vl_mpeg12_decoder.c revision 0121aae967d3d1366cccc8946cf89ad22818365e 
1/**************************************************************************
2 *
3 * Copyright 2009 Younes Manton.
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 TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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#include <math.h>
29#include <assert.h>
30
31#include <util/u_memory.h>
32#include <util/u_rect.h>
33#include <util/u_video.h>
34
35#include "vl_mpeg12_decoder.h"
36#include "vl_defines.h"
37
38#define SCALE_FACTOR_SNORM (32768.0f / 256.0f)
39#define SCALE_FACTOR_SSCALED (1.0f / 256.0f)
40
41static const enum pipe_format const_zscan_source_formats[] = {
42   PIPE_FORMAT_R16_SNORM,
43   PIPE_FORMAT_R16_SSCALED
44};
45
46static const unsigned num_zscan_source_formats =
47   sizeof(const_zscan_source_formats) / sizeof(enum pipe_format);
48
49static const enum pipe_format const_idct_source_formats[] = {
50   PIPE_FORMAT_R16G16B16A16_SNORM,
51   PIPE_FORMAT_R16G16B16A16_SSCALED
52};
53
54static const unsigned num_idct_source_formats =
55   sizeof(const_idct_source_formats) / sizeof(enum pipe_format);
56
57static const enum pipe_format const_idct_intermediate_formats[] = {
58   PIPE_FORMAT_R16G16B16A16_FLOAT,
59   PIPE_FORMAT_R16G16B16A16_SNORM,
60   PIPE_FORMAT_R16G16B16A16_SSCALED,
61   PIPE_FORMAT_R32G32B32A32_FLOAT
62};
63
64static const unsigned num_idct_intermediate_formats =
65   sizeof(const_idct_intermediate_formats) / sizeof(enum pipe_format);
66
67static const enum pipe_format const_mc_source_formats[] = {
68   PIPE_FORMAT_R16_SNORM,
69   PIPE_FORMAT_R16_SSCALED
70};
71
72static const unsigned num_mc_source_formats =
73   sizeof(const_mc_source_formats) / sizeof(enum pipe_format);
74
75static bool
76init_zscan_buffer(struct vl_mpeg12_buffer *buffer)
77{
78   enum pipe_format formats[3];
79
80   struct pipe_sampler_view **source;
81   struct pipe_surface **destination;
82
83   struct vl_mpeg12_decoder *dec;
84
85   unsigned i;
86
87   assert(buffer);
88
89   dec = (struct vl_mpeg12_decoder*)buffer->base.decoder;
90
91   formats[0] = formats[1] = formats[2] = dec->zscan_source_format;
92   buffer->zscan_source = vl_video_buffer_init(dec->base.context, dec->pipe,
93                                               dec->blocks_per_line * BLOCK_WIDTH * BLOCK_HEIGHT,
94                                               dec->max_blocks / dec->blocks_per_line,
95                                               1, PIPE_VIDEO_CHROMA_FORMAT_444,
96                                               formats, PIPE_USAGE_STATIC);
97   if (!buffer->zscan_source)
98      goto error_source;
99
100   source = buffer->zscan_source->get_sampler_view_planes(buffer->zscan_source);
101   if (!source)
102      goto error_sampler;
103
104   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
105      destination = dec->idct_source->get_surfaces(dec->idct_source);
106   else
107      destination = dec->mc_source->get_surfaces(dec->mc_source);
108
109   if (!destination)
110      goto error_surface;
111
112   for (i = 0; i < VL_MAX_PLANES; ++i)
113      if (!vl_zscan_init_buffer(i == 0 ? &dec->zscan_y : &dec->zscan_c,
114                                &buffer->zscan[i], source[i], destination[i]))
115         goto error_plane;
116
117   return true;
118
119error_plane:
120   for (; i > 0; --i)
121      vl_zscan_cleanup_buffer(&buffer->zscan[i - 1]);
122
123error_surface:
124error_sampler:
125   buffer->zscan_source->destroy(buffer->zscan_source);
126
127error_source:
128   return false;
129}
130
131static void
132cleanup_zscan_buffer(struct vl_mpeg12_buffer *buffer)
133{
134   unsigned i;
135
136   assert(buffer);
137
138   for (i = 0; i < VL_MAX_PLANES; ++i)
139      vl_zscan_cleanup_buffer(&buffer->zscan[i]);
140   buffer->zscan_source->destroy(buffer->zscan_source);
141}
142
143static bool
144init_idct_buffer(struct vl_mpeg12_buffer *buffer)
145{
146   struct pipe_sampler_view **idct_source_sv, **mc_source_sv;
147   struct pipe_surface **idct_surfaces;
148
149   struct vl_mpeg12_decoder *dec;
150
151   unsigned i;
152
153   assert(buffer);
154
155   dec = (struct vl_mpeg12_decoder*)buffer->base.decoder;
156
157   idct_source_sv = dec->idct_source->get_sampler_view_planes(dec->idct_source);
158   if (!idct_source_sv)
159      goto error_source_sv;
160
161   mc_source_sv = dec->mc_source->get_sampler_view_planes(dec->mc_source);
162   if (!mc_source_sv)
163      goto error_mc_source_sv;
164
165   idct_surfaces = dec->mc_source->get_surfaces(dec->mc_source);
166   if (!idct_surfaces)
167      goto error_surfaces;
168
169   for (i = 0; i < 3; ++i)
170      if (!vl_idct_init_buffer(i == 0 ? &dec->idct_y : &dec->idct_c,
171                               &buffer->idct[i], idct_source_sv[i],
172                               mc_source_sv[i], idct_surfaces[i]))
173         goto error_plane;
174
175   return true;
176
177error_plane:
178   for (; i > 0; --i)
179      vl_idct_cleanup_buffer(i == 1 ? &dec->idct_c : &dec->idct_y, &buffer->idct[i - 1]);
180
181error_surfaces:
182error_mc_source_sv:
183error_source_sv:
184   return false;
185}
186
187static void
188cleanup_idct_buffer(struct vl_mpeg12_buffer *buf)
189{
190   struct vl_mpeg12_decoder *dec;
191   assert(buf);
192
193   dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
194   assert(dec);
195
196   vl_idct_cleanup_buffer(&dec->idct_y, &buf->idct[0]);
197   vl_idct_cleanup_buffer(&dec->idct_c, &buf->idct[1]);
198   vl_idct_cleanup_buffer(&dec->idct_c, &buf->idct[2]);
199}
200
201static bool
202init_mc_buffer(struct vl_mpeg12_buffer *buf)
203{
204   struct vl_mpeg12_decoder *dec;
205
206   assert(buf);
207
208   dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
209   assert(dec);
210
211   if(!vl_mc_init_buffer(&dec->mc_y, &buf->mc[0]))
212      goto error_mc_y;
213
214   if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[1]))
215      goto error_mc_cb;
216
217   if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[2]))
218      goto error_mc_cr;
219
220   return true;
221
222error_mc_cr:
223   vl_mc_cleanup_buffer(&buf->mc[1]);
224
225error_mc_cb:
226   vl_mc_cleanup_buffer(&buf->mc[0]);
227
228error_mc_y:
229   return false;
230}
231
232static void
233cleanup_mc_buffer(struct vl_mpeg12_buffer *buf)
234{
235   unsigned i;
236
237   assert(buf);
238
239   for (i = 0; i < VL_MAX_PLANES; ++i)
240      vl_mc_cleanup_buffer(&buf->mc[i]);
241}
242
243static void
244vl_mpeg12_buffer_destroy(struct pipe_video_decode_buffer *buffer)
245{
246   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
247   struct vl_mpeg12_decoder *dec;
248
249   assert(buf);
250
251   dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
252   assert(dec);
253
254   cleanup_zscan_buffer(buf);
255
256   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
257      cleanup_idct_buffer(buf);
258
259   cleanup_mc_buffer(buf);
260
261   vl_vb_cleanup(&buf->vertex_stream);
262
263   FREE(buf);
264}
265
266static void
267vl_mpeg12_buffer_map(struct pipe_video_decode_buffer *buffer)
268{
269   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
270   struct vl_mpeg12_decoder *dec;
271
272   struct pipe_sampler_view **sampler_views;
273   unsigned i;
274
275   assert(buf);
276
277   dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
278   assert(dec);
279
280   vl_vb_map(&buf->vertex_stream, dec->pipe);
281
282   sampler_views = buf->zscan_source->get_sampler_view_planes(buf->zscan_source);
283
284   assert(sampler_views);
285
286   for (i = 0; i < VL_MAX_PLANES; ++i) {
287      struct pipe_resource *tex = sampler_views[i]->texture;
288      struct pipe_box rect =
289      {
290         0, 0, 0,
291         tex->width0,
292         tex->height0,
293         1
294      };
295
296      buf->tex_transfer[i] = dec->pipe->get_transfer
297      (
298         dec->pipe, tex,
299         0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
300         &rect
301      );
302
303      buf->texels[i] = dec->pipe->transfer_map(dec->pipe, buf->tex_transfer[i]);
304   }
305
306   if (dec->base.entrypoint == PIPE_VIDEO_ENTRYPOINT_BITSTREAM) {
307      struct pipe_ycbcr_block *ycbcr_stream[VL_MAX_PLANES];
308      struct pipe_motionvector *mv_stream[VL_MAX_REF_FRAMES];
309
310      for (i = 0; i < VL_MAX_PLANES; ++i)
311         ycbcr_stream[i] = vl_vb_get_ycbcr_stream(&buf->vertex_stream, i);
312
313      for (i = 0; i < VL_MAX_REF_FRAMES; ++i)
314         mv_stream[i] = vl_vb_get_mv_stream(&buf->vertex_stream, i);
315
316      vl_mpg12_bs_set_buffers(&buf->bs, ycbcr_stream, buf->texels, mv_stream);
317   } else {
318      for (i = 0; i < VL_MAX_PLANES; ++i)
319         vl_zscan_set_layout(&buf->zscan[i], dec->zscan_linear);
320   }
321}
322
323static struct pipe_ycbcr_block *
324vl_mpeg12_buffer_get_ycbcr_stream(struct pipe_video_decode_buffer *buffer, int component)
325{
326   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
327
328   assert(buf);
329
330   return vl_vb_get_ycbcr_stream(&buf->vertex_stream, component);
331}
332
333static short *
334vl_mpeg12_buffer_get_ycbcr_buffer(struct pipe_video_decode_buffer *buffer, int component)
335{
336   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
337
338   assert(buf);
339   assert(component < VL_MAX_PLANES);
340
341   return buf->texels[component];
342}
343
344static unsigned
345vl_mpeg12_buffer_get_mv_stream_stride(struct pipe_video_decode_buffer *buffer)
346{
347   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
348
349   assert(buf);
350
351   return vl_vb_get_mv_stream_stride(&buf->vertex_stream);
352}
353
354static struct pipe_motionvector *
355vl_mpeg12_buffer_get_mv_stream(struct pipe_video_decode_buffer *buffer, int ref_frame)
356{
357   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
358
359   assert(buf);
360
361   return vl_vb_get_mv_stream(&buf->vertex_stream, ref_frame);
362}
363
364static void
365vl_mpeg12_buffer_decode_bitstream(struct pipe_video_decode_buffer *buffer,
366                                  unsigned num_bytes, const void *data,
367                                  struct pipe_mpeg12_picture_desc *picture,
368                                  unsigned num_ycbcr_blocks[3])
369{
370   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
371   struct vl_mpeg12_decoder *dec;
372   unsigned i;
373
374   assert(buf);
375
376   dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
377   assert(dec);
378
379   for (i = 0; i < VL_MAX_PLANES; ++i)
380      vl_zscan_set_layout(&buf->zscan[i], picture->alternate_scan ? dec->zscan_alternate : dec->zscan_normal);
381
382   vl_mpg12_bs_decode(&buf->bs, num_bytes, data, picture, num_ycbcr_blocks);
383}
384
385static void
386vl_mpeg12_buffer_unmap(struct pipe_video_decode_buffer *buffer)
387{
388   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
389   struct vl_mpeg12_decoder *dec;
390   unsigned i;
391
392   assert(buf);
393
394   dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
395   assert(dec);
396
397   vl_vb_unmap(&buf->vertex_stream, dec->pipe);
398
399   for (i = 0; i < VL_MAX_PLANES; ++i) {
400      dec->pipe->transfer_unmap(dec->pipe, buf->tex_transfer[i]);
401      dec->pipe->transfer_destroy(dec->pipe, buf->tex_transfer[i]);
402   }
403}
404
405static void
406vl_mpeg12_destroy(struct pipe_video_decoder *decoder)
407{
408   struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
409
410   assert(decoder);
411
412   /* Asserted in softpipe_delete_fs_state() for some reason */
413   dec->pipe->bind_vs_state(dec->pipe, NULL);
414   dec->pipe->bind_fs_state(dec->pipe, NULL);
415
416   dec->pipe->delete_depth_stencil_alpha_state(dec->pipe, dec->dsa);
417   dec->pipe->delete_sampler_state(dec->pipe, dec->sampler_ycbcr);
418
419   vl_mc_cleanup(&dec->mc_y);
420   vl_mc_cleanup(&dec->mc_c);
421   dec->mc_source->destroy(dec->mc_source);
422
423   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
424      vl_idct_cleanup(&dec->idct_y);
425      vl_idct_cleanup(&dec->idct_c);
426      dec->idct_source->destroy(dec->idct_source);
427   }
428
429   vl_zscan_cleanup(&dec->zscan_y);
430   vl_zscan_cleanup(&dec->zscan_c);
431
432   dec->pipe->delete_vertex_elements_state(dec->pipe, dec->ves_ycbcr);
433   dec->pipe->delete_vertex_elements_state(dec->pipe, dec->ves_mv);
434
435   pipe_resource_reference(&dec->quads.buffer, NULL);
436   pipe_resource_reference(&dec->pos.buffer, NULL);
437
438   pipe_sampler_view_reference(&dec->zscan_linear, NULL);
439   pipe_sampler_view_reference(&dec->zscan_normal, NULL);
440   pipe_sampler_view_reference(&dec->zscan_alternate, NULL);
441
442   FREE(dec);
443}
444
445static struct pipe_video_decode_buffer *
446vl_mpeg12_create_buffer(struct pipe_video_decoder *decoder)
447{
448   struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
449   struct vl_mpeg12_buffer *buffer;
450
451   assert(dec);
452
453   buffer = CALLOC_STRUCT(vl_mpeg12_buffer);
454   if (buffer == NULL)
455      return NULL;
456
457   buffer->base.decoder = decoder;
458   buffer->base.destroy = vl_mpeg12_buffer_destroy;
459   buffer->base.map = vl_mpeg12_buffer_map;
460   buffer->base.get_ycbcr_stream = vl_mpeg12_buffer_get_ycbcr_stream;
461   buffer->base.get_ycbcr_buffer = vl_mpeg12_buffer_get_ycbcr_buffer;
462   buffer->base.get_mv_stream_stride = vl_mpeg12_buffer_get_mv_stream_stride;
463   buffer->base.get_mv_stream = vl_mpeg12_buffer_get_mv_stream;
464   buffer->base.decode_bitstream = vl_mpeg12_buffer_decode_bitstream;
465   buffer->base.unmap = vl_mpeg12_buffer_unmap;
466
467   if (!vl_vb_init(&buffer->vertex_stream, dec->pipe,
468                   dec->base.width / MACROBLOCK_WIDTH,
469                   dec->base.height / MACROBLOCK_HEIGHT))
470      goto error_vertex_buffer;
471
472   if (!init_mc_buffer(buffer))
473      goto error_mc;
474
475   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
476      if (!init_idct_buffer(buffer))
477         goto error_idct;
478
479   if (!init_zscan_buffer(buffer))
480      goto error_zscan;
481
482   if (dec->base.entrypoint == PIPE_VIDEO_ENTRYPOINT_BITSTREAM)
483      vl_mpg12_bs_init(&buffer->bs,
484                       dec->base.width / MACROBLOCK_WIDTH,
485                       dec->base.height / MACROBLOCK_HEIGHT);
486
487   return &buffer->base;
488
489error_zscan:
490   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
491      cleanup_idct_buffer(buffer);
492
493error_idct:
494   cleanup_mc_buffer(buffer);
495
496error_mc:
497   vl_vb_cleanup(&buffer->vertex_stream);
498
499error_vertex_buffer:
500   FREE(buffer);
501   return NULL;
502}
503
504static void
505vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
506                               unsigned num_ycbcr_blocks[3],
507                               struct pipe_video_buffer *refs[2],
508                               struct pipe_video_buffer *dst)
509{
510   struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer *)buffer;
511   struct vl_mpeg12_decoder *dec;
512
513   struct pipe_sampler_view **sv[VL_MAX_REF_FRAMES], **mc_source_sv;
514   struct pipe_surface **surfaces;
515
516   struct pipe_vertex_buffer vb[3];
517
518   unsigned i, j, component;
519   unsigned nr_components;
520
521   assert(buf);
522
523   dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
524   assert(dec);
525
526   for (i = 0; i < 2; ++i)
527      sv[i] = refs[i] ? refs[i]->get_sampler_view_planes(refs[i]) : NULL;
528
529   vb[0] = dec->quads;
530   vb[1] = dec->pos;
531
532   surfaces = dst->get_surfaces(dst);
533
534   dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_mv);
535   for (i = 0; i < VL_MAX_PLANES; ++i) {
536      if (!surfaces[i]) continue;
537
538      vl_mc_set_surface(&buf->mc[i], surfaces[i]);
539
540      for (j = 0; j < VL_MAX_REF_FRAMES; ++j) {
541         if (!sv[j]) continue;
542
543         vb[2] = vl_vb_get_mv(&buf->vertex_stream, j);;
544         dec->pipe->set_vertex_buffers(dec->pipe, 3, vb);
545
546         vl_mc_render_ref(&buf->mc[i], sv[j][i]);
547      }
548   }
549
550   dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_ycbcr);
551   for (i = 0; i < VL_MAX_PLANES; ++i) {
552      if (!num_ycbcr_blocks[i]) continue;
553
554      vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, i);
555      dec->pipe->set_vertex_buffers(dec->pipe, 2, vb);
556
557      vl_zscan_render(&buf->zscan[i] , num_ycbcr_blocks[i]);
558
559      if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
560         vl_idct_flush(i == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[i], num_ycbcr_blocks[i]);
561   }
562
563   mc_source_sv = dec->mc_source->get_sampler_view_planes(dec->mc_source);
564   for (i = 0, component = 0; i < VL_MAX_PLANES; ++i) {
565      if (!surfaces[i]) continue;
566
567      nr_components = util_format_get_nr_components(surfaces[i]->texture->format);
568      for (j = 0; j < nr_components; ++j, ++component) {
569         if (!num_ycbcr_blocks[i]) continue;
570
571         vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, component);
572         dec->pipe->set_vertex_buffers(dec->pipe, 2, vb);
573
574         if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
575            vl_idct_prepare_stage2(component == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[component]);
576         else {
577            dec->pipe->set_fragment_sampler_views(dec->pipe, 1, &mc_source_sv[component]);
578            dec->pipe->bind_fragment_sampler_states(dec->pipe, 1, &dec->sampler_ycbcr);
579         }
580         vl_mc_render_ycbcr(&buf->mc[i], j, num_ycbcr_blocks[component]);
581      }
582   }
583}
584
585static bool
586init_pipe_state(struct vl_mpeg12_decoder *dec)
587{
588   struct pipe_depth_stencil_alpha_state dsa;
589   struct pipe_sampler_state sampler;
590   unsigned i;
591
592   assert(dec);
593
594   memset(&dsa, 0, sizeof dsa);
595   dsa.depth.enabled = 0;
596   dsa.depth.writemask = 0;
597   dsa.depth.func = PIPE_FUNC_ALWAYS;
598   for (i = 0; i < 2; ++i) {
599      dsa.stencil[i].enabled = 0;
600      dsa.stencil[i].func = PIPE_FUNC_ALWAYS;
601      dsa.stencil[i].fail_op = PIPE_STENCIL_OP_KEEP;
602      dsa.stencil[i].zpass_op = PIPE_STENCIL_OP_KEEP;
603      dsa.stencil[i].zfail_op = PIPE_STENCIL_OP_KEEP;
604      dsa.stencil[i].valuemask = 0;
605      dsa.stencil[i].writemask = 0;
606   }
607   dsa.alpha.enabled = 0;
608   dsa.alpha.func = PIPE_FUNC_ALWAYS;
609   dsa.alpha.ref_value = 0;
610   dec->dsa = dec->pipe->create_depth_stencil_alpha_state(dec->pipe, &dsa);
611   dec->pipe->bind_depth_stencil_alpha_state(dec->pipe, dec->dsa);
612
613   memset(&sampler, 0, sizeof(sampler));
614   sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
615   sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
616   sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_BORDER;
617   sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
618   sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
619   sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
620   sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
621   sampler.compare_func = PIPE_FUNC_ALWAYS;
622   sampler.normalized_coords = 1;
623   dec->sampler_ycbcr = dec->pipe->create_sampler_state(dec->pipe, &sampler);
624   if (!dec->sampler_ycbcr)
625      return false;
626
627   return true;
628}
629
630static enum pipe_format
631find_first_supported_format(struct vl_mpeg12_decoder *dec,
632                            const enum pipe_format formats[],
633                            unsigned num_formats,
634                            enum pipe_texture_target target)
635{
636   struct pipe_screen *screen;
637   unsigned i;
638
639   assert(dec);
640
641   screen = dec->pipe->screen;
642
643   for (i = 0; i < num_formats; ++i)
644      if (screen->is_format_supported(dec->pipe->screen, formats[i], target, 1,
645                                      PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
646         return formats[i];
647
648   return PIPE_FORMAT_NONE;
649}
650
651static bool
652init_zscan(struct vl_mpeg12_decoder *dec)
653{
654   unsigned num_channels;
655
656   assert(dec);
657
658   dec->blocks_per_line = 4;
659   dec->max_blocks =
660      (dec->base.width * dec->base.height) /
661      (BLOCK_WIDTH * BLOCK_HEIGHT);
662
663   dec->zscan_source_format = find_first_supported_format(dec, const_zscan_source_formats,
664                                                          num_zscan_source_formats, PIPE_TEXTURE_2D);
665
666   if (dec->zscan_source_format == PIPE_FORMAT_NONE)
667      return false;
668
669   dec->zscan_linear = vl_zscan_layout(dec->pipe, vl_zscan_linear, dec->blocks_per_line);
670   dec->zscan_normal = vl_zscan_layout(dec->pipe, vl_zscan_normal, dec->blocks_per_line);
671   dec->zscan_alternate = vl_zscan_layout(dec->pipe, vl_zscan_alternate, dec->blocks_per_line);
672
673   num_channels = dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT ? 4 : 1;
674
675   if (!vl_zscan_init(&dec->zscan_y, dec->pipe, dec->base.width, dec->base.height,
676                      dec->blocks_per_line, dec->max_blocks, num_channels))
677      return false;
678
679   if (!vl_zscan_init(&dec->zscan_c, dec->pipe, dec->chroma_width, dec->chroma_height,
680                      dec->blocks_per_line, dec->max_blocks, num_channels))
681      return false;
682
683   return true;
684}
685
686static bool
687init_idct(struct vl_mpeg12_decoder *dec, float *mc_scale)
688{
689   unsigned nr_of_idct_render_targets;
690   enum pipe_format formats[3];
691
692   struct pipe_sampler_view *matrix, *transpose = NULL;
693   float matrix_scale, transpose_scale;
694
695   nr_of_idct_render_targets = dec->pipe->screen->get_param(dec->pipe->screen, PIPE_CAP_MAX_RENDER_TARGETS);
696
697   // more than 4 render targets usually doesn't makes any seens
698   nr_of_idct_render_targets = MIN2(nr_of_idct_render_targets, 4);
699
700   formats[0] = formats[1] = formats[2] = find_first_supported_format(dec, const_idct_source_formats,
701                                                                      num_idct_source_formats, PIPE_TEXTURE_2D);
702
703   switch (formats[0]) {
704   case PIPE_FORMAT_NONE:
705      goto error_idct_format;
706
707   case PIPE_FORMAT_R16G16B16A16_SSCALED:
708      matrix_scale = SCALE_FACTOR_SSCALED;
709      break;
710
711   case PIPE_FORMAT_R16G16B16A16_SNORM:
712      matrix_scale = SCALE_FACTOR_SNORM;
713      break;
714
715   default:
716      assert(0);
717      return false;
718   }
719
720   dec->idct_source = vl_video_buffer_init(dec->base.context, dec->pipe,
721                                           dec->base.width / 4, dec->base.height, 1,
722                                           dec->base.chroma_format,
723                                           formats, PIPE_USAGE_STATIC);
724   if (!dec->idct_source)
725      goto error_idct_source;
726
727   formats[0] = formats[1] = formats[2] = find_first_supported_format(dec, const_idct_intermediate_formats,
728                                                                      num_idct_intermediate_formats, PIPE_TEXTURE_3D);
729
730   switch (formats[0]) {
731   case PIPE_FORMAT_NONE:
732      goto error_mc_format;
733
734   case PIPE_FORMAT_R16G16B16A16_FLOAT:
735   case PIPE_FORMAT_R32G32B32A32_FLOAT:
736      transpose_scale = 1.0f;
737      *mc_scale = 1.0f;
738      break;
739
740   case PIPE_FORMAT_R16_SSCALED:
741      transpose_scale = matrix_scale = sqrt(matrix_scale);
742      transpose_scale /= SCALE_FACTOR_SSCALED;
743      *mc_scale = SCALE_FACTOR_SSCALED;
744      break;
745
746   default:
747      transpose_scale = matrix_scale = sqrt(matrix_scale);
748      *mc_scale = 1.0f;
749      break;
750   }
751
752   dec->mc_source = vl_video_buffer_init(dec->base.context, dec->pipe,
753                                         dec->base.width / nr_of_idct_render_targets,
754                                         dec->base.height / 4, nr_of_idct_render_targets,
755                                         dec->base.chroma_format,
756                                         formats, PIPE_USAGE_STATIC);
757
758   if (!dec->mc_source)
759      goto error_mc_source;
760
761   if (!(matrix = vl_idct_upload_matrix(dec->pipe, matrix_scale)))
762      goto error_matrix;
763
764   if (matrix_scale != transpose_scale) {
765      if (!(transpose = vl_idct_upload_matrix(dec->pipe, transpose_scale)))
766         goto error_transpose;
767   } else
768      pipe_sampler_view_reference(&transpose, matrix);
769
770   if (!vl_idct_init(&dec->idct_y, dec->pipe, dec->base.width, dec->base.height,
771                     nr_of_idct_render_targets, matrix, transpose))
772      goto error_y;
773
774   if(!vl_idct_init(&dec->idct_c, dec->pipe, dec->chroma_width, dec->chroma_height,
775                    nr_of_idct_render_targets, matrix, transpose))
776      goto error_c;
777
778   pipe_sampler_view_reference(&matrix, NULL);
779   pipe_sampler_view_reference(&transpose, NULL);
780
781   return true;
782
783error_c:
784   vl_idct_cleanup(&dec->idct_y);
785
786error_y:
787   pipe_sampler_view_reference(&transpose, NULL);
788
789error_transpose:
790   pipe_sampler_view_reference(&matrix, NULL);
791
792error_matrix:
793   dec->mc_source->destroy(dec->mc_source);
794
795error_mc_source:
796error_mc_format:
797   dec->idct_source->destroy(dec->idct_source);
798
799error_idct_source:
800error_idct_format:
801   return false;
802}
803
804static bool
805init_mc_source_widthout_idct(struct vl_mpeg12_decoder *dec, float *mc_scale)
806{
807   enum pipe_format formats[3];
808
809   formats[0] = formats[1] = formats[2] = find_first_supported_format(dec, const_mc_source_formats,
810                                                                      num_mc_source_formats, PIPE_TEXTURE_2D);
811
812   switch (formats[0]) {
813   case PIPE_FORMAT_NONE:
814      return false;
815
816   case PIPE_FORMAT_R16_SNORM:
817      *mc_scale = SCALE_FACTOR_SNORM;
818      break;
819
820   case PIPE_FORMAT_R16_SSCALED:
821      *mc_scale = SCALE_FACTOR_SSCALED;
822      break;
823
824   default:
825      assert(0);
826      return false;
827   }
828
829   dec->mc_source = vl_video_buffer_init(dec->base.context, dec->pipe,
830                                         dec->base.width, dec->base.height, 1,
831                                         dec->base.chroma_format,
832                                         formats, PIPE_USAGE_STATIC);
833
834   return dec->mc_source;
835}
836
837static void
838mc_vert_shader_callback(void *priv, struct vl_mc *mc,
839                        struct ureg_program *shader,
840                        unsigned first_output,
841                        struct ureg_dst tex)
842{
843   struct vl_mpeg12_decoder *dec = priv;
844   struct ureg_dst o_vtex;
845
846   assert(priv && mc);
847   assert(shader);
848
849   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
850      struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c;
851      vl_idct_stage2_vert_shader(idct, shader, first_output, tex);
852   } else {
853      o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, first_output);
854      ureg_MOV(shader, ureg_writemask(o_vtex, TGSI_WRITEMASK_XY), ureg_src(tex));
855   }
856}
857
858static void
859mc_frag_shader_callback(void *priv, struct vl_mc *mc,
860                        struct ureg_program *shader,
861                        unsigned first_input,
862                        struct ureg_dst dst)
863{
864   struct vl_mpeg12_decoder *dec = priv;
865   struct ureg_src src, sampler;
866
867   assert(priv && mc);
868   assert(shader);
869
870   if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
871      struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c;
872      vl_idct_stage2_frag_shader(idct, shader, first_input, dst);
873   } else {
874      src = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, first_input, TGSI_INTERPOLATE_LINEAR);
875      sampler = ureg_DECL_sampler(shader, 0);
876      ureg_TEX(shader, dst, TGSI_TEXTURE_2D, src, sampler);
877   }
878}
879
880struct pipe_video_decoder *
881vl_create_mpeg12_decoder(struct pipe_video_context *context,
882                         struct pipe_context *pipe,
883                         enum pipe_video_profile profile,
884                         enum pipe_video_entrypoint entrypoint,
885                         enum pipe_video_chroma_format chroma_format,
886                         unsigned width, unsigned height)
887{
888   struct vl_mpeg12_decoder *dec;
889   float mc_scale;
890
891   assert(u_reduce_video_profile(profile) == PIPE_VIDEO_CODEC_MPEG12);
892
893   dec = CALLOC_STRUCT(vl_mpeg12_decoder);
894
895   if (!dec)
896      return NULL;
897
898   dec->base.context = context;
899   dec->base.profile = profile;
900   dec->base.entrypoint = entrypoint;
901   dec->base.chroma_format = chroma_format;
902   dec->base.width = width;
903   dec->base.height = height;
904
905   dec->base.destroy = vl_mpeg12_destroy;
906   dec->base.create_buffer = vl_mpeg12_create_buffer;
907   dec->base.flush_buffer = vl_mpeg12_decoder_flush_buffer;
908
909   dec->pipe = pipe;
910
911   dec->quads = vl_vb_upload_quads(dec->pipe);
912   dec->pos = vl_vb_upload_pos(
913      dec->pipe,
914      dec->base.width / MACROBLOCK_WIDTH,
915      dec->base.height / MACROBLOCK_HEIGHT
916   );
917
918   dec->ves_ycbcr = vl_vb_get_ves_ycbcr(dec->pipe);
919   dec->ves_mv = vl_vb_get_ves_mv(dec->pipe);
920
921   /* TODO: Implement 422, 444 */
922   assert(dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420);
923
924   if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420) {
925      dec->chroma_width = dec->base.width / 2;
926      dec->chroma_height = dec->base.height / 2;
927   } else if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_422) {
928      dec->chroma_width = dec->base.width;
929      dec->chroma_height = dec->base.height / 2;
930   } else {
931      dec->chroma_width = dec->base.width;
932      dec->chroma_height = dec->base.height;
933   }
934
935   if (!init_zscan(dec))
936      goto error_zscan;
937
938   if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
939      if (!init_idct(dec, &mc_scale))
940         goto error_sources;
941   } else {
942      if (!init_mc_source_widthout_idct(dec, &mc_scale))
943         goto error_sources;
944   }
945
946   if (!vl_mc_init(&dec->mc_y, dec->pipe, dec->base.width, dec->base.height, MACROBLOCK_HEIGHT, mc_scale,
947                   mc_vert_shader_callback, mc_frag_shader_callback, dec))
948      goto error_mc_y;
949
950   // TODO
951   if (!vl_mc_init(&dec->mc_c, dec->pipe, dec->base.width, dec->base.height, BLOCK_HEIGHT, mc_scale,
952                   mc_vert_shader_callback, mc_frag_shader_callback, dec))
953      goto error_mc_c;
954
955   if (!init_pipe_state(dec))
956      goto error_pipe_state;
957
958   return &dec->base;
959
960error_pipe_state:
961   vl_mc_cleanup(&dec->mc_c);
962
963error_mc_c:
964   vl_mc_cleanup(&dec->mc_y);
965
966error_mc_y:
967   if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
968      vl_idct_cleanup(&dec->idct_y);
969      vl_idct_cleanup(&dec->idct_c);
970      dec->idct_source->destroy(dec->idct_source);
971   }
972   dec->mc_source->destroy(dec->mc_source);
973
974error_sources:
975   vl_zscan_cleanup(&dec->zscan_y);
976   vl_zscan_cleanup(&dec->zscan_c);
977
978error_zscan:
979   FREE(dec);
980   return NULL;
981}
982