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_drm_fourcc.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 unsigned calc_dpb_size(struct radeon_decoder *dec);
38 static unsigned calc_ctx_size_h264_perf(struct radeon_decoder *dec);
39 static unsigned calc_ctx_size_h265_main(struct radeon_decoder *dec);
40 static unsigned calc_ctx_size_h265_main10(struct radeon_decoder *dec,
41 struct pipe_h265_picture_desc *pic);
42
radeon_dec_destroy_associated_data(void * data)43 static void radeon_dec_destroy_associated_data(void *data)
44 {
45 /* NOOP, since we only use an intptr */
46 }
47
get_current_pic_index(struct radeon_decoder * dec,struct pipe_video_buffer * target,unsigned char * curr_pic_idx)48 static void get_current_pic_index(struct radeon_decoder *dec,
49 struct pipe_video_buffer *target,
50 unsigned char *curr_pic_idx)
51 {
52 for (int i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
53 if (dec->render_pic_list[i] && dec->render_pic_list[i] == target) {
54 if (target->codec != NULL)
55 *curr_pic_idx = (uintptr_t)vl_video_buffer_get_associated_data(target, &dec->base);
56 else {
57 *curr_pic_idx = i;
58 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
59 &radeon_dec_destroy_associated_data);
60 }
61 break;
62 } else if (!dec->render_pic_list[i]) {
63 dec->render_pic_list[i] = target;
64 *curr_pic_idx = i;
65 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
66 &radeon_dec_destroy_associated_data);
67 break;
68 }
69 }
70 }
71
get_h264_msg(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_h264_picture_desc * pic)72 static rvcn_dec_message_avc_t get_h264_msg(struct radeon_decoder *dec,
73 struct pipe_video_buffer *target,
74 struct pipe_h264_picture_desc *pic)
75 {
76 rvcn_dec_message_avc_t result;
77 unsigned i, j, k;
78
79 memset(&result, 0, sizeof(result));
80 switch (pic->base.profile) {
81 case PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE:
82 case PIPE_VIDEO_PROFILE_MPEG4_AVC_CONSTRAINED_BASELINE:
83 result.profile = RDECODE_H264_PROFILE_BASELINE;
84 break;
85
86 case PIPE_VIDEO_PROFILE_MPEG4_AVC_MAIN:
87 result.profile = RDECODE_H264_PROFILE_MAIN;
88 break;
89
90 case PIPE_VIDEO_PROFILE_MPEG4_AVC_HIGH:
91 result.profile = RDECODE_H264_PROFILE_HIGH;
92 break;
93
94 default:
95 assert(0);
96 break;
97 }
98
99 result.level = dec->base.level;
100
101 result.sps_info_flags = 0;
102 result.sps_info_flags |= pic->pps->sps->direct_8x8_inference_flag << 0;
103 result.sps_info_flags |= pic->pps->sps->mb_adaptive_frame_field_flag << 1;
104 result.sps_info_flags |= pic->pps->sps->frame_mbs_only_flag << 2;
105 result.sps_info_flags |= pic->pps->sps->delta_pic_order_always_zero_flag << 3;
106 result.sps_info_flags |= ((dec->dpb_type == DPB_DYNAMIC_TIER_2) ? 0 : 1)
107 << RDECODE_SPS_INFO_H264_EXTENSION_SUPPORT_FLAG_SHIFT;
108
109 result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
110 result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
111 result.log2_max_frame_num_minus4 = pic->pps->sps->log2_max_frame_num_minus4;
112 result.pic_order_cnt_type = pic->pps->sps->pic_order_cnt_type;
113 result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
114
115 switch (dec->base.chroma_format) {
116 case PIPE_VIDEO_CHROMA_FORMAT_NONE:
117 break;
118 case PIPE_VIDEO_CHROMA_FORMAT_400:
119 result.chroma_format = 0;
120 break;
121 case PIPE_VIDEO_CHROMA_FORMAT_420:
122 result.chroma_format = 1;
123 break;
124 case PIPE_VIDEO_CHROMA_FORMAT_422:
125 result.chroma_format = 2;
126 break;
127 case PIPE_VIDEO_CHROMA_FORMAT_444:
128 result.chroma_format = 3;
129 break;
130 case PIPE_VIDEO_CHROMA_FORMAT_440:
131 result.chroma_format = 4;
132 break;
133 }
134
135 result.pps_info_flags = 0;
136 result.pps_info_flags |= pic->pps->transform_8x8_mode_flag << 0;
137 result.pps_info_flags |= pic->pps->redundant_pic_cnt_present_flag << 1;
138 result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 2;
139 result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag << 3;
140 result.pps_info_flags |= pic->pps->weighted_bipred_idc << 4;
141 result.pps_info_flags |= pic->pps->weighted_pred_flag << 6;
142 result.pps_info_flags |= pic->pps->bottom_field_pic_order_in_frame_present_flag << 7;
143 result.pps_info_flags |= pic->pps->entropy_coding_mode_flag << 8;
144
145 result.num_slice_groups_minus1 = pic->pps->num_slice_groups_minus1;
146 result.slice_group_map_type = pic->pps->slice_group_map_type;
147 result.slice_group_change_rate_minus1 = pic->pps->slice_group_change_rate_minus1;
148 result.pic_init_qp_minus26 = pic->pps->pic_init_qp_minus26;
149 result.chroma_qp_index_offset = pic->pps->chroma_qp_index_offset;
150 result.second_chroma_qp_index_offset = pic->pps->second_chroma_qp_index_offset;
151
152 memcpy(result.scaling_list_4x4, pic->pps->ScalingList4x4, 6 * 16);
153 memcpy(result.scaling_list_8x8, pic->pps->ScalingList8x8, 2 * 64);
154
155 memcpy(dec->it, result.scaling_list_4x4, 6 * 16);
156 memcpy((dec->it + 96), result.scaling_list_8x8, 2 * 64);
157
158 result.num_ref_frames = pic->num_ref_frames;
159
160 result.num_ref_idx_l0_active_minus1 = pic->num_ref_idx_l0_active_minus1;
161 result.num_ref_idx_l1_active_minus1 = pic->num_ref_idx_l1_active_minus1;
162
163 result.frame_num = pic->frame_num;
164 memcpy(result.frame_num_list, pic->frame_num_list, 4 * 16);
165 result.curr_field_order_cnt_list[0] = pic->field_order_cnt[0];
166 result.curr_field_order_cnt_list[1] = pic->field_order_cnt[1];
167 memcpy(result.field_order_cnt_list, pic->field_order_cnt_list, 4 * 16 * 2);
168 result.non_existing_frame_flags = 0;
169 result.used_for_reference_flags = 0;
170
171 if (dec->dpb_type != DPB_DYNAMIC_TIER_2) {
172 result.decoded_pic_idx = pic->frame_num;
173 goto end;
174 }
175
176 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); i++) {
177 for (j = 0; (pic->ref[j] != NULL) && (j < ARRAY_SIZE(dec->render_pic_list)); j++) {
178 if (dec->render_pic_list[i] == pic->ref[j])
179 break;
180 if (j == ARRAY_SIZE(dec->render_pic_list) - 1)
181 dec->render_pic_list[i] = NULL;
182 else if (pic->ref[j + 1] == NULL)
183 dec->render_pic_list[i] = NULL;
184 }
185 }
186 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
187 if (dec->render_pic_list[i] && dec->render_pic_list[i] == target) {
188 if (target->codec != NULL){
189 result.decoded_pic_idx =
190 (uintptr_t)vl_video_buffer_get_associated_data(target, &dec->base);
191 } else {
192 result.decoded_pic_idx = i;
193 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
194 &radeon_dec_destroy_associated_data);
195 }
196 break;
197 }
198 }
199 if (i == ARRAY_SIZE(dec->render_pic_list)) {
200 for (i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
201 if (!dec->render_pic_list[i]) {
202 dec->render_pic_list[i] = target;
203 result.decoded_pic_idx = i;
204 vl_video_buffer_set_associated_data(target, &dec->base, (void *)(uintptr_t)i,
205 &radeon_dec_destroy_associated_data);
206 break;
207 }
208 }
209 }
210 for (i = 0; i < ARRAY_SIZE(result.ref_frame_list); i++) {
211 result.ref_frame_list[i] = pic->ref[i] ?
212 (uintptr_t)vl_video_buffer_get_associated_data(pic->ref[i], &dec->base) : 0xff;
213
214 if (result.ref_frame_list[i] != 0xff) {
215 if (pic->top_is_reference[i])
216 result.used_for_reference_flags |= (1 << (2 * i));
217 if (pic->bottom_is_reference[i])
218 result.used_for_reference_flags |= (1 << (2 * i + 1));
219
220 if (pic->is_long_term[i])
221 result.ref_frame_list[i] |= 0x80;
222
223 result.curr_pic_ref_frame_num++;
224
225 for (j = 0; j < ARRAY_SIZE(dec->h264_valid_ref_num); j++) {
226 if ((dec->h264_valid_ref_num[j] != (unsigned)-1)
227 && (dec->h264_valid_ref_num[j] == result.frame_num_list[i]))
228 break;
229 }
230
231 for (k = 0; k < ARRAY_SIZE(dec->h264_valid_poc_num); k++) {
232 if ((dec->h264_valid_poc_num[k] != (unsigned)-1)
233 && ((dec->h264_valid_poc_num[k] == result.field_order_cnt_list[i][0])
234 || dec->h264_valid_poc_num[k] == result.field_order_cnt_list[i][1]))
235 break;
236 }
237 }
238 if (result.ref_frame_list[i] != 0xff && (j == ARRAY_SIZE(dec->h264_valid_ref_num))
239 && (k == ARRAY_SIZE(dec->h264_valid_poc_num))) {
240 result.non_existing_frame_flags |= 1 << i;
241 result.curr_pic_ref_frame_num--;
242 result.ref_frame_list[i] = 0xff;
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 (pic->ref[j] != NULL) && (j < ARRAY_SIZE(dec->render_pic_list));
393 j++) {
394 if (dec->render_pic_list[i] == pic->ref[j])
395 break;
396 if (j == ARRAY_SIZE(dec->render_pic_list) - 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->seg, 0, 256);
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 || target->buffer_format == PIPE_FORMAT_P016) {
1002 result.p010_mode = 1;
1003 result.msb_mode = 1;
1004 } else {
1005 result.luma_10to8 = 1;
1006 result.chroma_10to8 = 1;
1007 }
1008 }
1009
1010 result.preskip_segid = 0;
1011 result.last_active_segid = 0;
1012 for (i = 0; i < 8; i++) {
1013 for (j = 0; j < 8; j++) {
1014 if (pic->picture_parameter.seg_info.feature_mask[i] & (1 << j)) {
1015 result.last_active_segid = i;
1016 if (j >= 5)
1017 result.preskip_segid = 1;
1018 }
1019 }
1020 }
1021
1022 result.seg_lossless_flag = 0;
1023 for (i = 0; i < 8; ++i) {
1024 int av1_get_qindex, qindex;
1025 int segfeature_active = pic->picture_parameter.seg_info.feature_mask[i] & (1 << 0);
1026 if (segfeature_active) {
1027 int seg_qindex = pic->picture_parameter.base_qindex +
1028 pic->picture_parameter.seg_info.feature_data[i][0];
1029 av1_get_qindex = seg_qindex < 0 ? 0 : (seg_qindex > 255 ? 255 : seg_qindex);
1030 } else {
1031 av1_get_qindex = pic->picture_parameter.base_qindex;
1032 }
1033 qindex = pic->picture_parameter.seg_info.segment_info_fields.enabled ?
1034 av1_get_qindex :
1035 pic->picture_parameter.base_qindex;
1036 result.seg_lossless_flag |= (((qindex == 0) && result.y_dc_delta_q == 0 &&
1037 result.u_dc_delta_q == 0 && result.v_dc_delta_q == 0 &&
1038 result.u_ac_delta_q == 0 && result.v_ac_delta_q == 0) << i);
1039 }
1040
1041 rvcn_dec_film_grain_params_t* fg_params = &result.film_grain;
1042 fg_params->apply_grain = pic->picture_parameter.film_grain_info.film_grain_info_fields.apply_grain;
1043 if (fg_params->apply_grain) {
1044 rvcn_dec_av1_fg_init_buf_t *fg_buf = (rvcn_dec_av1_fg_init_buf_t *)(dec->probs + 256);
1045
1046 fg_params->random_seed = pic->picture_parameter.film_grain_info.grain_seed;
1047 fg_params->grain_scale_shift =
1048 pic->picture_parameter.film_grain_info.film_grain_info_fields.grain_scale_shift;
1049 fg_params->scaling_shift =
1050 pic->picture_parameter.film_grain_info.film_grain_info_fields.grain_scaling_minus_8 + 8;
1051 fg_params->chroma_scaling_from_luma =
1052 pic->picture_parameter.film_grain_info.film_grain_info_fields.chroma_scaling_from_luma;
1053 fg_params->num_y_points = pic->picture_parameter.film_grain_info.num_y_points;
1054 fg_params->num_cb_points = pic->picture_parameter.film_grain_info.num_cb_points;
1055 fg_params->num_cr_points = pic->picture_parameter.film_grain_info.num_cr_points;
1056 fg_params->cb_mult = pic->picture_parameter.film_grain_info.cb_mult;
1057 fg_params->cb_luma_mult = pic->picture_parameter.film_grain_info.cb_luma_mult;
1058 fg_params->cb_offset = pic->picture_parameter.film_grain_info.cb_offset;
1059 fg_params->cr_mult = pic->picture_parameter.film_grain_info.cr_mult;
1060 fg_params->cr_luma_mult = pic->picture_parameter.film_grain_info.cr_luma_mult;
1061 fg_params->cr_offset = pic->picture_parameter.film_grain_info.cr_offset;
1062 fg_params->bit_depth_minus_8 = pic->picture_parameter.bit_depth_idx << 1;
1063
1064 for (i = 0; i < fg_params->num_y_points; ++i) {
1065 fg_params->scaling_points_y[i][0] = pic->picture_parameter.film_grain_info.point_y_value[i];
1066 fg_params->scaling_points_y[i][1] = pic->picture_parameter.film_grain_info.point_y_scaling[i];
1067 }
1068 for (i = 0; i < fg_params->num_cb_points; ++i) {
1069 fg_params->scaling_points_cb[i][0] = pic->picture_parameter.film_grain_info.point_cb_value[i];
1070 fg_params->scaling_points_cb[i][1] = pic->picture_parameter.film_grain_info.point_cb_scaling[i];
1071 }
1072 for (i = 0; i < fg_params->num_cr_points; ++i) {
1073 fg_params->scaling_points_cr[i][0] = pic->picture_parameter.film_grain_info.point_cr_value[i];
1074 fg_params->scaling_points_cr[i][1] = pic->picture_parameter.film_grain_info.point_cr_scaling[i];
1075 }
1076
1077 fg_params->ar_coeff_lag = pic->picture_parameter.film_grain_info.film_grain_info_fields.ar_coeff_lag;
1078 fg_params->ar_coeff_shift =
1079 pic->picture_parameter.film_grain_info.film_grain_info_fields.ar_coeff_shift_minus_6 + 6;
1080
1081 for (i = 0; i < 24; ++i)
1082 fg_params->ar_coeffs_y[i] = pic->picture_parameter.film_grain_info.ar_coeffs_y[i];
1083
1084 for (i = 0; i < 25; ++i) {
1085 fg_params->ar_coeffs_cb[i] = pic->picture_parameter.film_grain_info.ar_coeffs_cb[i];
1086 fg_params->ar_coeffs_cr[i] = pic->picture_parameter.film_grain_info.ar_coeffs_cr[i];
1087 }
1088
1089 fg_params->overlap_flag = pic->picture_parameter.film_grain_info.film_grain_info_fields.overlap_flag;
1090 fg_params->clip_to_restricted_range =
1091 pic->picture_parameter.film_grain_info.film_grain_info_fields.clip_to_restricted_range;
1092
1093 ac_vcn_av1_init_film_grain_buffer(fg_params, fg_buf);
1094 }
1095
1096 result.uncompressed_header_size = 0;
1097 for (i = 0; i < 7; ++i) {
1098 result.global_motion[i + 1].wmtype = (rvcn_dec_transformation_type_e)pic->picture_parameter.wm[i].wmtype;
1099 for (j = 0; j < 6; ++j)
1100 result.global_motion[i + 1].wmmat[j] = pic->picture_parameter.wm[i].wmmat[j];
1101 }
1102 for (i = 0; i < tile_count && i < 256; ++i) {
1103 result.tile_info[i].offset = pic->slice_parameter.slice_data_offset[i];
1104 result.tile_info[i].size = pic->slice_parameter.slice_data_size[i];
1105 }
1106
1107 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
1108 dec->ref_codec.bts = pic->picture_parameter.bit_depth_idx ? CODEC_10_BITS : CODEC_8_BITS;
1109 dec->ref_codec.index = result.curr_pic_idx;
1110 dec->ref_codec.ref_size = 8;
1111 dec->ref_codec.num_refs = num_refs;
1112 memset(dec->ref_codec.ref_list, 0x7f, sizeof(dec->ref_codec.ref_list));
1113 memcpy(dec->ref_codec.ref_list, result.ref_frame_map, sizeof(result.ref_frame_map));
1114 }
1115
1116 return result;
1117 }
1118
calc_ctx_size_h265_main(struct radeon_decoder * dec)1119 static unsigned calc_ctx_size_h265_main(struct radeon_decoder *dec)
1120 {
1121 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
1122 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
1123
1124 unsigned max_references = dec->base.max_references + 1;
1125
1126 if (dec->base.width * dec->base.height >= 4096 * 2000)
1127 max_references = MAX2(max_references, 8);
1128 else
1129 max_references = MAX2(max_references, 17);
1130
1131 width = align(width, 16);
1132 height = align(height, 16);
1133 return ((width + 255) / 16) * ((height + 255) / 16) * 16 * max_references + 52 * 1024;
1134 }
1135
calc_ctx_size_h265_main10(struct radeon_decoder * dec,struct pipe_h265_picture_desc * pic)1136 static unsigned calc_ctx_size_h265_main10(struct radeon_decoder *dec,
1137 struct pipe_h265_picture_desc *pic)
1138 {
1139 unsigned log2_ctb_size, width_in_ctb, height_in_ctb, num_16x16_block_per_ctb;
1140 unsigned context_buffer_size_per_ctb_row, cm_buffer_size, max_mb_address, db_left_tile_pxl_size;
1141 unsigned db_left_tile_ctx_size = 4096 / 16 * (32 + 16 * 4);
1142
1143 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
1144 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
1145 unsigned coeff_10bit =
1146 (pic->pps->sps->bit_depth_luma_minus8 || pic->pps->sps->bit_depth_chroma_minus8) ? 2 : 1;
1147
1148 unsigned max_references = dec->base.max_references + 1;
1149
1150 if (dec->base.width * dec->base.height >= 4096 * 2000)
1151 max_references = MAX2(max_references, 8);
1152 else
1153 max_references = MAX2(max_references, 17);
1154
1155 log2_ctb_size = pic->pps->sps->log2_min_luma_coding_block_size_minus3 + 3 +
1156 pic->pps->sps->log2_diff_max_min_luma_coding_block_size;
1157
1158 width_in_ctb = (width + ((1 << log2_ctb_size) - 1)) >> log2_ctb_size;
1159 height_in_ctb = (height + ((1 << log2_ctb_size) - 1)) >> log2_ctb_size;
1160
1161 num_16x16_block_per_ctb = ((1 << log2_ctb_size) >> 4) * ((1 << log2_ctb_size) >> 4);
1162 context_buffer_size_per_ctb_row = align(width_in_ctb * num_16x16_block_per_ctb * 16, 256);
1163 max_mb_address = (unsigned)ceil(height * 8 / 2048.0);
1164
1165 cm_buffer_size = max_references * context_buffer_size_per_ctb_row * height_in_ctb;
1166 db_left_tile_pxl_size = coeff_10bit * (max_mb_address * 2 * 2048 + 1024);
1167
1168 return cm_buffer_size + db_left_tile_ctx_size + db_left_tile_pxl_size;
1169 }
1170
get_vc1_msg(struct pipe_vc1_picture_desc * pic)1171 static rvcn_dec_message_vc1_t get_vc1_msg(struct pipe_vc1_picture_desc *pic)
1172 {
1173 rvcn_dec_message_vc1_t result;
1174
1175 memset(&result, 0, sizeof(result));
1176 switch (pic->base.profile) {
1177 case PIPE_VIDEO_PROFILE_VC1_SIMPLE:
1178 result.profile = RDECODE_VC1_PROFILE_SIMPLE;
1179 result.level = 1;
1180 break;
1181
1182 case PIPE_VIDEO_PROFILE_VC1_MAIN:
1183 result.profile = RDECODE_VC1_PROFILE_MAIN;
1184 result.level = 2;
1185 break;
1186
1187 case PIPE_VIDEO_PROFILE_VC1_ADVANCED:
1188 result.profile = RDECODE_VC1_PROFILE_ADVANCED;
1189 result.level = 4;
1190 break;
1191
1192 default:
1193 assert(0);
1194 }
1195
1196 result.sps_info_flags |= pic->postprocflag << 7;
1197 result.sps_info_flags |= pic->pulldown << 6;
1198 result.sps_info_flags |= pic->interlace << 5;
1199 result.sps_info_flags |= pic->tfcntrflag << 4;
1200 result.sps_info_flags |= pic->finterpflag << 3;
1201 result.sps_info_flags |= pic->psf << 1;
1202
1203 result.pps_info_flags |= pic->range_mapy_flag << 31;
1204 result.pps_info_flags |= pic->range_mapy << 28;
1205 result.pps_info_flags |= pic->range_mapuv_flag << 27;
1206 result.pps_info_flags |= pic->range_mapuv << 24;
1207 result.pps_info_flags |= pic->multires << 21;
1208 result.pps_info_flags |= pic->maxbframes << 16;
1209 result.pps_info_flags |= pic->overlap << 11;
1210 result.pps_info_flags |= pic->quantizer << 9;
1211 result.pps_info_flags |= pic->panscan_flag << 7;
1212 result.pps_info_flags |= pic->refdist_flag << 6;
1213 result.pps_info_flags |= pic->vstransform << 0;
1214
1215 if (pic->base.profile != PIPE_VIDEO_PROFILE_VC1_SIMPLE) {
1216 result.pps_info_flags |= pic->syncmarker << 20;
1217 result.pps_info_flags |= pic->rangered << 19;
1218 result.pps_info_flags |= pic->loopfilter << 5;
1219 result.pps_info_flags |= pic->fastuvmc << 4;
1220 result.pps_info_flags |= pic->extended_mv << 3;
1221 result.pps_info_flags |= pic->extended_dmv << 8;
1222 result.pps_info_flags |= pic->dquant << 1;
1223 }
1224
1225 result.chroma_format = 1;
1226
1227 return result;
1228 }
1229
get_ref_pic_idx(struct radeon_decoder * dec,struct pipe_video_buffer * ref)1230 static uint32_t get_ref_pic_idx(struct radeon_decoder *dec, struct pipe_video_buffer *ref)
1231 {
1232 uint32_t min = MAX2(dec->frame_number, NUM_MPEG2_REFS) - NUM_MPEG2_REFS;
1233 uint32_t max = MAX2(dec->frame_number, 1) - 1;
1234 uintptr_t frame;
1235
1236 /* seems to be the most sane fallback */
1237 if (!ref)
1238 return max;
1239
1240 /* get the frame number from the associated data */
1241 frame = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
1242
1243 /* limit the frame number to a valid range */
1244 return MAX2(MIN2(frame, max), min);
1245 }
1246
get_mpeg2_msg(struct radeon_decoder * dec,struct pipe_mpeg12_picture_desc * pic)1247 static rvcn_dec_message_mpeg2_vld_t get_mpeg2_msg(struct radeon_decoder *dec,
1248 struct pipe_mpeg12_picture_desc *pic)
1249 {
1250 const int *zscan = pic->alternate_scan ? vl_zscan_alternate : vl_zscan_normal;
1251 rvcn_dec_message_mpeg2_vld_t result;
1252 unsigned i;
1253
1254 memset(&result, 0, sizeof(result));
1255 result.decoded_pic_idx = dec->frame_number;
1256
1257 result.forward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[0]);
1258 result.backward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[1]);
1259
1260 if (pic->intra_matrix) {
1261 result.load_intra_quantiser_matrix = 1;
1262 for (i = 0; i < 64; ++i) {
1263 result.intra_quantiser_matrix[i] = pic->intra_matrix[zscan[i]];
1264 }
1265 }
1266 if (pic->non_intra_matrix) {
1267 result.load_nonintra_quantiser_matrix = 1;
1268 for (i = 0; i < 64; ++i) {
1269 result.nonintra_quantiser_matrix[i] = pic->non_intra_matrix[zscan[i]];
1270 }
1271 }
1272
1273 result.profile_and_level_indication = 0;
1274 result.chroma_format = 0x1;
1275
1276 result.picture_coding_type = pic->picture_coding_type;
1277 result.f_code[0][0] = pic->f_code[0][0] + 1;
1278 result.f_code[0][1] = pic->f_code[0][1] + 1;
1279 result.f_code[1][0] = pic->f_code[1][0] + 1;
1280 result.f_code[1][1] = pic->f_code[1][1] + 1;
1281 result.intra_dc_precision = pic->intra_dc_precision;
1282 result.pic_structure = pic->picture_structure;
1283 result.top_field_first = pic->top_field_first;
1284 result.frame_pred_frame_dct = pic->frame_pred_frame_dct;
1285 result.concealment_motion_vectors = pic->concealment_motion_vectors;
1286 result.q_scale_type = pic->q_scale_type;
1287 result.intra_vlc_format = pic->intra_vlc_format;
1288 result.alternate_scan = pic->alternate_scan;
1289
1290 return result;
1291 }
1292
get_mpeg4_msg(struct radeon_decoder * dec,struct pipe_mpeg4_picture_desc * pic)1293 static rvcn_dec_message_mpeg4_asp_vld_t get_mpeg4_msg(struct radeon_decoder *dec,
1294 struct pipe_mpeg4_picture_desc *pic)
1295 {
1296 rvcn_dec_message_mpeg4_asp_vld_t result;
1297 unsigned i;
1298
1299 memset(&result, 0, sizeof(result));
1300 result.decoded_pic_idx = dec->frame_number;
1301
1302 result.forward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[0]);
1303 result.backward_ref_pic_idx = get_ref_pic_idx(dec, pic->ref[1]);
1304
1305 result.variant_type = 0;
1306 result.profile_and_level_indication = 0xF0;
1307
1308 result.video_object_layer_verid = 0x5;
1309 result.video_object_layer_shape = 0x0;
1310
1311 result.video_object_layer_width = dec->base.width;
1312 result.video_object_layer_height = dec->base.height;
1313
1314 result.vop_time_increment_resolution = pic->vop_time_increment_resolution;
1315
1316 result.short_video_header = pic->short_video_header;
1317 result.interlaced = pic->interlaced;
1318 result.load_intra_quant_mat = 1;
1319 result.load_nonintra_quant_mat = 1;
1320 result.quarter_sample = pic->quarter_sample;
1321 result.complexity_estimation_disable = 1;
1322 result.resync_marker_disable = pic->resync_marker_disable;
1323 result.newpred_enable = 0;
1324 result.reduced_resolution_vop_enable = 0;
1325
1326 result.quant_type = pic->quant_type;
1327
1328 for (i = 0; i < 64; ++i) {
1329 result.intra_quant_mat[i] = pic->intra_matrix[vl_zscan_normal[i]];
1330 result.nonintra_quant_mat[i] = pic->non_intra_matrix[vl_zscan_normal[i]];
1331 }
1332
1333 return result;
1334 }
1335
rvcn_dec_message_create(struct radeon_decoder * dec)1336 static void rvcn_dec_message_create(struct radeon_decoder *dec)
1337 {
1338 rvcn_dec_message_header_t *header = dec->msg;
1339 rvcn_dec_message_create_t *create = dec->msg + sizeof(rvcn_dec_message_header_t);
1340 unsigned sizes = sizeof(rvcn_dec_message_header_t) + sizeof(rvcn_dec_message_create_t);
1341
1342 memset(dec->msg, 0, sizes);
1343 header->header_size = sizeof(rvcn_dec_message_header_t);
1344 header->total_size = sizes;
1345 header->num_buffers = 1;
1346 header->msg_type = RDECODE_MSG_CREATE;
1347 header->stream_handle = dec->stream_handle;
1348 header->status_report_feedback_number = 0;
1349
1350 header->index[0].message_id = RDECODE_MESSAGE_CREATE;
1351 header->index[0].offset = sizeof(rvcn_dec_message_header_t);
1352 header->index[0].size = sizeof(rvcn_dec_message_create_t);
1353 header->index[0].filled = 0;
1354
1355 create->stream_type = dec->stream_type;
1356 create->session_flags = 0;
1357 create->width_in_samples = dec->base.width;
1358 create->height_in_samples = dec->base.height;
1359 }
1360
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)1361 static unsigned rvcn_dec_dynamic_dpb_t2_message(struct radeon_decoder *dec, rvcn_dec_message_decode_t *decode,
1362 rvcn_dec_message_dynamic_dpb_t2_t *dynamic_dpb_t2, bool encrypted)
1363 {
1364 struct rvcn_dec_dynamic_dpb_t2 *dpb = NULL, *dummy = NULL;
1365 struct si_resource *res;
1366 unsigned width, height, size;
1367 uint64_t addr;
1368 int i;
1369
1370 width = align(decode->width_in_samples, dec->db_alignment);
1371 height = align(decode->height_in_samples, dec->db_alignment);
1372 size = align((width * height * 3) / 2, 256);
1373 if (dec->ref_codec.bts == CODEC_10_BITS)
1374 size = (((struct si_screen *)dec->screen)->info.vcn_ip_version == VCN_5_0_0) ? size * 2 : size * 3 / 2;
1375
1376 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
1377 bool found = false;
1378
1379 res = (dec->dpb_use_surf) ? &(((struct si_texture *)((struct vl_video_buffer *)d->vbuf)->resources[0])->buffer) :
1380 d->dpb.res;
1381 for (i = 0; i < dec->ref_codec.ref_size; ++i) {
1382 if (((dec->ref_codec.ref_list[i] & 0x7f) != 0x7f) && (d->index == (dec->ref_codec.ref_list[i] & 0x7f))) {
1383 if (!dummy)
1384 dummy = d;
1385 addr = dec->ws->buffer_get_virtual_address(res->buf);
1386 if (!addr && dummy) {
1387 struct si_resource *dummy_res;
1388 RVID_ERR("Ref list from application is incorrect, using dummy buffer instead.\n");
1389 dummy_res = (dec->dpb_use_surf) ? &(((struct si_texture *)((struct vl_video_buffer *)dummy->vbuf)->resources[0])->buffer) :
1390 dummy->dpb.res;
1391 addr = dec->ws->buffer_get_virtual_address(dummy_res->buf);
1392 }
1393 dec->ws->cs_add_buffer(&dec->cs, res->buf, RADEON_USAGE_READWRITE | RADEON_USAGE_SYNCHRONIZED, RADEON_DOMAIN_VRAM);
1394 dynamic_dpb_t2->dpbAddrLo[i] = addr;
1395 dynamic_dpb_t2->dpbAddrHi[i] = addr >> 32;
1396 ++dynamic_dpb_t2->dpbArraySize;
1397 found = true;
1398 }
1399 }
1400 if (!found) {
1401 if ((!dec->dpb_use_surf && (d->dpb.res->b.b.width0 * d->dpb.res->b.b.height0 != size)) ||
1402 ((dec->dpb_use_surf) && (d->vbuf->width != width || d->vbuf->height != height))) {
1403 list_del(&d->list);
1404 list_addtail(&d->list, &dec->dpb_unref_list);
1405 } else {
1406 d->index = 0x7f;
1407 }
1408 }
1409 }
1410
1411 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
1412 if ((!dec->dpb_use_surf && (d->dpb.res->b.b.width0 * d->dpb.res->b.b.height0 == size && d->index == dec->ref_codec.index)) ||
1413 (dec->dpb_use_surf && ((d->vbuf->width == width && d->vbuf->height == height) && d->index == dec->ref_codec.index))) {
1414 dpb = d;
1415 break;
1416 }
1417 }
1418
1419 if (!dpb) {
1420 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
1421 if (d->index == 0x7f) {
1422 d->index = dec->ref_codec.index;
1423 dpb = d;
1424 break;
1425 }
1426 }
1427 }
1428
1429 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_unref_list, list) {
1430 if (dec->prev_fence)
1431 dec->ws->fence_wait(dec->ws, dec->prev_fence, PIPE_DEFAULT_DECODER_FEEDBACK_TIMEOUT_NS);
1432 list_del(&d->list);
1433 if (dec->dpb_use_surf)
1434 d->vbuf->destroy(d->vbuf);
1435 else
1436 si_vid_destroy_buffer(&d->dpb);
1437 FREE(d);
1438 }
1439
1440 if (!dpb) {
1441 bool r;
1442 dpb = CALLOC_STRUCT(rvcn_dec_dynamic_dpb_t2);
1443 if (!dpb)
1444 return 1;
1445 dpb->index = dec->ref_codec.index;
1446
1447 if (dec->dpb_use_surf) {
1448 struct pipe_video_buffer templat;
1449 memset(&templat, 0, sizeof(templat));
1450 templat.buffer_format = (dec->ref_codec.bts != CODEC_10_BITS) ? PIPE_FORMAT_NV12 : PIPE_FORMAT_P010;
1451 templat.width = width;
1452 templat.height = height;
1453 if (encrypted)
1454 templat.bind |= PIPE_BIND_PROTECTED;
1455
1456 dpb->vbuf =
1457 dec->base.context->create_video_buffer_with_modifiers(dec->base.context, &templat, &dec->dpb_modifier, 1);
1458 r = (dpb->vbuf != NULL);
1459 } else {
1460 if (encrypted)
1461 r = si_vid_create_tmz_buffer(dec->screen, &dpb->dpb, size, PIPE_USAGE_DEFAULT);
1462 else
1463 r = si_vid_create_buffer(dec->screen, &dpb->dpb, size, PIPE_USAGE_DEFAULT);
1464 assert(encrypted == (bool)(dpb->dpb.res->flags & RADEON_FLAG_ENCRYPTED));
1465 }
1466
1467 if (!r) {
1468 RVID_ERR("Can't allocated dpb buffer.\n");
1469 FREE(dpb);
1470 return 1;
1471 }
1472 list_addtail(&dpb->list, &dec->dpb_ref_list);
1473 }
1474
1475 if (dynamic_dpb_t2->dpbArraySize < dec->ref_codec.num_refs) {
1476 struct rvcn_dec_dynamic_dpb_t2 *d =
1477 list_first_entry(&dec->dpb_ref_list, struct rvcn_dec_dynamic_dpb_t2, list);
1478
1479 res = (dec->dpb_use_surf) ? &(((struct si_texture *)((struct vl_video_buffer *)d->vbuf)->resources[0])->buffer) :
1480 d->dpb.res;
1481 addr = dec->ws->buffer_get_virtual_address(res->buf);
1482 if (!addr && dummy) {
1483 struct si_resource *dummy_res;
1484 dummy_res = (dec->dpb_use_surf) ? &(((struct si_texture *)((struct vl_video_buffer *)dummy->vbuf)->resources[0])->buffer) :
1485 dummy->dpb.res;
1486 addr = dec->ws->buffer_get_virtual_address(dummy_res->buf);
1487 }
1488 assert(addr);
1489 for (i = 0; i < dec->ref_codec.num_refs; ++i) {
1490 if (dynamic_dpb_t2->dpbAddrLo[i] || dynamic_dpb_t2->dpbAddrHi[i])
1491 continue;
1492 dynamic_dpb_t2->dpbAddrLo[i] = addr;
1493 dynamic_dpb_t2->dpbAddrHi[i] = addr >> 32;
1494 ++dynamic_dpb_t2->dpbArraySize;
1495 }
1496 assert(dynamic_dpb_t2->dpbArraySize == dec->ref_codec.num_refs);
1497 }
1498
1499 if (dec->dpb_use_surf) {
1500 struct si_texture *dpb_luma, *dpb_chroma;
1501
1502 dpb_luma = (struct si_texture *)((struct vl_video_buffer *)dpb->vbuf)->resources[0];
1503 dpb_chroma = (struct si_texture *)((struct vl_video_buffer *)dpb->vbuf)->resources[1];
1504
1505 decode->db_swizzle_mode = dpb_luma->surface.u.gfx9.swizzle_mode;
1506
1507 dec->ws->cs_add_buffer(&dec->cs, (dpb_luma->buffer).buf, RADEON_USAGE_READWRITE | RADEON_USAGE_SYNCHRONIZED, RADEON_DOMAIN_VRAM);
1508 addr = dec->ws->buffer_get_virtual_address((dpb_luma->buffer).buf);
1509
1510 dynamic_dpb_t2->dpbLumaPitch = dpb_luma->buffer.b.b.width0;
1511 dynamic_dpb_t2->dpbLumaAlignedHeight = dpb_luma->buffer.b.b.height0;
1512 dynamic_dpb_t2->dpbLumaAlignedSize = dpb_luma->surface.u.gfx9.surf_slice_size;
1513 dynamic_dpb_t2->dpbChromaPitch = dpb_chroma->buffer.b.b.width0;
1514 dynamic_dpb_t2->dpbChromaAlignedHeight = dpb_chroma->buffer.b.b.height0;
1515 dynamic_dpb_t2->dpbChromaAlignedSize = dpb_chroma->surface.u.gfx9.surf_slice_size;
1516 } else {
1517 dec->ws->cs_add_buffer(&dec->cs, dpb->dpb.res->buf, RADEON_USAGE_READWRITE | RADEON_USAGE_SYNCHRONIZED, RADEON_DOMAIN_VRAM);
1518 addr = dec->ws->buffer_get_virtual_address(dpb->dpb.res->buf);
1519
1520 if (((struct si_screen *)dec->screen)->info.vcn_ip_version == VCN_5_0_0)
1521 decode->db_swizzle_mode = RDECODE_VCN5_256B_D;
1522 dynamic_dpb_t2->dpbLumaPitch = align(decode->width_in_samples, dec->db_alignment);
1523 dynamic_dpb_t2->dpbLumaAlignedHeight = align(decode->height_in_samples, dec->db_alignment);
1524 dynamic_dpb_t2->dpbLumaAlignedSize = dynamic_dpb_t2->dpbLumaPitch *
1525 dynamic_dpb_t2->dpbLumaAlignedHeight;
1526 dynamic_dpb_t2->dpbChromaPitch = dynamic_dpb_t2->dpbLumaPitch >> 1;
1527 dynamic_dpb_t2->dpbChromaAlignedHeight = dynamic_dpb_t2->dpbLumaAlignedHeight >> 1;
1528 dynamic_dpb_t2->dpbChromaAlignedSize = dynamic_dpb_t2->dpbChromaPitch *
1529 dynamic_dpb_t2->dpbChromaAlignedHeight * 2;
1530
1531 if (dec->ref_codec.bts == CODEC_10_BITS) {
1532 if (((struct si_screen *)dec->screen)->info.vcn_ip_version == VCN_5_0_0) {
1533 dynamic_dpb_t2->dpbLumaAlignedSize = dynamic_dpb_t2->dpbLumaAlignedSize * 2;
1534 dynamic_dpb_t2->dpbChromaAlignedSize = dynamic_dpb_t2->dpbChromaAlignedSize * 2;
1535 } else {
1536 dynamic_dpb_t2->dpbLumaAlignedSize = dynamic_dpb_t2->dpbLumaAlignedSize * 3 / 2;
1537 dynamic_dpb_t2->dpbChromaAlignedSize = dynamic_dpb_t2->dpbChromaAlignedSize * 3 / 2;
1538 }
1539 }
1540 }
1541
1542 dynamic_dpb_t2->dpbCurrLo = addr;
1543 dynamic_dpb_t2->dpbCurrHi = addr >> 32;
1544 decode->decode_flags = 1;
1545 dynamic_dpb_t2->dpbConfigFlags = 0;
1546
1547 return 0;
1548 }
1549
rvcn_dec_message_decode(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)1550 static struct pb_buffer_lean *rvcn_dec_message_decode(struct radeon_decoder *dec,
1551 struct pipe_video_buffer *target,
1552 struct pipe_picture_desc *picture)
1553 {
1554 DECRYPT_PARAMETERS *decrypt = (DECRYPT_PARAMETERS *)picture->decrypt_key;
1555 bool encrypted = picture->protected_playback;
1556 struct si_texture *luma;
1557 struct si_texture *chroma;
1558 struct pipe_video_buffer *out_surf = target;
1559 ASSERTED struct si_screen *sscreen = (struct si_screen *)dec->screen;
1560 rvcn_dec_message_header_t *header;
1561 rvcn_dec_message_index_t *index_codec;
1562 rvcn_dec_message_index_t *index_drm = NULL;
1563 rvcn_dec_message_index_t *index_dynamic_dpb = NULL;
1564 rvcn_dec_message_index_t *index_hevc_direct_reflist = NULL;
1565 rvcn_dec_message_decode_t *decode;
1566 unsigned sizes = 0, offset_decode, offset_codec;
1567 unsigned offset_drm = 0, offset_dynamic_dpb = 0, offset_hevc_direct_reflist = 0;
1568 void *codec;
1569 rvcn_dec_message_drm_t *drm = NULL;
1570 rvcn_dec_message_dynamic_dpb_t *dynamic_dpb = NULL;
1571 rvcn_dec_message_dynamic_dpb_t2_t *dynamic_dpb_t2 = NULL;
1572 rvcn_dec_message_hevc_direct_ref_list_t *hevc_reflist = NULL;
1573 bool dpb_resize = false;
1574 header = dec->msg;
1575 sizes += sizeof(rvcn_dec_message_header_t);
1576
1577 index_codec = (void*)header + sizes;
1578 sizes += sizeof(rvcn_dec_message_index_t);
1579
1580 if (encrypted) {
1581 index_drm = (void*)header + sizes;
1582 sizes += sizeof(rvcn_dec_message_index_t);
1583 }
1584
1585 if (dec->dpb_type >= DPB_DYNAMIC_TIER_1) {
1586 index_dynamic_dpb = (void*)header + sizes;
1587 sizes += sizeof(rvcn_dec_message_index_t);
1588 }
1589
1590 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1591 index_hevc_direct_reflist = (void*)header + sizes;
1592 sizes += sizeof(rvcn_dec_message_index_t);
1593 }
1594
1595 offset_decode = sizes;
1596 decode = (void*)header + sizes;
1597 sizes += sizeof(rvcn_dec_message_decode_t);
1598
1599 if (encrypted) {
1600 offset_drm = sizes;
1601 drm = (void*)header + sizes;
1602 sizes += sizeof(rvcn_dec_message_drm_t);
1603 }
1604
1605 if (dec->dpb_type >= DPB_DYNAMIC_TIER_1) {
1606 offset_dynamic_dpb = sizes;
1607 if (dec->dpb_type == DPB_DYNAMIC_TIER_1) {
1608 dynamic_dpb = (void*)header + sizes;
1609 sizes += sizeof(rvcn_dec_message_dynamic_dpb_t);
1610 }
1611 else if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
1612 dynamic_dpb_t2 = (void*)header + sizes;
1613 sizes += sizeof(rvcn_dec_message_dynamic_dpb_t2_t);
1614 }
1615 }
1616
1617 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1618 offset_hevc_direct_reflist = sizes;
1619 hevc_reflist = (void*)header + sizes;
1620 sizes += align((4 + 2 * 15 * ((struct pipe_h265_picture_desc *)picture)->slice_parameter.slice_count), 4);
1621 }
1622
1623 offset_codec = sizes;
1624 codec = (void*)header + sizes;
1625
1626 memset(dec->msg, 0, sizes);
1627 header->header_size = sizeof(rvcn_dec_message_header_t);
1628 header->total_size = sizes;
1629 header->msg_type = RDECODE_MSG_DECODE;
1630 header->stream_handle = dec->stream_handle;
1631 header->status_report_feedback_number = dec->frame_number;
1632
1633 header->index[0].message_id = RDECODE_MESSAGE_DECODE;
1634 header->index[0].offset = offset_decode;
1635 header->index[0].size = sizeof(rvcn_dec_message_decode_t);
1636 header->index[0].filled = 0;
1637 header->num_buffers = 1;
1638
1639 index_codec->offset = offset_codec;
1640 index_codec->size = 0;
1641 index_codec->filled = 0;
1642 ++header->num_buffers;
1643
1644 if (encrypted) {
1645 index_drm->message_id = RDECODE_MESSAGE_DRM;
1646 index_drm->offset = offset_drm;
1647 index_drm->size = sizeof(rvcn_dec_message_drm_t);
1648 index_drm->filled = 0;
1649 ++header->num_buffers;
1650 }
1651
1652 if (dec->dpb_type >= DPB_DYNAMIC_TIER_1) {
1653 index_dynamic_dpb->message_id = RDECODE_MESSAGE_DYNAMIC_DPB;
1654 index_dynamic_dpb->offset = offset_dynamic_dpb;
1655 index_dynamic_dpb->filled = 0;
1656 ++header->num_buffers;
1657 if (dec->dpb_type == DPB_DYNAMIC_TIER_1)
1658 index_dynamic_dpb->size = sizeof(rvcn_dec_message_dynamic_dpb_t);
1659 else if (dec->dpb_type == DPB_DYNAMIC_TIER_2)
1660 index_dynamic_dpb->size = sizeof(rvcn_dec_message_dynamic_dpb_t2_t);
1661 }
1662
1663 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1664 index_hevc_direct_reflist->message_id = RDECODE_MESSAGE_HEVC_DIRECT_REF_LIST;
1665 index_hevc_direct_reflist->offset = offset_hevc_direct_reflist;
1666 index_hevc_direct_reflist->size = align((4 + 2 * 15 * ((struct pipe_h265_picture_desc *)picture)->slice_parameter.slice_count), 4);
1667 index_hevc_direct_reflist->filled = 0;
1668 ++header->num_buffers;
1669 }
1670
1671 decode->stream_type = dec->stream_type;
1672 decode->decode_flags = 0;
1673 decode->width_in_samples = dec->base.width;
1674 decode->height_in_samples = dec->base.height;
1675
1676 decode->bsd_size = align(dec->bs_size, 128);
1677
1678 if (dec->dpb_type != DPB_DYNAMIC_TIER_2) {
1679 bool r;
1680 if (!dec->dpb.res && dec->dpb_size) {
1681 if (encrypted) {
1682 r = si_vid_create_tmz_buffer(dec->screen, &dec->dpb, dec->dpb_size, PIPE_USAGE_DEFAULT);
1683 } else {
1684 r = si_vid_create_buffer(dec->screen, &dec->dpb, dec->dpb_size, PIPE_USAGE_DEFAULT);
1685 }
1686 assert(encrypted == (bool)(dec->dpb.res->flags & RADEON_FLAG_ENCRYPTED));
1687 if (!r) {
1688 RVID_ERR("Can't allocated dpb.\n");
1689 return NULL;
1690 }
1691 si_vid_clear_buffer(dec->base.context, &dec->dpb);
1692 } else if (dec->dpb_type == DPB_DYNAMIC_TIER_1 && dec->dpb.res &&
1693 (dec->max_width < dec->base.width || dec->max_height < dec->base.height)) {
1694 struct rvid_buf_offset_info buf_offset_info;
1695
1696 buf_offset_info.num_units = (NUM_VP9_REFS + 1);
1697 buf_offset_info.old_offset = (align(dec->max_width, dec->db_alignment) *
1698 align(dec->max_height, dec->db_alignment) * 3 / 2);
1699 buf_offset_info.new_offset = (align(dec->base.width, dec->db_alignment) *
1700 align(dec->base.height, dec->db_alignment) * 3 / 2);
1701
1702 dec->dpb_size = calc_dpb_size(dec);
1703 r = si_vid_resize_buffer(dec->base.context, &dec->cs, &dec->dpb, dec->dpb_size, &buf_offset_info);
1704 if (!r) {
1705 RVID_ERR("Can't resize dpb.\n");
1706 return NULL;
1707 }
1708 dec->max_width = dec->base.width;
1709 dec->max_height = dec->base.height;
1710 dpb_resize = true;
1711 }
1712 }
1713
1714 if (!dec->ctx.res) {
1715 enum pipe_video_format fmt = u_reduce_video_profile(picture->profile);
1716 if (dec->stream_type == RDECODE_CODEC_H264_PERF) {
1717 unsigned ctx_size = calc_ctx_size_h264_perf(dec);
1718 bool r;
1719 if (encrypted && dec->tmz_ctx) {
1720 r = si_vid_create_tmz_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1721 } else {
1722 r = si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1723 }
1724 assert((encrypted && dec->tmz_ctx) == (bool)(dec->ctx.res->flags & RADEON_FLAG_ENCRYPTED));
1725
1726 if (!r) {
1727 RVID_ERR("Can't allocated context buffer.\n");
1728 return NULL;
1729 }
1730 si_vid_clear_buffer(dec->base.context, &dec->ctx);
1731 } else if (fmt == PIPE_VIDEO_FORMAT_VP9) {
1732 unsigned ctx_size;
1733 uint8_t *ptr;
1734 bool r;
1735
1736 /* default probability + probability data */
1737 ctx_size = 2304 * 5;
1738
1739 if (((struct si_screen *)dec->screen)->info.vcn_ip_version >= VCN_2_0_0) {
1740 /* SRE collocated context data */
1741 ctx_size += 32 * 2 * 128 * 68;
1742 /* SMP collocated context data */
1743 ctx_size += 9 * 64 * 2 * 128 * 68;
1744 /* SDB left tile pixel */
1745 ctx_size += 8 * 2 * 2 * 8192;
1746 } else {
1747 ctx_size += 32 * 2 * 64 * 64;
1748 ctx_size += 9 * 64 * 2 * 64 * 64;
1749 ctx_size += 8 * 2 * 4096;
1750 }
1751
1752 if (dec->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2)
1753 ctx_size += 8 * 2 * 4096;
1754
1755 if (encrypted && dec->tmz_ctx) {
1756 r = si_vid_create_tmz_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1757 } else {
1758 r = si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1759 }
1760 if (!r) {
1761 RVID_ERR("Can't allocated context buffer.\n");
1762 return NULL;
1763 }
1764 si_vid_clear_buffer(dec->base.context, &dec->ctx);
1765
1766 /* ctx needs probs table */
1767 ptr = dec->ws->buffer_map(dec->ws, dec->ctx.res->buf, &dec->cs,
1768 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
1769 fill_probs_table(ptr);
1770 dec->ws->buffer_unmap(dec->ws, dec->ctx.res->buf);
1771 dec->bs_ptr = NULL;
1772 } else if (fmt == PIPE_VIDEO_FORMAT_HEVC) {
1773 unsigned ctx_size;
1774 bool r;
1775 if (dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)
1776 ctx_size = calc_ctx_size_h265_main10(dec, (struct pipe_h265_picture_desc *)picture);
1777 else
1778 ctx_size = calc_ctx_size_h265_main(dec);
1779
1780 if (encrypted && dec->tmz_ctx) {
1781 r = si_vid_create_tmz_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1782 } else {
1783 r = si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT);
1784 }
1785 if (!r) {
1786 RVID_ERR("Can't allocated context buffer.\n");
1787 return NULL;
1788 }
1789 si_vid_clear_buffer(dec->base.context, &dec->ctx);
1790 }
1791 }
1792 if (encrypted != dec->ws->cs_is_secure(&dec->cs)) {
1793 dec->ws->cs_flush(&dec->cs, RADEON_FLUSH_TOGGLE_SECURE_SUBMISSION, NULL);
1794 }
1795
1796 if (dec->stream_type == RDECODE_CODEC_AV1)
1797 rvcn_dec_av1_film_grain_surface(&out_surf, (struct pipe_av1_picture_desc *)picture);
1798
1799 luma = (struct si_texture *)((struct vl_video_buffer *)out_surf)->resources[0];
1800 chroma = (struct si_texture *)((struct vl_video_buffer *)out_surf)->resources[1];
1801
1802 decode->dpb_size = (dec->dpb_type != DPB_DYNAMIC_TIER_2) ? dec->dpb.res->buf->size : 0;
1803
1804 /* When texture being created, the bo will be created with total size of planes,
1805 * and all planes point to the same buffer */
1806 assert(si_resource(((struct vl_video_buffer *)out_surf)->resources[0])->buf->size ==
1807 si_resource(((struct vl_video_buffer *)out_surf)->resources[1])->buf->size);
1808
1809 decode->dt_size = si_resource(((struct vl_video_buffer *)out_surf)->resources[0])->buf->size;
1810
1811 decode->sct_size = 0;
1812 decode->sc_coeff_size = 0;
1813
1814 decode->sw_ctxt_size = RDECODE_SESSION_CONTEXT_SIZE;
1815 decode->db_pitch = align(dec->base.width, dec->db_alignment);
1816
1817 if ((((struct si_screen*)dec->screen)->info.vcn_ip_version >= VCN_3_0_0) &&
1818 (dec->stream_type == RDECODE_CODEC_VP9 || dec->stream_type == RDECODE_CODEC_AV1 ||
1819 dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10))
1820 decode->db_aligned_height = align(dec->base.height, 64);
1821
1822 decode->db_surf_tile_config = 0;
1823 decode->db_array_mode = dec->addr_gfx_mode;
1824
1825 decode->dt_pitch = luma->surface.u.gfx9.surf_pitch * luma->surface.blk_w;
1826 decode->dt_uv_pitch = chroma->surface.u.gfx9.surf_pitch * chroma->surface.blk_w;
1827
1828 if (luma->surface.meta_offset) {
1829 RVID_ERR("DCC surfaces not supported.\n");
1830 return NULL;
1831 }
1832
1833 decode->dt_tiling_mode = 0;
1834 decode->dt_swizzle_mode = luma->surface.u.gfx9.swizzle_mode;
1835 decode->dt_array_mode = dec->addr_gfx_mode;
1836 decode->dt_field_mode = ((struct vl_video_buffer *)out_surf)->base.interlaced;
1837 decode->dt_surf_tile_config = 0;
1838 decode->dt_uv_surf_tile_config = 0;
1839
1840 decode->dt_luma_top_offset = luma->surface.u.gfx9.surf_offset;
1841 decode->dt_chroma_top_offset = chroma->surface.u.gfx9.surf_offset;
1842 if (decode->dt_field_mode) {
1843 decode->dt_luma_bottom_offset =
1844 luma->surface.u.gfx9.surf_offset + luma->surface.u.gfx9.surf_slice_size;
1845 decode->dt_chroma_bottom_offset =
1846 chroma->surface.u.gfx9.surf_offset + chroma->surface.u.gfx9.surf_slice_size;
1847 } else {
1848 decode->dt_luma_bottom_offset = decode->dt_luma_top_offset;
1849 decode->dt_chroma_bottom_offset = decode->dt_chroma_top_offset;
1850 }
1851 if (dec->stream_type == RDECODE_CODEC_AV1)
1852 decode->db_pitch_uv = chroma->surface.u.gfx9.surf_pitch * chroma->surface.blk_w;
1853
1854 if (encrypted) {
1855 assert(sscreen->info.has_tmz_support);
1856 set_drm_keys(drm, decrypt);
1857 }
1858
1859 if (dec->dpb_type == DPB_DYNAMIC_TIER_1) {
1860 decode->decode_flags |= (RDECODE_FLAGS_USE_DYNAMIC_DPB_MASK | RDECODE_FLAGS_USE_PAL_MASK);
1861 // Add decode flag for RESIZE_DPB ,when we do resize
1862 if (dpb_resize == true)
1863 decode->decode_flags |= RDECODE_FLAGS_DPB_RESIZE_MASK;
1864
1865 dynamic_dpb->dpbArraySize = NUM_VP9_REFS + 1;
1866 dynamic_dpb->dpbLumaPitch = align(dec->max_width, dec->db_alignment);
1867 dynamic_dpb->dpbLumaAlignedHeight = align(dec->max_height, dec->db_alignment);
1868 dynamic_dpb->dpbLumaAlignedSize =
1869 dynamic_dpb->dpbLumaPitch * dynamic_dpb->dpbLumaAlignedHeight;
1870 dynamic_dpb->dpbChromaPitch = dynamic_dpb->dpbLumaPitch >> 1;
1871 dynamic_dpb->dpbChromaAlignedHeight = dynamic_dpb->dpbLumaAlignedHeight >> 1;
1872 dynamic_dpb->dpbChromaAlignedSize =
1873 dynamic_dpb->dpbChromaPitch * dynamic_dpb->dpbChromaAlignedHeight * 2;
1874 dynamic_dpb->dpbReserved0[0] = dec->db_alignment;
1875
1876 if (dec->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2) {
1877 dynamic_dpb->dpbLumaAlignedSize = dynamic_dpb->dpbLumaAlignedSize * 3 / 2;
1878 dynamic_dpb->dpbChromaAlignedSize = dynamic_dpb->dpbChromaAlignedSize * 3 / 2;
1879 }
1880 }
1881
1882 if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC)
1883 get_h265_reflist(hevc_reflist, (struct pipe_h265_picture_desc *)picture);
1884
1885 switch (u_reduce_video_profile(picture->profile)) {
1886 case PIPE_VIDEO_FORMAT_MPEG4_AVC: {
1887 rvcn_dec_message_avc_t avc = get_h264_msg(dec, target, (struct pipe_h264_picture_desc *)picture);
1888 memcpy(codec, (void *)&avc, sizeof(rvcn_dec_message_avc_t));
1889 index_codec->message_id = RDECODE_MESSAGE_AVC;
1890 index_codec->size = sizeof(rvcn_dec_message_avc_t);
1891 break;
1892 }
1893 case PIPE_VIDEO_FORMAT_HEVC: {
1894 rvcn_dec_message_hevc_t hevc =
1895 get_h265_msg(dec, target, (struct pipe_h265_picture_desc *)picture);
1896
1897 memcpy(codec, (void *)&hevc, sizeof(rvcn_dec_message_hevc_t));
1898 index_codec->message_id = RDECODE_MESSAGE_HEVC;
1899 index_codec->size = sizeof(rvcn_dec_message_hevc_t);
1900 break;
1901 }
1902 case PIPE_VIDEO_FORMAT_VC1: {
1903 rvcn_dec_message_vc1_t vc1 = get_vc1_msg((struct pipe_vc1_picture_desc *)picture);
1904
1905 memcpy(codec, (void *)&vc1, sizeof(rvcn_dec_message_vc1_t));
1906 if ((picture->profile == PIPE_VIDEO_PROFILE_VC1_SIMPLE) ||
1907 (picture->profile == PIPE_VIDEO_PROFILE_VC1_MAIN)) {
1908 decode->width_in_samples = align(decode->width_in_samples, 16) / 16;
1909 decode->height_in_samples = align(decode->height_in_samples, 16) / 16;
1910 }
1911 index_codec->message_id = RDECODE_MESSAGE_VC1;
1912 index_codec->size = sizeof(rvcn_dec_message_vc1_t);
1913 break;
1914 }
1915 case PIPE_VIDEO_FORMAT_MPEG12: {
1916 rvcn_dec_message_mpeg2_vld_t mpeg2 =
1917 get_mpeg2_msg(dec, (struct pipe_mpeg12_picture_desc *)picture);
1918
1919 memcpy(codec, (void *)&mpeg2, sizeof(rvcn_dec_message_mpeg2_vld_t));
1920 index_codec->message_id = RDECODE_MESSAGE_MPEG2_VLD;
1921 index_codec->size = sizeof(rvcn_dec_message_mpeg2_vld_t);
1922 break;
1923 }
1924 case PIPE_VIDEO_FORMAT_MPEG4: {
1925 rvcn_dec_message_mpeg4_asp_vld_t mpeg4 =
1926 get_mpeg4_msg(dec, (struct pipe_mpeg4_picture_desc *)picture);
1927
1928 memcpy(codec, (void *)&mpeg4, sizeof(rvcn_dec_message_mpeg4_asp_vld_t));
1929 index_codec->message_id = RDECODE_MESSAGE_MPEG4_ASP_VLD;
1930 index_codec->size = sizeof(rvcn_dec_message_mpeg4_asp_vld_t);
1931 break;
1932 }
1933 case PIPE_VIDEO_FORMAT_VP9: {
1934 rvcn_dec_message_vp9_t vp9 =
1935 get_vp9_msg(dec, target, (struct pipe_vp9_picture_desc *)picture);
1936
1937 memcpy(codec, (void *)&vp9, sizeof(rvcn_dec_message_vp9_t));
1938 index_codec->message_id = RDECODE_MESSAGE_VP9;
1939 index_codec->size = sizeof(rvcn_dec_message_vp9_t);
1940 break;
1941 }
1942 case PIPE_VIDEO_FORMAT_AV1: {
1943 rvcn_dec_message_av1_t av1 =
1944 get_av1_msg(dec, target, (struct pipe_av1_picture_desc *)picture);
1945
1946 memcpy(codec, (void *)&av1, sizeof(rvcn_dec_message_av1_t));
1947 index_codec->message_id = RDECODE_MESSAGE_AV1;
1948 index_codec->size = sizeof(rvcn_dec_message_av1_t);
1949
1950 if (dec->ctx.res == NULL) {
1951 unsigned ctx_size = ac_vcn_dec_calc_ctx_size_av1(dec->av1_version);
1952 uint8_t *ptr;
1953
1954 if (!si_vid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT))
1955 RVID_ERR("Can't allocated context buffer.\n");
1956 si_vid_clear_buffer(dec->base.context, &dec->ctx);
1957
1958 ptr = dec->ws->buffer_map(dec->ws, dec->ctx.res->buf, &dec->cs, PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
1959
1960 ac_vcn_av1_init_probs(dec->av1_version, ptr);
1961 dec->ws->buffer_unmap(dec->ws, dec->ctx.res->buf);
1962 }
1963
1964 break;
1965 }
1966 default:
1967 assert(0);
1968 return NULL;
1969 }
1970
1971 header->total_size += index_codec->size;
1972
1973 if (dec->ctx.res)
1974 decode->hw_ctxt_size = dec->ctx.res->buf->size;
1975
1976 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
1977 if (rvcn_dec_dynamic_dpb_t2_message(dec, decode, dynamic_dpb_t2, encrypted))
1978 return NULL;
1979 } else if (((struct si_screen *)dec->screen)->info.vcn_ip_version == VCN_5_0_0 &&
1980 dec->dpb_type == DPB_MAX_RES)
1981 decode->db_swizzle_mode = RDECODE_VCN5_256B_D;
1982
1983 return luma->buffer.buf;
1984 }
1985
rvcn_dec_message_destroy(struct radeon_decoder * dec)1986 static void rvcn_dec_message_destroy(struct radeon_decoder *dec)
1987 {
1988 rvcn_dec_message_header_t *header = dec->msg;
1989
1990 memset(dec->msg, 0, sizeof(rvcn_dec_message_header_t));
1991 header->header_size = sizeof(rvcn_dec_message_header_t);
1992 header->total_size = sizeof(rvcn_dec_message_header_t) - sizeof(rvcn_dec_message_index_t);
1993 header->num_buffers = 0;
1994 header->msg_type = RDECODE_MSG_DESTROY;
1995 header->stream_handle = dec->stream_handle;
1996 header->status_report_feedback_number = 0;
1997 }
1998
rvcn_dec_message_feedback(struct radeon_decoder * dec)1999 static void rvcn_dec_message_feedback(struct radeon_decoder *dec)
2000 {
2001 rvcn_dec_feedback_header_t *header = (void *)dec->fb;
2002
2003 header->header_size = sizeof(rvcn_dec_feedback_header_t);
2004 header->total_size = sizeof(rvcn_dec_feedback_header_t);
2005 header->num_buffers = 0;
2006 }
2007
rvcn_dec_sq_tail(struct radeon_decoder * dec)2008 static void rvcn_dec_sq_tail(struct radeon_decoder *dec)
2009 {
2010 if (dec->vcn_dec_sw_ring == false)
2011 return;
2012
2013 rvcn_sq_tail(&dec->cs, &dec->sq);
2014 }
2015 /* flush IB to the hardware */
flush(struct radeon_decoder * dec,unsigned flags,struct pipe_fence_handle ** fence)2016 static int flush(struct radeon_decoder *dec, unsigned flags,
2017 struct pipe_fence_handle **fence) {
2018 rvcn_dec_sq_tail(dec);
2019
2020 return dec->ws->cs_flush(&dec->cs, flags, fence);
2021 }
2022
2023 /* add a new set register command to the IB */
set_reg(struct radeon_decoder * dec,unsigned reg,uint32_t val)2024 static void set_reg(struct radeon_decoder *dec, unsigned reg, uint32_t val)
2025 {
2026 radeon_emit(&dec->cs, RDECODE_PKT0(reg >> 2, 0));
2027 radeon_emit(&dec->cs, val);
2028 }
2029
2030 /* 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)2031 static void send_cmd(struct radeon_decoder *dec, unsigned cmd, struct pb_buffer_lean *buf, uint32_t off,
2032 unsigned usage, enum radeon_bo_domain domain)
2033 {
2034 uint64_t addr;
2035
2036 dec->ws->cs_add_buffer(&dec->cs, buf, usage | RADEON_USAGE_SYNCHRONIZED, domain);
2037 addr = dec->ws->buffer_get_virtual_address(buf);
2038 addr = addr + off;
2039
2040 if (dec->vcn_dec_sw_ring == false) {
2041 set_reg(dec, dec->reg.data0, addr);
2042 set_reg(dec, dec->reg.data1, addr >> 32);
2043 set_reg(dec, dec->reg.cmd, cmd << 1);
2044 return;
2045 }
2046
2047 if (!dec->cs.current.cdw) {
2048 rvcn_sq_header(&dec->cs, &dec->sq, false);
2049 rvcn_decode_ib_package_t *ib_header =
2050 (rvcn_decode_ib_package_t *)&(dec->cs.current.buf[dec->cs.current.cdw]);
2051
2052 ib_header->package_size = sizeof(struct rvcn_decode_buffer_s) +
2053 sizeof(struct rvcn_decode_ib_package_s);
2054 dec->cs.current.cdw++;
2055 ib_header->package_type = (RDECODE_IB_PARAM_DECODE_BUFFER);
2056 dec->cs.current.cdw++;
2057
2058 dec->decode_buffer =
2059 (rvcn_decode_buffer_t *)&(dec->cs.current.buf[dec->cs.current.cdw]);
2060
2061 dec->cs.current.cdw += sizeof(struct rvcn_decode_buffer_s) / 4;
2062 memset(dec->decode_buffer, 0, sizeof(struct rvcn_decode_buffer_s));
2063 }
2064
2065 switch(cmd) {
2066 case RDECODE_CMD_MSG_BUFFER:
2067 dec->decode_buffer->valid_buf_flag |= RDECODE_CMDBUF_FLAGS_MSG_BUFFER;
2068 dec->decode_buffer->msg_buffer_address_hi = (addr >> 32);
2069 dec->decode_buffer->msg_buffer_address_lo = (addr);
2070 break;
2071 case RDECODE_CMD_DPB_BUFFER:
2072 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_DPB_BUFFER);
2073 dec->decode_buffer->dpb_buffer_address_hi = (addr >> 32);
2074 dec->decode_buffer->dpb_buffer_address_lo = (addr);
2075 break;
2076 case RDECODE_CMD_DECODING_TARGET_BUFFER:
2077 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_DECODING_TARGET_BUFFER);
2078 dec->decode_buffer->target_buffer_address_hi = (addr >> 32);
2079 dec->decode_buffer->target_buffer_address_lo = (addr);
2080 break;
2081 case RDECODE_CMD_FEEDBACK_BUFFER:
2082 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_FEEDBACK_BUFFER);
2083 dec->decode_buffer->feedback_buffer_address_hi = (addr >> 32);
2084 dec->decode_buffer->feedback_buffer_address_lo = (addr);
2085 break;
2086 case RDECODE_CMD_PROB_TBL_BUFFER:
2087 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_PROB_TBL_BUFFER);
2088 dec->decode_buffer->prob_tbl_buffer_address_hi = (addr >> 32);
2089 dec->decode_buffer->prob_tbl_buffer_address_lo = (addr);
2090 break;
2091 case RDECODE_CMD_SESSION_CONTEXT_BUFFER:
2092 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_SESSION_CONTEXT_BUFFER);
2093 dec->decode_buffer->session_contex_buffer_address_hi = (addr >> 32);
2094 dec->decode_buffer->session_contex_buffer_address_lo = (addr);
2095 break;
2096 case RDECODE_CMD_BITSTREAM_BUFFER:
2097 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_BITSTREAM_BUFFER);
2098 dec->decode_buffer->bitstream_buffer_address_hi = (addr >> 32);
2099 dec->decode_buffer->bitstream_buffer_address_lo = (addr);
2100 break;
2101 case RDECODE_CMD_IT_SCALING_TABLE_BUFFER:
2102 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_IT_SCALING_BUFFER);
2103 dec->decode_buffer->it_sclr_table_buffer_address_hi = (addr >> 32);
2104 dec->decode_buffer->it_sclr_table_buffer_address_lo = (addr);
2105 break;
2106 case RDECODE_CMD_CONTEXT_BUFFER:
2107 dec->decode_buffer->valid_buf_flag |= (RDECODE_CMDBUF_FLAGS_CONTEXT_BUFFER);
2108 dec->decode_buffer->context_buffer_address_hi = (addr >> 32);
2109 dec->decode_buffer->context_buffer_address_lo = (addr);
2110 break;
2111 default:
2112 printf("Not Support!");
2113 }
2114 }
2115
2116 /* do the codec needs an IT buffer ?*/
have_it(struct radeon_decoder * dec)2117 static bool have_it(struct radeon_decoder *dec)
2118 {
2119 return dec->stream_type == RDECODE_CODEC_H264_PERF || dec->stream_type == RDECODE_CODEC_H265;
2120 }
2121
2122 /* do the codec needs an probs buffer? */
have_probs(struct radeon_decoder * dec)2123 static bool have_probs(struct radeon_decoder *dec)
2124 {
2125 return (dec->stream_type == RDECODE_CODEC_VP9 || dec->stream_type == RDECODE_CODEC_AV1);
2126 }
2127
2128 /* map the next available message/feedback/itscaling buffer */
map_msg_fb_it_probs_buf(struct radeon_decoder * dec)2129 static void map_msg_fb_it_probs_buf(struct radeon_decoder *dec)
2130 {
2131 struct rvid_buffer *buf;
2132 uint8_t *ptr;
2133
2134 /* grab the current message/feedback buffer */
2135 buf = &dec->msg_fb_it_probs_buffers[dec->cur_buffer];
2136
2137 /* and map it for CPU access */
2138 ptr =
2139 dec->ws->buffer_map(dec->ws, buf->res->buf, &dec->cs, PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2140
2141 /* calc buffer offsets */
2142 dec->msg = ptr;
2143
2144 dec->fb = (uint32_t *)(ptr + FB_BUFFER_OFFSET);
2145 if (have_it(dec))
2146 dec->it = (uint8_t *)(ptr + FB_BUFFER_OFFSET + FB_BUFFER_SIZE);
2147 else if (have_probs(dec))
2148 dec->probs = (uint8_t *)(ptr + FB_BUFFER_OFFSET + FB_BUFFER_SIZE);
2149 }
2150
2151 /* unmap and send a message command to the VCPU */
send_msg_buf(struct radeon_decoder * dec)2152 static void send_msg_buf(struct radeon_decoder *dec)
2153 {
2154 struct rvid_buffer *buf;
2155
2156 /* ignore the request if message/feedback buffer isn't mapped */
2157 if (!dec->msg || !dec->fb)
2158 return;
2159
2160 /* grab the current message buffer */
2161 buf = &dec->msg_fb_it_probs_buffers[dec->cur_buffer];
2162
2163 /* unmap the buffer */
2164 dec->ws->buffer_unmap(dec->ws, buf->res->buf);
2165 dec->bs_ptr = NULL;
2166 dec->msg = NULL;
2167 dec->fb = NULL;
2168 dec->it = NULL;
2169 dec->probs = NULL;
2170
2171 if (dec->sessionctx.res)
2172 send_cmd(dec, RDECODE_CMD_SESSION_CONTEXT_BUFFER, dec->sessionctx.res->buf, 0,
2173 RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
2174
2175 /* and send it to the hardware */
2176 send_cmd(dec, RDECODE_CMD_MSG_BUFFER, buf->res->buf, 0, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
2177 }
2178
2179 /* cycle to the next set of buffers */
next_buffer(struct radeon_decoder * dec)2180 static void next_buffer(struct radeon_decoder *dec)
2181 {
2182 ++dec->cur_buffer;
2183 dec->cur_buffer %= dec->num_dec_bufs;
2184 }
2185
calc_ctx_size_h264_perf(struct radeon_decoder * dec)2186 static unsigned calc_ctx_size_h264_perf(struct radeon_decoder *dec)
2187 {
2188 unsigned width_in_mb, height_in_mb, ctx_size;
2189 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
2190 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
2191
2192 unsigned max_references = dec->base.max_references + 1;
2193
2194 // picture width & height in 16 pixel units
2195 width_in_mb = width / VL_MACROBLOCK_WIDTH;
2196 height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
2197
2198 unsigned fs_in_mb = width_in_mb * height_in_mb;
2199 unsigned num_dpb_buffer_lean;
2200 switch (dec->base.level) {
2201 case 30:
2202 num_dpb_buffer_lean = 8100 / fs_in_mb;
2203 break;
2204 case 31:
2205 num_dpb_buffer_lean = 18000 / fs_in_mb;
2206 break;
2207 case 32:
2208 num_dpb_buffer_lean = 20480 / fs_in_mb;
2209 break;
2210 case 41:
2211 num_dpb_buffer_lean = 32768 / fs_in_mb;
2212 break;
2213 case 42:
2214 num_dpb_buffer_lean = 34816 / fs_in_mb;
2215 break;
2216 case 50:
2217 num_dpb_buffer_lean = 110400 / fs_in_mb;
2218 break;
2219 case 51:
2220 num_dpb_buffer_lean = 184320 / fs_in_mb;
2221 break;
2222 default:
2223 num_dpb_buffer_lean = 184320 / fs_in_mb;
2224 break;
2225 }
2226 num_dpb_buffer_lean++;
2227 max_references = MAX2(MIN2(NUM_H264_REFS, num_dpb_buffer_lean), max_references);
2228 ctx_size = max_references * align(width_in_mb * height_in_mb * 192, 256);
2229
2230 return ctx_size;
2231 }
2232
2233 /* calculate size of reference picture buffer */
calc_dpb_size(struct radeon_decoder * dec)2234 static unsigned calc_dpb_size(struct radeon_decoder *dec)
2235 {
2236 unsigned width_in_mb, height_in_mb, image_size, dpb_size;
2237
2238 // always align them to MB size for dpb calculation
2239 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
2240 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
2241
2242 // always one more for currently decoded picture
2243 unsigned max_references = dec->base.max_references + 1;
2244
2245 // aligned size of a single frame
2246 image_size = align(width, dec->db_alignment) * align(height, dec->db_alignment);
2247 image_size += image_size / 2;
2248 image_size = align(image_size, 1024);
2249
2250 // picture width & height in 16 pixel units
2251 width_in_mb = width / VL_MACROBLOCK_WIDTH;
2252 height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
2253
2254 switch (u_reduce_video_profile(dec->base.profile)) {
2255 case PIPE_VIDEO_FORMAT_MPEG4_AVC: {
2256 unsigned fs_in_mb = width_in_mb * height_in_mb;
2257 unsigned num_dpb_buffer_lean;
2258
2259 switch (dec->base.level) {
2260 case 30:
2261 num_dpb_buffer_lean = 8100 / fs_in_mb;
2262 break;
2263 case 31:
2264 num_dpb_buffer_lean = 18000 / fs_in_mb;
2265 break;
2266 case 32:
2267 num_dpb_buffer_lean = 20480 / fs_in_mb;
2268 break;
2269 case 41:
2270 num_dpb_buffer_lean = 32768 / fs_in_mb;
2271 break;
2272 case 42:
2273 num_dpb_buffer_lean = 34816 / fs_in_mb;
2274 break;
2275 case 50:
2276 num_dpb_buffer_lean = 110400 / fs_in_mb;
2277 break;
2278 case 51:
2279 num_dpb_buffer_lean = 184320 / fs_in_mb;
2280 break;
2281 default:
2282 num_dpb_buffer_lean = 184320 / fs_in_mb;
2283 break;
2284 }
2285 num_dpb_buffer_lean++;
2286 max_references = MAX2(MIN2(NUM_H264_REFS, num_dpb_buffer_lean), max_references);
2287 dpb_size = image_size * max_references;
2288 break;
2289 }
2290
2291 case PIPE_VIDEO_FORMAT_HEVC:
2292 if (dec->base.width * dec->base.height >= 4096 * 2000)
2293 max_references = MAX2(max_references, 8);
2294 else
2295 max_references = MAX2(max_references, 17);
2296
2297 if (dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)
2298 dpb_size = align((align(width, dec->db_alignment) *
2299 align(height, dec->db_alignment) * 9) / 4, 256) * max_references;
2300 else
2301 dpb_size = align((align(width, dec->db_alignment) *
2302 align(height, dec->db_alignment) * 3) / 2, 256) * max_references;
2303 break;
2304
2305 case PIPE_VIDEO_FORMAT_VC1:
2306 // the firmware seems to always assume a minimum of ref frames
2307 max_references = MAX2(NUM_VC1_REFS, max_references);
2308
2309 // reference picture buffer
2310 dpb_size = image_size * max_references;
2311
2312 // CONTEXT_BUFFER
2313 dpb_size += width_in_mb * height_in_mb * 128;
2314
2315 // IT surface buffer
2316 dpb_size += width_in_mb * 64;
2317
2318 // DB surface buffer
2319 dpb_size += width_in_mb * 128;
2320
2321 // BP
2322 dpb_size += align(MAX2(width_in_mb, height_in_mb) * 7 * 16, 64);
2323 break;
2324
2325 case PIPE_VIDEO_FORMAT_MPEG12:
2326 // reference picture buffer, must be big enough for all frames
2327 dpb_size = image_size * NUM_MPEG2_REFS;
2328 break;
2329
2330 case PIPE_VIDEO_FORMAT_MPEG4:
2331 // reference picture buffer
2332 dpb_size = image_size * max_references;
2333
2334 // CM
2335 dpb_size += width_in_mb * height_in_mb * 64;
2336
2337 // IT surface buffer
2338 dpb_size += align(width_in_mb * height_in_mb * 32, 64);
2339
2340 dpb_size = MAX2(dpb_size, 30 * 1024 * 1024);
2341 break;
2342
2343 case PIPE_VIDEO_FORMAT_VP9:
2344 max_references = MAX2(max_references, 9);
2345
2346 if (dec->dpb_type == DPB_MAX_RES)
2347 dpb_size = (((struct si_screen *)dec->screen)->info.vcn_ip_version >= VCN_2_0_0)
2348 ? (8192 * 4320 * 3 / 2) * max_references
2349 : (4096 * 3000 * 3 / 2) * max_references;
2350 else
2351 dpb_size = (align(dec->base.width, dec->db_alignment) *
2352 align(dec->base.height, dec->db_alignment) * 3 / 2) * max_references;
2353
2354 if (dec->base.profile == PIPE_VIDEO_PROFILE_VP9_PROFILE2)
2355 dpb_size = dpb_size * 3 / 2;
2356 break;
2357
2358 case PIPE_VIDEO_FORMAT_AV1:
2359 max_references = MAX2(max_references, 9);
2360 dpb_size = 8192 * 4320 * 3 / 2 * max_references * 3 / 2;
2361 break;
2362
2363 case PIPE_VIDEO_FORMAT_JPEG:
2364 dpb_size = 0;
2365 break;
2366
2367 default:
2368 // something is missing here
2369 assert(0);
2370
2371 // at least use a sane default value
2372 dpb_size = 32 * 1024 * 1024;
2373 break;
2374 }
2375 return dpb_size;
2376 }
2377
2378 /**
2379 * destroy this video decoder
2380 */
radeon_dec_destroy(struct pipe_video_codec * decoder)2381 static void radeon_dec_destroy(struct pipe_video_codec *decoder)
2382 {
2383 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2384 unsigned i;
2385
2386 assert(decoder);
2387
2388 if (dec->stream_type != RDECODE_CODEC_JPEG) {
2389 map_msg_fb_it_probs_buf(dec);
2390 rvcn_dec_message_destroy(dec);
2391 send_msg_buf(dec);
2392 flush(dec, 0, &dec->destroy_fence);
2393 dec->ws->fence_wait(dec->ws, dec->destroy_fence, PIPE_DEFAULT_DECODER_FEEDBACK_TIMEOUT_NS);
2394 dec->ws->fence_reference(dec->ws, &dec->destroy_fence, NULL);
2395 }
2396
2397 dec->ws->fence_reference(dec->ws, &dec->prev_fence, NULL);
2398 dec->ws->cs_destroy(&dec->cs);
2399
2400 if (dec->stream_type == RDECODE_CODEC_JPEG) {
2401 for (i = 0; i < dec->njctx; i++) {
2402 dec->ws->cs_destroy(&dec->jcs[i]);
2403 dec->ws->ctx_destroy(dec->jctx[i]);
2404 }
2405 }
2406
2407 if (dec->msg_fb_it_probs_buffers && dec->bs_buffers) {
2408 for (i = 0; i < dec->num_dec_bufs; ++i) {
2409 si_vid_destroy_buffer(&dec->msg_fb_it_probs_buffers[i]);
2410 si_vid_destroy_buffer(&dec->bs_buffers[i]);
2411 }
2412 FREE(dec->msg_fb_it_probs_buffers);
2413 FREE(dec->bs_buffers);
2414 }
2415 dec->num_dec_bufs = 0;
2416
2417 if (dec->dpb_type != DPB_DYNAMIC_TIER_2) {
2418 si_vid_destroy_buffer(&dec->dpb);
2419 } else {
2420 list_for_each_entry_safe(struct rvcn_dec_dynamic_dpb_t2, d, &dec->dpb_ref_list, list) {
2421 list_del(&d->list);
2422 if (dec->dpb_use_surf)
2423 d->vbuf->destroy(d->vbuf);
2424 else
2425 si_vid_destroy_buffer(&d->dpb);
2426 FREE(d);
2427 }
2428 }
2429 si_vid_destroy_buffer(&dec->ctx);
2430 si_vid_destroy_buffer(&dec->sessionctx);
2431
2432 FREE(dec->jcs);
2433 FREE(dec->jctx);
2434 FREE(dec);
2435 }
2436
2437 /**
2438 * start decoding of a new frame
2439 */
radeon_dec_begin_frame(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2440 static void radeon_dec_begin_frame(struct pipe_video_codec *decoder,
2441 struct pipe_video_buffer *target,
2442 struct pipe_picture_desc *picture)
2443 {
2444 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2445 uintptr_t frame;
2446
2447 assert(decoder);
2448
2449 frame = ++dec->frame_number;
2450 if (dec->stream_type != RDECODE_CODEC_VP9 && dec->stream_type != RDECODE_CODEC_AV1
2451 && dec->stream_type != RDECODE_CODEC_H264_PERF)
2452 vl_video_buffer_set_associated_data(target, decoder, (void *)frame,
2453 &radeon_dec_destroy_associated_data);
2454
2455 dec->bs_size = 0;
2456 dec->bs_ptr = dec->ws->buffer_map(dec->ws, dec->bs_buffers[dec->cur_buffer].res->buf, &dec->cs,
2457 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2458 }
2459
2460 /**
2461 * decode a macroblock
2462 */
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)2463 static void radeon_dec_decode_macroblock(struct pipe_video_codec *decoder,
2464 struct pipe_video_buffer *target,
2465 struct pipe_picture_desc *picture,
2466 const struct pipe_macroblock *macroblocks,
2467 unsigned num_macroblocks)
2468 {
2469 /* not supported (yet) */
2470 assert(0);
2471 }
2472
2473 /**
2474 * decode a bitstream
2475 */
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)2476 static void radeon_dec_decode_bitstream(struct pipe_video_codec *decoder,
2477 struct pipe_video_buffer *target,
2478 struct pipe_picture_desc *picture, unsigned num_buffers,
2479 const void *const *buffers, const unsigned *sizes)
2480 {
2481 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2482 unsigned i;
2483
2484 assert(decoder);
2485
2486 if (!dec->bs_ptr)
2487 return;
2488
2489 unsigned long total_bs_size = dec->bs_size;
2490 for (i = 0; i < num_buffers; ++i)
2491 total_bs_size += sizes[i];
2492
2493 struct rvid_buffer *buf = &dec->bs_buffers[dec->cur_buffer];
2494
2495 if (total_bs_size > buf->res->buf->size) {
2496 dec->ws->buffer_unmap(dec->ws, buf->res->buf);
2497 dec->bs_ptr = NULL;
2498
2499 if (!dec->bs_size) {
2500 struct rvid_buffer old_buf = *buf;
2501 if (!si_vid_create_buffer(dec->screen, buf, total_bs_size, buf->usage)) {
2502 RVID_ERR("Can't create bitstream buffer!");
2503 return;
2504 }
2505 si_vid_destroy_buffer(&old_buf);
2506 } else if (!si_vid_resize_buffer(dec->base.context, &dec->cs, buf, total_bs_size, NULL)) {
2507 RVID_ERR("Can't resize bitstream buffer!");
2508 return;
2509 }
2510
2511 dec->bs_ptr = dec->ws->buffer_map(dec->ws, buf->res->buf, &dec->cs,
2512 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2513 if (!dec->bs_ptr)
2514 return;
2515
2516 dec->bs_ptr += dec->bs_size;
2517 }
2518
2519 for (i = 0; i < num_buffers; ++i) {
2520 memcpy(dec->bs_ptr, buffers[i], sizes[i]);
2521 dec->bs_size += sizes[i];
2522 dec->bs_ptr += sizes[i];
2523 }
2524 }
2525
2526 /**
2527 * send cmd for vcn dec
2528 */
send_cmd_dec(struct radeon_decoder * dec,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2529 void send_cmd_dec(struct radeon_decoder *dec, struct pipe_video_buffer *target,
2530 struct pipe_picture_desc *picture)
2531 {
2532 struct pb_buffer_lean *dt;
2533 struct rvid_buffer *msg_fb_it_probs_buf, *bs_buf;
2534
2535 msg_fb_it_probs_buf = &dec->msg_fb_it_probs_buffers[dec->cur_buffer];
2536 bs_buf = &dec->bs_buffers[dec->cur_buffer];
2537
2538 memset(dec->bs_ptr, 0, align(dec->bs_size, 128) - dec->bs_size);
2539 dec->ws->buffer_unmap(dec->ws, bs_buf->res->buf);
2540 dec->bs_ptr = NULL;
2541
2542 map_msg_fb_it_probs_buf(dec);
2543 dt = rvcn_dec_message_decode(dec, target, picture);
2544 rvcn_dec_message_feedback(dec);
2545 send_msg_buf(dec);
2546
2547 if (dec->dpb_type != DPB_DYNAMIC_TIER_2)
2548 send_cmd(dec, RDECODE_CMD_DPB_BUFFER, dec->dpb.res->buf, 0, RADEON_USAGE_READWRITE,
2549 RADEON_DOMAIN_VRAM);
2550 if (dec->ctx.res)
2551 send_cmd(dec, RDECODE_CMD_CONTEXT_BUFFER, dec->ctx.res->buf, 0, RADEON_USAGE_READWRITE,
2552 RADEON_DOMAIN_VRAM);
2553 send_cmd(dec, RDECODE_CMD_BITSTREAM_BUFFER, bs_buf->res->buf, 0, RADEON_USAGE_READ,
2554 RADEON_DOMAIN_GTT);
2555 send_cmd(dec, RDECODE_CMD_DECODING_TARGET_BUFFER, dt, 0, RADEON_USAGE_WRITE, RADEON_DOMAIN_VRAM);
2556 send_cmd(dec, RDECODE_CMD_FEEDBACK_BUFFER, msg_fb_it_probs_buf->res->buf, FB_BUFFER_OFFSET,
2557 RADEON_USAGE_WRITE, RADEON_DOMAIN_GTT);
2558 if (have_it(dec))
2559 send_cmd(dec, RDECODE_CMD_IT_SCALING_TABLE_BUFFER, msg_fb_it_probs_buf->res->buf,
2560 FB_BUFFER_OFFSET + FB_BUFFER_SIZE, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
2561 else if (have_probs(dec))
2562 send_cmd(dec, RDECODE_CMD_PROB_TBL_BUFFER, msg_fb_it_probs_buf->res->buf,
2563 FB_BUFFER_OFFSET + FB_BUFFER_SIZE, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
2564
2565 if (dec->vcn_dec_sw_ring == false)
2566 set_reg(dec, dec->reg.cntl, 1);
2567 }
2568
2569 /**
2570 * end decoding of the current frame
2571 */
radeon_dec_end_frame(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2572 static int radeon_dec_end_frame(struct pipe_video_codec *decoder, struct pipe_video_buffer *target,
2573 struct pipe_picture_desc *picture)
2574 {
2575 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2576
2577 assert(decoder);
2578
2579 if (!dec->bs_ptr)
2580 return 1;
2581
2582 dec->send_cmd(dec, target, picture);
2583 flush(dec, picture->flush_flags, picture->fence);
2584 if (picture->fence)
2585 dec->ws->fence_reference(dec->ws, &dec->prev_fence, *picture->fence);
2586
2587 next_buffer(dec);
2588 return 0;
2589 }
2590
2591 /**
2592 * end decoding of the current jpeg frame
2593 */
radeon_dec_jpeg_end_frame(struct pipe_video_codec * decoder,struct pipe_video_buffer * target,struct pipe_picture_desc * picture)2594 static int radeon_dec_jpeg_end_frame(struct pipe_video_codec *decoder, struct pipe_video_buffer *target,
2595 struct pipe_picture_desc *picture)
2596 {
2597 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2598 struct pipe_mjpeg_picture_desc *pic = (struct pipe_mjpeg_picture_desc *)picture;
2599
2600 assert(decoder);
2601
2602 if (!dec->bs_ptr)
2603 return 1;
2604
2605 dec->jpg.crop_x = ROUND_DOWN_TO(pic->picture_parameter.crop_x, VL_MACROBLOCK_WIDTH);
2606 dec->jpg.crop_y = ROUND_DOWN_TO(pic->picture_parameter.crop_y, VL_MACROBLOCK_HEIGHT);
2607 dec->jpg.crop_width = align(pic->picture_parameter.crop_width, VL_MACROBLOCK_WIDTH);
2608 dec->jpg.crop_height = align(pic->picture_parameter.crop_height, VL_MACROBLOCK_HEIGHT);
2609 if (dec->jpg.crop_x + dec->jpg.crop_width > pic->picture_parameter.picture_width)
2610 dec->jpg.crop_width = 0;
2611 if (dec->jpg.crop_y + dec->jpg.crop_height > pic->picture_parameter.picture_height)
2612 dec->jpg.crop_height = 0;
2613 dec->send_cmd(dec, target, picture);
2614 dec->ws->cs_flush(&dec->jcs[dec->cb_idx], picture->flush_flags, NULL);
2615 next_buffer(dec);
2616 dec->cb_idx = (dec->cb_idx+1) % dec->njctx;
2617 return 0;
2618 }
2619
2620 /**
2621 * flush any outstanding command buffers to the hardware
2622 */
radeon_dec_flush(struct pipe_video_codec * decoder)2623 static void radeon_dec_flush(struct pipe_video_codec *decoder)
2624 {
2625 }
2626
radeon_dec_get_decoder_fence(struct pipe_video_codec * decoder,struct pipe_fence_handle * fence,uint64_t timeout)2627 static int radeon_dec_get_decoder_fence(struct pipe_video_codec *decoder,
2628 struct pipe_fence_handle *fence,
2629 uint64_t timeout) {
2630
2631 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2632
2633 return dec->ws->fence_wait(dec->ws, fence, timeout);
2634 }
2635
radeon_dec_destroy_fence(struct pipe_video_codec * decoder,struct pipe_fence_handle * fence)2636 static void radeon_dec_destroy_fence(struct pipe_video_codec *decoder,
2637 struct pipe_fence_handle *fence)
2638 {
2639 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2640
2641 dec->ws->fence_reference(dec->ws, &fence, NULL);
2642 }
2643
radeon_has_video_preferred_gfx12_swizzle_mode(struct radeon_decoder * dec,uint64_t * modifier)2644 static bool radeon_has_video_preferred_gfx12_swizzle_mode(struct radeon_decoder *dec, uint64_t* modifier)
2645 {
2646 uint64_t *allowed_modifiers;
2647 int allowed_mod_count, i, j;
2648 enum pipe_format format[2] = { PIPE_FORMAT_NV12, PIPE_FORMAT_P010 };
2649
2650 for (i = 0; i < 2; ++i) {
2651 dec->screen->query_dmabuf_modifiers(dec->screen, format[i], 0, NULL, NULL, &allowed_mod_count);
2652 allowed_modifiers = (uint64_t *)calloc(allowed_mod_count, sizeof(uint64_t));
2653 if (!allowed_modifiers)
2654 return false;
2655
2656 dec->screen->query_dmabuf_modifiers(dec->screen, format[i], allowed_mod_count,
2657 allowed_modifiers, NULL, &allowed_mod_count);
2658 for (j = 0; j < allowed_mod_count; ++j) {
2659 if (AMD_FMT_MOD_GET(TILE, allowed_modifiers[j]) == AMD_FMT_MOD_TILE_GFX12_256B_2D)
2660 break;
2661 }
2662 if (j == allowed_mod_count) {
2663 free(allowed_modifiers);
2664 return false;;
2665 }
2666 *modifier = allowed_modifiers[j];
2667 free(allowed_modifiers);
2668 }
2669
2670 return true;
2671 }
2672 /**
2673 * update render list when target buffer got updated, use the existing
2674 * index and update the new buffer to associate with it.
2675 */
radeon_dec_update_render_list(struct pipe_video_codec * decoder,struct pipe_video_buffer * old,struct pipe_video_buffer * updated)2676 static void radeon_dec_update_render_list(struct pipe_video_codec *decoder,
2677 struct pipe_video_buffer *old,
2678 struct pipe_video_buffer *updated)
2679 {
2680 struct radeon_decoder *dec = (struct radeon_decoder *)decoder;
2681 void *index = vl_video_buffer_get_associated_data(old, decoder);
2682
2683 vl_video_buffer_set_associated_data(updated, decoder, index,
2684 old->destroy_associated_data);
2685 for (int i = 0; i < ARRAY_SIZE(dec->render_pic_list); ++i) {
2686 if (dec->render_pic_list[i] == old) {
2687 dec->render_pic_list[i] = updated;
2688 break;
2689 }
2690 }
2691 }
2692 /**
2693 * create and HW decoder
2694 */
radeon_create_decoder(struct pipe_context * context,const struct pipe_video_codec * templ)2695 struct pipe_video_codec *radeon_create_decoder(struct pipe_context *context,
2696 const struct pipe_video_codec *templ)
2697 {
2698 struct si_context *sctx = (struct si_context *)context;
2699 struct radeon_winsys *ws = sctx->ws;
2700 unsigned width = templ->width, height = templ->height;
2701 unsigned bs_buf_size, stream_type = 0, ring = AMD_IP_VCN_DEC;
2702 struct radeon_decoder *dec;
2703 int r, i;
2704
2705 switch (u_reduce_video_profile(templ->profile)) {
2706 case PIPE_VIDEO_FORMAT_MPEG12:
2707 if (templ->entrypoint > PIPE_VIDEO_ENTRYPOINT_BITSTREAM)
2708 return vl_create_mpeg12_decoder(context, templ);
2709 stream_type = RDECODE_CODEC_MPEG2_VLD;
2710 break;
2711 case PIPE_VIDEO_FORMAT_MPEG4:
2712 width = align(width, VL_MACROBLOCK_WIDTH);
2713 height = align(height, VL_MACROBLOCK_HEIGHT);
2714 stream_type = RDECODE_CODEC_MPEG4;
2715 break;
2716 case PIPE_VIDEO_FORMAT_VC1:
2717 stream_type = RDECODE_CODEC_VC1;
2718 break;
2719 case PIPE_VIDEO_FORMAT_MPEG4_AVC:
2720 width = align(width, VL_MACROBLOCK_WIDTH);
2721 height = align(height, VL_MACROBLOCK_HEIGHT);
2722 stream_type = RDECODE_CODEC_H264_PERF;
2723 break;
2724 case PIPE_VIDEO_FORMAT_HEVC:
2725 stream_type = RDECODE_CODEC_H265;
2726 break;
2727 case PIPE_VIDEO_FORMAT_VP9:
2728 stream_type = RDECODE_CODEC_VP9;
2729 break;
2730 case PIPE_VIDEO_FORMAT_AV1:
2731 stream_type = RDECODE_CODEC_AV1;
2732 break;
2733 case PIPE_VIDEO_FORMAT_JPEG:
2734 stream_type = RDECODE_CODEC_JPEG;
2735 ring = AMD_IP_VCN_JPEG;
2736 break;
2737 default:
2738 assert(0);
2739 break;
2740 }
2741
2742 dec = CALLOC_STRUCT(radeon_decoder);
2743
2744 if (!dec)
2745 return NULL;
2746
2747 dec->base = *templ;
2748 dec->base.context = context;
2749 dec->base.width = width;
2750 dec->base.height = height;
2751 dec->max_width = width;
2752 dec->max_height = height;
2753 dec->base.destroy = radeon_dec_destroy;
2754 dec->base.begin_frame = radeon_dec_begin_frame;
2755 dec->base.decode_macroblock = radeon_dec_decode_macroblock;
2756 dec->base.decode_bitstream = radeon_dec_decode_bitstream;
2757 dec->base.end_frame = radeon_dec_end_frame;
2758 dec->base.flush = radeon_dec_flush;
2759 dec->base.get_decoder_fence = radeon_dec_get_decoder_fence;
2760 dec->base.destroy_fence = radeon_dec_destroy_fence;
2761 dec->base.update_decoder_target = radeon_dec_update_render_list;
2762
2763 dec->stream_type = stream_type;
2764 dec->stream_handle = si_vid_alloc_stream_handle();
2765 dec->screen = context->screen;
2766 dec->ws = ws;
2767
2768 if (u_reduce_video_profile(templ->profile) != PIPE_VIDEO_FORMAT_JPEG &&
2769 (sctx->vcn_ip_ver >= VCN_4_0_0)) {
2770 dec->vcn_dec_sw_ring = true;
2771 ring = AMD_IP_VCN_UNIFIED;
2772 }
2773
2774 dec->sq.ib_total_size_in_dw = NULL;
2775 dec->sq.ib_checksum = NULL;
2776
2777 if (!ws->cs_create(&dec->cs, sctx->ctx, ring, NULL, NULL)) {
2778 RVID_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 RVID_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 RVID_ERR("Can't allocated 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 RVID_ERR("Can't allocated bitstream buffers.\n");
2852 goto error;
2853 }
2854
2855 si_vid_clear_buffer(context, &dec->msg_fb_it_probs_buffers[i]);
2856 si_vid_clear_buffer(context, &dec->bs_buffers[i]);
2857
2858 if (have_probs(dec) && dec->stream_type == RDECODE_CODEC_VP9) {
2859 struct rvid_buffer *buf;
2860 void *ptr;
2861
2862 buf = &dec->msg_fb_it_probs_buffers[i];
2863 ptr = dec->ws->buffer_map(dec->ws, buf->res->buf, &dec->cs,
2864 PIPE_MAP_WRITE | RADEON_MAP_TEMPORARY);
2865 ptr += FB_BUFFER_OFFSET + FB_BUFFER_SIZE;
2866 fill_probs_table(ptr);
2867 dec->ws->buffer_unmap(dec->ws, buf->res->buf);
2868 dec->bs_ptr = NULL;
2869 }
2870 }
2871
2872 if ((sctx->vcn_ip_ver >= VCN_3_0_0) &&
2873 (stream_type == RDECODE_CODEC_VP9 ||
2874 stream_type == RDECODE_CODEC_AV1 ||
2875 ((stream_type == RDECODE_CODEC_H265) && templ->expect_chunked_decode) ||
2876 ((stream_type == RDECODE_CODEC_H264_PERF) && templ->expect_chunked_decode)))
2877 dec->dpb_type = DPB_DYNAMIC_TIER_2;
2878 else if (sctx->vcn_ip_ver <= VCN_2_6_0 && stream_type == RDECODE_CODEC_VP9)
2879 dec->dpb_type = DPB_DYNAMIC_TIER_1;
2880 else
2881 dec->dpb_type = DPB_MAX_RES;
2882
2883 dec->db_alignment = (sctx->vcn_ip_ver >= VCN_2_0_0 &&
2884 dec->base.width > 32 && (dec->stream_type == RDECODE_CODEC_VP9 ||
2885 dec->stream_type == RDECODE_CODEC_AV1 ||
2886 dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)) ? 64 : 32;
2887
2888 if (sctx->vcn_ip_ver >= VCN_5_0_0) {
2889 if (stream_type == RDECODE_CODEC_VP9 ||
2890 stream_type == RDECODE_CODEC_AV1 ||
2891 stream_type == RDECODE_CODEC_H265 ||
2892 stream_type == RDECODE_CODEC_H264_PERF)
2893 dec->db_alignment = 64;
2894 }
2895
2896 if (dec->dpb_type != DPB_DYNAMIC_TIER_2)
2897 dec->dpb_size = calc_dpb_size(dec);
2898
2899 if (!si_vid_create_buffer(dec->screen, &dec->sessionctx, RDECODE_SESSION_CONTEXT_SIZE,
2900 PIPE_USAGE_DEFAULT)) {
2901 RVID_ERR("Can't allocated session ctx.\n");
2902 goto error;
2903 }
2904 si_vid_clear_buffer(context, &dec->sessionctx);
2905
2906 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_LINEAR;
2907 dec->av1_version = RDECODE_AV1_VER_0;
2908
2909 switch (sctx->vcn_ip_ver) {
2910 case VCN_1_0_0:
2911 case VCN_1_0_1:
2912 dec->reg.data0 = RDECODE_VCN1_GPCOM_VCPU_DATA0;
2913 dec->reg.data1 = RDECODE_VCN1_GPCOM_VCPU_DATA1;
2914 dec->reg.cmd = RDECODE_VCN1_GPCOM_VCPU_CMD;
2915 dec->reg.cntl = RDECODE_VCN1_ENGINE_CNTL;
2916 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V1;
2917 break;
2918 case VCN_2_0_0:
2919 case VCN_2_0_2:
2920 case VCN_2_0_3:
2921 case VCN_2_2_0:
2922 dec->reg.data0 = RDECODE_VCN2_GPCOM_VCPU_DATA0;
2923 dec->reg.data1 = RDECODE_VCN2_GPCOM_VCPU_DATA1;
2924 dec->reg.cmd = RDECODE_VCN2_GPCOM_VCPU_CMD;
2925 dec->reg.cntl = RDECODE_VCN2_ENGINE_CNTL;
2926 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V2;
2927 break;
2928 case VCN_2_5_0:
2929 case VCN_2_6_0:
2930 case VCN_3_0_0:
2931 case VCN_3_0_2:
2932 case VCN_3_0_16:
2933 case VCN_3_0_33:
2934 case VCN_3_1_1:
2935 case VCN_3_1_2:
2936 dec->reg.data0 = RDECODE_VCN2_5_GPCOM_VCPU_DATA0;
2937 dec->reg.data1 = RDECODE_VCN2_5_GPCOM_VCPU_DATA1;
2938 dec->reg.cmd = RDECODE_VCN2_5_GPCOM_VCPU_CMD;
2939 dec->reg.cntl = RDECODE_VCN2_5_ENGINE_CNTL;
2940 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V2;
2941 break;
2942 case VCN_4_0_3:
2943 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V3;
2944 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX9;
2945 dec->av1_version = RDECODE_AV1_VER_1;
2946 break;
2947 case VCN_4_0_0:
2948 case VCN_4_0_2:
2949 case VCN_4_0_4:
2950 case VCN_4_0_5:
2951 case VCN_4_0_6:
2952 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V2;
2953 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX11;
2954 dec->av1_version = RDECODE_AV1_VER_1;
2955 break;
2956 case VCN_5_0_0:
2957 dec->jpg_reg.version = RDECODE_JPEG_REG_VER_V3;
2958 dec->addr_gfx_mode = RDECODE_ARRAY_MODE_ADDRLIB_SEL_GFX11;
2959 dec->av1_version = RDECODE_AV1_VER_1;
2960 if (radeon_has_video_preferred_gfx12_swizzle_mode(dec, &dec->dpb_modifier))
2961 dec->dpb_use_surf = true;
2962 break;
2963 default:
2964 RVID_ERR("VCN is not supported.\n");
2965 goto error;
2966 }
2967
2968 if (dec->stream_type != RDECODE_CODEC_JPEG) {
2969 map_msg_fb_it_probs_buf(dec);
2970 rvcn_dec_message_create(dec);
2971 send_msg_buf(dec);
2972 r = flush(dec, 0, NULL);
2973 if (r)
2974 goto error;
2975 } else if (dec->jpg_reg.version != RDECODE_JPEG_REG_VER_V1) {
2976 dec->jpg_reg.jrbc_ib_cond_rd_timer = vcnipUVD_JRBC_IB_COND_RD_TIMER;
2977 dec->jpg_reg.jrbc_ib_ref_data = vcnipUVD_JRBC_IB_REF_DATA;
2978 dec->jpg_reg.jpeg_rb_base = vcnipUVD_JPEG_RB_BASE;
2979 dec->jpg_reg.jpeg_rb_size = vcnipUVD_JPEG_RB_SIZE;
2980 dec->jpg_reg.jpeg_rb_wptr = vcnipUVD_JPEG_RB_WPTR;
2981 dec->jpg_reg.jpeg_int_en = vcnipUVD_JPEG_INT_EN;
2982 dec->jpg_reg.jpeg_cntl = vcnipUVD_JPEG_CNTL;
2983 dec->jpg_reg.jpeg_rb_rptr = vcnipUVD_JPEG_RB_RPTR;
2984 if (dec->jpg_reg.version == RDECODE_JPEG_REG_VER_V2) {
2985 dec->jpg_reg.jpeg_dec_soft_rst = vcnipUVD_JPEG_DEC_SOFT_RST;
2986 dec->jpg_reg.lmi_jpeg_read_64bit_bar_high = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_HIGH;
2987 dec->jpg_reg.lmi_jpeg_read_64bit_bar_low = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_LOW;
2988 dec->jpg_reg.jpeg_pitch = vcnipUVD_JPEG_PITCH;
2989 dec->jpg_reg.jpeg_uv_pitch = vcnipUVD_JPEG_UV_PITCH;
2990 dec->jpg_reg.dec_addr_mode = vcnipJPEG_DEC_ADDR_MODE;
2991 dec->jpg_reg.dec_y_gfx10_tiling_surface = vcnipJPEG_DEC_Y_GFX10_TILING_SURFACE;
2992 dec->jpg_reg.dec_uv_gfx10_tiling_surface = vcnipJPEG_DEC_UV_GFX10_TILING_SURFACE;
2993 dec->jpg_reg.lmi_jpeg_write_64bit_bar_high = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_HIGH;
2994 dec->jpg_reg.lmi_jpeg_write_64bit_bar_low = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_LOW;
2995 dec->jpg_reg.jpeg_tier_cntl2 = vcnipUVD_JPEG_TIER_CNTL2;
2996 dec->jpg_reg.jpeg_outbuf_cntl = vcnipUVD_JPEG_OUTBUF_CNTL;
2997 dec->jpg_reg.jpeg_outbuf_rptr = vcnipUVD_JPEG_OUTBUF_RPTR;
2998 dec->jpg_reg.jpeg_outbuf_wptr = vcnipUVD_JPEG_OUTBUF_WPTR;
2999 dec->jpg_reg.jpeg_index = vcnipUVD_JPEG_INDEX;
3000 dec->jpg_reg.jpeg_data = vcnipUVD_JPEG_DATA;
3001 } else {
3002 dec->jpg_reg.jpeg_dec_soft_rst = vcnipUVD_JPEG_DEC_SOFT_RST_1;
3003 dec->jpg_reg.lmi_jpeg_read_64bit_bar_high = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_HIGH_1;
3004 dec->jpg_reg.lmi_jpeg_read_64bit_bar_low = vcnipUVD_LMI_JPEG_READ_64BIT_BAR_LOW_1;
3005 dec->jpg_reg.jpeg_pitch = vcnipUVD_JPEG_PITCH_1;
3006 dec->jpg_reg.jpeg_uv_pitch = vcnipUVD_JPEG_UV_PITCH_1;
3007 dec->jpg_reg.dec_addr_mode = vcnipJPEG_DEC_ADDR_MODE_1;
3008 dec->jpg_reg.dec_y_gfx10_tiling_surface = vcnipJPEG_DEC_Y_GFX10_TILING_SURFACE_1;
3009 dec->jpg_reg.dec_uv_gfx10_tiling_surface = vcnipJPEG_DEC_UV_GFX10_TILING_SURFACE_1;
3010 dec->jpg_reg.lmi_jpeg_write_64bit_bar_high = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_HIGH_1;
3011 dec->jpg_reg.lmi_jpeg_write_64bit_bar_low = vcnipUVD_LMI_JPEG_WRITE_64BIT_BAR_LOW_1;
3012 dec->jpg_reg.jpeg_tier_cntl2 = vcnipUVD_JPEG_TIER_CNTL2_1;
3013 dec->jpg_reg.jpeg_outbuf_cntl = vcnipUVD_JPEG_OUTBUF_CNTL_1;
3014 dec->jpg_reg.jpeg_outbuf_rptr = vcnipUVD_JPEG_OUTBUF_RPTR_1;
3015 dec->jpg_reg.jpeg_outbuf_wptr = vcnipUVD_JPEG_OUTBUF_WPTR_1;
3016 dec->jpg_reg.jpeg_luma_base0_0 = vcnipUVD_JPEG_LUMA_BASE0_0;
3017 dec->jpg_reg.jpeg_chroma_base0_0 = vcnipUVD_JPEG_CHROMA_BASE0_0;
3018 dec->jpg_reg.jpeg_chromav_base0_0 = vcnipUVD_JPEG_CHROMAV_BASE0_0;
3019 }
3020 }
3021
3022 next_buffer(dec);
3023
3024 if (stream_type == RDECODE_CODEC_JPEG)
3025 dec->send_cmd = send_cmd_jpeg;
3026 else
3027 dec->send_cmd = send_cmd_dec;
3028
3029
3030 if (dec->dpb_type == DPB_DYNAMIC_TIER_2) {
3031 list_inithead(&dec->dpb_ref_list);
3032 list_inithead(&dec->dpb_unref_list);
3033 }
3034
3035 dec->tmz_ctx = sctx->vcn_ip_ver < VCN_2_2_0 && sctx->vcn_ip_ver != VCN_UNKNOWN;
3036
3037 return &dec->base;
3038
3039 error:
3040 dec->ws->cs_destroy(&dec->cs);
3041
3042 if (dec->stream_type == RDECODE_CODEC_JPEG) {
3043 for (i = 0; i < dec->njctx; i++) {
3044 dec->ws->cs_destroy(&dec->jcs[i]);
3045 dec->ws->ctx_destroy(dec->jctx[i]);
3046 }
3047 }
3048
3049 if (dec->msg_fb_it_probs_buffers && dec->bs_buffers) {
3050 for (i = 0; i < dec->num_dec_bufs; ++i) {
3051 si_vid_destroy_buffer(&dec->msg_fb_it_probs_buffers[i]);
3052 si_vid_destroy_buffer(&dec->bs_buffers[i]);
3053 }
3054 FREE(dec->msg_fb_it_probs_buffers);
3055 FREE(dec->bs_buffers);
3056 }
3057
3058 if (dec->dpb_type != DPB_DYNAMIC_TIER_2)
3059 si_vid_destroy_buffer(&dec->dpb);
3060 si_vid_destroy_buffer(&dec->ctx);
3061 si_vid_destroy_buffer(&dec->sessionctx);
3062
3063 err:
3064 if (dec->jcs)
3065 FREE(dec->jcs);
3066 if (dec->jctx)
3067 FREE(dec->jctx);
3068 FREE(dec);
3069
3070 return NULL;
3071 }
3072