projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi:optimizing SET_CONTEXT_REG for shaders GS
[mesa.git] / src / gallium / drivers / r600 / radeon_uvd.c
1 /**************************************************************************
2 *
3 * Copyright 2011 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 /*
29 * Authors:
30 * Christian König <christian.koenig@amd.com>
31 *
32 */
33
34 #include <sys/types.h>
35 #include <assert.h>
36 #include <errno.h>
37 #include <unistd.h>
38 #include <stdio.h>
39
40 #include "pipe/p_video_codec.h"
41
42 #include "util/u_memory.h"
43 #include "util/u_video.h"
44
45 #include "vl/vl_defines.h"
46 #include "vl/vl_mpeg12_decoder.h"
47
48 #include "r600_pipe_common.h"
49 #include "radeon_video.h"
50 #include "radeon_uvd.h"
51
52 #define NUM_BUFFERS 4
53
54 #define NUM_MPEG2_REFS 6
55 #define NUM_H264_REFS 17
56 #define NUM_VC1_REFS 5
57
58 #define FB_BUFFER_OFFSET 0x1000
59 #define FB_BUFFER_SIZE 2048
60 #define FB_BUFFER_SIZE_TONGA (2048 * 64)
61 #define IT_SCALING_TABLE_SIZE 992
62 #define UVD_SESSION_CONTEXT_SIZE (128 * 1024)
63
64 /* UVD decoder representation */
65 struct ruvd_decoder {
66 struct pipe_video_codec base;
67
68 ruvd_set_dtb set_dtb;
69
70 unsigned stream_handle;
71 unsigned stream_type;
72 unsigned frame_number;
73
74 struct pipe_screen *screen;
75 struct radeon_winsys* ws;
76 struct radeon_cmdbuf* cs;
77
78 unsigned cur_buffer;
79
80 struct rvid_buffer msg_fb_it_buffers[NUM_BUFFERS];
81 struct ruvd_msg *msg;
82 uint32_t *fb;
83 unsigned fb_size;
84 uint8_t *it;
85
86 struct rvid_buffer bs_buffers[NUM_BUFFERS];
87 void* bs_ptr;
88 unsigned bs_size;
89
90 struct rvid_buffer dpb;
91 bool use_legacy;
92 struct rvid_buffer ctx;
93 struct rvid_buffer sessionctx;
94 struct {
95 unsigned data0;
96 unsigned data1;
97 unsigned cmd;
98 unsigned cntl;
99 } reg;
100 };
101
102 /* flush IB to the hardware */
103 static int flush(struct ruvd_decoder *dec, unsigned flags)
104 {
105 return dec->ws->cs_flush(dec->cs, flags, NULL);
106 }
107
108 /* add a new set register command to the IB */
109 static void set_reg(struct ruvd_decoder *dec, unsigned reg, uint32_t val)
110 {
111 radeon_emit(dec->cs, RUVD_PKT0(reg >> 2, 0));
112 radeon_emit(dec->cs, val);
113 }
114
115 /* send a command to the VCPU through the GPCOM registers */
116 static void send_cmd(struct ruvd_decoder *dec, unsigned cmd,
117 struct pb_buffer* buf, uint32_t off,
118 enum radeon_bo_usage usage, enum radeon_bo_domain domain)
119 {
120 int reloc_idx;
121
122 reloc_idx = dec->ws->cs_add_buffer(dec->cs, buf, usage | RADEON_USAGE_SYNCHRONIZED,
123 domain, 0);
124 if (!dec->use_legacy) {
125 uint64_t addr;
126 addr = dec->ws->buffer_get_virtual_address(buf);
127 addr = addr + off;
128 set_reg(dec, dec->reg.data0, addr);
129 set_reg(dec, dec->reg.data1, addr >> 32);
130 } else {
131 off += dec->ws->buffer_get_reloc_offset(buf);
132 set_reg(dec, RUVD_GPCOM_VCPU_DATA0, off);
133 set_reg(dec, RUVD_GPCOM_VCPU_DATA1, reloc_idx * 4);
134 }
135 set_reg(dec, dec->reg.cmd, cmd << 1);
136 }
137
138 /* do the codec needs an IT buffer ?*/
139 static bool have_it(struct ruvd_decoder *dec)
140 {
141 return dec->stream_type == RUVD_CODEC_H264_PERF ||
142 dec->stream_type == RUVD_CODEC_H265;
143 }
144
145 /* map the next available message/feedback/itscaling buffer */
146 static void map_msg_fb_it_buf(struct ruvd_decoder *dec)
147 {
148 struct rvid_buffer* buf;
149 uint8_t *ptr;
150
151 /* grab the current message/feedback buffer */
152 buf = &dec->msg_fb_it_buffers[dec->cur_buffer];
153
154 /* and map it for CPU access */
155 ptr = dec->ws->buffer_map(buf->res->buf, dec->cs, PIPE_TRANSFER_WRITE);
156
157 /* calc buffer offsets */
158 dec->msg = (struct ruvd_msg *)ptr;
159 memset(dec->msg, 0, sizeof(*dec->msg));
160
161 dec->fb = (uint32_t *)(ptr + FB_BUFFER_OFFSET);
162 if (have_it(dec))
163 dec->it = (uint8_t *)(ptr + FB_BUFFER_OFFSET + dec->fb_size);
164 }
165
166 /* unmap and send a message command to the VCPU */
167 static void send_msg_buf(struct ruvd_decoder *dec)
168 {
169 struct rvid_buffer* buf;
170
171 /* ignore the request if message/feedback buffer isn't mapped */
172 if (!dec->msg || !dec->fb)
173 return;
174
175 /* grab the current message buffer */
176 buf = &dec->msg_fb_it_buffers[dec->cur_buffer];
177
178 /* unmap the buffer */
179 dec->ws->buffer_unmap(buf->res->buf);
180 dec->msg = NULL;
181 dec->fb = NULL;
182 dec->it = NULL;
183
184
185 if (dec->sessionctx.res)
186 send_cmd(dec, RUVD_CMD_SESSION_CONTEXT_BUFFER,
187 dec->sessionctx.res->buf, 0, RADEON_USAGE_READWRITE,
188 RADEON_DOMAIN_VRAM);
189
190 /* and send it to the hardware */
191 send_cmd(dec, RUVD_CMD_MSG_BUFFER, buf->res->buf, 0,
192 RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
193 }
194
195 /* cycle to the next set of buffers */
196 static void next_buffer(struct ruvd_decoder *dec)
197 {
198 ++dec->cur_buffer;
199 dec->cur_buffer %= NUM_BUFFERS;
200 }
201
202 /* convert the profile into something UVD understands */
203 static uint32_t profile2stream_type(struct ruvd_decoder *dec, unsigned family)
204 {
205 switch (u_reduce_video_profile(dec->base.profile)) {
206 case PIPE_VIDEO_FORMAT_MPEG4_AVC:
207 return RUVD_CODEC_H264;
208
209 case PIPE_VIDEO_FORMAT_VC1:
210 return RUVD_CODEC_VC1;
211
212 case PIPE_VIDEO_FORMAT_MPEG12:
213 return RUVD_CODEC_MPEG2;
214
215 case PIPE_VIDEO_FORMAT_MPEG4:
216 return RUVD_CODEC_MPEG4;
217
218 case PIPE_VIDEO_FORMAT_HEVC:
219 return RUVD_CODEC_H265;
220
221 case PIPE_VIDEO_FORMAT_JPEG:
222 return RUVD_CODEC_MJPEG;
223
224 default:
225 assert(0);
226 return 0;
227 }
228 }
229
230 static unsigned calc_ctx_size_h265_main(struct ruvd_decoder *dec)
231 {
232 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
233 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
234
235 unsigned max_references = dec->base.max_references + 1;
236
237 if (dec->base.width * dec->base.height >= 4096*2000)
238 max_references = MAX2(max_references, 8);
239 else
240 max_references = MAX2(max_references, 17);
241
242 width = align (width, 16);
243 height = align (height, 16);
244 return ((width + 255) / 16) * ((height + 255) / 16) * 16 * max_references + 52 * 1024;
245 }
246
247 static unsigned calc_ctx_size_h265_main10(struct ruvd_decoder *dec, struct pipe_h265_picture_desc *pic)
248 {
249 unsigned block_size, log2_ctb_size, width_in_ctb, height_in_ctb, num_16x16_block_per_ctb;
250 unsigned context_buffer_size_per_ctb_row, cm_buffer_size, max_mb_address, db_left_tile_pxl_size;
251 unsigned db_left_tile_ctx_size = 4096 / 16 * (32 + 16 * 4);
252
253 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
254 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
255 unsigned coeff_10bit = (pic->pps->sps->bit_depth_luma_minus8 || pic->pps->sps->bit_depth_chroma_minus8) ? 2 : 1;
256
257 unsigned max_references = dec->base.max_references + 1;
258
259 if (dec->base.width * dec->base.height >= 4096*2000)
260 max_references = MAX2(max_references, 8);
261 else
262 max_references = MAX2(max_references, 17);
263
264 block_size = (1 << (pic->pps->sps->log2_min_luma_coding_block_size_minus3 + 3));
265 log2_ctb_size = block_size + pic->pps->sps->log2_diff_max_min_luma_coding_block_size;
266
267 width_in_ctb = (width + ((1 << log2_ctb_size) - 1)) >> log2_ctb_size;
268 height_in_ctb = (height + ((1 << log2_ctb_size) - 1)) >> log2_ctb_size;
269
270 num_16x16_block_per_ctb = ((1 << log2_ctb_size) >> 4) * ((1 << log2_ctb_size) >> 4);
271 context_buffer_size_per_ctb_row = align(width_in_ctb * num_16x16_block_per_ctb * 16, 256);
272 max_mb_address = (unsigned) ceil(height * 8 / 2048.0);
273
274 cm_buffer_size = max_references * context_buffer_size_per_ctb_row * height_in_ctb;
275 db_left_tile_pxl_size = coeff_10bit * (max_mb_address * 2 * 2048 + 1024);
276
277 return cm_buffer_size + db_left_tile_ctx_size + db_left_tile_pxl_size;
278 }
279
280 static unsigned get_db_pitch_alignment(struct ruvd_decoder *dec)
281 {
282 return 16;
283 }
284
285 /* calculate size of reference picture buffer */
286 static unsigned calc_dpb_size(struct ruvd_decoder *dec)
287 {
288 unsigned width_in_mb, height_in_mb, image_size, dpb_size;
289
290 // always align them to MB size for dpb calculation
291 unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
292 unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
293
294 // always one more for currently decoded picture
295 unsigned max_references = dec->base.max_references + 1;
296
297 // aligned size of a single frame
298 image_size = align(width, get_db_pitch_alignment(dec)) * height;
299 image_size += image_size / 2;
300 image_size = align(image_size, 1024);
301
302 // picture width & height in 16 pixel units
303 width_in_mb = width / VL_MACROBLOCK_WIDTH;
304 height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
305
306 switch (u_reduce_video_profile(dec->base.profile)) {
307 case PIPE_VIDEO_FORMAT_MPEG4_AVC: {
308 if (!dec->use_legacy) {
309 unsigned fs_in_mb = width_in_mb * height_in_mb;
310 unsigned alignment = 64, num_dpb_buffer;
311
312 if (dec->stream_type == RUVD_CODEC_H264_PERF)
313 alignment = 256;
314 switch(dec->base.level) {
315 case 30:
316 num_dpb_buffer = 8100 / fs_in_mb;
317 break;
318 case 31:
319 num_dpb_buffer = 18000 / fs_in_mb;
320 break;
321 case 32:
322 num_dpb_buffer = 20480 / fs_in_mb;
323 break;
324 case 41:
325 num_dpb_buffer = 32768 / fs_in_mb;
326 break;
327 case 42:
328 num_dpb_buffer = 34816 / fs_in_mb;
329 break;
330 case 50:
331 num_dpb_buffer = 110400 / fs_in_mb;
332 break;
333 case 51:
334 num_dpb_buffer = 184320 / fs_in_mb;
335 break;
336 default:
337 num_dpb_buffer = 184320 / fs_in_mb;
338 break;
339 }
340 num_dpb_buffer++;
341 max_references = MAX2(MIN2(NUM_H264_REFS, num_dpb_buffer), max_references);
342 dpb_size = image_size * max_references;
343 if ((dec->stream_type != RUVD_CODEC_H264_PERF)) {
344 dpb_size += max_references * align(width_in_mb * height_in_mb * 192, alignment);
345 dpb_size += align(width_in_mb * height_in_mb * 32, alignment);
346 }
347 } else {
348 // the firmware seems to allways assume a minimum of ref frames
349 max_references = MAX2(NUM_H264_REFS, max_references);
350 // reference picture buffer
351 dpb_size = image_size * max_references;
352 if ((dec->stream_type != RUVD_CODEC_H264_PERF)) {
353 // macroblock context buffer
354 dpb_size += width_in_mb * height_in_mb * max_references * 192;
355 // IT surface buffer
356 dpb_size += width_in_mb * height_in_mb * 32;
357 }
358 }
359 break;
360 }
361
362 case PIPE_VIDEO_FORMAT_HEVC:
363 if (dec->base.width * dec->base.height >= 4096*2000)
364 max_references = MAX2(max_references, 8);
365 else
366 max_references = MAX2(max_references, 17);
367
368 width = align (width, 16);
369 height = align (height, 16);
370 if (dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)
371 dpb_size = align((align(width, get_db_pitch_alignment(dec)) * height * 9) / 4, 256) * max_references;
372 else
373 dpb_size = align((align(width, get_db_pitch_alignment(dec)) * height * 3) / 2, 256) * max_references;
374 break;
375
376 case PIPE_VIDEO_FORMAT_VC1:
377 // the firmware seems to allways assume a minimum of ref frames
378 max_references = MAX2(NUM_VC1_REFS, max_references);
379
380 // reference picture buffer
381 dpb_size = image_size * max_references;
382
383 // CONTEXT_BUFFER
384 dpb_size += width_in_mb * height_in_mb * 128;
385
386 // IT surface buffer
387 dpb_size += width_in_mb * 64;
388
389 // DB surface buffer
390 dpb_size += width_in_mb * 128;
391
392 // BP
393 dpb_size += align(MAX2(width_in_mb, height_in_mb) * 7 * 16, 64);
394 break;
395
396 case PIPE_VIDEO_FORMAT_MPEG12:
397 // reference picture buffer, must be big enough for all frames
398 dpb_size = image_size * NUM_MPEG2_REFS;
399 break;
400
401 case PIPE_VIDEO_FORMAT_MPEG4:
402 // reference picture buffer
403 dpb_size = image_size * max_references;
404
405 // CM
406 dpb_size += width_in_mb * height_in_mb * 64;
407
408 // IT surface buffer
409 dpb_size += align(width_in_mb * height_in_mb * 32, 64);
410
411 dpb_size = MAX2(dpb_size, 30 * 1024 * 1024);
412 break;
413
414 case PIPE_VIDEO_FORMAT_JPEG:
415 dpb_size = 0;
416 break;
417
418 default:
419 // something is missing here
420 assert(0);
421
422 // at least use a sane default value
423 dpb_size = 32 * 1024 * 1024;
424 break;
425 }
426 return dpb_size;
427 }
428
429 /* free associated data in the video buffer callback */
430 static void ruvd_destroy_associated_data(void *data)
431 {
432 /* NOOP, since we only use an intptr */
433 }
434
435 /* get h264 specific message bits */
436 static struct ruvd_h264 get_h264_msg(struct ruvd_decoder *dec, struct pipe_h264_picture_desc *pic)
437 {
438 struct ruvd_h264 result;
439
440 memset(&result, 0, sizeof(result));
441 switch (pic->base.profile) {
442 case PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE:
443 case PIPE_VIDEO_PROFILE_MPEG4_AVC_CONSTRAINED_BASELINE:
444 result.profile = RUVD_H264_PROFILE_BASELINE;
445 break;
446
447 case PIPE_VIDEO_PROFILE_MPEG4_AVC_MAIN:
448 result.profile = RUVD_H264_PROFILE_MAIN;
449 break;
450
451 case PIPE_VIDEO_PROFILE_MPEG4_AVC_HIGH:
452 result.profile = RUVD_H264_PROFILE_HIGH;
453 break;
454
455 default:
456 assert(0);
457 break;
458 }
459
460 result.level = dec->base.level;
461
462 result.sps_info_flags = 0;
463 result.sps_info_flags |= pic->pps->sps->direct_8x8_inference_flag << 0;
464 result.sps_info_flags |= pic->pps->sps->mb_adaptive_frame_field_flag << 1;
465 result.sps_info_flags |= pic->pps->sps->frame_mbs_only_flag << 2;
466 result.sps_info_flags |= pic->pps->sps->delta_pic_order_always_zero_flag << 3;
467
468 result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
469 result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
470 result.log2_max_frame_num_minus4 = pic->pps->sps->log2_max_frame_num_minus4;
471 result.pic_order_cnt_type = pic->pps->sps->pic_order_cnt_type;
472 result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
473
474 switch (dec->base.chroma_format) {
475 case PIPE_VIDEO_CHROMA_FORMAT_NONE:
476 /* TODO: assert? */
477 break;
478 case PIPE_VIDEO_CHROMA_FORMAT_400:
479 result.chroma_format = 0;
480 break;
481 case PIPE_VIDEO_CHROMA_FORMAT_420:
482 result.chroma_format = 1;
483 break;
484 case PIPE_VIDEO_CHROMA_FORMAT_422:
485 result.chroma_format = 2;
486 break;
487 case PIPE_VIDEO_CHROMA_FORMAT_444:
488 result.chroma_format = 3;
489 break;
490 }
491
492 result.pps_info_flags = 0;
493 result.pps_info_flags |= pic->pps->transform_8x8_mode_flag << 0;
494 result.pps_info_flags |= pic->pps->redundant_pic_cnt_present_flag << 1;
495 result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 2;
496 result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag << 3;
497 result.pps_info_flags |= pic->pps->weighted_bipred_idc << 4;
498 result.pps_info_flags |= pic->pps->weighted_pred_flag << 6;
499 result.pps_info_flags |= pic->pps->bottom_field_pic_order_in_frame_present_flag << 7;
500 result.pps_info_flags |= pic->pps->entropy_coding_mode_flag << 8;
501
502 result.num_slice_groups_minus1 = pic->pps->num_slice_groups_minus1;
503 result.slice_group_map_type = pic->pps->slice_group_map_type;
504 result.slice_group_change_rate_minus1 = pic->pps->slice_group_change_rate_minus1;
505 result.pic_init_qp_minus26 = pic->pps->pic_init_qp_minus26;
506 result.chroma_qp_index_offset = pic->pps->chroma_qp_index_offset;
507 result.second_chroma_qp_index_offset = pic->pps->second_chroma_qp_index_offset;
508
509 memcpy(result.scaling_list_4x4, pic->pps->ScalingList4x4, 6*16);
510 memcpy(result.scaling_list_8x8, pic->pps->ScalingList8x8, 2*64);
511
512 if (dec->stream_type == RUVD_CODEC_H264_PERF) {
513 memcpy(dec->it, result.scaling_list_4x4, 6*16);
514 memcpy((dec->it + 96), result.scaling_list_8x8, 2*64);
515 }
516
517 result.num_ref_frames = pic->num_ref_frames;
518
519 result.num_ref_idx_l0_active_minus1 = pic->num_ref_idx_l0_active_minus1;
520 result.num_ref_idx_l1_active_minus1 = pic->num_ref_idx_l1_active_minus1;
521
522 result.frame_num = pic->frame_num;
523 memcpy(result.frame_num_list, pic->frame_num_list, 4*16);
524 result.curr_field_order_cnt_list[0] = pic->field_order_cnt[0];
525 result.curr_field_order_cnt_list[1] = pic->field_order_cnt[1];
526 memcpy(result.field_order_cnt_list, pic->field_order_cnt_list, 4*16*2);
527
528 result.decoded_pic_idx = pic->frame_num;
529
530 return result;
531 }
532
533 /* get h265 specific message bits */
534 static struct ruvd_h265 get_h265_msg(struct ruvd_decoder *dec, struct pipe_video_buffer *target,
535 struct pipe_h265_picture_desc *pic)
536 {
537 struct ruvd_h265 result;
538 unsigned i;
539
540 memset(&result, 0, sizeof(result));
541
542 result.sps_info_flags = 0;
543 result.sps_info_flags |= pic->pps->sps->scaling_list_enabled_flag << 0;
544 result.sps_info_flags |= pic->pps->sps->amp_enabled_flag << 1;
545 result.sps_info_flags |= pic->pps->sps->sample_adaptive_offset_enabled_flag << 2;
546 result.sps_info_flags |= pic->pps->sps->pcm_enabled_flag << 3;
547 result.sps_info_flags |= pic->pps->sps->pcm_loop_filter_disabled_flag << 4;
548 result.sps_info_flags |= pic->pps->sps->long_term_ref_pics_present_flag << 5;
549 result.sps_info_flags |= pic->pps->sps->sps_temporal_mvp_enabled_flag << 6;
550 result.sps_info_flags |= pic->pps->sps->strong_intra_smoothing_enabled_flag << 7;
551 result.sps_info_flags |= pic->pps->sps->separate_colour_plane_flag << 8;
552 if (pic->UseRefPicList == true)
553 result.sps_info_flags |= 1 << 10;
554
555 result.chroma_format = pic->pps->sps->chroma_format_idc;
556 result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
557 result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
558 result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
559 result.sps_max_dec_pic_buffering_minus1 = pic->pps->sps->sps_max_dec_pic_buffering_minus1;
560 result.log2_min_luma_coding_block_size_minus3 = pic->pps->sps->log2_min_luma_coding_block_size_minus3;
561 result.log2_diff_max_min_luma_coding_block_size = pic->pps->sps->log2_diff_max_min_luma_coding_block_size;
562 result.log2_min_transform_block_size_minus2 = pic->pps->sps->log2_min_transform_block_size_minus2;
563 result.log2_diff_max_min_transform_block_size = pic->pps->sps->log2_diff_max_min_transform_block_size;
564 result.max_transform_hierarchy_depth_inter = pic->pps->sps->max_transform_hierarchy_depth_inter;
565 result.max_transform_hierarchy_depth_intra = pic->pps->sps->max_transform_hierarchy_depth_intra;
566 result.pcm_sample_bit_depth_luma_minus1 = pic->pps->sps->pcm_sample_bit_depth_luma_minus1;
567 result.pcm_sample_bit_depth_chroma_minus1 = pic->pps->sps->pcm_sample_bit_depth_chroma_minus1;
568 result.log2_min_pcm_luma_coding_block_size_minus3 = pic->pps->sps->log2_min_pcm_luma_coding_block_size_minus3;
569 result.log2_diff_max_min_pcm_luma_coding_block_size = pic->pps->sps->log2_diff_max_min_pcm_luma_coding_block_size;
570 result.num_short_term_ref_pic_sets = pic->pps->sps->num_short_term_ref_pic_sets;
571
572 result.pps_info_flags = 0;
573 result.pps_info_flags |= pic->pps->dependent_slice_segments_enabled_flag << 0;
574 result.pps_info_flags |= pic->pps->output_flag_present_flag << 1;
575 result.pps_info_flags |= pic->pps->sign_data_hiding_enabled_flag << 2;
576 result.pps_info_flags |= pic->pps->cabac_init_present_flag << 3;
577 result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 4;
578 result.pps_info_flags |= pic->pps->transform_skip_enabled_flag << 5;
579 result.pps_info_flags |= pic->pps->cu_qp_delta_enabled_flag << 6;
580 result.pps_info_flags |= pic->pps->pps_slice_chroma_qp_offsets_present_flag << 7;
581 result.pps_info_flags |= pic->pps->weighted_pred_flag << 8;
582 result.pps_info_flags |= pic->pps->weighted_bipred_flag << 9;
583 result.pps_info_flags |= pic->pps->transquant_bypass_enabled_flag << 10;
584 result.pps_info_flags |= pic->pps->tiles_enabled_flag << 11;
585 result.pps_info_flags |= pic->pps->entropy_coding_sync_enabled_flag << 12;
586 result.pps_info_flags |= pic->pps->uniform_spacing_flag << 13;
587 result.pps_info_flags |= pic->pps->loop_filter_across_tiles_enabled_flag << 14;
588 result.pps_info_flags |= pic->pps->pps_loop_filter_across_slices_enabled_flag << 15;
589 result.pps_info_flags |= pic->pps->deblocking_filter_override_enabled_flag << 16;
590 result.pps_info_flags |= pic->pps->pps_deblocking_filter_disabled_flag << 17;
591 result.pps_info_flags |= pic->pps->lists_modification_present_flag << 18;
592 result.pps_info_flags |= pic->pps->slice_segment_header_extension_present_flag << 19;
593 //result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag; ???
594
595 result.num_extra_slice_header_bits = pic->pps->num_extra_slice_header_bits;
596 result.num_long_term_ref_pic_sps = pic->pps->sps->num_long_term_ref_pics_sps;
597 result.num_ref_idx_l0_default_active_minus1 = pic->pps->num_ref_idx_l0_default_active_minus1;
598 result.num_ref_idx_l1_default_active_minus1 = pic->pps->num_ref_idx_l1_default_active_minus1;
599 result.pps_cb_qp_offset = pic->pps->pps_cb_qp_offset;
600 result.pps_cr_qp_offset = pic->pps->pps_cr_qp_offset;
601 result.pps_beta_offset_div2 = pic->pps->pps_beta_offset_div2;
602 result.pps_tc_offset_div2 = pic->pps->pps_tc_offset_div2;
603 result.diff_cu_qp_delta_depth = pic->pps->diff_cu_qp_delta_depth;
604 result.num_tile_columns_minus1 = pic->pps->num_tile_columns_minus1;
605 result.num_tile_rows_minus1 = pic->pps->num_tile_rows_minus1;
606 result.log2_parallel_merge_level_minus2 = pic->pps->log2_parallel_merge_level_minus2;
607 result.init_qp_minus26 = pic->pps->init_qp_minus26;
608
609 for (i = 0; i < 19; ++i)
610 result.column_width_minus1[i] = pic->pps->column_width_minus1[i];
611
612 for (i = 0; i < 21; ++i)
613 result.row_height_minus1[i] = pic->pps->row_height_minus1[i];
614
615 result.num_delta_pocs_ref_rps_idx = pic->NumDeltaPocsOfRefRpsIdx;
616 result.curr_idx = pic->CurrPicOrderCntVal;
617 result.curr_poc = pic->CurrPicOrderCntVal;
618
619 vl_video_buffer_set_associated_data(target, &dec->base,
620 (void *)(uintptr_t)pic->CurrPicOrderCntVal,
621 &ruvd_destroy_associated_data);
622
623 for (i = 0; i < 16; ++i) {
624 struct pipe_video_buffer *ref = pic->ref[i];
625 uintptr_t ref_pic = 0;
626
627 result.poc_list[i] = pic->PicOrderCntVal[i];
628
629 if (ref)
630 ref_pic = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
631 else
632 ref_pic = 0x7F;
633 result.ref_pic_list[i] = ref_pic;
634 }
635
636 for (i = 0; i < 8; ++i) {
637 result.ref_pic_set_st_curr_before[i] = 0xFF;
638 result.ref_pic_set_st_curr_after[i] = 0xFF;
639 result.ref_pic_set_lt_curr[i] = 0xFF;
640 }
641
642 for (i = 0; i < pic->NumPocStCurrBefore; ++i)
643 result.ref_pic_set_st_curr_before[i] = pic->RefPicSetStCurrBefore[i];
644
645 for (i = 0; i < pic->NumPocStCurrAfter; ++i)
646 result.ref_pic_set_st_curr_after[i] = pic->RefPicSetStCurrAfter[i];
647
648 for (i = 0; i < pic->NumPocLtCurr; ++i)
649 result.ref_pic_set_lt_curr[i] = pic->RefPicSetLtCurr[i];
650
651 for (i = 0; i < 6; ++i)
652 result.ucScalingListDCCoefSizeID2[i] = pic->pps->sps->ScalingListDCCoeff16x16[i];
653
654 for (i = 0; i < 2; ++i)
655 result.ucScalingListDCCoefSizeID3[i] = pic->pps->sps->ScalingListDCCoeff32x32[i];
656
657 memcpy(dec->it, pic->pps->sps->ScalingList4x4, 6 * 16);
658 memcpy(dec->it + 96, pic->pps->sps->ScalingList8x8, 6 * 64);
659 memcpy(dec->it + 480, pic->pps->sps->ScalingList16x16, 6 * 64);
660 memcpy(dec->it + 864, pic->pps->sps->ScalingList32x32, 2 * 64);
661
662 for (i = 0 ; i < 2 ; i++) {
663 for (int j = 0 ; j < 15 ; j++)
664 result.direct_reflist[i][j] = pic->RefPicList[i][j];
665 }
666
667 if (pic->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10) {
668 if (target->buffer_format == PIPE_FORMAT_P016) {
669 result.p010_mode = 1;
670 result.msb_mode = 1;
671 } else {
672 result.luma_10to8 = 5;
673 result.chroma_10to8 = 5;
674 result.sclr_luma10to8 = 4;
675 result.sclr_chroma10to8 = 4;
676 }
677 }
678
679 /* TODO
680 result.highestTid;
681 result.isNonRef;
682
683 IDRPicFlag;
684 RAPPicFlag;
685 NumPocTotalCurr;
686 NumShortTermPictureSliceHeaderBits;
687 NumLongTermPictureSliceHeaderBits;
688
689 IsLongTerm[16];
690 */
691
692 return result;
693 }
694
695 /* get vc1 specific message bits */
696 static struct ruvd_vc1 get_vc1_msg(struct pipe_vc1_picture_desc *pic)
697 {
698 struct ruvd_vc1 result;
699
700 memset(&result, 0, sizeof(result));
701
702 switch(pic->base.profile) {
703 case PIPE_VIDEO_PROFILE_VC1_SIMPLE:
704 result.profile = RUVD_VC1_PROFILE_SIMPLE;
705 result.level = 1;
706 break;
707
708 case PIPE_VIDEO_PROFILE_VC1_MAIN:
709 result.profile = RUVD_VC1_PROFILE_MAIN;
710 result.level = 2;
711 break;
712
713 case PIPE_VIDEO_PROFILE_VC1_ADVANCED:
714 result.profile = RUVD_VC1_PROFILE_ADVANCED;
715 result.level = 4;
716 break;
717
718 default:
719 assert(0);
720 }
721
722 /* fields common for all profiles */
723 result.sps_info_flags |= pic->postprocflag << 7;
724 result.sps_info_flags |= pic->pulldown << 6;
725 result.sps_info_flags |= pic->interlace << 5;
726 result.sps_info_flags |= pic->tfcntrflag << 4;
727 result.sps_info_flags |= pic->finterpflag << 3;
728 result.sps_info_flags |= pic->psf << 1;
729
730 result.pps_info_flags |= pic->range_mapy_flag << 31;
731 result.pps_info_flags |= pic->range_mapy << 28;
732 result.pps_info_flags |= pic->range_mapuv_flag << 27;
733 result.pps_info_flags |= pic->range_mapuv << 24;
734 result.pps_info_flags |= pic->multires << 21;
735 result.pps_info_flags |= pic->maxbframes << 16;
736 result.pps_info_flags |= pic->overlap << 11;
737 result.pps_info_flags |= pic->quantizer << 9;
738 result.pps_info_flags |= pic->panscan_flag << 7;
739 result.pps_info_flags |= pic->refdist_flag << 6;
740 result.pps_info_flags |= pic->vstransform << 0;
741
742 /* some fields only apply to main/advanced profile */
743 if (pic->base.profile != PIPE_VIDEO_PROFILE_VC1_SIMPLE) {
744 result.pps_info_flags |= pic->syncmarker << 20;
745 result.pps_info_flags |= pic->rangered << 19;
746 result.pps_info_flags |= pic->loopfilter << 5;
747 result.pps_info_flags |= pic->fastuvmc << 4;
748 result.pps_info_flags |= pic->extended_mv << 3;
749 result.pps_info_flags |= pic->extended_dmv << 8;
750 result.pps_info_flags |= pic->dquant << 1;
751 }
752
753 result.chroma_format = 1;
754
755 #if 0
756 //(((unsigned int)(pPicParams->advance.reserved1)) << SPS_INFO_VC1_RESERVED_SHIFT)
757 uint32_t slice_count
758 uint8_t picture_type
759 uint8_t frame_coding_mode
760 uint8_t deblockEnable
761 uint8_t pquant
762 #endif
763
764 return result;
765 }
766
767 /* extract the frame number from a referenced video buffer */
768 static uint32_t get_ref_pic_idx(struct ruvd_decoder *dec, struct pipe_video_buffer *ref)
769 {
770 uint32_t min = MAX2(dec->frame_number, NUM_MPEG2_REFS) - NUM_MPEG2_REFS;
771 uint32_t max = MAX2(dec->frame_number, 1) - 1;
772 uintptr_t frame;
773
774 /* seems to be the most sane fallback */
775 if (!ref)
776 return max;
777
778 /* get the frame number from the associated data */
779 frame = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
780
781 /* limit the frame number to a valid range */
782 return MAX2(MIN2(frame, max), min);
783 }
784
785 /* get mpeg2 specific msg bits */
786 static struct ruvd_mpeg2 get_mpeg2_msg(struct ruvd_decoder *dec,
787 struct pipe_mpeg12_picture_desc *pic)
788 {
789 const int *zscan = pic->alternate_scan ? vl_zscan_alternate : vl_zscan_normal;
790 struct ruvd_mpeg2 result;
791 unsigned i;
792
793 memset(&result, 0, sizeof(result));
794 result.decoded_pic_idx = dec->frame_number;
795 for (i = 0; i < 2; ++i)
796 result.ref_pic_idx[i] = get_ref_pic_idx(dec, pic->ref[i]);
797
798 result.load_intra_quantiser_matrix = 1;
799 result.load_nonintra_quantiser_matrix = 1;
800
801 for (i = 0; i < 64; ++i) {
802 result.intra_quantiser_matrix[i] = pic->intra_matrix[zscan[i]];
803 result.nonintra_quantiser_matrix[i] = pic->non_intra_matrix[zscan[i]];
804 }
805
806 result.profile_and_level_indication = 0;
807 result.chroma_format = 0x1;
808
809 result.picture_coding_type = pic->picture_coding_type;
810 result.f_code[0][0] = pic->f_code[0][0] + 1;
811 result.f_code[0][1] = pic->f_code[0][1] + 1;
812 result.f_code[1][0] = pic->f_code[1][0] + 1;
813 result.f_code[1][1] = pic->f_code[1][1] + 1;
814 result.intra_dc_precision = pic->intra_dc_precision;
815 result.pic_structure = pic->picture_structure;
816 result.top_field_first = pic->top_field_first;
817 result.frame_pred_frame_dct = pic->frame_pred_frame_dct;
818 result.concealment_motion_vectors = pic->concealment_motion_vectors;
819 result.q_scale_type = pic->q_scale_type;
820 result.intra_vlc_format = pic->intra_vlc_format;
821 result.alternate_scan = pic->alternate_scan;
822
823 return result;
824 }
825
826 /* get mpeg4 specific msg bits */
827 static struct ruvd_mpeg4 get_mpeg4_msg(struct ruvd_decoder *dec,
828 struct pipe_mpeg4_picture_desc *pic)
829 {
830 struct ruvd_mpeg4 result;
831 unsigned i;
832
833 memset(&result, 0, sizeof(result));
834 result.decoded_pic_idx = dec->frame_number;
835 for (i = 0; i < 2; ++i)
836 result.ref_pic_idx[i] = get_ref_pic_idx(dec, pic->ref[i]);
837
838 result.variant_type = 0;
839 result.profile_and_level_indication = 0xF0; // ASP Level0
840
841 result.video_object_layer_verid = 0x5; // advanced simple
842 result.video_object_layer_shape = 0x0; // rectangular
843
844 result.video_object_layer_width = dec->base.width;
845 result.video_object_layer_height = dec->base.height;
846
847 result.vop_time_increment_resolution = pic->vop_time_increment_resolution;
848
849 result.flags |= pic->short_video_header << 0;
850 //result.flags |= obmc_disable << 1;
851 result.flags |= pic->interlaced << 2;
852 result.flags |= 1 << 3; // load_intra_quant_mat
853 result.flags |= 1 << 4; // load_nonintra_quant_mat
854 result.flags |= pic->quarter_sample << 5;
855 result.flags |= 1 << 6; // complexity_estimation_disable
856 result.flags |= pic->resync_marker_disable << 7;
857 //result.flags |= data_partitioned << 8;
858 //result.flags |= reversible_vlc << 9;
859 result.flags |= 0 << 10; // newpred_enable
860 result.flags |= 0 << 11; // reduced_resolution_vop_enable
861 //result.flags |= scalability << 12;
862 //result.flags |= is_object_layer_identifier << 13;
863 //result.flags |= fixed_vop_rate << 14;
864 //result.flags |= newpred_segment_type << 15;
865
866 result.quant_type = pic->quant_type;
867
868 for (i = 0; i < 64; ++i) {
869 result.intra_quant_mat[i] = pic->intra_matrix[vl_zscan_normal[i]];
870 result.nonintra_quant_mat[i] = pic->non_intra_matrix[vl_zscan_normal[i]];
871 }
872
873 /*
874 int32_t trd [2]
875 int32_t trb [2]
876 uint8_t vop_coding_type
877 uint8_t vop_fcode_forward
878 uint8_t vop_fcode_backward
879 uint8_t rounding_control
880 uint8_t alternate_vertical_scan_flag
881 uint8_t top_field_first
882 */
883
884 return result;
885 }
886
887 static void get_mjpeg_slice_header(struct ruvd_decoder *dec, struct pipe_mjpeg_picture_desc *pic)
888 {
889 int size = 0, saved_size, len_pos, i;
890 uint16_t *bs;
891 uint8_t *buf = dec->bs_ptr;
892
893 /* SOI */
894 buf[size++] = 0xff;
895 buf[size++] = 0xd8;
896
897 /* DQT */
898 buf[size++] = 0xff;
899 buf[size++] = 0xdb;
900
901 len_pos = size++;
902 size++;
903
904 for (i = 0; i < 4; ++i) {
905 if (pic->quantization_table.load_quantiser_table[i] == 0)
906 continue;
907
908 buf[size++] = i;
909 memcpy((buf + size), &pic->quantization_table.quantiser_table[i], 64);
910 size += 64;
911 }
912
913 bs = (uint16_t*)&buf[len_pos];
914 *bs = util_bswap16(size - 4);
915
916 saved_size = size;
917
918 /* DHT */
919 buf[size++] = 0xff;
920 buf[size++] = 0xc4;
921
922 len_pos = size++;
923 size++;
924
925 for (i = 0; i < 2; ++i) {
926 if (pic->huffman_table.load_huffman_table[i] == 0)
927 continue;
928
929 buf[size++] = 0x00 | i;
930 memcpy((buf + size), &pic->huffman_table.table[i].num_dc_codes, 16);
931 size += 16;
932 memcpy((buf + size), &pic->huffman_table.table[i].dc_values, 12);
933 size += 12;
934 }
935
936 for (i = 0; i < 2; ++i) {
937 if (pic->huffman_table.load_huffman_table[i] == 0)
938 continue;
939
940 buf[size++] = 0x10 | i;
941 memcpy((buf + size), &pic->huffman_table.table[i].num_ac_codes, 16);
942 size += 16;
943 memcpy((buf + size), &pic->huffman_table.table[i].ac_values, 162);
944 size += 162;
945 }
946
947 bs = (uint16_t*)&buf[len_pos];
948 *bs = util_bswap16(size - saved_size - 2);
949
950 saved_size = size;
951
952 /* DRI */
953 if (pic->slice_parameter.restart_interval) {
954 buf[size++] = 0xff;
955 buf[size++] = 0xdd;
956 buf[size++] = 0x00;
957 buf[size++] = 0x04;
958 bs = (uint16_t*)&buf[size++];
959 *bs = util_bswap16(pic->slice_parameter.restart_interval);
960 saved_size = ++size;
961 }
962
963 /* SOF */
964 buf[size++] = 0xff;
965 buf[size++] = 0xc0;
966
967 len_pos = size++;
968 size++;
969
970 buf[size++] = 0x08;
971
972 bs = (uint16_t*)&buf[size++];
973 *bs = util_bswap16(pic->picture_parameter.picture_height);
974 size++;
975
976 bs = (uint16_t*)&buf[size++];
977 *bs = util_bswap16(pic->picture_parameter.picture_width);
978 size++;
979
980 buf[size++] = pic->picture_parameter.num_components;
981
982 for (i = 0; i < pic->picture_parameter.num_components; ++i) {
983 buf[size++] = pic->picture_parameter.components[i].component_id;
984 buf[size++] = pic->picture_parameter.components[i].h_sampling_factor << 4 |
985 pic->picture_parameter.components[i].v_sampling_factor;
986 buf[size++] = pic->picture_parameter.components[i].quantiser_table_selector;
987 }
988
989 bs = (uint16_t*)&buf[len_pos];
990 *bs = util_bswap16(size - saved_size - 2);
991
992 saved_size = size;
993
994 /* SOS */
995 buf[size++] = 0xff;
996 buf[size++] = 0xda;
997
998 len_pos = size++;
999 size++;
1000
1001 buf[size++] = pic->slice_parameter.num_components;
1002
1003 for (i = 0; i < pic->slice_parameter.num_components; ++i) {
1004 buf[size++] = pic->slice_parameter.components[i].component_selector;
1005 buf[size++] = pic->slice_parameter.components[i].dc_table_selector << 4 |
1006 pic->slice_parameter.components[i].ac_table_selector;
1007 }
1008
1009 buf[size++] = 0x00;
1010 buf[size++] = 0x3f;
1011 buf[size++] = 0x00;
1012
1013 bs = (uint16_t*)&buf[len_pos];
1014 *bs = util_bswap16(size - saved_size - 2);
1015
1016 dec->bs_ptr += size;
1017 dec->bs_size += size;
1018 }
1019
1020 /**
1021 * destroy this video decoder
1022 */
1023 static void ruvd_destroy(struct pipe_video_codec *decoder)
1024 {
1025 struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
1026 unsigned i;
1027
1028 assert(decoder);
1029
1030 map_msg_fb_it_buf(dec);
1031 dec->msg->size = sizeof(*dec->msg);
1032 dec->msg->msg_type = RUVD_MSG_DESTROY;
1033 dec->msg->stream_handle = dec->stream_handle;
1034 send_msg_buf(dec);
1035
1036 flush(dec, 0);
1037
1038 dec->ws->cs_destroy(dec->cs);
1039
1040 for (i = 0; i < NUM_BUFFERS; ++i) {
1041 rvid_destroy_buffer(&dec->msg_fb_it_buffers[i]);
1042 rvid_destroy_buffer(&dec->bs_buffers[i]);
1043 }
1044
1045 rvid_destroy_buffer(&dec->dpb);
1046 rvid_destroy_buffer(&dec->ctx);
1047 rvid_destroy_buffer(&dec->sessionctx);
1048
1049 FREE(dec);
1050 }
1051
1052 /**
1053 * start decoding of a new frame
1054 */
1055 static void ruvd_begin_frame(struct pipe_video_codec *decoder,
1056 struct pipe_video_buffer *target,
1057 struct pipe_picture_desc *picture)
1058 {
1059 struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
1060 uintptr_t frame;
1061
1062 assert(decoder);
1063
1064 frame = ++dec->frame_number;
1065 vl_video_buffer_set_associated_data(target, decoder, (void *)frame,
1066 &ruvd_destroy_associated_data);
1067
1068 dec->bs_size = 0;
1069 dec->bs_ptr = dec->ws->buffer_map(
1070 dec->bs_buffers[dec->cur_buffer].res->buf,
1071 dec->cs, PIPE_TRANSFER_WRITE);
1072 }
1073
1074 /**
1075 * decode a macroblock
1076 */
1077 static void ruvd_decode_macroblock(struct pipe_video_codec *decoder,
1078 struct pipe_video_buffer *target,
1079 struct pipe_picture_desc *picture,
1080 const struct pipe_macroblock *macroblocks,
1081 unsigned num_macroblocks)
1082 {
1083 /* not supported (yet) */
1084 assert(0);
1085 }
1086
1087 /**
1088 * decode a bitstream
1089 */
1090 static void ruvd_decode_bitstream(struct pipe_video_codec *decoder,
1091 struct pipe_video_buffer *target,
1092 struct pipe_picture_desc *picture,
1093 unsigned num_buffers,
1094 const void * const *buffers,
1095 const unsigned *sizes)
1096 {
1097 struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
1098 enum pipe_video_format format = u_reduce_video_profile(picture->profile);
1099 unsigned i;
1100
1101 assert(decoder);
1102
1103 if (!dec->bs_ptr)
1104 return;
1105
1106 if (format == PIPE_VIDEO_FORMAT_JPEG)
1107 get_mjpeg_slice_header(dec, (struct pipe_mjpeg_picture_desc*)picture);
1108
1109 for (i = 0; i < num_buffers; ++i) {
1110 struct rvid_buffer *buf = &dec->bs_buffers[dec->cur_buffer];
1111 unsigned new_size = dec->bs_size + sizes[i];
1112
1113 if (format == PIPE_VIDEO_FORMAT_JPEG)
1114 new_size += 2; /* save for EOI */
1115
1116 if (new_size > buf->res->buf->size) {
1117 dec->ws->buffer_unmap(buf->res->buf);
1118 if (!rvid_resize_buffer(dec->screen, dec->cs, buf, new_size)) {
1119 RVID_ERR("Can't resize bitstream buffer!");
1120 return;
1121 }
1122
1123 dec->bs_ptr = dec->ws->buffer_map(buf->res->buf, dec->cs,
1124 PIPE_TRANSFER_WRITE);
1125 if (!dec->bs_ptr)
1126 return;
1127
1128 dec->bs_ptr += dec->bs_size;
1129 }
1130
1131 memcpy(dec->bs_ptr, buffers[i], sizes[i]);
1132 dec->bs_size += sizes[i];
1133 dec->bs_ptr += sizes[i];
1134 }
1135
1136 if (format == PIPE_VIDEO_FORMAT_JPEG) {
1137 ((uint8_t *)dec->bs_ptr)[0] = 0xff; /* EOI */
1138 ((uint8_t *)dec->bs_ptr)[1] = 0xd9;
1139 dec->bs_size += 2;
1140 dec->bs_ptr += 2;
1141 }
1142 }
1143
1144 /**
1145 * end decoding of the current frame
1146 */
1147 static void ruvd_end_frame(struct pipe_video_codec *decoder,
1148 struct pipe_video_buffer *target,
1149 struct pipe_picture_desc *picture)
1150 {
1151 struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
1152 struct pb_buffer *dt;
1153 struct rvid_buffer *msg_fb_it_buf, *bs_buf;
1154 unsigned bs_size;
1155
1156 assert(decoder);
1157
1158 if (!dec->bs_ptr)
1159 return;
1160
1161 msg_fb_it_buf = &dec->msg_fb_it_buffers[dec->cur_buffer];
1162 bs_buf = &dec->bs_buffers[dec->cur_buffer];
1163
1164 bs_size = align(dec->bs_size, 128);
1165 memset(dec->bs_ptr, 0, bs_size - dec->bs_size);
1166 dec->ws->buffer_unmap(bs_buf->res->buf);
1167
1168 map_msg_fb_it_buf(dec);
1169 dec->msg->size = sizeof(*dec->msg);
1170 dec->msg->msg_type = RUVD_MSG_DECODE;
1171 dec->msg->stream_handle = dec->stream_handle;
1172 dec->msg->status_report_feedback_number = dec->frame_number;
1173
1174 dec->msg->body.decode.stream_type = dec->stream_type;
1175 dec->msg->body.decode.decode_flags = 0x1;
1176 dec->msg->body.decode.width_in_samples = dec->base.width;
1177 dec->msg->body.decode.height_in_samples = dec->base.height;
1178
1179 if ((picture->profile == PIPE_VIDEO_PROFILE_VC1_SIMPLE) ||
1180 (picture->profile == PIPE_VIDEO_PROFILE_VC1_MAIN)) {
1181 dec->msg->body.decode.width_in_samples = align(dec->msg->body.decode.width_in_samples, 16) / 16;
1182 dec->msg->body.decode.height_in_samples = align(dec->msg->body.decode.height_in_samples, 16) / 16;
1183 }
1184
1185 if (dec->dpb.res)
1186 dec->msg->body.decode.dpb_size = dec->dpb.res->buf->size;
1187 dec->msg->body.decode.bsd_size = bs_size;
1188 dec->msg->body.decode.db_pitch = align(dec->base.width, get_db_pitch_alignment(dec));
1189
1190 dt = dec->set_dtb(dec->msg, (struct vl_video_buffer *)target);
1191
1192 switch (u_reduce_video_profile(picture->profile)) {
1193 case PIPE_VIDEO_FORMAT_MPEG4_AVC:
1194 dec->msg->body.decode.codec.h264 = get_h264_msg(dec, (struct pipe_h264_picture_desc*)picture);
1195 break;
1196
1197 case PIPE_VIDEO_FORMAT_HEVC:
1198 dec->msg->body.decode.codec.h265 = get_h265_msg(dec, target, (struct pipe_h265_picture_desc*)picture);
1199 if (dec->ctx.res == NULL) {
1200 unsigned ctx_size;
1201 if (dec->base.profile == PIPE_VIDEO_PROFILE_HEVC_MAIN_10)
1202 ctx_size = calc_ctx_size_h265_main10(dec, (struct pipe_h265_picture_desc*)picture);
1203 else
1204 ctx_size = calc_ctx_size_h265_main(dec);
1205 if (!rvid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT)) {
1206 RVID_ERR("Can't allocated context buffer.\n");
1207 }
1208 rvid_clear_buffer(decoder->context, &dec->ctx);
1209 }
1210
1211 if (dec->ctx.res)
1212 dec->msg->body.decode.dpb_reserved = dec->ctx.res->buf->size;
1213 break;
1214
1215 case PIPE_VIDEO_FORMAT_VC1:
1216 dec->msg->body.decode.codec.vc1 = get_vc1_msg((struct pipe_vc1_picture_desc*)picture);
1217 break;
1218
1219 case PIPE_VIDEO_FORMAT_MPEG12:
1220 dec->msg->body.decode.codec.mpeg2 = get_mpeg2_msg(dec, (struct pipe_mpeg12_picture_desc*)picture);
1221 break;
1222
1223 case PIPE_VIDEO_FORMAT_MPEG4:
1224 dec->msg->body.decode.codec.mpeg4 = get_mpeg4_msg(dec, (struct pipe_mpeg4_picture_desc*)picture);
1225 break;
1226
1227 case PIPE_VIDEO_FORMAT_JPEG:
1228 break;
1229
1230 default:
1231 assert(0);
1232 return;
1233 }
1234
1235 dec->msg->body.decode.db_surf_tile_config = dec->msg->body.decode.dt_surf_tile_config;
1236 dec->msg->body.decode.extension_support = 0x1;
1237
1238 /* set at least the feedback buffer size */
1239 dec->fb[0] = dec->fb_size;
1240
1241 send_msg_buf(dec);
1242
1243 if (dec->dpb.res)
1244 send_cmd(dec, RUVD_CMD_DPB_BUFFER, dec->dpb.res->buf, 0,
1245 RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
1246
1247 if (dec->ctx.res)
1248 send_cmd(dec, RUVD_CMD_CONTEXT_BUFFER, dec->ctx.res->buf, 0,
1249 RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
1250 send_cmd(dec, RUVD_CMD_BITSTREAM_BUFFER, bs_buf->res->buf,
1251 0, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
1252 send_cmd(dec, RUVD_CMD_DECODING_TARGET_BUFFER, dt, 0,
1253 RADEON_USAGE_WRITE, RADEON_DOMAIN_VRAM);
1254 send_cmd(dec, RUVD_CMD_FEEDBACK_BUFFER, msg_fb_it_buf->res->buf,
1255 FB_BUFFER_OFFSET, RADEON_USAGE_WRITE, RADEON_DOMAIN_GTT);
1256 if (have_it(dec))
1257 send_cmd(dec, RUVD_CMD_ITSCALING_TABLE_BUFFER, msg_fb_it_buf->res->buf,
1258 FB_BUFFER_OFFSET + dec->fb_size, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
1259 set_reg(dec, dec->reg.cntl, 1);
1260
1261 flush(dec, PIPE_FLUSH_ASYNC);
1262 next_buffer(dec);
1263 }
1264
1265 /**
1266 * flush any outstanding command buffers to the hardware
1267 */
1268 static void ruvd_flush(struct pipe_video_codec *decoder)
1269 {
1270 }
1271
1272 /**
1273 * create and UVD decoder
1274 */
1275 struct pipe_video_codec *ruvd_create_decoder(struct pipe_context *context,
1276 const struct pipe_video_codec *templ,
1277 ruvd_set_dtb set_dtb)
1278 {
1279 struct radeon_winsys* ws = ((struct r600_common_context *)context)->ws;
1280 struct r600_common_context *rctx = (struct r600_common_context*)context;
1281 unsigned dpb_size;
1282 unsigned width = templ->width, height = templ->height;
1283 unsigned bs_buf_size;
1284 struct radeon_info info;
1285 struct ruvd_decoder *dec;
1286 int r, i;
1287
1288 ws->query_info(ws, &info);
1289
1290 switch(u_reduce_video_profile(templ->profile)) {
1291 case PIPE_VIDEO_FORMAT_MPEG12:
1292 if (templ->entrypoint > PIPE_VIDEO_ENTRYPOINT_BITSTREAM || info.family < CHIP_PALM)
1293 return vl_create_mpeg12_decoder(context, templ);
1294
1295 /* fall through */
1296 case PIPE_VIDEO_FORMAT_MPEG4:
1297 width = align(width, VL_MACROBLOCK_WIDTH);
1298 height = align(height, VL_MACROBLOCK_HEIGHT);
1299 break;
1300 case PIPE_VIDEO_FORMAT_MPEG4_AVC:
1301 width = align(width, VL_MACROBLOCK_WIDTH);
1302 height = align(height, VL_MACROBLOCK_HEIGHT);
1303 break;
1304
1305 default:
1306 break;
1307 }
1308
1309
1310 dec = CALLOC_STRUCT(ruvd_decoder);
1311
1312 if (!dec)
1313 return NULL;
1314
1315 if (info.drm_major < 3)
1316 dec->use_legacy = true;
1317
1318 dec->base = *templ;
1319 dec->base.context = context;
1320 dec->base.width = width;
1321 dec->base.height = height;
1322
1323 dec->base.destroy = ruvd_destroy;
1324 dec->base.begin_frame = ruvd_begin_frame;
1325 dec->base.decode_macroblock = ruvd_decode_macroblock;
1326 dec->base.decode_bitstream = ruvd_decode_bitstream;
1327 dec->base.end_frame = ruvd_end_frame;
1328 dec->base.flush = ruvd_flush;
1329
1330 dec->stream_type = profile2stream_type(dec, info.family);
1331 dec->set_dtb = set_dtb;
1332 dec->stream_handle = rvid_alloc_stream_handle();
1333 dec->screen = context->screen;
1334 dec->ws = ws;
1335 dec->cs = ws->cs_create(rctx->ctx, RING_UVD, NULL, NULL);
1336 if (!dec->cs) {
1337 RVID_ERR("Can't get command submission context.\n");
1338 goto error;
1339 }
1340
1341 dec->fb_size = FB_BUFFER_SIZE;
1342 bs_buf_size = width * height * (512 / (16 * 16));
1343 for (i = 0; i < NUM_BUFFERS; ++i) {
1344 unsigned msg_fb_it_size = FB_BUFFER_OFFSET + dec->fb_size;
1345 STATIC_ASSERT(sizeof(struct ruvd_msg) <= FB_BUFFER_OFFSET);
1346 if (have_it(dec))
1347 msg_fb_it_size += IT_SCALING_TABLE_SIZE;
1348 if (!rvid_create_buffer(dec->screen, &dec->msg_fb_it_buffers[i],
1349 msg_fb_it_size, PIPE_USAGE_STAGING)) {
1350 RVID_ERR("Can't allocated message buffers.\n");
1351 goto error;
1352 }
1353
1354 if (!rvid_create_buffer(dec->screen, &dec->bs_buffers[i],
1355 bs_buf_size, PIPE_USAGE_STAGING)) {
1356 RVID_ERR("Can't allocated bitstream buffers.\n");
1357 goto error;
1358 }
1359
1360 rvid_clear_buffer(context, &dec->msg_fb_it_buffers[i]);
1361 rvid_clear_buffer(context, &dec->bs_buffers[i]);
1362 }
1363
1364 dpb_size = calc_dpb_size(dec);
1365 if (dpb_size) {
1366 if (!rvid_create_buffer(dec->screen, &dec->dpb, dpb_size, PIPE_USAGE_DEFAULT)) {
1367 RVID_ERR("Can't allocated dpb.\n");
1368 goto error;
1369 }
1370 rvid_clear_buffer(context, &dec->dpb);
1371 }
1372
1373 dec->reg.data0 = RUVD_GPCOM_VCPU_DATA0;
1374 dec->reg.data1 = RUVD_GPCOM_VCPU_DATA1;
1375 dec->reg.cmd = RUVD_GPCOM_VCPU_CMD;
1376 dec->reg.cntl = RUVD_ENGINE_CNTL;
1377
1378 map_msg_fb_it_buf(dec);
1379 dec->msg->size = sizeof(*dec->msg);
1380 dec->msg->msg_type = RUVD_MSG_CREATE;
1381 dec->msg->stream_handle = dec->stream_handle;
1382 dec->msg->body.create.stream_type = dec->stream_type;
1383 dec->msg->body.create.width_in_samples = dec->base.width;
1384 dec->msg->body.create.height_in_samples = dec->base.height;
1385 dec->msg->body.create.dpb_size = dpb_size;
1386 send_msg_buf(dec);
1387 r = flush(dec, 0);
1388 if (r)
1389 goto error;
1390
1391 next_buffer(dec);
1392
1393 return &dec->base;
1394
1395 error:
1396 if (dec->cs) dec->ws->cs_destroy(dec->cs);
1397
1398 for (i = 0; i < NUM_BUFFERS; ++i) {
1399 rvid_destroy_buffer(&dec->msg_fb_it_buffers[i]);
1400 rvid_destroy_buffer(&dec->bs_buffers[i]);
1401 }
1402
1403 rvid_destroy_buffer(&dec->dpb);
1404 rvid_destroy_buffer(&dec->ctx);
1405 rvid_destroy_buffer(&dec->sessionctx);
1406
1407 FREE(dec);
1408
1409 return NULL;
1410 }
1411
1412 /* calculate top/bottom offset */
1413 static unsigned texture_offset(struct radeon_surf *surface, unsigned layer)
1414 {
1415 return surface->u.legacy.level[0].offset +
1416 layer * (uint64_t)surface->u.legacy.level[0].slice_size_dw * 4;
1417 }
1418
1419 /* hw encode the aspect of macro tiles */
1420 static unsigned macro_tile_aspect(unsigned macro_tile_aspect)
1421 {
1422 switch (macro_tile_aspect) {
1423 default:
1424 case 1: macro_tile_aspect = 0; break;
1425 case 2: macro_tile_aspect = 1; break;
1426 case 4: macro_tile_aspect = 2; break;
1427 case 8: macro_tile_aspect = 3; break;
1428 }
1429 return macro_tile_aspect;
1430 }
1431
1432 /* hw encode the bank width and height */
1433 static unsigned bank_wh(unsigned bankwh)
1434 {
1435 switch (bankwh) {
1436 default:
1437 case 1: bankwh = 0; break;
1438 case 2: bankwh = 1; break;
1439 case 4: bankwh = 2; break;
1440 case 8: bankwh = 3; break;
1441 }
1442 return bankwh;
1443 }
1444
1445 /**
1446 * fill decoding target field from the luma and chroma surfaces
1447 */
1448 void ruvd_set_dt_surfaces(struct ruvd_msg *msg, struct radeon_surf *luma,
1449 struct radeon_surf *chroma)
1450 {
1451 msg->body.decode.dt_pitch = luma->u.legacy.level[0].nblk_x * luma->blk_w;
1452 switch (luma->u.legacy.level[0].mode) {
1453 case RADEON_SURF_MODE_LINEAR_ALIGNED:
1454 msg->body.decode.dt_tiling_mode = RUVD_TILE_LINEAR;
1455 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_LINEAR;
1456 break;
1457 case RADEON_SURF_MODE_1D:
1458 msg->body.decode.dt_tiling_mode = RUVD_TILE_8X8;
1459 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_1D_THIN;
1460 break;
1461 case RADEON_SURF_MODE_2D:
1462 msg->body.decode.dt_tiling_mode = RUVD_TILE_8X8;
1463 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_2D_THIN;
1464 break;
1465 default:
1466 assert(0);
1467 break;
1468 }
1469
1470 msg->body.decode.dt_luma_top_offset = texture_offset(luma, 0);
1471 if (chroma)
1472 msg->body.decode.dt_chroma_top_offset = texture_offset(chroma, 0);
1473 if (msg->body.decode.dt_field_mode) {
1474 msg->body.decode.dt_luma_bottom_offset = texture_offset(luma, 1);
1475 if (chroma)
1476 msg->body.decode.dt_chroma_bottom_offset = texture_offset(chroma, 1);
1477 } else {
1478 msg->body.decode.dt_luma_bottom_offset = msg->body.decode.dt_luma_top_offset;
1479 msg->body.decode.dt_chroma_bottom_offset = msg->body.decode.dt_chroma_top_offset;
1480 }
1481
1482 if (chroma) {
1483 assert(luma->u.legacy.bankw == chroma->u.legacy.bankw);
1484 assert(luma->u.legacy.bankh == chroma->u.legacy.bankh);
1485 assert(luma->u.legacy.mtilea == chroma->u.legacy.mtilea);
1486 }
1487
1488 msg->body.decode.dt_surf_tile_config |= RUVD_BANK_WIDTH(bank_wh(luma->u.legacy.bankw));
1489 msg->body.decode.dt_surf_tile_config |= RUVD_BANK_HEIGHT(bank_wh(luma->u.legacy.bankh));
1490 msg->body.decode.dt_surf_tile_config |= RUVD_MACRO_TILE_ASPECT_RATIO(macro_tile_aspect(luma->u.legacy.mtilea));
1491 }