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