parameter.pic_fields.segmentation_enabled
570 << RDECODE_FRAME_HDR_INFO_VP9_SEGMENTATION_ENABLED_SHIFT) &
571 RDECODE_FRAME_HDR_INFO_VP9_SEGMENTATION_ENABLED_MASK;
572
573 result.frame_header_flags |= (pic->picture_parameter.pic_fields.segmentation_update_map
574 << RDECODE_FRAME_HDR_INFO_VP9_SEGMENTATION_UPDATE_MAP_SHIFT) &
575 RDECODE_FRAME_HDR_INFO_VP9_SEGMENTATION_UPDATE_MAP_MASK;
576
577 result.frame_header_flags |= (pic->picture_parameter.pic_fields.segmentation_temporal_update
578 << RDECODE_FRAME_HDR_INFO_VP9_SEGMENTATION_TEMPORAL_UPDATE_SHIFT) &
579 RDECODE_FRAME_HDR_INFO_VP9_SEGMENTATION_TEMPORAL_UPDATE_MASK;
580
581 result.frame_header_flags |= (pic->picture_parameter.mode_ref_delta_enabled
582 << RDECODE_FRAME_HDR_INFO_VP9_MODE_REF_DELTA_ENABLED_SHIFT) &
583 RDECODE_FRAME_HDR_INFO_VP9_MODE_REF_DELTA_ENABLED_MASK;
584
585 result.frame_header_flags |= (pic->picture_parameter.mode_ref_delta_update
586 << RDECODE_FRAME_HDR_INFO_VP9_MODE_REF_DELTA_UPDATE_SHIFT) &
587 RDECODE_FRAME_HDR_INFO_VP9_MODE_REF_DELTA_UPDATE_MASK;
588
589 result.frame_header_flags |=
590 ((dec->show_frame && !pic->picture_parameter.pic_fields.error_resilient_mode &&
591 dec->last_width == dec->base.width && dec->last_height == dec->base.height)
592 << RDECODE_FRAME_HDR_INFO_VP9_USE_PREV_IN_FIND_MV_REFS_SHIFT) &
593 RDECODE_FRAME_HDR_INFO_VP9_USE_PREV_IN_FIND_MV_REFS_MASK;
594 dec->show_frame = pic->picture_parameter.pic_fields.show_frame;
595
596 result.frame_header_flags |= (1 << RDECODE_FRAME_HDR_INFO_VP9_USE_UNCOMPRESSED_HEADER_SHIFT) &
597 RDECODE_FRAME_HDR_INFO_VP9_USE_UNCOMPRESSED_HEADER_MASK;
598
599 result.interp_filter = pic->picture_parameter.pic_fields.mcomp_filter_type;
600
601 result.frame_context_idx = pic->picture_parameter.pic_fields.frame_context_idx;
602 result.reset_frame_context = pic->picture_parameter.pic_fields.reset_frame_context;
603
604 result.filter_level = pic->picture_parameter.filter_level;
605 result.sharpness_level = pic->picture_parameter.sharpness_level;
606
607 for (i = 0; i < 8; ++i)
608 memcpy(result.lf_adj_level[i], pic->slice_parameter.seg_param[i].filter_level, 4 * 2);
609
610 if (pic->picture_parameter.pic_fields.lossless_flag) {
611 result.base_qindex = 0;
612 result.y_dc_delta_q = 0;
613 result.uv_ac_delta_q = 0;
614 result.uv_dc_delta_q = 0;
615 } else {
616 result.base_qindex = pic->picture_parameter.base_qindex;
617 result.y_dc_delta_q = pic->picture_parameter.y_dc_delta_q;
618 result.uv_ac_delta_q = pic->picture_parameter.uv_ac_delta_q;
619 result.uv_dc_delta_q = pic->picture_parameter.uv_dc_delta_q;
620 }
621
622 result.log2_tile_cols = pic->picture_parameter.log2_tile_columns;
623 result.log2_tile_rows = pic->picture_parameter.log2_tile_rows;
624 result.chroma_format = 1;
625 result.bit_depth_luma_minus8 = result.bit_depth_chroma_minus8 =
626 (pic->picture_parameter.bit_depth - 8);
627
628 result.vp9_frame_size = align(dec->bs_size, 128);
629 result.uncompressed_header_size = pic->picture_parameter.frame_header_length_in_bytes;
630 result.compressed_header_size = pic->picture_parameter.first_partition_size;
631
632 assert(dec->base.max_references + 1 <= ARRAY_SIZE(dec->render_pic_list));
633
634 //clear the dec->render list if it is not used as a reference
635 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); i++) {
636 if (dec->render_pic_list[i]) {
637 for (j=0;j<8;j++) {
638 if (dec->render_pic_list[i] == pic->ref[j])
639 break;
640 }
641 if (j == 8)
642 dec->render_pic_list[i] = NULL;
643 }
644 }
645
646 get_current_pic_index(dec, target, &result.curr_pic_idx);
647
648 for (i = 0; i < 8; i++) {
649 uintptr_t ref_frame;
650 if (pic->ref[i]) {
651 ref_frame = (uintptr_t)vl_video_buffer_get_associated_data(pic->ref[i], &dec->base);
652 num_refs++;
653 } else
654 ref_frame = 0x7f;
655 result.ref_frame_map[i] = ref_frame;
656 }
657
658 result.frame_refs[0] = result.ref_frame_map[pic->picture_parameter.pic_fields.last_ref_frame];
659 result.ref_frame_sign_bias[0] = pic->picture_parameter.pic_fields.last_ref_frame_sign_bias;
660 result.frame_refs[1] = result.ref_frame_map[pic->picture_parameter.pic_fields.golden_ref_frame];
661 result.ref_frame_sign_bias[1] = pic->picture_parameter.pic_fields.golden_ref_frame_sign_bias;
662 result.frame_refs[2] = result.ref_frame_map[pic->picture_parameter.pic_fields.alt_ref_frame];
663 result.ref_frame_sign_bias[2] = pic->picture_parameter.pic_fields.alt_ref_frame_sign_bias;
664
665 if (pic->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2) {
666 if (target->buffer_format == PIPE_FORMAT_P010 || target->buffer_format == PIPE_FORMAT_P016) {
667 result.p010_mode = 1;
668 result.msb_mode = 1;
669 } else {
670 result.p010_mode = 0;
671 result.luma_10to8 = 1;
672 result.chroma_10to8 = 1;
673 }
674 }
675
676 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
677 dec->ref_codec.bts = (pic->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2) ?
678 CODEC_10_BITS : CODEC_8_BITS;
679 dec->ref_codec.index = result.curr_pic_idx;
680 dec->ref_codec.ref_size = 8;
681 dec->ref_codec.num_refs = num_refs;
682 memset(dec->ref_codec.ref_list, 0x7f, sizeof(dec->ref_codec.ref_list));
683 memcpy(dec->ref_codec.ref_list, result.ref_frame_map, sizeof(result.ref_frame_map));
684 }
685
686 dec->last_width = dec->base.width;
687 dec->last_height = dec->base.height;
688
689 return result;
690 }
691
get_h265_reflist(rvcn_dec_message_hevc_direct_ref_list_t * hevc_reflist,struct pipe_h265_picture_desc * pic)692 static void get_h265_reflist(rvcn_dec_message_hevc_direct_ref_list_t *hevc_reflist,
693 struct pipe_h265_picture_desc *pic)
694 {
695 hevc_reflist->num_direct_reflist = pic->slice_parameter.slice_count;
696 for (int i = 0; i <hevc_reflist->num_direct_reflist; i++) {
697 for (int j = 0; j < 2; j++) {
698 for (int k = 0; k < 15; k++)
699 hevc_reflist->multi_direct_reflist[i][j][k] = pic->RefPicList[i][j][k];
700 }
701 }
702 }
703
set_drm_keys(rvcn_dec_message_drm_t * drm,DECRYPT_PARAMETERS * decrypted)704 static void set_drm_keys(rvcn_dec_message_drm_t *drm, DECRYPT_PARAMETERS *decrypted)
705 {
706 int cbc = decrypted->u.s.cbc;
707 int ctr = decrypted->u.s.ctr;
708 int id = decrypted->u.s.drm_id;
709 int ekc = 1;
710 int data1 = 1;
711 int data2 = 1;
712
713 drm->drm_cmd = 0;
714 drm->drm_cntl = 0;
715
716 drm->drm_cntl = 1 << DRM_CNTL_BYPASS_SHIFT;
717
718 if (cbc || ctr) {
719 drm->drm_cntl = 0 << DRM_CNTL_BYPASS_SHIFT;
720 drm->drm_cmd |= 0xff << DRM_CMD_BYTE_MASK_SHIFT;
721
722 if (ctr)
723 drm->drm_cmd |= 0x00 << DRM_CMD_ALGORITHM_SHIFT;
724 else if (cbc)
725 drm->drm_cmd |= 0x02 << DRM_CMD_ALGORITHM_SHIFT;
726
727 drm->drm_cmd |= 1 << DRM_CMD_GEN_MASK_SHIFT;
728 drm->drm_cmd |= ekc << DRM_CMD_UNWRAP_KEY_SHIFT;
729 drm->drm_cmd |= 0 << DRM_CMD_OFFSET_SHIFT;
730 drm->drm_cmd |= data2 << DRM_CMD_CNT_DATA_SHIFT;
731 drm->drm_cmd |= data1 << DRM_CMD_CNT_KEY_SHIFT;
732 drm->drm_cmd |= ekc << DRM_CMD_KEY_SHIFT;
733 drm->drm_cmd |= id << DRM_CMD_SESSION_SEL_SHIFT;
734
735 if (ekc)
736 memcpy(drm->drm_wrapped_key, decrypted->encrypted_key, 16);
737 if (data1)
738 memcpy(drm->drm_key, decrypted->session_iv, 16);
739 if (data2)
740 memcpy(drm->drm_counter, decrypted->encrypted_iv, 16);
741 drm->drm_offset = 0;
742 }
743 }
744
rvcn_dec_av1_film_grain_surface(struct pipe_video_buffer ** target,struct pipe_av1_picture_desc * pic)745 static void rvcn_dec_av1_film_grain_surface(struct pipe_video_buffer **target,
746 struct pipe_av1_picture_desc *pic)
747 {
748 if (!pic->picture_parameter.film_grain_info.film_grain_info_fields.apply_grain ||
749 !pic->film_grain_target)
750 return;
751
752 *target = pic->film_grain_target;
753 }
754
get_av1_msg(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_av1_picture_desc * pic)755 static rvcn_dec_message_av1_t get_av1_msg(struct radeon_decoder *dec,
756 struct pipe_video_buffer *target,
757 struct pipe_av1_picture_desc *pic)
758 {
759 rvcn_dec_message_av1_t result;
760 unsigned i, j, num_refs = 0;
761 uint16_t tile_count = pic->picture_parameter.tile_cols * pic->picture_parameter.tile_rows;
762
763 memset(&result, 0, sizeof(result));
764
765 result.frame_header_flags = (pic->picture_parameter.pic_info_fields.show_frame
766 << RDECODE_FRAME_HDR_INFO_AV1_SHOW_FRAME_SHIFT) &
767 RDECODE_FRAME_HDR_INFO_AV1_SHOW_FRAME_MASK;
768
769 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.disable_cdf_update
770 << RDECODE_FRAME_HDR_INFO_AV1_DISABLE_CDF_UPDATE_SHIFT) &
771 RDECODE_FRAME_HDR_INFO_AV1_DISABLE_CDF_UPDATE_MASK;
772
773 result.frame_header_flags |= ((!pic->picture_parameter.pic_info_fields.disable_frame_end_update_cdf)
774 << RDECODE_FRAME_HDR_INFO_AV1_REFRESH_FRAME_CONTEXT_SHIFT) &
775 RDECODE_FRAME_HDR_INFO_AV1_REFRESH_FRAME_CONTEXT_MASK;
776
777 result.frame_header_flags |= ((pic->picture_parameter.pic_info_fields.frame_type ==
778 2 /* INTRA_ONLY_FRAME */) << RDECODE_FRAME_HDR_INFO_AV1_INTRA_ONLY_SHIFT) &
779 RDECODE_FRAME_HDR_INFO_AV1_INTRA_ONLY_MASK;
780
781 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.allow_intrabc
782 << RDECODE_FRAME_HDR_INFO_AV1_ALLOW_INTRABC_SHIFT) &
783 RDECODE_FRAME_HDR_INFO_AV1_ALLOW_INTRABC_MASK;
784
785 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.allow_high_precision_mv
786 << RDECODE_FRAME_HDR_INFO_AV1_ALLOW_HIGH_PRECISION_MV_SHIFT) &
787 RDECODE_FRAME_HDR_INFO_AV1_ALLOW_HIGH_PRECISION_MV_MASK;
788
789 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.mono_chrome
790 << RDECODE_FRAME_HDR_INFO_AV1_MONOCHROME_SHIFT) &
791 RDECODE_FRAME_HDR_INFO_AV1_MONOCHROME_MASK;
792
793 result.frame_header_flags |= (pic->picture_parameter.mode_control_fields.skip_mode_present
794 << RDECODE_FRAME_HDR_INFO_AV1_SKIP_MODE_FLAG_SHIFT) &
795 RDECODE_FRAME_HDR_INFO_AV1_SKIP_MODE_FLAG_MASK;
796
797 result.frame_header_flags |= (((pic->picture_parameter.qmatrix_fields.qm_y == 0xf) ? 0 : 1)
798 << RDECODE_FRAME_HDR_INFO_AV1_USING_QMATRIX_SHIFT) &
799 RDECODE_FRAME_HDR_INFO_AV1_USING_QMATRIX_MASK;
800
801 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_filter_intra
802 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_FILTER_INTRA_SHIFT) &
803 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_FILTER_INTRA_MASK;
804
805 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_intra_edge_filter
806 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_INTRA_EDGE_FILTER_SHIFT) &
807 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_INTRA_EDGE_FILTER_MASK;
808
809 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_interintra_compound
810 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_INTERINTRA_COMPOUND_SHIFT) &
811 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_INTERINTRA_COMPOUND_MASK;
812
813 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_masked_compound
814 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_MASKED_COMPOUND_SHIFT) &
815 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_MASKED_COMPOUND_MASK;
816
817 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.allow_warped_motion
818 << RDECODE_FRAME_HDR_INFO_AV1_ALLOW_WARPED_MOTION_SHIFT) &
819 RDECODE_FRAME_HDR_INFO_AV1_ALLOW_WARPED_MOTION_MASK;
820
821 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_dual_filter
822 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_DUAL_FILTER_SHIFT) &
823 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_DUAL_FILTER_MASK;
824
825 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_order_hint
826 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_ORDER_HINT_SHIFT) &
827 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_ORDER_HINT_MASK;
828
829 result.frame_header_flags |= (pic->picture_parameter.seq_info_fields.enable_jnt_comp
830 << RDECODE_FRAME_HDR_INFO_AV1_ENABLE_JNT_COMP_SHIFT) &
831 RDECODE_FRAME_HDR_INFO_AV1_ENABLE_JNT_COMP_MASK;
832
833 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.use_ref_frame_mvs
834 << RDECODE_FRAME_HDR_INFO_AV1_ALLOW_REF_FRAME_MVS_SHIFT) &
835 RDECODE_FRAME_HDR_INFO_AV1_ALLOW_REF_FRAME_MVS_MASK;
836
837 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.allow_screen_content_tools
838 << RDECODE_FRAME_HDR_INFO_AV1_ALLOW_SCREEN_CONTENT_TOOLS_SHIFT) &
839 RDECODE_FRAME_HDR_INFO_AV1_ALLOW_SCREEN_CONTENT_TOOLS_MASK;
840
841 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.force_integer_mv
842 << RDECODE_FRAME_HDR_INFO_AV1_CUR_FRAME_FORCE_INTEGER_MV_SHIFT) &
843 RDECODE_FRAME_HDR_INFO_AV1_CUR_FRAME_FORCE_INTEGER_MV_MASK;
844
845 result.frame_header_flags |= (pic->picture_parameter.loop_filter_info_fields.mode_ref_delta_enabled
846 << RDECODE_FRAME_HDR_INFO_AV1_MODE_REF_DELTA_ENABLED_SHIFT) &
847 RDECODE_FRAME_HDR_INFO_AV1_MODE_REF_DELTA_ENABLED_MASK;
848
849 result.frame_header_flags |= (pic->picture_parameter.loop_filter_info_fields.mode_ref_delta_update
850 << RDECODE_FRAME_HDR_INFO_AV1_MODE_REF_DELTA_UPDATE_SHIFT) &
851 RDECODE_FRAME_HDR_INFO_AV1_MODE_REF_DELTA_UPDATE_MASK;
852
853 result.frame_header_flags |= (pic->picture_parameter.mode_control_fields.delta_q_present_flag
854 << RDECODE_FRAME_HDR_INFO_AV1_DELTA_Q_PRESENT_FLAG_SHIFT) &
855 RDECODE_FRAME_HDR_INFO_AV1_DELTA_Q_PRESENT_FLAG_MASK;
856
857 result.frame_header_flags |= (pic->picture_parameter.mode_control_fields.delta_lf_present_flag
858 << RDECODE_FRAME_HDR_INFO_AV1_DELTA_LF_PRESENT_FLAG_SHIFT) &
859 RDECODE_FRAME_HDR_INFO_AV1_DELTA_LF_PRESENT_FLAG_MASK;
860
861 result.frame_header_flags |= (pic->picture_parameter.mode_control_fields.reduced_tx_set_used
862 << RDECODE_FRAME_HDR_INFO_AV1_REDUCED_TX_SET_USED_SHIFT) &
863 RDECODE_FRAME_HDR_INFO_AV1_REDUCED_TX_SET_USED_MASK;
864
865 result.frame_header_flags |= (pic->picture_parameter.seg_info.segment_info_fields.enabled
866 << RDECODE_FRAME_HDR_INFO_AV1_SEGMENTATION_ENABLED_SHIFT) &
867 RDECODE_FRAME_HDR_INFO_AV1_SEGMENTATION_ENABLED_MASK;
868
869 result.frame_header_flags |= (pic->picture_parameter.seg_info.segment_info_fields.update_map
870 << RDECODE_FRAME_HDR_INFO_AV1_SEGMENTATION_UPDATE_MAP_SHIFT) &
871 RDECODE_FRAME_HDR_INFO_AV1_SEGMENTATION_UPDATE_MAP_MASK;
872
873 result.frame_header_flags |= (pic->picture_parameter.seg_info.segment_info_fields.temporal_update
874 << RDECODE_FRAME_HDR_INFO_AV1_SEGMENTATION_TEMPORAL_UPDATE_SHIFT) &
875 RDECODE_FRAME_HDR_INFO_AV1_SEGMENTATION_TEMPORAL_UPDATE_MASK;
876
877 result.frame_header_flags |= (pic->picture_parameter.mode_control_fields.delta_lf_multi
878 << RDECODE_FRAME_HDR_INFO_AV1_DELTA_LF_MULTI_SHIFT) &
879 RDECODE_FRAME_HDR_INFO_AV1_DELTA_LF_MULTI_MASK;
880
881 result.frame_header_flags |= (pic->picture_parameter.pic_info_fields.is_motion_mode_switchable
882 << RDECODE_FRAME_HDR_INFO_AV1_SWITCHABLE_SKIP_MODE_SHIFT) &
883 RDECODE_FRAME_HDR_INFO_AV1_SWITCHABLE_SKIP_MODE_MASK;
884
885 result.frame_header_flags |= ((!pic->picture_parameter.refresh_frame_flags)
886 << RDECODE_FRAME_HDR_INFO_AV1_SKIP_REFERENCE_UPDATE_SHIFT) &
887 RDECODE_FRAME_HDR_INFO_AV1_SKIP_REFERENCE_UPDATE_MASK;
888
889 result.frame_header_flags |= ((!pic->picture_parameter.seq_info_fields.ref_frame_mvs)
890 << RDECODE_FRAME_HDR_INFO_AV1_DISABLE_REF_FRAME_MVS_SHIFT) &
891 RDECODE_FRAME_HDR_INFO_AV1_DISABLE_REF_FRAME_MVS_MASK;
892
893 result.current_frame_id = pic->picture_parameter.current_frame_id;
894 result.frame_offset = pic->picture_parameter.order_hint;
895
896 result.profile = pic->picture_parameter.profile;
897 result.is_annexb = 0;
898 result.frame_type = pic->picture_parameter.pic_info_fields.frame_type;
899 result.primary_ref_frame = pic->picture_parameter.primary_ref_frame;
900
901 get_current_pic_index(dec, target, &result.curr_pic_idx);
902
903 result.sb_size = pic->picture_parameter.seq_info_fields.use_128x128_superblock;
904 result.interp_filter = pic->picture_parameter.interp_filter;
905 for (i = 0; i < 2; ++i)
906 result.filter_level[i] = pic->picture_parameter.filter_level[i];
907 result.filter_level_u = pic->picture_parameter.filter_level_u;
908 result.filter_level_v = pic->picture_parameter.filter_level_v;
909 result.sharpness_level = pic->picture_parameter.loop_filter_info_fields.sharpness_level;
910 for (i = 0; i < 8; ++i)
911 result.ref_deltas[i] = pic->picture_parameter.ref_deltas[i];
912 for (i = 0; i < 2; ++i)
913 result.mode_deltas[i] = pic->picture_parameter.mode_deltas[i];
914 result.base_qindex = pic->picture_parameter.base_qindex;
915 result.y_dc_delta_q = pic->picture_parameter.y_dc_delta_q;
916 result.u_dc_delta_q = pic->picture_parameter.u_dc_delta_q;
917 result.v_dc_delta_q = pic->picture_parameter.v_dc_delta_q;
918 result.u_ac_delta_q = pic->picture_parameter.u_ac_delta_q;
919 result.v_ac_delta_q = pic->picture_parameter.v_ac_delta_q;
920 result.qm_y = pic->picture_parameter.qmatrix_fields.qm_y | 0xf0;
921 result.qm_u = pic->picture_parameter.qmatrix_fields.qm_u | 0xf0;
922 result.qm_v = pic->picture_parameter.qmatrix_fields.qm_v | 0xf0;
923 result.delta_q_res = 1 << pic->picture_parameter.mode_control_fields.log2_delta_q_res;
924 result.delta_lf_res = 1 << pic->picture_parameter.mode_control_fields.log2_delta_lf_res;
925
926 result.tile_cols = pic->picture_parameter.tile_cols;
927 result.tile_rows = pic->picture_parameter.tile_rows;
928 result.tx_mode = pic->picture_parameter.mode_control_fields.tx_mode;
929 result.reference_mode = (pic->picture_parameter.mode_control_fields.reference_select == 1) ? 2 : 0;
930 result.chroma_format = pic->picture_parameter.seq_info_fields.mono_chrome ? 0 : 1;
931 result.tile_size_bytes = 0xff;
932 result.context_update_tile_id = pic->picture_parameter.context_update_tile_id;
933 for (i = 0; i < 65; ++i) {
934 result.tile_col_start_sb[i] = pic->picture_parameter.tile_col_start_sb[i];
935 result.tile_row_start_sb[i] = pic->picture_parameter.tile_row_start_sb[i];
936 }
937 result.max_width = pic->picture_parameter.max_width;
938 result.max_height = pic->picture_parameter.max_height;
939 if (pic->picture_parameter.pic_info_fields.use_superres) {
940 result.width = (pic->picture_parameter.frame_width * 8 + pic->picture_parameter.superres_scale_denominator / 2) /
941 pic->picture_parameter.superres_scale_denominator;
942 result.superres_scale_denominator = pic->picture_parameter.superres_scale_denominator;
943 } else {
944 result.width = pic->picture_parameter.frame_width;
945 result.superres_scale_denominator = pic->picture_parameter.superres_scale_denominator;
946 }
947 result.height = pic->picture_parameter.frame_height;
948 result.superres_upscaled_width = pic->picture_parameter.frame_width;
949 result.order_hint_bits = pic->picture_parameter.order_hint_bits_minus_1 + 1;
950
951 /* Limit to target size in case applications try to decode into smaller
952 * target buffer. */
953 result.width = MIN2(target->width, result.width);
954 result.height = MIN2(target->height, result.height);
955 result.superres_upscaled_width = MIN2(target->width, result.superres_upscaled_width);
956
957 for (i = 0; i < NUM_AV1_REFS; ++i) {
958 uintptr_t ref_frame;
959 if (pic->ref[i]) {
960 ref_frame = (uintptr_t)vl_video_buffer_get_associated_data(pic->ref[i], &dec->base);
961 num_refs++;
962 } else
963 ref_frame = 0x7f;
964 result.ref_frame_map[i] = ref_frame;
965 }
966 for (i = 0; i < NUM_AV1_REFS_PER_FRAME; ++i)
967 result.frame_refs[i] = result.ref_frame_map[pic->picture_parameter.ref_frame_idx[i]];
968
969 result.bit_depth_luma_minus8 = result.bit_depth_chroma_minus8 = pic->picture_parameter.bit_depth_idx << 1;
970
971 for (i = 0; i < 8; ++i) {
972 for (j = 0; j < 8; ++j)
973 result.feature_data[i][j] = pic->picture_parameter.seg_info.feature_data[i][j];
974 result.feature_mask[i] = pic->picture_parameter.seg_info.feature_mask[i];
975 }
976 memcpy(dec->probs, &pic->picture_parameter.seg_info.feature_data, 128);
977 memcpy((dec->probs + 128), &pic->picture_parameter.seg_info.feature_mask, 8);
978
979 result.cdef_damping = pic->picture_parameter.cdef_damping_minus_3 + 3;
980 result.cdef_bits = pic->picture_parameter.cdef_bits;
981 for (i = 0; i < 8; ++i) {
982 result.cdef_strengths[i] = pic->picture_parameter.cdef_y_strengths[i];
983 result.cdef_uv_strengths[i] = pic->picture_parameter.cdef_uv_strengths[i];
984 }
985 result.frame_restoration_type[0] = pic->picture_parameter.loop_restoration_fields.yframe_restoration_type;
986 result.frame_restoration_type[1] = pic->picture_parameter.loop_restoration_fields.cbframe_restoration_type;
987 result.frame_restoration_type[2] = pic->picture_parameter.loop_restoration_fields.crframe_restoration_type;
988 for (i = 0; i < 3; ++i) {
989 int log2_num = 0;
990 int unit_size = pic->picture_parameter.lr_unit_size[i];
991 if (unit_size) {
992 while (unit_size >>= 1)
993 log2_num++;
994 result.log2_restoration_unit_size_minus5[i] = log2_num - 5;
995 } else {
996 result.log2_restoration_unit_size_minus5[i] = 0;
997 }
998 }
999
1000 if (pic->picture_parameter.bit_depth_idx) {
1001 if (target->buffer_format == PIPE_FORMAT_P010 ||
1002 target->buffer_format == PIPE_FORMAT_P012 ||
1003 target->buffer_format == PIPE_FORMAT_P016) {
1004 result.p010_mode = 1;
1005 result.msb_mode = 1;
1006 } else {
1007 result.luma_10to8 = 1;
1008 result.chroma_10to8 = 1;
1009 }
1010 }
1011
1012 result.preskip_segid = 0;
1013 result.last_active_segid = 0;
1014 for (i = 0; i < 8; i++) {
1015 for (j = 0; j < 8; j++) {
1016 if (pic->picture_parameter.seg_info.feature_mask[i] & (1 << j)) {
1017 result.last_active_segid = i;
1018 if (j >= 5)
1019 result.preskip_segid = 1;
1020 }
1021 }
1022 }
1023
1024 result.seg_lossless_flag = 0;
1025 for (i = 0; i < 8; ++i) {
1026 int av1_get_qindex, qindex;
1027 int segfeature_active = pic->picture_parameter.seg_info.feature_mask[i] & (1 << 0);
1028 if (segfeature_active) {
1029 int seg_qindex = pic->picture_parameter.base_qindex +
1030 pic->picture_parameter.seg_info.feature_data[i][0];
1031 av1_get_qindex = seg_qindex < 0 ? 0 : (seg_qindex > 255 ? 255 : seg_qindex);
1032 } else {
1033 av1_get_qindex = pic->picture_parameter.base_qindex;
1034 }
1035 qindex = pic->picture_parameter.seg_info.segment_info_fields.enabled ?
1036 av1_get_qindex :
1037 pic->picture_parameter.base_qindex;
1038 result.seg_lossless_flag |= (((qindex == 0) && result.y_dc_delta_q == 0 &&
1039 result.u_dc_delta_q == 0 && result.v_dc_delta_q == 0 &&
1040 result.u_ac_delta_q == 0 && result.v_ac_delta_q == 0) << i);
1041 }
1042
1043 rvcn_dec_film_grain_params_t* fg_params = &result.film_grain;
1044 fg_params->apply_grain = pic->picture_parameter.film_grain_info.film_grain_info_fields.apply_grain;
1045 if (fg_params->apply_grain) {
1046 rvcn_dec_av1_fg_init_buf_t *fg_buf = (rvcn_dec_av1_fg_init_buf_t *)(dec->probs + 256);
1047
1048 fg_params->random_seed = pic->picture_parameter.film_grain_info.grain_seed;
1049 fg_params->grain_scale_shift =
1050 pic->picture_parameter.film_grain_info.film_grain_info_fields.grain_scale_shift;
1051 fg_params->scaling_shift =
1052 pic->picture_parameter.film_grain_info.film_grain_info_fields.grain_scaling_minus_8 + 8;
1053 fg_params->chroma_scaling_from_luma =
1054 pic->picture_parameter.film_grain_info.film_grain_info_fields.chroma_scaling_from_luma;
1055 fg_params->num_y_points = pic->picture_parameter.film_grain_info.num_y_points;
1056 fg_params->num_cb_points = pic->picture_parameter.film_grain_info.num_cb_points;
1057 fg_params->num_cr_points = pic->picture_parameter.film_grain_info.num_cr_points;
1058 fg_params->cb_mult = pic->picture_parameter.film_grain_info.cb_mult;
1059 fg_params->cb_luma_mult = pic->picture_parameter.film_grain_info.cb_luma_mult;
1060 fg_params->cb_offset = pic->picture_parameter.film_grain_info.cb_offset;
1061 fg_params->cr_mult = pic->picture_parameter.film_grain_info.cr_mult;
1062 fg_params->cr_luma_mult = pic->picture_parameter.film_grain_info.cr_luma_mult;
1063 fg_params->cr_offset = pic->picture_parameter.film_grain_info.cr_offset;
1064 fg_params->bit_depth_minus_8 = pic->picture_parameter.bit_depth_idx << 1;
1065
1066 for (i = 0; i < fg_params->num_y_points; ++i) {
1067 fg_params->scaling_points_y[i][0] = pic->picture_parameter.film_grain_info.point_y_value[i];
1068 fg_params->scaling_points_y[i][1] = pic->picture_parameter.film_grain_info.point_y_scaling[i];
1069 }
1070 for (i = 0; i < fg_params->num_cb_points; ++i) {
1071 fg_params->scaling_points_cb[i][0] = pic->picture_parameter.film_grain_info.point_cb_value[i];
1072 fg_params->scaling_points_cb[i][1] = pic->picture_parameter.film_grain_info.point_cb_scaling[i];
1073 }
1074 for (i = 0; i < fg_params->num_cr_points; ++i) {
1075 fg_params->scaling_points_cr[i][0] = pic->picture_parameter.film_grain_info.point_cr_value[i];
1076 fg_params->scaling_points_cr[i][1] = pic->picture_parameter.film_grain_info.point_cr_scaling[i];
1077 }
1078
1079 fg_params->ar_coeff_lag = pic->picture_parameter.film_grain_info.film_grain_info_fields.ar_coeff_lag;
1080 fg_params->ar_coeff_shift =
1081 pic->picture_parameter.film_grain_info.film_grain_info_fields.ar_coeff_shift_minus_6 + 6;
1082
1083 for (i = 0; i < 24; ++i)
1084 fg_params->ar_coeffs_y[i] = pic->picture_parameter.film_grain_info.ar_coeffs_y[i];
1085
1086 for (i = 0; i < 25; ++i) {
1087 fg_params->ar_coeffs_cb[i] = pic->picture_parameter.film_grain_info.ar_coeffs_cb[i];
1088 fg_params->ar_coeffs_cr[i] = pic->picture_parameter.film_grain_info.ar_coeffs_cr[i];
1089 }
1090
1091 fg_params->overlap_flag = pic->picture_parameter.film_grain_info.film_grain_info_fields.overlap_flag;
1092 fg_params->clip_to_restricted_range =
1093 pic->picture_parameter.film_grain_info.film_grain_info_fields.clip_to_restricted_range;
1094
1095 ac_vcn_av1_init_film_grain_buffer(fg_params, fg_buf);
1096 }
1097
1098 result.uncompressed_header_size = 0;
1099 for (i = 0; i < 7; ++i) {
1100 result.global_motion[i + 1].wmtype = (rvcn_dec_transformation_type_e)pic->picture_parameter.wm[i].wmtype;
1101 for (j = 0; j < 6; ++j)
1102 result.global_motion[i + 1].wmmat[j] = pic->picture_parameter.wm[i].wmmat[j];
1103 }
1104 for (i = 0; i < tile_count && i < 256; ++i) {
1105 result.tile_info[i].offset = pic->slice_parameter.slice_data_offset[i];
1106 result.tile_info[i].size = pic->slice_parameter.slice_data_size[i];
1107 }
1108
1109 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
1110 dec->ref_codec.bts = pic->picture_parameter.bit_depth_idx;
1111 dec->ref_codec.index = result.curr_pic_idx;
1112 dec->ref_codec.ref_size = 8;
1113 dec->ref_codec.num_refs = num_refs;
1114 memset(dec->ref_codec.ref_list, 0x7f, sizeof(dec->ref_codec.ref_list));
1115 memcpy(dec->ref_codec.ref_list, result.ref_frame_map, sizeof(result.ref_frame_map));
1116 }
1117
1118 return result;
1119 }
1120
calc_ctx_size_h265_main(struct radeon_decoder * dec)1121 static unsigned calc_ctx_size_h265_main(struct radeon_decoder *dec)
1122 {
1123 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
1124 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
1125
1126 unsigned max_references = dec->base.max_references + 1;
1127
1128 if (dec->base.width * dec->base.height >= 4096 * 2000)
1129 max_references = MAX2(max_references, 8);
1130 else
1131 max_references = MAX2(max_references, 17);
1132
1133 width = align(width, 16);
1134 height = align(height, 16);
1135 return ((width + 255) / 16) * ((height + 255) / 16) * 16 * max_references + 52 * 1024;
1136 }
1137
calc_ctx_size_h265_main10(struct radeon_decoder * dec,struct pipe_h265_picture_desc * pic)1138 static unsigned calc_ctx_size_h265_main10(struct radeon_decoder *dec,
1139 struct pipe_h265_picture_desc *pic)
1140 {
1141 unsigned log2_ctb_size, width_in_ctb, height_in_ctb, num_16x16_block_per_ctb;
1142 unsigned context_buffer_size_per_ctb_row, cm_buffer_size, max_mb_address, db_left_tile_pxl_size;
1143 unsigned db_left_tile_ctx_size = 4096 / 16 * (32 + 16 * 4);
1144
1145 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
1146 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
1147 unsigned coeff_10bit =
1148 (pic->pps->sps->bit_depth_luma_minus8 || pic->pps->sps->bit_depth_chroma_minus8) ? 2 : 1;
1149
1150 unsigned max_references = dec->base.max_references + 1;
1151
1152 if (dec->base.width * dec->base.height >= 4096 * 2000)
1153 max_references = MAX2(max_references, 8);
1154 else
1155 max_references = MAX2(max_references, 17);
1156
1157 log2_ctb_size = pic->pps->sps->log2_min_luma_coding_block_size_minus3 + 3 +
1158 pic->pps->sps->log2_diff_max_min_luma_coding_block_size;
1159
1160 width_in_ctb = (width + ((1 << log2_ctb_size) - 1)) >> log2_ctb_size;
1161 height_in_ctb = (height + ((1 << log2_ctb_size) - 1)) >> log2_ctb_size;
1162
1163 num_16x16_block_per_ctb = ((1 << log2_ctb_size) >> 4) * ((1 << log2_ctb_size) >> 4);
1164 context_buffer_size_per_ctb_row = align(width_in_ctb * num_16x16_block_per_ctb * 16, 256);
1165 max_mb_address = (unsigned)ceil(height * 8 / 2048.0);
1166
1167 cm_buffer_size = max_references * context_buffer_size_per_ctb_row * height_in_ctb;
1168 db_left_tile_pxl_size = coeff_10bit * (max_mb_address * 2 * 2048 + 1024);
1169
1170 return cm_buffer_size + db_left_tile_ctx_size + db_left_tile_pxl_size;
1171 }
1172
get_vc1_msg(struct pipe_vc1_picture_desc * pic)1173 static rvcn_dec_message_vc1_t get_vc1_msg(struct pipe_vc1_picture_desc *pic)
1174 {
1175 rvcn_dec_message_vc1_t result;
1176
1177 memset(&result, 0, sizeof(result));
1178 switch (pic->base.profile) {
1179 case PIPE_VIDEO_PROFILE_VC1_SIMPLE:
1180 result.profile = RDECODE_VC1_PROFILE_SIMPLE;
1181 result.level = 1;
1182 break;
1183
1184 case PIPE_VIDEO_PROFILE_VC1_MAIN:
1185 result.profile = RDECODE_VC1_PROFILE_MAIN;
1186 result.level = 2;
1187 break;
1188
1189 case PIPE_VIDEO_PROFILE_VC1_ADVANCED:
1190 result.profile = RDECODE_VC1_PROFILE_ADVANCED;
1191 result.level = 4;
1192 break;
1193
1194 default:
1195 assert(0);
1196 }
1197
1198 result.sps_info_flags |= pic->postprocflag << 7;
1199 result.sps_info_flags |= pic->pulldown << 6;
1200 result.sps_info_flags |= pic->interlace << 5;
1201 result.sps_info_flags |= pic->tfcntrflag << 4;
1202 result.sps_info_flags |= pic->finterpflag << 3;
1203 result.sps_info_flags |= pic->psf << 1;
1204
1205 result.pps_info_flags |= pic->range_mapy_flag << 31;
1206 result.pps_info_flags |= pic->range_mapy << 28;
1207 result.pps_info_flags |= pic->range_mapuv_flag << 27;
1208 result.pps_info_flags |= pic->range_mapuv << 24;
1209 result.pps_info_flags |= pic->multires << 21;
1210 result.pps_info_flags |= pic->maxbframes << 16;
1211 result.pps_info_flags |= pic->overlap << 11;
1212 result.pps_info_flags |= pic->quantizer << 9;
1213 result.pps_info_flags |= pic->panscan_flag << 7;
1214 result.pps_info_flags |= pic->refdist_flag << 6;
1215 result.pps_info_flags |= pic->vstransform << 0;
1216
1217 if (pic->base.profile != PIPE_VIDEO_PROFILE_VC1_SIMPLE) {
1218 result.pps_info_flags |= pic->syncmarker << 20;
1219 result.pps_info_flags |= pic->rangered << 19;
1220 result.pps_info_flags |= pic->loopfilter << 5;
1221 result.pps_info_flags |= pic->fastuvmc << 4;
1222 result.pps_info_flags |= pic->extended_mv << 3;
1223 result.pps_info_flags |= pic->extended_dmv << 8;
1224 result.pps_info_flags |= pic->dquant << 1;
1225 }
1226
1227 result.chroma_format = 1;
1228
1229 return result;
1230 }
1231
get_ref_pic_idx(struct radeon_decoder * dec,struct pipe_video_buffer * ref)1232 static uint32_t get_ref_pic_idx(struct radeon_decoder *dec, struct pipe_video_buffer *ref)
1233 {
1234 uint32_t min = MAX2(dec->frame_number, NUM_MPEG2_REFS) - NUM_MPEG2_REFS;
1235 uint32_t max = MAX2(dec->frame_number, 1) - 1;
1236 uintptr_t frame;
1237
1238 /* seems to be the most sane fallback */
1239 if (!ref)
1240 return max;
1241
1242 /* get the frame number from the associated data */
1243 frame = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
1244
1245 /* limit the frame number to a valid range */
1246 return MAX2(MIN2(frame, max), min);
1247 }
1248
get_mpeg2_msg(struct radeon_decoder * dec,struct pipe_mpeg12_picture_desc * pic)1249 static rvcn_dec_message_mpeg2_vld_t get_mpeg2_msg(struct radeon_decoder *dec,
1250 struct pipe_mpeg12_picture_desc *pic)
1251 {
1252 const int *zscan = pic->alternate_scan ? vl_zscan_alternate : vl_zscan_normal;
1253 rvcn_dec_message_mpeg2_vld_t result;
1254 unsigned i;
1255
1256 memset(&result, 0, sizeof(result));
1257 result.decoded_pic_idx = dec->frame_number;
1258
1259 result.forward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[0]);
1260 result.backward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[1]);
1261
1262 if (pic->intra_matrix) {
1263 result.load_intra_quantiser_matrix = 1;
1264 for (i = 0; i < 64; ++i) {
1265 result.intra_quantiser_matrix[i] = pic->intra_matrix[zscan[i]];
1266 }
1267 }
1268 if (pic->non_intra_matrix) {
1269 result.load_nonintra_quantiser_matrix = 1;
1270 for (i = 0; i < 64; ++i) {
1271 result.nonintra_quantiser_matrix[i] = pic->non_intra_matrix[zscan[i]];
1272 }
1273 }
1274
1275 result.profile_and_level_indication = 0;
1276 result.chroma_format = 0x1;
1277
1278 result.picture_coding_type = pic->picture_coding_type;
1279 result.f_code[0][0] = pic->f_code[0][0] + 1;
1280 result.f_code[0][1] = pic->f_code[0][1] + 1;
1281 result.f_code[1][0] = pic->f_code[1][0] + 1;
1282 result.f_code[1][1] = pic->f_code[1][1] + 1;
1283 result.intra_dc_precision = pic->intra_dc_precision;
1284 result.pic_structure = pic->picture_structure;
1285 result.top_field_first = pic->top_field_first;
1286 result.frame_pred_frame_dct = pic->frame_pred_frame_dct;
1287 result.concealment_motion_vectors = pic->concealment_motion_vectors;
1288 result.q_scale_type = pic->q_scale_type;
1289 result.intra_vlc_format = pic->intra_vlc_format;
1290 result.alternate_scan = pic->alternate_scan;
1291
1292 return result;
1293 }
1294
get_mpeg4_msg(struct radeon_decoder * dec,struct pipe_mpeg4_picture_desc * pic)1295 static rvcn_dec_message_mpeg4_asp_vld_t get_mpeg4_msg(struct radeon_decoder *dec,
1296 struct pipe_mpeg4_picture_desc *pic)
1297 {
1298 rvcn_dec_message_mpeg4_asp_vld_t result;
1299 unsigned i;
1300
1301 memset(&result, 0, sizeof(result));
1302 result.decoded_pic_idx = dec->frame_number;
1303
1304 result.forward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[0]);
1305 result.backward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[1]);
1306
1307 result.variant_type = 0;
1308 result.profile_and_level_indication = 0xF0;
1309
1310 result.video_object_layer_verid = 0x5;
1311 result.video_object_layer_shape = 0x0;
1312
1313 result.video_object_layer_width = dec->base.width;
1314 result.video_object_layer_height = dec->base.height;
1315
1316 result.vop_time_increment_resolution = pic->vop_time_increment_resolution;
1317
1318 result.short_video_header = pic->short_video_header;
1319 result.interlaced = pic->interlaced;
1320 result.load_intra_quant_mat = 1;
1321 result.load_nonintra_quant_mat = 1;
1322 result.quarter_sample = pic->quarter_sample;
1323 result.complexity_estimation_disable = 1;
1324 result.resync_marker_disable = pic->resync_marker_disable;
1325 result.newpred_enable = 0;
1326 result.reduced_resolution_vop_enable = 0;
1327
1328 result.quant_type = pic->quant_type;
1329
1330 for (i = 0; i < 64; ++i) {
1331 result.intra_quant_mat[i] = pic->intra_matrix[vl_zscan_normal[i]];
1332 result.nonintra_quant_mat[i] = pic->non_intra_matrix[vl_zscan_normal[i]];
1333 }
1334
1335 return result;
1336 }
1337
rvcn_dec_message_create(struct radeon_decoder * dec)1338 static void rvcn_dec_message_create(struct radeon_decoder *dec)
1339 {
1340 rvcn_dec_message_header_t *header = dec->msg;
1341 rvcn_dec_message_create_t *create = dec->msg + sizeof(rvcn_dec_message_header_t);
1342 unsigned sizes = sizeof(rvcn_dec_message_header_t) + sizeof(rvcn_dec_message_create_t);
1343
1344 memset(dec->msg, 0, sizes);
1345 header->header_size = sizeof(rvcn_dec_message_header_t);
1346 header->total_size = sizes;
1347 header->num_buffers = 1;
1348 header->msg_type = RDECODE_MSG_CREATE;
1349 header->stream_handle = dec->stream_handle;
1350 header->status_report_feedback_number = 0;
1351
1352 header->index[0].message_id = RDECODE_MESSAGE_CREATE;
1353 header->index[0].offset = sizeof(rvcn_dec_message_header_t);
1354 header->index[0].size = sizeof(rvcn_dec_message_create_t);
1355 header->index[0].filled = 0;
1356
1357 create->stream_type = dec->stream_type;
1358 create->session_flags = 0;
1359 create->width_in_samples = dec->base.width;
1360 create->height_in_samples = dec->base.height;
1361 }
1362
rvcn_dec_dynamic_dpb_t2_message(struct radeon_decoder * dec,rvcn_dec_message_decode_t * decode,rvcn_dec_message_dynamic_dpb_t2_t * dynamic_dpb_t2,bool encrypted)1363 static unsigned rvcn_dec_dynamic_dpb_t2_message(struct radeon_decoder *dec, rvcn_dec_message_decode_t *decode,
1364 rvcn_dec_message_dynamic_dpb_t2_t *dynamic_dpb_t2, bool encrypted)
1365 {
1366 struct rvcn_dec_dynamic_dpb_t2 *dpb = NULL;
1367 struct si_resource *res;
1368 unsigned width, height;
1369 uint64_t addr;
1370 int i;
1371
1372 width = align(decode->width_in_samples, dec->db_alignment);
1373 height = align(decode->height_in_samples, dec->db_alignment);
1374
1375 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
1376 bool found = false;
1377
1378 res = &(((struct si_texture *)((struct vl_video_buffer *)d->vbuf)->resources[0])->buffer);
1379 for (i = 0; i < dec->ref_codec.ref_size; ++i) {
1380 if (((dec->ref_codec.ref_list[i] & 0x7f) != 0x7f) && (d->index == (dec->ref_codec.ref_list[i] & 0x7f))) {
1381 addr = dec->ws->buffer_get_virtual_address(res->buf);
1382 dec->ws->cs_add_buffer(&dec->cs, res->buf, RADEON_USAGE_READWRITE | RADEON_USAGE_SYNCHRONIZED, RADEON_DOMAIN_VRAM);
1383 dynamic_dpb_t2->dpbAddrLo[i] = addr;
1384 dynamic_dpb_t2->dpbAddrHi[i] = addr >> 32;
1385 ++dynamic_dpb_t2->dpbArraySize;
1386 found = true;
1387 }
1388 }
1389 if (!found) {
1390 if (d->vbuf->width != width || d->vbuf->height != height) {
1391 list_del(&d->list);
1392 list_addtail(&d->list, &dec->dpb_unref_list);
1393 } else {
1394 d->index = 0x7f;
1395 }
1396 }
1397 }
1398
1399 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
1400 if (d->vbuf->width == width && d->vbuf->height == height && d->index == dec->ref_codec.index) {
1401 dpb = d;
1402 break;
1403 }
1404 }
1405
1406 if (!dpb) {
1407 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
1408 if (d->index == 0x7f) {
1409 d->index = dec->ref_codec.index;
1410 dpb = d;
1411 break;
1412 }
1413 }
1414 }
1415
1416 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_unref_list, list) {
1417 list_del(&d->list);
1418 d->vbuf->destroy(d->vbuf);
1419 FREE(d);
1420 }
1421
1422 if (!dpb) {
1423 dpb = CALLOC_STRUCT(rvcn_dec_dynamic_dpb_t2);
1424 if (!dpb)
1425 return 1;
1426 dpb->index = dec->ref_codec.index;
1427
1428 struct pipe_video_buffer templat;
1429 memset(&templat, 0, sizeof(templat));
1430 templat.buffer_format = get_buffer_format(dec);
1431 templat.width = width;
1432 templat.height = height;
1433 templat.bind = PIPE_BIND_VIDEO_DECODE_DPB;
1434 if (encrypted)
1435 templat.bind |= PIPE_BIND_PROTECTED;
1436 dpb->vbuf = dec->base.context->create_video_buffer(dec->base.context, &templat);
1437
1438 if (!dpb->vbuf) {
1439 RADEON_DEC_ERR("Can't allocate dpb buffer.\n");
1440 FREE(dpb);
1441 return 1;
1442 }
1443 list_addtail(&dpb->list, &dec->dpb_ref_list);
1444 }
1445
1446 if (dynamic_dpb_t2->dpbArraySize < dec->ref_codec.num_refs) {
1447 struct rvcn_dec_dynamic_dpb_t2 *d =
1448 list_first_entry(&dec->dpb_ref_list, struct rvcn_dec_dynamic_dpb_t2, list);
1449
1450 res = &(((struct si_texture *)((struct vl_video_buffer *)d->vbuf)->resources[0])->buffer);
1451 addr = dec->ws->buffer_get_virtual_address(res->buf);
1452 for (i = 0; i < dec->ref_codec.num_refs; ++i) {
1453 if (dynamic_dpb_t2->dpbAddrLo[i] || dynamic_dpb_t2->dpbAddrHi[i])
1454 continue;
1455 dynamic_dpb_t2->dpbAddrLo[i] = addr;
1456 dynamic_dpb_t2->dpbAddrHi[i] = addr >> 32;
1457 ++dynamic_dpb_t2->dpbArraySize;
1458 }
1459 assert(dynamic_dpb_t2->dpbArraySize == dec->ref_codec.num_refs);
1460 }
1461
1462 struct si_texture *dpb_luma, *dpb_chroma;
1463
1464 dpb_luma = (struct si_texture *)((struct vl_video_buffer *)dpb->vbuf)->resources[0];
1465 dpb_chroma = (struct si_texture *)((struct vl_video_buffer *)dpb->vbuf)->resources[1];
1466
1467 decode->db_swizzle_mode = dpb_luma->surface.u.gfx9.swizzle_mode;
1468
1469 dec->ws->cs_add_buffer(&dec->cs, (dpb_luma->buffer).buf, RADEON_USAGE_READWRITE | RADEON_USAGE_SYNCHRONIZED, RADEON_DOMAIN_VRAM);
1470 addr = dec->ws->buffer_get_virtual_address((dpb_luma->buffer).buf);
1471
1472 dynamic_dpb_t2->dpbLumaPitch = dpb_luma->surface.u.gfx9.surf_pitch;
1473 dynamic_dpb_t2->dpbLumaAlignedHeight = dpb_luma->surface.u.gfx9.surf_height;
1474 dynamic_dpb_t2->dpbLumaAlignedSize = dpb_luma->surface.u.gfx9.surf_slice_size;
1475 dynamic_dpb_t2->dpbChromaPitch = dpb_chroma->surface.u.gfx9.surf_pitch;
1476 dynamic_dpb_t2->dpbChromaAlignedHeight = dpb_chroma->surface.u.gfx9.surf_height;
1477 dynamic_dpb_t2->dpbChromaAlignedSize = dpb_chroma->surface.u.gfx9.surf_slice_size;
1478
1479 dynamic_dpb_t2->dpbCurrLo = addr;
1480 dynamic_dpb_t2->dpbCurrHi = addr >> 32;
1481 decode->decode_flags = 1;
1482 dynamic_dpb_t2->dpbConfigFlags = 0;
1483
1484 return 0;
1485 }
1486
rvcn_dec_message_decode(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)1487 static struct pb_buffer_lean *rvcn_dec_message_decode(struct radeon_decoder *dec,
1488 struct pipe_video_buffer *target,
1489 struct pipe_picture_desc *picture)
1490 {
1491 DECRYPT_PARAMETERS *decrypt = (DECRYPT_PARAMETERS *)picture->decrypt_key;
1492 bool encrypted = picture->protected_playback;
1493 struct si_texture *luma;
1494 struct si_texture *chroma;
1495 struct pipe_video_buffer *out_surf = target;
1496 ASSERTED struct si_screen *sscreen = (struct si_screen *)dec->screen;
1497 rvcn_dec_message_header_t *header;
1498 rvcn_dec_message_index_t *index_codec;
1499 rvcn_dec_message_index_t *index_drm = NULL;
1500 rvcn_dec_message_index_t *index_dynamic_dpb = NULL;
1501 rvcn_dec_message_index_t *index_hevc_direct_reflist = NULL;
1502 rvcn_dec_message_decode_t *decode;
1503 unsigned sizes = 0, offset_decode, offset_codec;
1504 unsigned offset_drm = 0, offset_dynamic_dpb = 0, offset_hevc_direct_reflist = 0;
1505 void *codec;
1506 rvcn_dec_message_drm_t *drm = NULL;
1507 rvcn_dec_message_dynamic_dpb_t *dynamic_dpb = NULL;
1508 rvcn_dec_message_dynamic_dpb_t2_t *dynamic_dpb_t2 = NULL;
1509 rvcn_dec_message_hevc_direct_ref_list_t *hevc_reflist = NULL;
1510 bool dpb_resize = false;
1511 header = dec->msg;
1512 sizes += sizeof(rvcn_dec_message_header_t);
1513
1514 index_codec = (void*)header + sizes;
1515 sizes += sizeof(rvcn_dec_message_index_t);
1516
1517 if (encrypted) {
1518 index_drm = (void*)header + sizes;
1519 sizes += sizeof(rvcn_dec_message_index_t);
1520 }
1521
1522 if (dec->dpb_type >= DPB_DYNAMIC_TIER_1) {
1523 index_dynamic_dpb = (void*)header + sizes;
1524 sizes += sizeof(rvcn_dec_message_index_t);
1525 }
1526
1527 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1528 index_hevc_direct_reflist = (void*)header + sizes;
1529 sizes += sizeof(rvcn_dec_message_index_t);
1530 }
1531
1532 offset_decode = sizes;
1533 decode = (void*)header + sizes;
1534 sizes += sizeof(rvcn_dec_message_decode_t);
1535
1536 if (encrypted) {
1537 offset_drm = sizes;
1538 drm = (void*)header + sizes;
1539 sizes += sizeof(rvcn_dec_message_drm_t);
1540 }
1541
1542 if (dec->dpb_type >= DPB_DYNAMIC_TIER_1) {
1543 offset_dynamic_dpb = sizes;
1544 if (dec->dpb_type == DPB_DYNAMIC_TIER_1) {
1545 dynamic_dpb = (void*)header + sizes;
1546 sizes += sizeof(rvcn_dec_message_dynamic_dpb_t);
1547 }
1548 else if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
1549 dynamic_dpb_t2 = (void*)header + sizes;
1550 sizes += sizeof(rvcn_dec_message_dynamic_dpb_t2_t);
1551 }
1552 }
1553
1554 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1555 offset_hevc_direct_reflist = sizes;
1556 hevc_reflist = (void*)header + sizes;
1557 sizes += align((4 + 2 * 15 * ((struct pipe_h265_picture_desc *)picture)->slice_parameter.slice_count), 4);
1558 }
1559
1560 offset_codec = sizes;
1561 codec = (void*)header + sizes;
1562
1563 memset(dec->msg, 0, sizes);
1564 header->header_size = sizeof(rvcn_dec_message_header_t);
1565 header->total_size = sizes;
1566 header->msg_type = RDECODE_MSG_DECODE;
1567 header->stream_handle = dec->stream_handle;
1568 header->status_report_feedback_number = dec->frame_number;
1569
1570 header->index[0].message_id = RDECODE_MESSAGE_DECODE;
1571 header->index[0].offset = offset_decode;
1572 header->index[0].size = sizeof(rvcn_dec_message_decode_t);
1573 header->index[0].filled = 0;
1574 header->num_buffers = 1;
1575
1576 index_codec->offset = offset_codec;
1577 index_codec->size = 0;
1578 index_codec->filled = 0;
1579 ++header->num_buffers;
1580
1581 if (encrypted) {
1582 index_drm->message_id = RDECODE_MESSAGE_DRM;
1583 index_drm->offset = offset_drm;
1584 index_drm->size = sizeof(rvcn_dec_message_drm_t);
1585 index_drm->filled = 0;
1586 ++header->num_buffers;
1587 }
1588
1589 if (dec->dpb_type >= DPB_DYNAMIC_TIER_1) {
1590 index_dynamic_dpb->message_id = RDECODE_MESSAGE_DYNAMIC_DPB;
1591 index_dynamic_dpb->offset = offset_dynamic_dpb;
1592 index_dynamic_dpb->filled = 0;
1593 ++header->num_buffers;
1594 if (dec->dpb_type == DPB_DYNAMIC_TIER_1)
1595 index_dynamic_dpb->size = sizeof(rvcn_dec_message_dynamic_dpb_t);
1596 else if (dec->dpb_type == DPB_DYNAMIC_TIER_2)
1597 index_dynamic_dpb->size = sizeof(rvcn_dec_message_dynamic_dpb_t2_t);
1598 }
1599
1600 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1601 index_hevc_direct_reflist->message_id = RDECODE_MESSAGE_HEVC_DIRECT_REF_LIST;
1602 index_hevc_direct_reflist->offset = offset_hevc_direct_reflist;
1603 index_hevc_direct_reflist->size = align((4 + 2 * 15 * ((struct pipe_h265_picture_desc *)picture)->slice_parameter.slice_count), 4);
1604 index_hevc_direct_reflist->filled = 0;
1605 ++header->num_buffers;
1606 }
1607
1608 decode->stream_type = dec->stream_type;
1609 decode->decode_flags = 0;
1610 decode->width_in_samples = dec->base.width;
1611 decode->height_in_samples = dec->base.height;
1612
1613 decode->bsd_size = align(dec->bs_size, 128);
1614
1615 if (dec->dpb_type != DPB_DYNAMIC_TIER_2) {
1616 bool r;
1617 if (!dec->dpb.res && dec->dpb_size) {
1618 if (encrypted) {
1619 r = si_vid_create_tmz_buffer(dec->screen, &dec->dpb, dec->dpb_size, PIPE_USAGE_DEFAULT);
1620 } else {
1621 r = si_vid_create_buffer(dec->screen, &dec->dpb, dec->dpb_size, PIPE_USAGE_DEFAULT);
1622 }
1623 assert(encrypted == (bool)(dec->dpb.res->flags & RADEON_FLAG_ENCRYPTED));
1624 if (!r) {
1625 RADEON_DEC_ERR("Can't allocate dpb.\n");
1626 return NULL;
1627 }
1628 } else if (dec->dpb_type == DPB_DYNAMIC_TIER_1 && dec->dpb.res &&
1629 (dec->max_width < dec->base.width || dec->max_height < dec->base.height)) {
1630 struct rvid_buf_offset_info buf_offset_info;
1631
1632 buf_offset_info.num_units = (NUM_VP9_REFS + 1);
1633 buf_offset_info.old_offset = (align(dec->max_width, dec->db_alignment) *
1634 align(dec->max_height, dec->db_alignment) * 3 / 2);
1635 buf_offset_info.new_offset = (align(dec->base.width, dec->db_alignment) *
1636 align(dec->base.height, dec->db_alignment) * 3 / 2);
1637
1638 dec->dpb_size = calc_dpb_size(dec);
1639 r = si_vid_resize_buffer(dec->base.context, &dec->cs, &dec->dpb, dec->dpb_size, &buf_offset_info);
1640 if (!r) {
1641 RADEON_DEC_ERR("Can't resize dpb.\n");
1642 return NULL;
1643 }
1644 dec->max_width = dec->base.width;
1645 dec->max_height = dec->base.height;
1646 dpb_resize = true;
1647 }
1648 }
1649
1650 if (!dec->ctx.res) {
1651 enum pipe_video_format fmt = u_reduce_video_profile(picture->profile);
1652 if (dec->stream_type == RDECODE_CODEC_H264_PERF) {
1653 unsigned ctx_size = calc_ctx_size_h264_perf(dec);
1654 bool r;
1655 if (encrypted && dec->tmz_ctx) {
1656 r = si_vid_create_tmz_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1657 } else {
1658 r = si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1659 }
1660 assert((encrypted && dec->tmz_ctx) == (bool)(dec->ctx.res->flags & RADEON_FLAG_ENCRYPTED));
1661
1662 if (!r) {
1663 RADEON_DEC_ERR("Can't allocate context buffer.\n");
1664 return NULL;
1665 }
1666 } else if (fmt == PIPE_VIDEO_FORMAT_VP9) {
1667 unsigned ctx_size;
1668 uint8_t *ptr;
1669 bool r;
1670
1671 /* default probability + probability data */
1672 ctx_size = 2304 * 5;
1673
1674 if (((struct si_screen *)dec->screen)->info.vcn_ip_version >= VCN_2_0_0) {
1675 /* SRE collocated context data */
1676 ctx_size += 32 * 2 * 128 * 68;
1677 /* SMP collocated context data */
1678 ctx_size += 9 * 64 * 2 * 128 * 68;
1679 /* SDB left tile pixel */
1680 ctx_size += 8 * 2 * 2 * 8192;
1681 } else {
1682 ctx_size += 32 * 2 * 64 * 64;
1683 ctx_size += 9 * 64 * 2 * 64 * 64;
1684 ctx_size += 8 * 2 * 4096;
1685 }
1686
1687 if (dec->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2)
1688 ctx_size += 8 * 2 * 4096;
1689
1690 if (encrypted && dec->tmz_ctx) {
1691 r = si_vid_create_tmz_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1692 } else {
1693 r = si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1694 }
1695 if (!r) {
1696 RADEON_DEC_ERR("Can't allocate context buffer.\n");
1697 return NULL;
1698 }
1699
1700 /* ctx needs probs table */
1701 ptr = dec->ws->buffer_map(dec->ws, dec->ctx.res->buf, &dec->cs,
1702 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
1703 fill_probs_table(ptr);
1704 dec->ws->buffer_unmap(dec->ws, dec->ctx.res->buf);
1705 dec->bs_ptr = NULL;
1706 } else if (fmt == PIPE_VIDEO_FORMAT_HEVC) {
1707 unsigned ctx_size;
1708 bool r;
1709 if (dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)
1710 ctx_size = calc_ctx_size_h265_main10(dec, (struct pipe_h265_picture_desc *)picture);
1711 else
1712 ctx_size = calc_ctx_size_h265_main(dec);
1713
1714 if (encrypted && dec->tmz_ctx) {
1715 r = si_vid_create_tmz_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1716 } else {
1717 r = si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1718 }
1719 if (!r) {
1720 RADEON_DEC_ERR("Can't allocate context buffer.\n");
1721 return NULL;
1722 }
1723 }
1724 }
1725 if (encrypted != dec->ws->cs_is_secure(&dec->cs)) {
1726 dec->ws->cs_flush(&dec->cs, RADEON_FLUSH_TOGGLE_SECURE_SUBMISSION, NULL);
1727 }
1728
1729 if (dec->stream_type == RDECODE_CODEC_AV1)
1730 rvcn_dec_av1_film_grain_surface(&out_surf, (struct pipe_av1_picture_desc *)picture);
1731
1732 luma = (struct si_texture *)((struct vl_video_buffer *)out_surf)->resources[0];
1733 chroma = (struct si_texture *)((struct vl_video_buffer *)out_surf)->resources[1];
1734
1735 decode->dpb_size = (dec->dpb_type != DPB_DYNAMIC_TIER_2) ? dec->dpb.res->buf->size : 0;
1736
1737 /* When texture being created, the bo will be created with total size of planes,
1738 * and all planes point to the same buffer */
1739 assert(si_resource(((struct vl_video_buffer *)out_surf)->resources[0])->buf->size ==
1740 si_resource(((struct vl_video_buffer *)out_surf)->resources[1])->buf->size);
1741
1742 decode->dt_size = si_resource(((struct vl_video_buffer *)out_surf)->resources[0])->buf->size;
1743
1744 decode->sct_size = 0;
1745 decode->sc_coeff_size = 0;
1746
1747 decode->sw_ctxt_size = RDECODE_SESSION_CONTEXT_SIZE;
1748 decode->db_pitch = align(dec->base.width, dec->db_alignment);
1749
1750 if ((((struct si_screen*)dec->screen)->info.vcn_ip_version >= VCN_3_0_0) &&
1751 (dec->stream_type == RDECODE_CODEC_VP9 || dec->stream_type == RDECODE_CODEC_AV1 ||
1752 dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10))
1753 decode->db_aligned_height = align(dec->base.height, 64);
1754
1755 decode->db_surf_tile_config = 0;
1756 decode->db_array_mode = dec->addr_gfx_mode;
1757
1758 decode->dt_pitch = luma->surface.u.gfx9.surf_pitch * luma->surface.blk_w;
1759 decode->dt_uv_pitch = chroma->surface.u.gfx9.surf_pitch * chroma->surface.blk_w;
1760
1761 if (luma->surface.meta_offset) {
1762 RADEON_DEC_ERR("DCC surfaces not supported.\n");
1763 return NULL;
1764 }
1765
1766 decode->dt_tiling_mode = 0;
1767 decode->dt_swizzle_mode = luma->surface.u.gfx9.swizzle_mode;
1768 decode->dt_array_mode = dec->addr_gfx_mode;
1769 decode->dt_field_mode = ((struct vl_video_buffer *)out_surf)->base.interlaced;
1770 decode->dt_surf_tile_config = 0;
1771 decode->dt_uv_surf_tile_config = 0;
1772
1773 decode->dt_luma_top_offset = luma->surface.u.gfx9.surf_offset;
1774 decode->dt_chroma_top_offset = chroma->surface.u.gfx9.surf_offset;
1775 if (decode->dt_field_mode) {
1776 decode->dt_luma_bottom_offset =
1777 luma->surface.u.gfx9.surf_offset + luma->surface.u.gfx9.surf_slice_size;
1778 decode->dt_chroma_bottom_offset =
1779 chroma->surface.u.gfx9.surf_offset + chroma->surface.u.gfx9.surf_slice_size;
1780 } else {
1781 decode->dt_luma_bottom_offset = decode->dt_luma_top_offset;
1782 decode->dt_chroma_bottom_offset = decode->dt_chroma_top_offset;
1783 }
1784 decode->mif_wrc_en = sscreen->info.vcn_ip_version >= VCN_3_0_0;
1785 if (dec->stream_type == RDECODE_CODEC_AV1)
1786 decode->db_pitch_uv = chroma->surface.u.gfx9.surf_pitch * chroma->surface.blk_w;
1787
1788 if (encrypted) {
1789 assert(sscreen->info.has_tmz_support);
1790 set_drm_keys(drm, decrypt);
1791 }
1792
1793 if (dec->dpb_type == DPB_DYNAMIC_TIER_1) {
1794 decode->decode_flags |= (RDECODE_FLAGS_USE_DYNAMIC_DPB_MASK | RDECODE_FLAGS_USE_PAL_MASK);
1795 // Add decode flag for RESIZE_DPB ,when we do resize
1796 if (dpb_resize == true)
1797 decode->decode_flags |= RDECODE_FLAGS_DPB_RESIZE_MASK;
1798
1799 dynamic_dpb->dpbArraySize = NUM_VP9_REFS + 1;
1800 dynamic_dpb->dpbLumaPitch = align(dec->max_width, dec->db_alignment);
1801 dynamic_dpb->dpbLumaAlignedHeight = align(dec->max_height, dec->db_alignment);
1802 dynamic_dpb->dpbLumaAlignedSize =
1803 dynamic_dpb->dpbLumaPitch * dynamic_dpb->dpbLumaAlignedHeight;
1804 dynamic_dpb->dpbChromaPitch = dynamic_dpb->dpbLumaPitch >> 1;
1805 dynamic_dpb->dpbChromaAlignedHeight = dynamic_dpb->dpbLumaAlignedHeight >> 1;
1806 dynamic_dpb->dpbChromaAlignedSize =
1807 dynamic_dpb->dpbChromaPitch * dynamic_dpb->dpbChromaAlignedHeight * 2;
1808 dynamic_dpb->dpbReserved0[0] = dec->db_alignment;
1809
1810 if (dec->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2) {
1811 dynamic_dpb->dpbLumaAlignedSize = dynamic_dpb->dpbLumaAlignedSize * 3 / 2;
1812 dynamic_dpb->dpbChromaAlignedSize = dynamic_dpb->dpbChromaAlignedSize * 3 / 2;
1813 }
1814 }
1815
1816 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC)
1817 get_h265_reflist(hevc_reflist, (struct pipe_h265_picture_desc *)picture);
1818
1819 switch (u_reduce_video_profile(picture->profile)) {
1820 case PIPE_VIDEO_FORMAT_MPEG4_AVC: {
1821 rvcn_dec_message_avc_t avc = get_h264_msg(dec, target, (struct pipe_h264_picture_desc *)picture);
1822 memcpy(codec, (void *)&avc, sizeof(rvcn_dec_message_avc_t));
1823 index_codec->message_id = RDECODE_MESSAGE_AVC;
1824 index_codec->size = sizeof(rvcn_dec_message_avc_t);
1825 break;
1826 }
1827 case PIPE_VIDEO_FORMAT_HEVC: {
1828 rvcn_dec_message_hevc_t hevc =
1829 get_h265_msg(dec, target, (struct pipe_h265_picture_desc *)picture);
1830
1831 memcpy(codec, (void *)&hevc, sizeof(rvcn_dec_message_hevc_t));
1832 index_codec->message_id = RDECODE_MESSAGE_HEVC;
1833 index_codec->size = sizeof(rvcn_dec_message_hevc_t);
1834 break;
1835 }
1836 case PIPE_VIDEO_FORMAT_VC1: {
1837 rvcn_dec_message_vc1_t vc1 = get_vc1_msg((struct pipe_vc1_picture_desc *)picture);
1838
1839 memcpy(codec, (void *)&vc1, sizeof(rvcn_dec_message_vc1_t));
1840 if ((picture->profile == PIPE_VIDEO_PROFILE_VC1_SIMPLE) ||
1841 (picture->profile == PIPE_VIDEO_PROFILE_VC1_MAIN)) {
1842 decode->width_in_samples = align(decode->width_in_samples, 16) / 16;
1843 decode->height_in_samples = align(decode->height_in_samples, 16) / 16;
1844 }
1845 index_codec->message_id = RDECODE_MESSAGE_VC1;
1846 index_codec->size = sizeof(rvcn_dec_message_vc1_t);
1847 break;
1848 }
1849 case PIPE_VIDEO_FORMAT_MPEG12: {
1850 rvcn_dec_message_mpeg2_vld_t mpeg2 =
1851 get_mpeg2_msg(dec, (struct pipe_mpeg12_picture_desc *)picture);
1852
1853 memcpy(codec, (void *)&mpeg2, sizeof(rvcn_dec_message_mpeg2_vld_t));
1854 index_codec->message_id = RDECODE_MESSAGE_MPEG2_VLD;
1855 index_codec->size = sizeof(rvcn_dec_message_mpeg2_vld_t);
1856 break;
1857 }
1858 case PIPE_VIDEO_FORMAT_MPEG4: {
1859 rvcn_dec_message_mpeg4_asp_vld_t mpeg4 =
1860 get_mpeg4_msg(dec, (struct pipe_mpeg4_picture_desc *)picture);
1861
1862 memcpy(codec, (void *)&mpeg4, sizeof(rvcn_dec_message_mpeg4_asp_vld_t));
1863 index_codec->message_id = RDECODE_MESSAGE_MPEG4_ASP_VLD;
1864 index_codec->size = sizeof(rvcn_dec_message_mpeg4_asp_vld_t);
1865 break;
1866 }
1867 case PIPE_VIDEO_FORMAT_VP9: {
1868 rvcn_dec_message_vp9_t vp9 =
1869 get_vp9_msg(dec, target, (struct pipe_vp9_picture_desc *)picture);
1870
1871 memcpy(codec, (void *)&vp9, sizeof(rvcn_dec_message_vp9_t));
1872 index_codec->message_id = RDECODE_MESSAGE_VP9;
1873 index_codec->size = sizeof(rvcn_dec_message_vp9_t);
1874 break;
1875 }
1876 case PIPE_VIDEO_FORMAT_AV1: {
1877 rvcn_dec_message_av1_t av1 =
1878 get_av1_msg(dec, target, (struct pipe_av1_picture_desc *)picture);
1879
1880 memcpy(codec, (void *)&av1, sizeof(rvcn_dec_message_av1_t));
1881 index_codec->message_id = RDECODE_MESSAGE_AV1;
1882 index_codec->size = sizeof(rvcn_dec_message_av1_t);
1883
1884 if (dec->ctx.res == NULL) {
1885 unsigned ctx_size = ac_vcn_dec_calc_ctx_size_av1(dec->av1_version);
1886 uint8_t *ptr;
1887
1888 if (!si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT))
1889 RADEON_DEC_ERR("Can't allocate context buffer.\n");
1890
1891 ptr = dec->ws->buffer_map(dec->ws, dec->ctx.res->buf, &dec->cs, PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
1892
1893 ac_vcn_av1_init_probs(dec->av1_version, ptr);
1894 dec->ws->buffer_unmap(dec->ws, dec->ctx.res->buf);
1895 }
1896
1897 break;
1898 }
1899 default:
1900 assert(0);
1901 return NULL;
1902 }
1903
1904 header->total_size += index_codec->size;
1905
1906 if (dec->ctx.res)
1907 decode->hw_ctxt_size = dec->ctx.res->buf->size;
1908
1909 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
1910 if (rvcn_dec_dynamic_dpb_t2_message(dec, decode, dynamic_dpb_t2, encrypted))
1911 return NULL;
1912 } else if (((struct si_screen *)dec->screen)->info.vcn_ip_version == VCN_5_0_0 &&
1913 dec->dpb_type == DPB_MAX_RES)
1914 decode->db_swizzle_mode = RDECODE_VCN5_256B_D;
1915
1916 return luma->buffer.buf;
1917 }
1918
rvcn_dec_message_destroy(struct radeon_decoder * dec)1919 static void rvcn_dec_message_destroy(struct radeon_decoder *dec)
1920 {
1921 rvcn_dec_message_header_t *header = dec->msg;
1922
1923 memset(dec->msg, 0, sizeof(rvcn_dec_message_header_t));
1924 header->header_size = sizeof(rvcn_dec_message_header_t);
1925 header->total_size = sizeof(rvcn_dec_message_header_t) - sizeof(rvcn_dec_message_index_t);
1926 header->num_buffers = 0;
1927 header->msg_type = RDECODE_MSG_DESTROY;
1928 header->stream_handle = dec->stream_handle;
1929 header->status_report_feedback_number = 0;
1930 }
1931
rvcn_dec_message_feedback(struct radeon_decoder * dec)1932 static void rvcn_dec_message_feedback(struct radeon_decoder *dec)
1933 {
1934 rvcn_dec_feedback_header_t *header = (void *)dec->fb;
1935
1936 header->header_size = sizeof(rvcn_dec_feedback_header_t);
1937 header->total_size = sizeof(rvcn_dec_feedback_header_t);
1938 header->num_buffers = 0;
1939 }
1940
rvcn_dec_sq_tail(struct radeon_decoder * dec)1941 static void rvcn_dec_sq_tail(struct radeon_decoder *dec)
1942 {
1943 if (dec->vcn_dec_sw_ring == false)
1944 return;
1945
1946 rvcn_sq_tail(&dec->cs, &dec->sq);
1947 }
1948 /* flush IB to the hardware */
flush(struct radeon_decoder * dec,unsigned flags,struct pipe_fence_handle ** fence)1949 static int flush(struct radeon_decoder *dec, unsigned flags,
1950 struct pipe_fence_handle **fence)
1951 {
1952 struct si_screen *sscreen = (struct si_screen *)dec->screen;
1953
1954 rvcn_dec_sq_tail(dec);
1955
1956 if (sscreen->debug_flags & DBG(IB)) {
1957 struct ac_ib_parser ib_parser = {
1958 .f = stderr,
1959 .ib = dec->cs.current.buf,
1960 .num_dw = dec->cs.current.cdw,
1961 .gfx_level = sscreen->info.gfx_level,
1962 .vcn_version = sscreen->info.vcn_ip_version,
1963 .family = sscreen->info.family,
1964 .ip_type = dec->stream_type == RDECODE_CODEC_JPEG ? AMD_IP_VCN_JPEG :
1965 dec->vcn_dec_sw_ring ? AMD_IP_VCN_ENC : AMD_IP_VCN_DEC,
1966 };
1967 ac_parse_ib(&ib_parser, "IB");
1968 }
1969
1970 return dec->ws->cs_flush(&dec->cs, flags, fence);
1971 }
1972
1973 /* add a new set register command to the IB */
set_reg(struct radeon_decoder * dec,unsigned reg,uint32_t val)1974 static void set_reg(struct radeon_decoder *dec, unsigned reg, uint32_t val)
1975 {
1976 radeon_emit(&dec->cs, RDECODE_PKT0(reg >> 2, 0));
1977 radeon_emit(&dec->cs, val);
1978 }
1979
1980 /* send a command to the VCPU through the GPCOM registers */
send_cmd(struct radeon_decoder * dec,unsigned cmd,struct pb_buffer_lean * buf,uint32_t off,unsigned usage,enum radeon_bo_domain domain)1981 static void send_cmd(struct radeon_decoder *dec, unsigned cmd, struct pb_buffer_lean *buf, uint32_t off,
1982 unsigned usage, enum radeon_bo_domain domain)
1983 {
1984 uint64_t addr;
1985
1986 dec->ws->cs_add_buffer(&dec->cs, buf, usage | RADEON_USAGE_SYNCHRONIZED, domain);
1987 addr = dec->ws->buffer_get_virtual_address(buf);
1988 addr = addr + off;
1989
1990 if (dec->vcn_dec_sw_ring == false) {
1991 set_reg(dec, dec->reg.data0, addr);
1992 set_reg(dec, dec->reg.data1, addr >> 32);
1993 set_reg(dec, dec->reg.cmd, cmd << 1);
1994 return;
1995 }
1996
1997 if (!dec->cs.current.cdw) {
1998 rvcn_sq_header(&dec->cs, &dec->sq, false);
1999 rvcn_decode_ib_package_t *ib_header =
2000 (rvcn_decode_ib_package_t *)&(dec->cs.current.buf[dec->cs.current.cdw]);
2001
2002 ib_header->package_size = sizeof(struct rvcn_decode_buffer_s) +
2003 sizeof(struct rvcn_decode_ib_package_s);
2004 dec->cs.current.cdw++;
2005 ib_header->package_type = (RDECODE_IB_PARAM_DECODE_BUFFER);
2006 dec->cs.current.cdw++;
2007
2008 dec->decode_buffer =
2009 (rvcn_decode_buffer_t *)&(dec->cs.current.buf[dec->cs.current.cdw]);
2010
2011 dec->cs.current.cdw += sizeof(struct rvcn_decode_buffer_s) / 4;
2012 memset(dec->decode_buffer, 0, sizeof(struct rvcn_decode_buffer_s));
2013 }
2014
2015 switch(cmd) {
2016 case RDECODE_CMD_MSG_BUFFER:
2017 dec->decode_buffer->valid_buf_flag |= RDECODE_CMDBUF_FLAGS_MSG_BUFFER;
2018 dec->decode_buffer->msg_buffer_address_hi = (addr >> 32);
2019 dec->decode_buffer->msg_buffer_address_lo = (addr);
2020 break;
2021 case RDECODE_CMD_DPB_BUFFER:
2022 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_DPB_BUFFER);
2023 dec->decode_buffer->dpb_buffer_address_hi = (addr >> 32);
2024 dec->decode_buffer->dpb_buffer_address_lo = (addr);
2025 break;
2026 case RDECODE_CMD_DECODING_TARGET_BUFFER:
2027 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_DECODING_TARGET_BUFFER);
2028 dec->decode_buffer->target_buffer_address_hi = (addr >> 32);
2029 dec->decode_buffer->target_buffer_address_lo = (addr);
2030 break;
2031 case RDECODE_CMD_FEEDBACK_BUFFER:
2032 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_FEEDBACK_BUFFER);
2033 dec->decode_buffer->feedback_buffer_address_hi = (addr >> 32);
2034 dec->decode_buffer->feedback_buffer_address_lo = (addr);
2035 break;
2036 case RDECODE_CMD_PROB_TBL_BUFFER:
2037 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_PROB_TBL_BUFFER);
2038 dec->decode_buffer->prob_tbl_buffer_address_hi = (addr >> 32);
2039 dec->decode_buffer->prob_tbl_buffer_address_lo = (addr);
2040 break;
2041 case RDECODE_CMD_SESSION_CONTEXT_BUFFER:
2042 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_SESSION_CONTEXT_BUFFER);
2043 dec->decode_buffer->session_contex_buffer_address_hi = (addr >> 32);
2044 dec->decode_buffer->session_contex_buffer_address_lo = (addr);
2045 break;
2046 case RDECODE_CMD_BITSTREAM_BUFFER:
2047 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_BITSTREAM_BUFFER);
2048 dec->decode_buffer->bitstream_buffer_address_hi = (addr >> 32);
2049 dec->decode_buffer->bitstream_buffer_address_lo = (addr);
2050 break;
2051 case RDECODE_CMD_IT_SCALING_TABLE_BUFFER:
2052 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_IT_SCALING_BUFFER);
2053 dec->decode_buffer->it_sclr_table_buffer_address_hi = (addr >> 32);
2054 dec->decode_buffer->it_sclr_table_buffer_address_lo = (addr);
2055 break;
2056 case RDECODE_CMD_CONTEXT_BUFFER:
2057 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_CONTEXT_BUFFER);
2058 dec->decode_buffer->context_buffer_address_hi = (addr >> 32);
2059 dec->decode_buffer->context_buffer_address_lo = (addr);
2060 break;
2061 default:
2062 printf("Not Support!");
2063 }
2064 }
2065
2066 /* do the codec needs an IT buffer ?*/
have_it(struct radeon_decoder * dec)2067 static bool have_it(struct radeon_decoder *dec)
2068 {
2069 return dec->stream_type == RDECODE_CODEC_H264_PERF || dec->stream_type == RDECODE_CODEC_H265;
2070 }
2071
2072 /* do the codec needs an probs buffer? */
have_probs(struct radeon_decoder * dec)2073 static bool have_probs(struct radeon_decoder *dec)
2074 {
2075 return (dec->stream_type == RDECODE_CODEC_VP9 || dec->stream_type == RDECODE_CODEC_AV1);
2076 }
2077
2078 /* map the next available message/feedback/itscaling buffer */
map_msg_fb_it_probs_buf(struct radeon_decoder * dec)2079 static void map_msg_fb_it_probs_buf(struct radeon_decoder *dec)
2080 {
2081 struct rvid_buffer *buf;
2082 uint8_t *ptr;
2083
2084 /* grab the current message/feedback buffer */
2085 buf = &dec->msg_fb_it_probs_buffers[dec->cur_buffer];
2086
2087 /* and map it for CPU access */
2088 ptr =
2089 dec->ws->buffer_map(dec->ws, buf->res->buf, &dec->cs, PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2090
2091 /* calc buffer offsets */
2092 dec->msg = ptr;
2093
2094 dec->fb = (uint32_t *)(ptr + FB_BUFFER_OFFSET);
2095 if (have_it(dec))
2096 dec->it = (uint8_t *)(ptr + FB_BUFFER_OFFSET + FB_BUFFER_SIZE);
2097 else if (have_probs(dec))
2098 dec->probs = (uint8_t *)(ptr + FB_BUFFER_OFFSET + FB_BUFFER_SIZE);
2099 }
2100
2101 /* unmap and send a message command to the VCPU */
send_msg_buf(struct radeon_decoder * dec)2102 static void send_msg_buf(struct radeon_decoder *dec)
2103 {
2104 struct rvid_buffer *buf;
2105
2106 /* ignore the request if message/feedback buffer isn't mapped */
2107 if (!dec->msg || !dec->fb)
2108 return;
2109
2110 /* grab the current message buffer */
2111 buf = &dec->msg_fb_it_probs_buffers[dec->cur_buffer];
2112
2113 /* unmap the buffer */
2114 dec->ws->buffer_unmap(dec->ws, buf->res->buf);
2115 dec->bs_ptr = NULL;
2116 dec->msg = NULL;
2117 dec->fb = NULL;
2118 dec->it = NULL;
2119 dec->probs = NULL;
2120
2121 if (dec->sessionctx.res)
2122 send_cmd(dec, RDECODE_CMD_SESSION_CONTEXT_BUFFER, dec->sessionctx.res->buf, 0,
2123 RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
2124
2125 /* and send it to the hardware */
2126 send_cmd(dec, RDECODE_CMD_MSG_BUFFER, buf->res->buf, 0, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
2127 }
2128
2129 /* cycle to the next set of buffers */
next_buffer(struct radeon_decoder * dec)2130 static void next_buffer(struct radeon_decoder *dec)
2131 {
2132 ++dec->cur_buffer;
2133 dec->cur_buffer %= dec->num_dec_bufs;
2134 }
2135
calc_ctx_size_h264_perf(struct radeon_decoder * dec)2136 static unsigned calc_ctx_size_h264_perf(struct radeon_decoder *dec)
2137 {
2138 unsigned width_in_mb, height_in_mb, ctx_size;
2139 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
2140 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
2141
2142 unsigned max_references = dec->base.max_references + 1;
2143
2144 // picture width & height in 16 pixel units
2145 width_in_mb = width / VL_MACROBLOCK_WIDTH;
2146 height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
2147
2148 unsigned fs_in_mb = width_in_mb * height_in_mb;
2149 unsigned num_dpb_buffer_lean;
2150 switch (dec->base.level) {
2151 case 30:
2152 num_dpb_buffer_lean = 8100 / fs_in_mb;
2153 break;
2154 case 31:
2155 num_dpb_buffer_lean = 18000 / fs_in_mb;
2156 break;
2157 case 32:
2158 num_dpb_buffer_lean = 20480 / fs_in_mb;
2159 break;
2160 case 41:
2161 num_dpb_buffer_lean = 32768 / fs_in_mb;
2162 break;
2163 case 42:
2164 num_dpb_buffer_lean = 34816 / fs_in_mb;
2165 break;
2166 case 50:
2167 num_dpb_buffer_lean = 110400 / fs_in_mb;
2168 break;
2169 case 51:
2170 num_dpb_buffer_lean = 184320 / fs_in_mb;
2171 break;
2172 default:
2173 num_dpb_buffer_lean = 184320 / fs_in_mb;
2174 break;
2175 }
2176 num_dpb_buffer_lean++;
2177 max_references = MAX2(MIN2(NUM_H264_REFS, num_dpb_buffer_lean), max_references);
2178 ctx_size = max_references * align(width_in_mb * height_in_mb * 192, 256);
2179
2180 return ctx_size;
2181 }
2182
2183 /* calculate size of reference picture buffer */
calc_dpb_size(struct radeon_decoder * dec)2184 static unsigned calc_dpb_size(struct radeon_decoder *dec)
2185 {
2186 unsigned width_in_mb, height_in_mb, image_size, dpb_size;
2187
2188 // always align them to MB size for dpb calculation
2189 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
2190 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
2191
2192 // always one more for currently decoded picture
2193 unsigned max_references = dec->base.max_references + 1;
2194
2195 // aligned size of a single frame
2196 image_size = align(width, dec->db_alignment) * align(height, dec->db_alignment);
2197 image_size += image_size / 2;
2198 image_size = align(image_size, 1024);
2199
2200 // picture width & height in 16 pixel units
2201 width_in_mb = width / VL_MACROBLOCK_WIDTH;
2202 height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
2203
2204 switch (u_reduce_video_profile(dec->base.profile)) {
2205 case PIPE_VIDEO_FORMAT_MPEG4_AVC: {
2206 unsigned fs_in_mb = width_in_mb * height_in_mb;
2207 unsigned num_dpb_buffer_lean;
2208
2209 switch (dec->base.level) {
2210 case 30:
2211 num_dpb_buffer_lean = 8100 / fs_in_mb;
2212 break;
2213 case 31:
2214 num_dpb_buffer_lean = 18000 / fs_in_mb;
2215 break;
2216 case 32:
2217 num_dpb_buffer_lean = 20480 / fs_in_mb;
2218 break;
2219 case 41:
2220 num_dpb_buffer_lean = 32768 / fs_in_mb;
2221 break;
2222 case 42:
2223 num_dpb_buffer_lean = 34816 / fs_in_mb;
2224 break;
2225 case 50:
2226 num_dpb_buffer_lean = 110400 / fs_in_mb;
2227 break;
2228 case 51:
2229 num_dpb_buffer_lean = 184320 / fs_in_mb;
2230 break;
2231 default:
2232 num_dpb_buffer_lean = 184320 / fs_in_mb;
2233 break;
2234 }
2235 num_dpb_buffer_lean++;
2236 max_references = MAX2(MIN2(NUM_H264_REFS, num_dpb_buffer_lean), max_references);
2237 dpb_size = image_size * max_references;
2238 break;
2239 }
2240
2241 case PIPE_VIDEO_FORMAT_HEVC:
2242 if (dec->base.width * dec->base.height >= 4096 * 2000)
2243 max_references = MAX2(max_references, 8);
2244 else
2245 max_references = MAX2(max_references, 17);
2246
2247 if (dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)
2248 dpb_size = align((align(width, dec->db_alignment) *
2249 align(height, dec->db_alignment) * 9) / 4, 256) * max_references;
2250 else
2251 dpb_size = align((align(width, dec->db_alignment) *
2252 align(height, dec->db_alignment) * 3) / 2, 256) * max_references;
2253 break;
2254
2255 case PIPE_VIDEO_FORMAT_VC1:
2256 // the firmware seems to always assume a minimum of ref frames
2257 max_references = MAX2(NUM_VC1_REFS, max_references);
2258
2259 // reference picture buffer
2260 dpb_size = image_size * max_references;
2261
2262 // CONTEXT_BUFFER
2263 dpb_size += width_in_mb * height_in_mb * 128;
2264
2265 // IT surface buffer
2266 dpb_size += width_in_mb * 64;
2267
2268 // DB surface buffer
2269 dpb_size += width_in_mb * 128;
2270
2271 // BP
2272 dpb_size += align(MAX2(width_in_mb, height_in_mb) * 7 * 16, 64);
2273 break;
2274
2275 case PIPE_VIDEO_FORMAT_MPEG12:
2276 // reference picture buffer, must be big enough for all frames
2277 dpb_size = image_size * NUM_MPEG2_REFS;
2278 break;
2279
2280 case PIPE_VIDEO_FORMAT_MPEG4:
2281 // reference picture buffer
2282 dpb_size = image_size * max_references;
2283
2284 // CM
2285 dpb_size += width_in_mb * height_in_mb * 64;
2286
2287 // IT surface buffer
2288 dpb_size += align(width_in_mb * height_in_mb * 32, 64);
2289
2290 dpb_size = MAX2(dpb_size, 30 * 1024 * 1024);
2291 break;
2292
2293 case PIPE_VIDEO_FORMAT_VP9:
2294 max_references = MAX2(max_references, 9);
2295
2296 if (dec->dpb_type == DPB_MAX_RES)
2297 dpb_size = (((struct si_screen *)dec->screen)->info.vcn_ip_version >= VCN_2_0_0)
2298 ? (8192 * 4320 * 3 / 2) * max_references
2299 : (4096 * 3000 * 3 / 2) * max_references;
2300 else
2301 dpb_size = (align(dec->base.width, dec->db_alignment) *
2302 align(dec->base.height, dec->db_alignment) * 3 / 2) * max_references;
2303
2304 if (dec->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2)
2305 dpb_size = dpb_size * 3 / 2;
2306 break;
2307
2308 case PIPE_VIDEO_FORMAT_AV1:
2309 max_references = MAX2(max_references, 9);
2310 dpb_size = 8192 * 4320 * 3 / 2 * max_references * 3 / 2;
2311 break;
2312
2313 case PIPE_VIDEO_FORMAT_JPEG:
2314 dpb_size = 0;
2315 break;
2316
2317 default:
2318 // something is missing here
2319 assert(0);
2320
2321 // at least use a sane default value
2322 dpb_size = 32 * 1024 * 1024;
2323 break;
2324 }
2325 return dpb_size;
2326 }
2327
2328 /**
2329 * destroy this video decoder
2330 */
radeon_dec_destroy(struct pipe_video_codec * decoder)2331 static void radeon_dec_destroy(struct pipe_video_codec *decoder)
2332 {
2333 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2334 unsigned i;
2335
2336 assert(decoder);
2337
2338 if (dec->bs_ptr) {
2339 dec->ws->buffer_unmap(dec->ws, dec->bs_buffers[dec->cur_buffer].res->buf);
2340 dec->bs_ptr = NULL;
2341 }
2342
2343 if (dec->msg) {
2344 dec->ws->buffer_unmap(dec->ws, dec->msg_fb_it_probs_buffers[dec->cur_buffer].res->buf);
2345 dec->msg = NULL;
2346 }
2347
2348 if (dec->stream_type != RDECODE_CODEC_JPEG) {
2349 struct pipe_fence_handle *fence = NULL;
2350 map_msg_fb_it_probs_buf(dec);
2351 rvcn_dec_message_destroy(dec);
2352 send_msg_buf(dec);
2353 flush(dec, 0, &fence);
2354 dec->ws->fence_wait(dec->ws, fence, OS_TIMEOUT_INFINITE);
2355 dec->ws->fence_reference(dec->ws, &fence, NULL);
2356 }
2357
2358 dec->ws->cs_destroy(&dec->cs);
2359 if (dec->ectx)
2360 dec->ectx->destroy(dec->ectx);
2361
2362 if (dec->stream_type == RDECODE_CODEC_JPEG) {
2363 for (i = 0; i < dec->njctx; i++) {
2364 dec->ws->cs_destroy(&dec->jcs[i]);
2365 dec->ws->ctx_destroy(dec->jctx[i]);
2366 }
2367 }
2368
2369 if (dec->msg_fb_it_probs_buffers && dec->bs_buffers) {
2370 for (i = 0; i < dec->num_dec_bufs; ++i) {
2371 si_vid_destroy_buffer(&dec->msg_fb_it_probs_buffers[i]);
2372 si_vid_destroy_buffer(&dec->bs_buffers[i]);
2373 }
2374 FREE(dec->msg_fb_it_probs_buffers);
2375 FREE(dec->bs_buffers);
2376 }
2377 dec->num_dec_bufs = 0;
2378
2379 if (dec->dpb_type != DPB_DYNAMIC_TIER_2) {
2380 si_vid_destroy_buffer(&dec->dpb);
2381 } else {
2382 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
2383 list_del(&d->list);
2384 d->vbuf->destroy(d->vbuf);
2385 FREE(d);
2386 }
2387 }
2388 si_vid_destroy_buffer(&dec->ctx);
2389 si_vid_destroy_buffer(&dec->sessionctx);
2390
2391 FREE(dec->jcs);
2392 FREE(dec->jctx);
2393 FREE(dec);
2394 }
2395
2396 /**
2397 * start decoding of a new frame
2398 */
radeon_dec_begin_frame(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2399 static void radeon_dec_begin_frame(struct pipe_video_codec *decoder,
2400 struct pipe_video_buffer *target,
2401 struct pipe_picture_desc *picture)
2402 {
2403 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2404 uintptr_t frame;
2405
2406 assert(decoder);
2407
2408 if (dec->error)
2409 return;
2410
2411 switch (dec->stream_type) {
2412 case RDECODE_CODEC_VP9: {
2413 struct pipe_vp9_picture_desc *pic = (struct pipe_vp9_picture_desc *)picture;
2414 /* Only 10 bit is supported for Profile 2 */
2415 if (pic->picture_parameter.bit_depth > 10) {
2416 dec->error = true;
2417 return;
2418 }
2419 break;
2420 }
2421 case RDECODE_CODEC_AV1: {
2422 struct pipe_av1_picture_desc *pic = (struct pipe_av1_picture_desc *)picture;
2423 /* Only 4:2:0 is supported for Profile 2 */
2424 if (!pic->picture_parameter.seq_info_fields.subsampling_x ||
2425 !pic->picture_parameter.seq_info_fields.subsampling_y) {
2426 dec->error = true;
2427 return;
2428 }
2429 break;
2430 }
2431 default:
2432 break;
2433 }
2434
2435 frame = ++dec->frame_number;
2436 if (dec->stream_type != RDECODE_CODEC_VP9 && dec->stream_type != RDECODE_CODEC_AV1
2437 && dec->stream_type != RDECODE_CODEC_H264_PERF)
2438 vl_video_buffer_set_associated_data(target, decoder, (void *)frame,
2439 &radeon_dec_destroy_associated_data);
2440
2441 dec->bs_size = 0;
2442 dec->bs_ptr = dec->ws->buffer_map(dec->ws, dec->bs_buffers[dec->cur_buffer].res->buf, &dec->cs,
2443 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2444 }
2445
2446 /**
2447 * decode a macroblock
2448 */
radeon_dec_decode_macroblock(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture,const struct pipe_macroblock * macroblocks,unsigned num_macroblocks)2449 static void radeon_dec_decode_macroblock(struct pipe_video_codec *decoder,
2450 struct pipe_video_buffer *target,
2451 struct pipe_picture_desc *picture,
2452 const struct pipe_macroblock *macroblocks,
2453 unsigned num_macroblocks)
2454 {
2455 /* not supported (yet) */
2456 assert(0);
2457 }
2458
2459 /**
2460 * decode a bitstream
2461 */
radeon_dec_decode_bitstream(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture,unsigned num_buffers,const void * const * buffers,const unsigned * sizes)2462 static void radeon_dec_decode_bitstream(struct pipe_video_codec *decoder,
2463 struct pipe_video_buffer *target,
2464 struct pipe_picture_desc *picture, unsigned num_buffers,
2465 const void *const *buffers, const unsigned *sizes)
2466 {
2467 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2468 unsigned i;
2469
2470 assert(decoder);
2471
2472 if (dec->error)
2473 return;
2474
2475 if (!dec->bs_ptr) {
2476 RADEON_DEC_ERR("Invalid bitstream ptr!\n");
2477 return;
2478 }
2479
2480 unsigned long total_bs_size = dec->bs_size;
2481 for (i = 0; i < num_buffers; ++i)
2482 total_bs_size += sizes[i];
2483
2484 struct rvid_buffer *buf = &dec->bs_buffers[dec->cur_buffer];
2485
2486 if (total_bs_size > buf->res->buf->size) {
2487 dec->ws->buffer_unmap(dec->ws, buf->res->buf);
2488 dec->bs_ptr = NULL;
2489
2490 total_bs_size = align(total_bs_size, 128);
2491
2492 if (!dec->bs_size) {
2493 struct rvid_buffer old_buf = *buf;
2494 if (!si_vid_create_buffer(dec->screen, buf, total_bs_size, buf->usage)) {
2495 RADEON_DEC_ERR("Can't create bitstream buffer!");
2496 return;
2497 }
2498 si_vid_destroy_buffer(&old_buf);
2499 } else if (!si_vid_resize_buffer(dec->base.context, &dec->cs, buf, total_bs_size, NULL)) {
2500 RADEON_DEC_ERR("Can't resize bitstream buffer!");
2501 return;
2502 }
2503
2504 dec->bs_ptr = dec->ws->buffer_map(dec->ws, buf->res->buf, &dec->cs,
2505 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2506 if (!dec->bs_ptr)
2507 return;
2508
2509 dec->bs_ptr += dec->bs_size;
2510 }
2511
2512 for (i = 0; i < num_buffers; ++i) {
2513 memcpy(dec->bs_ptr, buffers[i], sizes[i]);
2514 dec->bs_size += sizes[i];
2515 dec->bs_ptr += sizes[i];
2516 }
2517 }
2518
2519 /**
2520 * send cmd for vcn dec
2521 */
send_cmd_dec(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2522 bool send_cmd_dec(struct radeon_decoder *dec, struct pipe_video_buffer *target,
2523 struct pipe_picture_desc *picture)
2524 {
2525 struct pb_buffer_lean *dt;
2526 struct rvid_buffer *msg_fb_it_probs_buf, *bs_buf;
2527
2528 msg_fb_it_probs_buf = &dec->msg_fb_it_probs_buffers[dec->cur_buffer];
2529 bs_buf = &dec->bs_buffers[dec->cur_buffer];
2530
2531 memset(dec->bs_ptr, 0, align(dec->bs_size, 128) - dec->bs_size);
2532 dec->ws->buffer_unmap(dec->ws, bs_buf->res->buf);
2533 dec->bs_ptr = NULL;
2534
2535 map_msg_fb_it_probs_buf(dec);
2536 dt = rvcn_dec_message_decode(dec, target, picture);
2537 if (!dt)
2538 return false;
2539 rvcn_dec_message_feedback(dec);
2540 send_msg_buf(dec);
2541
2542 if (dec->dpb_type != DPB_DYNAMIC_TIER_2)
2543 send_cmd(dec, RDECODE_CMD_DPB_BUFFER, dec->dpb.res->buf, 0, RADEON_USAGE_READWRITE,
2544 RADEON_DOMAIN_VRAM);
2545 if (dec->ctx.res)
2546 send_cmd(dec, RDECODE_CMD_CONTEXT_BUFFER, dec->ctx.res->buf, 0, RADEON_USAGE_READWRITE,
2547 RADEON_DOMAIN_VRAM);
2548 send_cmd(dec, RDECODE_CMD_BITSTREAM_BUFFER, bs_buf->res->buf, 0, RADEON_USAGE_READ,
2549 RADEON_DOMAIN_GTT);
2550 send_cmd(dec, RDECODE_CMD_DECODING_TARGET_BUFFER, dt, 0, RADEON_USAGE_WRITE, RADEON_DOMAIN_VRAM);
2551 send_cmd(dec, RDECODE_CMD_FEEDBACK_BUFFER, msg_fb_it_probs_buf->res->buf, FB_BUFFER_OFFSET,
2552 RADEON_USAGE_WRITE, RADEON_DOMAIN_GTT);
2553 if (have_it(dec))
2554 send_cmd(dec, RDECODE_CMD_IT_SCALING_TABLE_BUFFER, msg_fb_it_probs_buf->res->buf,
2555 FB_BUFFER_OFFSET + FB_BUFFER_SIZE, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
2556 else if (have_probs(dec))
2557 send_cmd(dec, RDECODE_CMD_PROB_TBL_BUFFER, msg_fb_it_probs_buf->res->buf,
2558 FB_BUFFER_OFFSET + FB_BUFFER_SIZE, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
2559
2560 if (dec->vcn_dec_sw_ring == false)
2561 set_reg(dec, dec->reg.cntl, 1);
2562
2563 return true;
2564 }
2565
2566 /**
2567 * end decoding of the current frame
2568 */
radeon_dec_end_frame(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2569 static int radeon_dec_end_frame(struct pipe_video_codec *decoder, struct pipe_video_buffer *target,
2570 struct pipe_picture_desc *picture)
2571 {
2572 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2573
2574 assert(decoder);
2575
2576 if (dec->error)
2577 return 1;
2578
2579 if (!dec->send_cmd(dec, target, picture))
2580 return 1;
2581
2582 flush(dec, picture->flush_flags, picture->fence);
2583
2584 next_buffer(dec);
2585 return 0;
2586 }
2587
radeon_dec_jpeg_check_format(struct radeon_decoder * dec,enum pipe_format format,unsigned sampling_factor)2588 static bool radeon_dec_jpeg_check_format(struct radeon_decoder *dec, enum pipe_format format,
2589 unsigned sampling_factor)
2590 {
2591 enum pipe_format expected_format;
2592
2593 switch (sampling_factor) {
2594 case 0x221111:
2595 expected_format = PIPE_FORMAT_NV12;
2596 break;
2597 case 0x211111:
2598 case 0x221212:
2599 case 0x222121:
2600 expected_format = PIPE_FORMAT_YUYV;
2601 break;
2602 case 0x111111:
2603 case 0x222222:
2604 case 0x444444:
2605 expected_format = PIPE_FORMAT_Y8_U8_V8_444_UNORM;
2606 break;
2607 case 0x121111:
2608 expected_format = PIPE_FORMAT_Y8_U8_V8_440_UNORM;
2609 break;
2610 case 0x11:
2611 case 0x44:
2612 expected_format = PIPE_FORMAT_Y8_400_UNORM;
2613 break;
2614 default:
2615 RADEON_DEC_ERR("Unsupported sampling factor 0x%x\n", sampling_factor);
2616 return false;
2617 }
2618
2619 /* Format conversion */
2620 if (format == PIPE_FORMAT_R8G8B8A8_UNORM ||
2621 format == PIPE_FORMAT_A8R8G8B8_UNORM ||
2622 format == PIPE_FORMAT_R8_G8_B8_UNORM)
2623 return true;
2624
2625 return expected_format == format;
2626 }
2627
2628 /**
2629 * end decoding of the current jpeg frame
2630 */
radeon_dec_jpeg_end_frame(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2631 static int radeon_dec_jpeg_end_frame(struct pipe_video_codec *decoder, struct pipe_video_buffer *target,
2632 struct pipe_picture_desc *picture)
2633 {
2634 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2635 struct pipe_mjpeg_picture_desc *pic = (struct pipe_mjpeg_picture_desc *)picture;
2636
2637 assert(decoder);
2638
2639 if (!radeon_dec_jpeg_check_format(dec, target->buffer_format, pic->picture_parameter.sampling_factor))
2640 RADEON_DEC_ERR("Decode format check failed\n");
2641
2642 if (dec->error)
2643 return 1;
2644
2645 dec->jpg.crop_x = ROUND_DOWN_TO(pic->picture_parameter.crop_x, VL_MACROBLOCK_WIDTH);
2646 dec->jpg.crop_y = ROUND_DOWN_TO(pic->picture_parameter.crop_y, VL_MACROBLOCK_HEIGHT);
2647 dec->jpg.crop_width = align(pic->picture_parameter.crop_width, VL_MACROBLOCK_WIDTH);
2648 dec->jpg.crop_height = align(pic->picture_parameter.crop_height, VL_MACROBLOCK_HEIGHT);
2649 if (dec->jpg.crop_x + dec->jpg.crop_width > pic->picture_parameter.picture_width)
2650 dec->jpg.crop_width = 0;
2651 if (dec->jpg.crop_y + dec->jpg.crop_height > pic->picture_parameter.picture_height)
2652 dec->jpg.crop_height = 0;
2653 dec->send_cmd(dec, target, picture);
2654 dec->ws->cs_flush(&dec->jcs[dec->cb_idx], picture->flush_flags, NULL);
2655 next_buffer(dec);
2656 dec->cb_idx = (dec->cb_idx+1) % dec->njctx;
2657 return 0;
2658 }
2659
2660 /**
2661 * flush any outstanding command buffers to the hardware
2662 */
radeon_dec_flush(struct pipe_video_codec * decoder)2663 static void radeon_dec_flush(struct pipe_video_codec *decoder)
2664 {
2665 }
2666
radeon_dec_fence_wait(struct pipe_video_codec * decoder,struct pipe_fence_handle * fence,uint64_t timeout)2667 static int radeon_dec_fence_wait(struct pipe_video_codec *decoder,
2668 struct pipe_fence_handle *fence,
2669 uint64_t timeout)
2670 {
2671 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2672
2673 return dec->ws->fence_wait(dec->ws, fence, timeout);
2674 }
2675
radeon_dec_destroy_fence(struct pipe_video_codec * decoder,struct pipe_fence_handle * fence)2676 static void radeon_dec_destroy_fence(struct pipe_video_codec *decoder,
2677 struct pipe_fence_handle *fence)
2678 {
2679 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2680
2681 dec->ws->fence_reference(dec->ws, &fence, NULL);
2682 }
2683
2684 /**
2685 * create and HW decoder
2686 */
radeon_create_decoder(struct pipe_context * context,const struct pipe_video_codec * templ)2687 struct pipe_video_codec *radeon_create_decoder(struct pipe_context *context,
2688 const struct pipe_video_codec *templ)
2689 {
2690 struct si_context *sctx = (struct si_context *)context;
2691 struct radeon_winsys *ws = sctx->ws;
2692 unsigned width = templ->width, height = templ->height;
2693 unsigned bs_buf_size, stream_type = 0, ring = AMD_IP_VCN_DEC;
2694 struct radeon_decoder *dec;
2695 int r, i;
2696
2697 switch (u_reduce_video_profile(templ->profile)) {
2698 case PIPE_VIDEO_FORMAT_MPEG12:
2699 if (templ->entrypoint > PIPE_VIDEO_ENTRYPOINT_BITSTREAM)
2700 return vl_create_mpeg12_decoder(context, templ);
2701 stream_type = RDECODE_CODEC_MPEG2_VLD;
2702 break;
2703 case PIPE_VIDEO_FORMAT_MPEG4:
2704 width = align(width, VL_MACROBLOCK_WIDTH);
2705 height = align(height, VL_MACROBLOCK_HEIGHT);
2706 stream_type = RDECODE_CODEC_MPEG4;
2707 break;
2708 case PIPE_VIDEO_FORMAT_VC1:
2709 stream_type = RDECODE_CODEC_VC1;
2710 break;
2711 case PIPE_VIDEO_FORMAT_MPEG4_AVC:
2712 width = align(width, VL_MACROBLOCK_WIDTH);
2713 height = align(height, VL_MACROBLOCK_HEIGHT);
2714 stream_type = RDECODE_CODEC_H264_PERF;
2715 break;
2716 case PIPE_VIDEO_FORMAT_HEVC:
2717 stream_type = RDECODE_CODEC_H265;
2718 break;
2719 case PIPE_VIDEO_FORMAT_VP9:
2720 stream_type = RDECODE_CODEC_VP9;
2721 break;
2722 case PIPE_VIDEO_FORMAT_AV1:
2723 stream_type = RDECODE_CODEC_AV1;
2724 break;
2725 case PIPE_VIDEO_FORMAT_JPEG:
2726 stream_type = RDECODE_CODEC_JPEG;
2727 ring = AMD_IP_VCN_JPEG;
2728 break;
2729 default:
2730 assert(0);
2731 break;
2732 }
2733
2734 dec = CALLOC_STRUCT(radeon_decoder);
2735
2736 if (!dec)
2737 return NULL;
2738
2739 if (sctx->vcn_has_ctx) {
2740 dec->ectx = context->screen->context_create(context->screen, NULL, PIPE_CONTEXT_COMPUTE_ONLY);
2741 if (!dec->ectx)
2742 sctx->vcn_has_ctx = false;
2743 }
2744
2745 dec->base = *templ;
2746 dec->base.context = (sctx->vcn_has_ctx) ? dec->ectx : context;
2747 dec->base.width = width;
2748 dec->base.height = height;
2749 dec->max_width = width;
2750 dec->max_height = height;
2751 dec->base.destroy = radeon_dec_destroy;
2752 dec->base.begin_frame = radeon_dec_begin_frame;
2753 dec->base.decode_macroblock = radeon_dec_decode_macroblock;
2754 dec->base.decode_bitstream = radeon_dec_decode_bitstream;
2755 dec->base.end_frame = radeon_dec_end_frame;
2756 dec->base.flush = radeon_dec_flush;
2757 dec->base.fence_wait = radeon_dec_fence_wait;
2758 dec->base.destroy_fence = radeon_dec_destroy_fence;
2759
2760 dec->stream_type = stream_type;
2761 dec->stream_handle = si_vid_alloc_stream_handle();
2762 dec->screen = context->screen;
2763 dec->ws = ws;
2764
2765 if (u_reduce_video_profile(templ->profile) != PIPE_VIDEO_FORMAT_JPEG &&
2766 (sctx->vcn_ip_ver >= VCN_4_0_0)) {
2767 dec->vcn_dec_sw_ring = true;
2768 ring = AMD_IP_VCN_UNIFIED;
2769 }
2770
2771 dec->sq.signature_ib_total_size_in_dw = NULL;
2772 dec->sq.signature_ib_checksum = NULL;
2773 dec->sq.engine_ib_size_of_packages = NULL;
2774
2775 if (!ws->cs_create(&dec->cs,
2776 (sctx->vcn_has_ctx) ? ((struct si_context *)dec->ectx)->ctx : sctx->ctx,
2777 ring, NULL, NULL)) {
2778 RADEON_DEC_ERR("Can't get command submission context.\n");
2779 goto error;
2780 }
2781
2782 if (dec->stream_type == RDECODE_CODEC_JPEG) {
2783
2784 if (((struct si_screen*)dec->screen)->info.ip[AMD_IP_VCN_JPEG].num_instances > 1 &&
2785 ((struct si_screen*)dec->screen)->info.ip[AMD_IP_VCN_JPEG].num_instances <= MAX_JPEG_INST)
2786 dec->njctx = ((struct si_screen*)dec->screen)->info.ip[AMD_IP_VCN_JPEG].num_instances;
2787 else
2788 dec->njctx = 1;
2789
2790 dec->jctx = (struct radeon_winsys_ctx **) CALLOC(dec->njctx,
2791 sizeof(struct radeon_winsys_ctx *));
2792 dec->jcs = (struct radeon_cmdbuf *) CALLOC(dec->njctx, sizeof(struct radeon_cmdbuf));
2793 if(!dec->jctx || !dec->jcs)
2794 goto err;
2795 for (i = 0; i < dec->njctx; i++) {
2796 /* Initialize the context handle and the command stream. */
2797 dec->jctx[i] = dec->ws->ctx_create(dec->ws, RADEON_CTX_PRIORITY_MEDIUM,
2798 sctx->context_flags & PIPE_CONTEXT_LOSE_CONTEXT_ON_RESET);
2799 if (!sctx->ctx)
2800 goto error;
2801 if (!dec->ws->cs_create(&dec->jcs[i], dec->jctx[i], ring, NULL, NULL)) {
2802 RADEON_DEC_ERR("Can't get additional command submission context for mJPEG.\n");
2803 goto error;
2804 }
2805 }
2806 dec->base.end_frame = radeon_dec_jpeg_end_frame;
2807 dec->cb_idx = 0;
2808 }
2809
2810 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); i++)
2811 dec->render_pic_list[i] = NULL;
2812
2813 if ((sctx->vcn_ip_ver >= VCN_3_0_0) && (stream_type == RDECODE_CODEC_H264_PERF)) {
2814 for (i = 0; i < ARRAY_SIZE(dec->h264_valid_ref_num); i++)
2815 dec->h264_valid_ref_num[i] = (unsigned) -1;
2816 for (i = 0; i < ARRAY_SIZE(dec->h264_valid_poc_num); i++)
2817 dec->h264_valid_poc_num[i] = (unsigned) -1;
2818 }
2819
2820 if (dec->stream_type == RDECODE_CODEC_JPEG) {
2821 if (sctx->vcn_ip_ver == VCN_4_0_3)
2822 dec->num_dec_bufs = dec->njctx;
2823 else
2824 dec->num_dec_bufs = dec->njctx * NUM_BUFFERS;
2825 } else
2826 dec->num_dec_bufs = NUM_BUFFERS;
2827
2828 bs_buf_size = align(width * height / 32, 128);
2829 dec->msg_fb_it_probs_buffers = (struct rvid_buffer *) CALLOC(dec->num_dec_bufs, sizeof(struct rvid_buffer));
2830 dec->bs_buffers = (struct rvid_buffer *) CALLOC(dec->num_dec_bufs, sizeof(struct rvid_buffer));
2831 if(!dec->msg_fb_it_probs_buffers || !dec->bs_buffers)
2832 goto error;
2833
2834 for (i = 0; i < dec->num_dec_bufs; ++i) {
2835 unsigned msg_fb_it_probs_size = FB_BUFFER_OFFSET + FB_BUFFER_SIZE;
2836 if (have_it(dec))
2837 msg_fb_it_probs_size += IT_SCALING_TABLE_SIZE;
2838 else if (have_probs(dec))
2839 msg_fb_it_probs_size += (dec->stream_type == RDECODE_CODEC_VP9) ?
2840 VP9_PROBS_TABLE_SIZE :
2841 sizeof(rvcn_dec_av1_segment_fg_t);
2842 /* use vram to improve performance, workaround an unknown bug */
2843 if (!si_vid_create_buffer(dec->screen, &dec->msg_fb_it_probs_buffers[i], msg_fb_it_probs_size,
2844 PIPE_USAGE_DEFAULT)) {
2845 RADEON_DEC_ERR("Can't allocate message buffers.\n");
2846 goto error;
2847 }
2848
2849 if (!si_vid_create_buffer(dec->screen, &dec->bs_buffers[i], bs_buf_size,
2850 PIPE_USAGE_STAGING)) {
2851 RADEON_DEC_ERR("Can't allocate bitstream buffers.\n");
2852 goto error;
2853 }
2854
2855 if (have_probs(dec) && dec->stream_type == RDECODE_CODEC_VP9) {
2856 struct rvid_buffer *buf;
2857 void *ptr;
2858
2859 buf = &dec->msg_fb_it_probs_buffers[i];
2860 ptr = dec->ws->buffer_map(dec->ws, buf->res->buf, &dec->cs,
2861 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2862 ptr += FB_BUFFER_OFFSET + FB_BUFFER_SIZE;
2863 fill_probs_table(ptr);
2864 dec->ws->buffer_unmap(dec->ws, buf->res->buf);
2865 dec->bs_ptr = NULL;
2866 }
2867 }
2868
2869 if ((sctx->vcn_ip_ver >= VCN_3_0_0) &&
2870 (stream_type == RDECODE_CODEC_VP9 ||
2871 stream_type == RDECODE_CODEC_AV1 ||
2872 ((stream_type == RDECODE_CODEC_H265) && templ->expect_chunked_decode) ||
2873 ((stream_type == RDECODE_CODEC_H264_PERF) && templ->expect_chunked_decode)))
2874 dec->dpb_type = DPB_DYNAMIC_TIER_2;
2875 else if (sctx->vcn_ip_ver <= VCN_2_6_0 && stream_type == RDECODE_CODEC_VP9)
2876 dec->dpb_type = DPB_DYNAMIC_TIER_1;
2877 else
2878 dec->dpb_type = DPB_MAX_RES;
2879
2880 dec->db_alignment = (sctx->vcn_ip_ver >= VCN_2_0_0 &&
2881 dec->base.width > 32 && (dec->stream_type == RDECODE_CODEC_VP9 ||
2882 dec->stream_type == RDECODE_CODEC_AV1 ||
2883 dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)) ? 64 : 32;
2884
2885 if (sctx->vcn_ip_ver >= VCN_5_0_0) {
2886 if (stream_type == RDECODE_CODEC_VP9 ||
2887 stream_type == RDECODE_CODEC_AV1 ||
2888 stream_type == RDECODE_CODEC_H265 ||
2889 stream_type == RDECODE_CODEC_H264_PERF)
2890 dec->db_alignment = 64;
2891 }
2892
2893 if (dec->dpb_type != DPB_DYNAMIC_TIER_2)
2894 dec->dpb_size = calc_dpb_size(dec);
2895
2896 if (!si_vid_create_buffer(dec->screen, &dec->sessionctx, RDECODE_SESSION_CONTEXT_SIZE,
2897 PIPE_USAGE_DEFAULT)) {
2898 RADEON_DEC_ERR("Can't allocate session ctx.\n");
2899 goto error;
2900 }
2901
2902 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_LINEAR;
2903 dec->av1_version = RDECODE_AV1_VER_0;
2904
2905 switch (sctx->vcn_ip_ver) {
2906 case VCN_1_0_0:
2907 case VCN_1_0_1:
2908 dec->reg.data0 = RDECODE_VCN1_GPCOM_VCPU_DATA0;
2909 dec->reg.data1 = RDECODE_VCN1_GPCOM_VCPU_DATA1;
2910 dec->reg.cmd = RDECODE_VCN1_GPCOM_VCPU_CMD;
2911 dec->reg.cntl = RDECODE_VCN1_ENGINE_CNTL;
2912 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V1;
2913 break;
2914 case VCN_2_0_0:
2915 case VCN_2_0_2:
2916 case VCN_2_0_3:
2917 case VCN_2_2_0:
2918 dec->reg.data0 = RDECODE_VCN2_GPCOM_VCPU_DATA0;
2919 dec->reg.data1 = RDECODE_VCN2_GPCOM_VCPU_DATA1;
2920 dec->reg.cmd = RDECODE_VCN2_GPCOM_VCPU_CMD;
2921 dec->reg.cntl = RDECODE_VCN2_ENGINE_CNTL;
2922 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V2;
2923 break;
2924 case VCN_2_5_0:
2925 case VCN_2_6_0:
2926 case VCN_3_0_0:
2927 case VCN_3_0_2:
2928 case VCN_3_0_16:
2929 case VCN_3_0_33:
2930 case VCN_3_1_1:
2931 case VCN_3_1_2:
2932 dec->reg.data0 = RDECODE_VCN2_5_GPCOM_VCPU_DATA0;
2933 dec->reg.data1 = RDECODE_VCN2_5_GPCOM_VCPU_DATA1;
2934 dec->reg.cmd = RDECODE_VCN2_5_GPCOM_VCPU_CMD;
2935 dec->reg.cntl = RDECODE_VCN2_5_ENGINE_CNTL;
2936 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V2;
2937 break;
2938 case VCN_4_0_3:
2939 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V3;
2940 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX9;
2941 dec->av1_version = RDECODE_AV1_VER_1;
2942 break;
2943 case VCN_4_0_0:
2944 case VCN_4_0_2:
2945 case VCN_4_0_4:
2946 case VCN_4_0_5:
2947 case VCN_4_0_6:
2948 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V2;
2949 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX11;
2950 dec->av1_version = RDECODE_AV1_VER_1;
2951 break;
2952 case VCN_5_0_0:
2953 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V3;
2954 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX11;
2955 dec->av1_version = RDECODE_AV1_VER_1;
2956 break;
2957 case VCN_5_0_1:
2958 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V3;
2959 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX9;
2960 dec->av1_version = RDECODE_AV1_VER_1;
2961 break;
2962 default:
2963 RADEON_DEC_ERR("VCN is not supported.\n");
2964 goto error;
2965 }
2966
2967 if (dec->stream_type != RDECODE_CODEC_JPEG) {
2968 map_msg_fb_it_probs_buf(dec);
2969 rvcn_dec_message_create(dec);
2970 send_msg_buf(dec);
2971 r = flush(dec, 0, NULL);
2972 if (r)
2973 goto error;
2974 } else if (dec->jpg_reg.version != RDECODE_JPEG_REG_VER_V1) {
2975 dec->jpg_reg.jrbc_ib_cond_rd_timer = vcnipUVD_JRBC_IB_COND_RD_TIMER;
2976 dec->jpg_reg.jrbc_ib_ref_data = vcnipUVD_JRBC_IB_REF_DATA;
2977 dec->jpg_reg.jpeg_rb_base = vcnipUVD_JPEG_RB_BASE;
2978 dec->jpg_reg.jpeg_rb_size = vcnipUVD_JPEG_RB_SIZE;
2979 dec->jpg_reg.jpeg_rb_wptr = vcnipUVD_JPEG_RB_WPTR;
2980 dec->jpg_reg.jpeg_int_en = vcnipUVD_JPEG_INT_EN;
2981 dec->jpg_reg.jpeg_cntl = vcnipUVD_JPEG_CNTL;
2982 dec->jpg_reg.jpeg_rb_rptr = vcnipUVD_JPEG_RB_RPTR;
2983 if (dec->jpg_reg.version == RDECODE_JPEG_REG_VER_V2) {
2984 dec->jpg_reg.jpeg_dec_soft_rst = vcnipUVD_JPEG_DEC_SOFT_RST;
2985 dec->jpg_reg.lmi_jpeg_read_64bit_bar_high = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_HIGH;
2986 dec->jpg_reg.lmi_jpeg_read_64bit_bar_low = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_LOW;
2987 dec->jpg_reg.jpeg_pitch = vcnipUVD_JPEG_PITCH;
2988 dec->jpg_reg.jpeg_uv_pitch = vcnipUVD_JPEG_UV_PITCH;
2989 dec->jpg_reg.dec_addr_mode = vcnipJPEG_DEC_ADDR_MODE;
2990 dec->jpg_reg.dec_y_gfx10_tiling_surface = vcnipJPEG_DEC_Y_GFX10_TILING_SURFACE;
2991 dec->jpg_reg.dec_uv_gfx10_tiling_surface = vcnipJPEG_DEC_UV_GFX10_TILING_SURFACE;
2992 dec->jpg_reg.lmi_jpeg_write_64bit_bar_high = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_HIGH;
2993 dec->jpg_reg.lmi_jpeg_write_64bit_bar_low = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_LOW;
2994 dec->jpg_reg.jpeg_tier_cntl2 = vcnipUVD_JPEG_TIER_CNTL2;
2995 dec->jpg_reg.jpeg_outbuf_cntl = vcnipUVD_JPEG_OUTBUF_CNTL;
2996 dec->jpg_reg.jpeg_outbuf_rptr = vcnipUVD_JPEG_OUTBUF_RPTR;
2997 dec->jpg_reg.jpeg_outbuf_wptr = vcnipUVD_JPEG_OUTBUF_WPTR;
2998 dec->jpg_reg.jpeg_index = vcnipUVD_JPEG_INDEX;
2999 dec->jpg_reg.jpeg_data = vcnipUVD_JPEG_DATA;
3000 } else {
3001 dec->jpg_reg.jpeg_dec_soft_rst = vcnipUVD_JPEG_DEC_SOFT_RST_1;
3002 dec->jpg_reg.lmi_jpeg_read_64bit_bar_high = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_HIGH_1;
3003 dec->jpg_reg.lmi_jpeg_read_64bit_bar_low = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_LOW_1;
3004 dec->jpg_reg.jpeg_pitch = vcnipUVD_JPEG_PITCH_1;
3005 dec->jpg_reg.jpeg_uv_pitch = vcnipUVD_JPEG_UV_PITCH_1;
3006 dec->jpg_reg.dec_addr_mode = vcnipJPEG_DEC_ADDR_MODE_1;
3007 dec->jpg_reg.dec_y_gfx10_tiling_surface = vcnipJPEG_DEC_Y_GFX10_TILING_SURFACE_1;
3008 dec->jpg_reg.dec_uv_gfx10_tiling_surface = vcnipJPEG_DEC_UV_GFX10_TILING_SURFACE_1;
3009 dec->jpg_reg.lmi_jpeg_write_64bit_bar_high = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_HIGH_1;
3010 dec->jpg_reg.lmi_jpeg_write_64bit_bar_low = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_LOW_1;
3011 dec->jpg_reg.jpeg_tier_cntl2 = vcnipUVD_JPEG_TIER_CNTL2_1;
3012 dec->jpg_reg.jpeg_outbuf_cntl = vcnipUVD_JPEG_OUTBUF_CNTL_1;
3013 dec->jpg_reg.jpeg_outbuf_rptr = vcnipUVD_JPEG_OUTBUF_RPTR_1;
3014 dec->jpg_reg.jpeg_outbuf_wptr = vcnipUVD_JPEG_OUTBUF_WPTR_1;
3015 dec->jpg_reg.jpeg_luma_base0_0 = vcnipUVD_JPEG_LUMA_BASE0_0;
3016 dec->jpg_reg.jpeg_chroma_base0_0 = vcnipUVD_JPEG_CHROMA_BASE0_0;
3017 dec->jpg_reg.jpeg_chromav_base0_0 = vcnipUVD_JPEG_CHROMAV_BASE0_0;
3018 }
3019 }
3020
3021 next_buffer(dec);
3022
3023 if (stream_type == RDECODE_CODEC_JPEG)
3024 dec->send_cmd = send_cmd_jpeg;
3025 else
3026 dec->send_cmd = send_cmd_dec;
3027
3028
3029 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
3030 list_inithead(&dec->dpb_ref_list);
3031 list_inithead(&dec->dpb_unref_list);
3032 }
3033
3034 dec->tmz_ctx = sctx->vcn_ip_ver < VCN_2_2_0 && sctx->vcn_ip_ver != VCN_UNKNOWN;
3035
3036 return &dec->base;
3037
3038 error:
3039 dec->ws->cs_destroy(&dec->cs);
3040 if (dec->ectx)
3041 dec->ectx->destroy(dec->ectx);
3042
3043 if (dec->stream_type == RDECODE_CODEC_JPEG) {
3044 for (i = 0; i < dec->njctx; i++) {
3045 dec->ws->cs_destroy(&dec->jcs[i]);
3046 dec->ws->ctx_destroy(dec->jctx[i]);
3047 }
3048 }
3049
3050 if (dec->msg_fb_it_probs_buffers && dec->bs_buffers) {
3051 for (i = 0; i < dec->num_dec_bufs; ++i) {
3052 si_vid_destroy_buffer(&dec->msg_fb_it_probs_buffers[i]);
3053 si_vid_destroy_buffer(&dec->bs_buffers[i]);
3054 }
3055 FREE(dec->msg_fb_it_probs_buffers);
3056 FREE(dec->bs_buffers);
3057 }
3058
3059 if (dec->dpb_type != DPB_DYNAMIC_TIER_2)
3060 si_vid_destroy_buffer(&dec->dpb);
3061 si_vid_destroy_buffer(&dec->ctx);
3062 si_vid_destroy_buffer(&dec->sessionctx);
3063
3064 err:
3065 if (dec->jcs)
3066 FREE(dec->jcs);
3067 if (dec->jctx)
3068 FREE(dec->jctx);
3069 FREE(dec);
3070
3071 return NULL;
3072 }
3073
get_buffer_format(struct radeon_decoder * dec)3074 static enum pipe_format get_buffer_format(struct radeon_decoder *dec)
3075 {
3076 switch (dec->ref_codec.bts) {
3077 case CODEC_10_BITS:
3078 return PIPE_FORMAT_P010;
3079 case CODEC_12_BITS:
3080 return PIPE_FORMAT_P012;
3081 default:
3082 return PIPE_FORMAT_NV12;
3083 }
3084 }
3085