src/gallium/auxiliary/vl/vl_mpeg12_decoder.c

   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
  41 struct format_config {
  42    enum pipe_format zscan_source_format;
  43    enum pipe_format idct_source_format;
  44    enum pipe_format mc_source_format;
  45
  46    float idct_scale;
  47    float mc_scale;
  48 };
  49
  50 static const struct format_config bitstream_format_config[] = {
  51    { PIPE_FORMAT_R16_SSCALED, PIPE_FORMAT_R16G16B16A16_SSCALED, PIPE_FORMAT_R16G16B16A16_FLOAT, 1.0f, SCALE_FACTOR_SSCALED },
  52    { PIPE_FORMAT_R16_SSCALED, PIPE_FORMAT_R16G16B16A16_SSCALED, PIPE_FORMAT_R16G16B16A16_SSCALED, 1.0f, SCALE_FACTOR_SSCALED },
  53    { PIPE_FORMAT_R16_SNORM, PIPE_FORMAT_R16G16B16A16_SNORM, PIPE_FORMAT_R16G16B16A16_FLOAT, 1.0f, SCALE_FACTOR_SNORM },
  54    { PIPE_FORMAT_R16_SNORM, PIPE_FORMAT_R16G16B16A16_SNORM, PIPE_FORMAT_R16G16B16A16_SNORM, 1.0f, SCALE_FACTOR_SNORM }
  55 };
  56
  57 static const unsigned num_bitstream_format_configs =
  58    sizeof(bitstream_format_config) / sizeof(struct format_config);
  59
  60 static const struct format_config idct_format_config[] = {
  61    { PIPE_FORMAT_R16_SSCALED, PIPE_FORMAT_R16G16B16A16_SSCALED, PIPE_FORMAT_R16G16B16A16_FLOAT, 1.0f, SCALE_FACTOR_SSCALED },
  62    { PIPE_FORMAT_R16_SSCALED, PIPE_FORMAT_R16G16B16A16_SSCALED, PIPE_FORMAT_R16G16B16A16_SSCALED, 1.0f, SCALE_FACTOR_SSCALED },
  63    { PIPE_FORMAT_R16_SNORM, PIPE_FORMAT_R16G16B16A16_SNORM, PIPE_FORMAT_R16G16B16A16_FLOAT, 1.0f, SCALE_FACTOR_SNORM },
  64    { PIPE_FORMAT_R16_SNORM, PIPE_FORMAT_R16G16B16A16_SNORM, PIPE_FORMAT_R16G16B16A16_SNORM, 1.0f, SCALE_FACTOR_SNORM }
  65 };
  66
  67 static const unsigned num_idct_format_configs =
  68    sizeof(idct_format_config) / sizeof(struct format_config);
  69
  70 static const struct format_config mc_format_config[] = {
  71    //{ PIPE_FORMAT_R16_SSCALED, PIPE_FORMAT_NONE, PIPE_FORMAT_R16_SSCALED, 0.0f, SCALE_FACTOR_SSCALED },
  72    { PIPE_FORMAT_R16_SNORM, PIPE_FORMAT_NONE, PIPE_FORMAT_R16_SNORM, 0.0f, SCALE_FACTOR_SNORM }
  73 };
  74
  75 static const unsigned num_mc_format_configs =
  76    sizeof(mc_format_config) / sizeof(struct format_config);
  77
  78 static bool
  79 init_zscan_buffer(struct vl_mpeg12_buffer *buffer)
  80 {
  81    enum pipe_format formats[3];
  82
  83    struct pipe_sampler_view **source;
  84    struct pipe_surface **destination;
  85
  86    struct vl_mpeg12_decoder *dec;
  87
  88    unsigned i;
  89
  90    assert(buffer);
  91
  92    dec = (struct vl_mpeg12_decoder*)buffer->base.decoder;
  93
  94    formats[0] = formats[1] = formats[2] = dec->zscan_source_format;
  95    buffer->zscan_source = vl_video_buffer_create_ex
  96    (
  97       dec->base.context,
  98       dec->blocks_per_line * BLOCK_WIDTH * BLOCK_HEIGHT,
  99       align(dec->num_blocks, dec->blocks_per_line) / dec->blocks_per_line,
 100       1, PIPE_VIDEO_CHROMA_FORMAT_444, formats, PIPE_USAGE_STATIC
 101    );
 102
 103    if (!buffer->zscan_source)
 104       goto error_source;
 105
 106    source = buffer->zscan_source->get_sampler_view_planes(buffer->zscan_source);
 107    if (!source)
 108       goto error_sampler;
 109
 110    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 111       destination = dec->idct_source->get_surfaces(dec->idct_source);
 112    else
 113       destination = dec->mc_source->get_surfaces(dec->mc_source);
 114
 115    if (!destination)
 116       goto error_surface;
 117
 118    for (i = 0; i < VL_MAX_PLANES; ++i)
 119       if (!vl_zscan_init_buffer(i == 0 ? &dec->zscan_y : &dec->zscan_c,
 120                                 &buffer->zscan[i], source[i], destination[i]))
 121          goto error_plane;
 122
 123    return true;
 124
 125 error_plane:
 126    for (; i > 0; --i)
 127       vl_zscan_cleanup_buffer(&buffer->zscan[i - 1]);
 128
 129 error_surface:
 130 error_sampler:
 131    buffer->zscan_source->destroy(buffer->zscan_source);
 132
 133 error_source:
 134    return false;
 135 }
 136
 137 static void
 138 cleanup_zscan_buffer(struct vl_mpeg12_buffer *buffer)
 139 {
 140    unsigned i;
 141
 142    assert(buffer);
 143
 144    for (i = 0; i < VL_MAX_PLANES; ++i)
 145       vl_zscan_cleanup_buffer(&buffer->zscan[i]);
 146    buffer->zscan_source->destroy(buffer->zscan_source);
 147 }
 148
 149 static bool
 150 init_idct_buffer(struct vl_mpeg12_buffer *buffer)
 151 {
 152    struct pipe_sampler_view **idct_source_sv, **mc_source_sv;
 153
 154    struct vl_mpeg12_decoder *dec;
 155
 156    unsigned i;
 157
 158    assert(buffer);
 159
 160    dec = (struct vl_mpeg12_decoder*)buffer->base.decoder;
 161
 162    idct_source_sv = dec->idct_source->get_sampler_view_planes(dec->idct_source);
 163    if (!idct_source_sv)
 164       goto error_source_sv;
 165
 166    mc_source_sv = dec->mc_source->get_sampler_view_planes(dec->mc_source);
 167    if (!mc_source_sv)
 168       goto error_mc_source_sv;
 169
 170    for (i = 0; i < 3; ++i)
 171       if (!vl_idct_init_buffer(i == 0 ? &dec->idct_y : &dec->idct_c,
 172                                &buffer->idct[i], idct_source_sv[i],
 173                                mc_source_sv[i]))
 174          goto error_plane;
 175
 176    return true;
 177
 178 error_plane:
 179    for (; i > 0; --i)
 180       vl_idct_cleanup_buffer(&buffer->idct[i - 1]);
 181
 182 error_mc_source_sv:
 183 error_source_sv:
 184    return false;
 185 }
 186
 187 static void
 188 cleanup_idct_buffer(struct vl_mpeg12_buffer *buf)
 189 {
 190    struct vl_mpeg12_decoder *dec;
 191    unsigned i;
 192
 193    assert(buf);
 194
 195    dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
 196    assert(dec);
 197
 198    for (i = 0; i < 3; ++i)
 199       vl_idct_cleanup_buffer(&buf->idct[0]);
 200 }
 201
 202 static bool
 203 init_mc_buffer(struct vl_mpeg12_buffer *buf)
 204 {
 205    struct vl_mpeg12_decoder *dec;
 206
 207    assert(buf);
 208
 209    dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
 210    assert(dec);
 211
 212    if(!vl_mc_init_buffer(&dec->mc_y, &buf->mc[0]))
 213       goto error_mc_y;
 214
 215    if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[1]))
 216       goto error_mc_cb;
 217
 218    if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[2]))
 219       goto error_mc_cr;
 220
 221    return true;
 222
 223 error_mc_cr:
 224    vl_mc_cleanup_buffer(&buf->mc[1]);
 225
 226 error_mc_cb:
 227    vl_mc_cleanup_buffer(&buf->mc[0]);
 228
 229 error_mc_y:
 230    return false;
 231 }
 232
 233 static void
 234 cleanup_mc_buffer(struct vl_mpeg12_buffer *buf)
 235 {
 236    unsigned i;
 237
 238    assert(buf);
 239
 240    for (i = 0; i < VL_MAX_PLANES; ++i)
 241       vl_mc_cleanup_buffer(&buf->mc[i]);
 242 }
 243
 244 static void
 245 vl_mpeg12_buffer_destroy(struct pipe_video_decode_buffer *buffer)
 246 {
 247    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 248    struct vl_mpeg12_decoder *dec;
 249
 250    assert(buf);
 251
 252    dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
 253    assert(dec);
 254
 255    cleanup_zscan_buffer(buf);
 256
 257    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 258       cleanup_idct_buffer(buf);
 259
 260    cleanup_mc_buffer(buf);
 261
 262    vl_vb_cleanup(&buf->vertex_stream);
 263
 264    FREE(buf);
 265 }
 266
 267 static void
 268 vl_mpeg12_buffer_begin_frame(struct pipe_video_decode_buffer *buffer)
 269 {
 270    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 271    struct vl_mpeg12_decoder *dec;
 272
 273    struct pipe_sampler_view **sampler_views;
 274    unsigned i;
 275
 276    assert(buf);
 277
 278    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 279    assert(dec);
 280
 281    vl_vb_map(&buf->vertex_stream, dec->base.context);
 282
 283    sampler_views = buf->zscan_source->get_sampler_view_planes(buf->zscan_source);
 284
 285    assert(sampler_views);
 286
 287    for (i = 0; i < VL_MAX_PLANES; ++i) {
 288       struct pipe_resource *tex = sampler_views[i]->texture;
 289       struct pipe_box rect =
 290       {
 291          0, 0, 0,
 292          tex->width0,
 293          tex->height0,
 294          1
 295       };
 296
 297       buf->tex_transfer[i] = dec->base.context->get_transfer
 298       (
 299          dec->base.context, tex,
 300          0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
 301          &rect
 302       );
 303
 304       buf->texels[i] = dec->base.context->transfer_map(dec->base.context, buf->tex_transfer[i]);
 305    }
 306
 307    if (dec->base.entrypoint == PIPE_VIDEO_ENTRYPOINT_BITSTREAM) {
 308       struct pipe_ycbcr_block *ycbcr_stream[VL_MAX_PLANES];
 309       struct pipe_motionvector *mv_stream[VL_MAX_REF_FRAMES];
 310
 311       for (i = 0; i < VL_MAX_PLANES; ++i)
 312          ycbcr_stream[i] = vl_vb_get_ycbcr_stream(&buf->vertex_stream, i);
 313
 314       for (i = 0; i < VL_MAX_REF_FRAMES; ++i)
 315          mv_stream[i] = vl_vb_get_mv_stream(&buf->vertex_stream, i);
 316
 317       vl_mpg12_bs_set_buffers(&buf->bs, ycbcr_stream, buf->texels, mv_stream);
 318    } else {
 319
 320       for (i = 0; i < VL_MAX_PLANES; ++i)
 321          vl_zscan_set_layout(&buf->zscan[i], dec->zscan_linear);
 322    }
 323 }
 324
 325 static void
 326 vl_mpeg12_buffer_set_quant_matrix(struct pipe_video_decode_buffer *buffer,
 327                                   const uint8_t intra_matrix[64],
 328                                   const uint8_t non_intra_matrix[64])
 329 {
 330    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 331    unsigned i;
 332
 333    for (i = 0; i < VL_MAX_PLANES; ++i) {
 334       vl_zscan_upload_quant(&buf->zscan[i], intra_matrix, true);
 335       vl_zscan_upload_quant(&buf->zscan[i], non_intra_matrix, false);
 336    }
 337 }
 338
 339 static struct pipe_ycbcr_block *
 340 vl_mpeg12_buffer_get_ycbcr_stream(struct pipe_video_decode_buffer *buffer, int component)
 341 {
 342    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 343
 344    assert(buf);
 345
 346    return vl_vb_get_ycbcr_stream(&buf->vertex_stream, component);
 347 }
 348
 349 static short *
 350 vl_mpeg12_buffer_get_ycbcr_buffer(struct pipe_video_decode_buffer *buffer, int component)
 351 {
 352    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 353
 354    assert(buf);
 355    assert(component < VL_MAX_PLANES);
 356
 357    return buf->texels[component];
 358 }
 359
 360 static unsigned
 361 vl_mpeg12_buffer_get_mv_stream_stride(struct pipe_video_decode_buffer *buffer)
 362 {
 363    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 364
 365    assert(buf);
 366
 367    return vl_vb_get_mv_stream_stride(&buf->vertex_stream);
 368 }
 369
 370 static struct pipe_motionvector *
 371 vl_mpeg12_buffer_get_mv_stream(struct pipe_video_decode_buffer *buffer, int ref_frame)
 372 {
 373    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 374
 375    assert(buf);
 376
 377    return vl_vb_get_mv_stream(&buf->vertex_stream, ref_frame);
 378 }
 379
 380 static void
 381 vl_mpeg12_buffer_decode_bitstream(struct pipe_video_decode_buffer *buffer,
 382                                   unsigned num_bytes, const void *data,
 383                                   struct pipe_picture_desc *picture,
 384                                   unsigned num_ycbcr_blocks[3])
 385 {
 386    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 387    struct pipe_mpeg12_picture_desc *pic = (struct pipe_mpeg12_picture_desc *)picture;
 388
 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    for (i = 0; i < VL_MAX_PLANES; ++i)
 398       vl_zscan_set_layout(&buf->zscan[i], pic->alternate_scan ? dec->zscan_alternate : dec->zscan_normal);
 399
 400    vl_mpg12_bs_decode(&buf->bs, num_bytes, data, pic, num_ycbcr_blocks);
 401 }
 402
 403 static void
 404 vl_mpeg12_buffer_end_frame(struct pipe_video_decode_buffer *buffer)
 405 {
 406    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 407    struct vl_mpeg12_decoder *dec;
 408    unsigned i;
 409
 410    assert(buf);
 411
 412    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 413    assert(dec);
 414
 415    vl_vb_unmap(&buf->vertex_stream, dec->base.context);
 416
 417    for (i = 0; i < VL_MAX_PLANES; ++i) {
 418       dec->base.context->transfer_unmap(dec->base.context, buf->tex_transfer[i]);
 419       dec->base.context->transfer_destroy(dec->base.context, buf->tex_transfer[i]);
 420    }
 421 }
 422
 423 static void
 424 vl_mpeg12_destroy(struct pipe_video_decoder *decoder)
 425 {
 426    struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
 427
 428    assert(decoder);
 429
 430    /* Asserted in softpipe_delete_fs_state() for some reason */
 431    dec->base.context->bind_vs_state(dec->base.context, NULL);
 432    dec->base.context->bind_fs_state(dec->base.context, NULL);
 433
 434    dec->base.context->delete_depth_stencil_alpha_state(dec->base.context, dec->dsa);
 435    dec->base.context->delete_sampler_state(dec->base.context, dec->sampler_ycbcr);
 436
 437    vl_mc_cleanup(&dec->mc_y);
 438    vl_mc_cleanup(&dec->mc_c);
 439    dec->mc_source->destroy(dec->mc_source);
 440
 441    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 442       vl_idct_cleanup(&dec->idct_y);
 443       vl_idct_cleanup(&dec->idct_c);
 444       dec->idct_source->destroy(dec->idct_source);
 445    }
 446
 447    vl_zscan_cleanup(&dec->zscan_y);
 448    vl_zscan_cleanup(&dec->zscan_c);
 449
 450    dec->base.context->delete_vertex_elements_state(dec->base.context, dec->ves_ycbcr);
 451    dec->base.context->delete_vertex_elements_state(dec->base.context, dec->ves_mv);
 452
 453    pipe_resource_reference(&dec->quads.buffer, NULL);
 454    pipe_resource_reference(&dec->pos.buffer, NULL);
 455    pipe_resource_reference(&dec->block_num.buffer, NULL);
 456
 457    pipe_sampler_view_reference(&dec->zscan_linear, NULL);
 458    pipe_sampler_view_reference(&dec->zscan_normal, NULL);
 459    pipe_sampler_view_reference(&dec->zscan_alternate, NULL);
 460
 461    FREE(dec);
 462 }
 463
 464 static struct pipe_video_decode_buffer *
 465 vl_mpeg12_create_buffer(struct pipe_video_decoder *decoder)
 466 {
 467    struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
 468    struct vl_mpeg12_buffer *buffer;
 469
 470    assert(dec);
 471
 472    buffer = CALLOC_STRUCT(vl_mpeg12_buffer);
 473    if (buffer == NULL)
 474       return NULL;
 475
 476    buffer->base.decoder = decoder;
 477    buffer->base.destroy = vl_mpeg12_buffer_destroy;
 478    buffer->base.begin_frame = vl_mpeg12_buffer_begin_frame;
 479    buffer->base.set_quant_matrix = vl_mpeg12_buffer_set_quant_matrix;
 480    buffer->base.get_ycbcr_stream = vl_mpeg12_buffer_get_ycbcr_stream;
 481    buffer->base.get_ycbcr_buffer = vl_mpeg12_buffer_get_ycbcr_buffer;
 482    buffer->base.get_mv_stream_stride = vl_mpeg12_buffer_get_mv_stream_stride;
 483    buffer->base.get_mv_stream = vl_mpeg12_buffer_get_mv_stream;
 484    buffer->base.decode_bitstream = vl_mpeg12_buffer_decode_bitstream;
 485    buffer->base.end_frame = vl_mpeg12_buffer_end_frame;
 486
 487    if (!vl_vb_init(&buffer->vertex_stream, dec->base.context,
 488                    dec->base.width / MACROBLOCK_WIDTH,
 489                    dec->base.height / MACROBLOCK_HEIGHT))
 490       goto error_vertex_buffer;
 491
 492    if (!init_mc_buffer(buffer))
 493       goto error_mc;
 494
 495    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 496       if (!init_idct_buffer(buffer))
 497          goto error_idct;
 498
 499    if (!init_zscan_buffer(buffer))
 500       goto error_zscan;
 501
 502    if (dec->base.entrypoint == PIPE_VIDEO_ENTRYPOINT_BITSTREAM)
 503       vl_mpg12_bs_init(&buffer->bs,
 504                        dec->base.width / MACROBLOCK_WIDTH,
 505                        dec->base.height / MACROBLOCK_HEIGHT);
 506
 507    return &buffer->base;
 508
 509 error_zscan:
 510    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 511       cleanup_idct_buffer(buffer);
 512
 513 error_idct:
 514    cleanup_mc_buffer(buffer);
 515
 516 error_mc:
 517    vl_vb_cleanup(&buffer->vertex_stream);
 518
 519 error_vertex_buffer:
 520    FREE(buffer);
 521    return NULL;
 522 }
 523
 524 static void
 525 vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
 526                                unsigned num_ycbcr_blocks[3],
 527                                struct pipe_video_buffer *refs[2],
 528                                struct pipe_video_buffer *dst)
 529 {
 530    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer *)buffer;
 531    struct vl_mpeg12_decoder *dec;
 532
 533    struct pipe_sampler_view **sv[VL_MAX_REF_FRAMES], **mc_source_sv;
 534    struct pipe_surface **surfaces;
 535
 536    struct pipe_vertex_buffer vb[3];
 537
 538    unsigned i, j, component;
 539    unsigned nr_components;
 540
 541    assert(buf);
 542
 543    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 544    assert(dec);
 545
 546    for (i = 0; i < 2; ++i)
 547       sv[i] = refs[i] ? refs[i]->get_sampler_view_planes(refs[i]) : NULL;
 548
 549    vb[0] = dec->quads;
 550    vb[1] = dec->pos;
 551
 552    surfaces = dst->get_surfaces(dst);
 553
 554    dec->base.context->bind_vertex_elements_state(dec->base.context, dec->ves_mv);
 555    for (i = 0; i < VL_MAX_PLANES; ++i) {
 556       if (!surfaces[i]) continue;
 557
 558       vl_mc_set_surface(&buf->mc[i], surfaces[i]);
 559
 560       for (j = 0; j < VL_MAX_REF_FRAMES; ++j) {
 561          if (!sv[j]) continue;
 562
 563          vb[2] = vl_vb_get_mv(&buf->vertex_stream, j);;
 564          dec->base.context->set_vertex_buffers(dec->base.context, 3, vb);
 565
 566          vl_mc_render_ref(&buf->mc[i], sv[j][i]);
 567       }
 568    }
 569
 570    vb[2] = dec->block_num;
 571
 572    dec->base.context->bind_vertex_elements_state(dec->base.context, dec->ves_ycbcr);
 573    for (i = 0; i < VL_MAX_PLANES; ++i) {
 574       if (!num_ycbcr_blocks[i]) continue;
 575
 576       vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, i);
 577       dec->base.context->set_vertex_buffers(dec->base.context, 3, vb);
 578
 579       vl_zscan_render(&buf->zscan[i] , num_ycbcr_blocks[i]);
 580
 581       if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 582          vl_idct_flush(&buf->idct[i], num_ycbcr_blocks[i]);
 583    }
 584
 585    mc_source_sv = dec->mc_source->get_sampler_view_planes(dec->mc_source);
 586    for (i = 0, component = 0; i < VL_MAX_PLANES; ++i) {
 587       if (!surfaces[i]) continue;
 588
 589       nr_components = util_format_get_nr_components(surfaces[i]->texture->format);
 590       for (j = 0; j < nr_components; ++j, ++component) {
 591          if (!num_ycbcr_blocks[i]) continue;
 592
 593          vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, component);
 594          dec->base.context->set_vertex_buffers(dec->base.context, 3, vb);
 595
 596          if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 597             vl_idct_prepare_stage2(&buf->idct[component]);
 598          else {
 599             dec->base.context->set_fragment_sampler_views(dec->base.context, 1, &mc_source_sv[component]);
 600             dec->base.context->bind_fragment_sampler_states(dec->base.context, 1, &dec->sampler_ycbcr);
 601          }
 602          vl_mc_render_ycbcr(&buf->mc[i], j, num_ycbcr_blocks[component]);
 603       }
 604    }
 605 }
 606
 607 static bool
 608 init_pipe_state(struct vl_mpeg12_decoder *dec)
 609 {
 610    struct pipe_depth_stencil_alpha_state dsa;
 611    struct pipe_sampler_state sampler;
 612    unsigned i;
 613
 614    assert(dec);
 615
 616    memset(&dsa, 0, sizeof dsa);
 617    dsa.depth.enabled = 0;
 618    dsa.depth.writemask = 0;
 619    dsa.depth.func = PIPE_FUNC_ALWAYS;
 620    for (i = 0; i < 2; ++i) {
 621       dsa.stencil[i].enabled = 0;
 622       dsa.stencil[i].func = PIPE_FUNC_ALWAYS;
 623       dsa.stencil[i].fail_op = PIPE_STENCIL_OP_KEEP;
 624       dsa.stencil[i].zpass_op = PIPE_STENCIL_OP_KEEP;
 625       dsa.stencil[i].zfail_op = PIPE_STENCIL_OP_KEEP;
 626       dsa.stencil[i].valuemask = 0;
 627       dsa.stencil[i].writemask = 0;
 628    }
 629    dsa.alpha.enabled = 0;
 630    dsa.alpha.func = PIPE_FUNC_ALWAYS;
 631    dsa.alpha.ref_value = 0;
 632    dec->dsa = dec->base.context->create_depth_stencil_alpha_state(dec->base.context, &dsa);
 633    dec->base.context->bind_depth_stencil_alpha_state(dec->base.context, dec->dsa);
 634
 635    memset(&sampler, 0, sizeof(sampler));
 636    sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 637    sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 638    sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_BORDER;
 639    sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
 640    sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 641    sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
 642    sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
 643    sampler.compare_func = PIPE_FUNC_ALWAYS;
 644    sampler.normalized_coords = 1;
 645    dec->sampler_ycbcr = dec->base.context->create_sampler_state(dec->base.context, &sampler);
 646    if (!dec->sampler_ycbcr)
 647       return false;
 648
 649    return true;
 650 }
 651
 652 static const struct format_config*
 653 find_format_config(struct vl_mpeg12_decoder *dec, const struct format_config configs[], unsigned num_configs)
 654 {
 655    struct pipe_screen *screen;
 656    unsigned i;
 657
 658    assert(dec);
 659
 660    screen = dec->base.context->screen;
 661
 662    for (i = 0; i < num_configs; ++i) {
 663       if (!screen->is_format_supported(screen, configs[i].zscan_source_format, PIPE_TEXTURE_2D,
 664                                        1, PIPE_BIND_SAMPLER_VIEW))
 665          continue;
 666
 667       if (configs[i].idct_source_format != PIPE_FORMAT_NONE) {
 668          if (!screen->is_format_supported(screen, configs[i].idct_source_format, PIPE_TEXTURE_2D,
 669                                           1, PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 670             continue;
 671
 672          if (!screen->is_format_supported(screen, configs[i].mc_source_format, PIPE_TEXTURE_3D,
 673                                           1, PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 674             continue;
 675       } else {
 676          if (!screen->is_format_supported(screen, configs[i].mc_source_format, PIPE_TEXTURE_2D,
 677                                           1, PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 678             continue;
 679       }
 680       return &configs[i];
 681    }
 682
 683    return NULL;
 684 }
 685
 686 static bool
 687 init_zscan(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 688 {
 689    unsigned num_channels;
 690
 691    assert(dec);
 692
 693    dec->zscan_source_format = format_config->zscan_source_format;
 694    dec->zscan_linear = vl_zscan_layout(dec->base.context, vl_zscan_linear, dec->blocks_per_line);
 695    dec->zscan_normal = vl_zscan_layout(dec->base.context, vl_zscan_normal, dec->blocks_per_line);
 696    dec->zscan_alternate = vl_zscan_layout(dec->base.context, vl_zscan_alternate, dec->blocks_per_line);
 697
 698    num_channels = dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT ? 4 : 1;
 699
 700    if (!vl_zscan_init(&dec->zscan_y, dec->base.context, dec->base.width, dec->base.height,
 701                       dec->blocks_per_line, dec->num_blocks, num_channels))
 702       return false;
 703
 704    if (!vl_zscan_init(&dec->zscan_c, dec->base.context, dec->chroma_width, dec->chroma_height,
 705                       dec->blocks_per_line, dec->num_blocks, num_channels))
 706       return false;
 707
 708    return true;
 709 }
 710
 711 static bool
 712 init_idct(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 713 {
 714    unsigned nr_of_idct_render_targets, max_inst;
 715    enum pipe_format formats[3];
 716
 717    struct pipe_sampler_view *matrix = NULL;
 718
 719    nr_of_idct_render_targets = dec->base.context->screen->get_param
 720    (
 721       dec->base.context->screen, PIPE_CAP_MAX_RENDER_TARGETS
 722    );
 723
 724    max_inst = dec->base.context->screen->get_shader_param
 725    (
 726       dec->base.context->screen, PIPE_SHADER_FRAGMENT, PIPE_SHADER_CAP_MAX_INSTRUCTIONS
 727    );
 728
 729    // Just assume we need 32 inst per render target, not 100% true, but should work in most cases
 730    if (nr_of_idct_render_targets >= 4 && max_inst >= 32*4)
 731       // more than 4 render targets usually doesn't makes any seens
 732       nr_of_idct_render_targets = 4;
 733    else
 734       nr_of_idct_render_targets = 1;
 735
 736    formats[0] = formats[1] = formats[2] = format_config->idct_source_format;
 737    dec->idct_source = vl_video_buffer_create_ex
 738    (
 739       dec->base.context, dec->base.width / 4, dec->base.height, 1,
 740       dec->base.chroma_format, formats, PIPE_USAGE_STATIC
 741    );
 742
 743    if (!dec->idct_source)
 744       goto error_idct_source;
 745
 746    formats[0] = formats[1] = formats[2] = format_config->mc_source_format;
 747    dec->mc_source = vl_video_buffer_create_ex
 748    (
 749       dec->base.context, dec->base.width / nr_of_idct_render_targets,
 750       dec->base.height / 4, nr_of_idct_render_targets,
 751       dec->base.chroma_format, formats, PIPE_USAGE_STATIC
 752    );
 753
 754    if (!dec->mc_source)
 755       goto error_mc_source;
 756
 757    if (!(matrix = vl_idct_upload_matrix(dec->base.context, format_config->idct_scale)))
 758       goto error_matrix;
 759
 760    if (!vl_idct_init(&dec->idct_y, dec->base.context, dec->base.width, dec->base.height,
 761                      nr_of_idct_render_targets, matrix, matrix))
 762       goto error_y;
 763
 764    if(!vl_idct_init(&dec->idct_c, dec->base.context, dec->chroma_width, dec->chroma_height,
 765                     nr_of_idct_render_targets, matrix, matrix))
 766       goto error_c;
 767
 768    pipe_sampler_view_reference(&matrix, NULL);
 769
 770    return true;
 771
 772 error_c:
 773    vl_idct_cleanup(&dec->idct_y);
 774
 775 error_y:
 776    pipe_sampler_view_reference(&matrix, NULL);
 777
 778 error_matrix:
 779    dec->mc_source->destroy(dec->mc_source);
 780
 781 error_mc_source:
 782    dec->idct_source->destroy(dec->idct_source);
 783
 784 error_idct_source:
 785    return false;
 786 }
 787
 788 static bool
 789 init_mc_source_widthout_idct(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 790 {
 791    enum pipe_format formats[3];
 792
 793    formats[0] = formats[1] = formats[2] = format_config->mc_source_format;
 794    dec->mc_source = vl_video_buffer_create_ex
 795    (
 796       dec->base.context, dec->base.width, dec->base.height, 1,
 797       dec->base.chroma_format, formats, PIPE_USAGE_STATIC
 798    );
 799
 800    return dec->mc_source != NULL;
 801 }
 802
 803 static void
 804 mc_vert_shader_callback(void *priv, struct vl_mc *mc,
 805                         struct ureg_program *shader,
 806                         unsigned first_output,
 807                         struct ureg_dst tex)
 808 {
 809    struct vl_mpeg12_decoder *dec = priv;
 810    struct ureg_dst o_vtex;
 811
 812    assert(priv && mc);
 813    assert(shader);
 814
 815    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 816       struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c;
 817       vl_idct_stage2_vert_shader(idct, shader, first_output, tex);
 818    } else {
 819       o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, first_output);
 820       ureg_MOV(shader, ureg_writemask(o_vtex, TGSI_WRITEMASK_XY), ureg_src(tex));
 821    }
 822 }
 823
 824 static void
 825 mc_frag_shader_callback(void *priv, struct vl_mc *mc,
 826                         struct ureg_program *shader,
 827                         unsigned first_input,
 828                         struct ureg_dst dst)
 829 {
 830    struct vl_mpeg12_decoder *dec = priv;
 831    struct ureg_src src, sampler;
 832
 833    assert(priv && mc);
 834    assert(shader);
 835
 836    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 837       struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c;
 838       vl_idct_stage2_frag_shader(idct, shader, first_input, dst);
 839    } else {
 840       src = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, first_input, TGSI_INTERPOLATE_LINEAR);
 841       sampler = ureg_DECL_sampler(shader, 0);
 842       ureg_TEX(shader, dst, TGSI_TEXTURE_2D, src, sampler);
 843    }
 844 }
 845
 846 struct pipe_video_decoder *
 847 vl_create_mpeg12_decoder(struct pipe_context *context,
 848                          enum pipe_video_profile profile,
 849                          enum pipe_video_entrypoint entrypoint,
 850                          enum pipe_video_chroma_format chroma_format,
 851                          unsigned width, unsigned height)
 852 {
 853    const unsigned block_size_pixels = BLOCK_WIDTH * BLOCK_HEIGHT;
 854    const struct format_config *format_config;
 855    struct vl_mpeg12_decoder *dec;
 856
 857    assert(u_reduce_video_profile(profile) == PIPE_VIDEO_CODEC_MPEG12);
 858
 859    dec = CALLOC_STRUCT(vl_mpeg12_decoder);
 860
 861    if (!dec)
 862       return NULL;
 863
 864    dec->base.context = context;
 865    dec->base.profile = profile;
 866    dec->base.entrypoint = entrypoint;
 867    dec->base.chroma_format = chroma_format;
 868    dec->base.width = width;
 869    dec->base.height = height;
 870
 871    dec->base.destroy = vl_mpeg12_destroy;
 872    dec->base.create_buffer = vl_mpeg12_create_buffer;
 873    dec->base.flush_buffer = vl_mpeg12_decoder_flush_buffer;
 874
 875    dec->blocks_per_line = MAX2(util_next_power_of_two(dec->base.width) / block_size_pixels, 4);
 876    dec->num_blocks = (dec->base.width * dec->base.height) / block_size_pixels;
 877
 878    dec->quads = vl_vb_upload_quads(dec->base.context);
 879    dec->pos = vl_vb_upload_pos(
 880       dec->base.context,
 881       dec->base.width / MACROBLOCK_WIDTH,
 882       dec->base.height / MACROBLOCK_HEIGHT
 883    );
 884    dec->block_num = vl_vb_upload_block_num(dec->base.context, dec->num_blocks);
 885
 886    dec->ves_ycbcr = vl_vb_get_ves_ycbcr(dec->base.context);
 887    dec->ves_mv = vl_vb_get_ves_mv(dec->base.context);
 888
 889    /* TODO: Implement 422, 444 */
 890    assert(dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420);
 891
 892    if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420) {
 893       dec->chroma_width = dec->base.width / 2;
 894       dec->chroma_height = dec->base.height / 2;
 895    } else if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_422) {
 896       dec->chroma_width = dec->base.width;
 897       dec->chroma_height = dec->base.height / 2;
 898    } else {
 899       dec->chroma_width = dec->base.width;
 900       dec->chroma_height = dec->base.height;
 901    }
 902
 903    switch (entrypoint) {
 904    case PIPE_VIDEO_ENTRYPOINT_BITSTREAM:
 905       format_config = find_format_config(dec, bitstream_format_config, num_bitstream_format_configs);
 906       break;
 907
 908    case PIPE_VIDEO_ENTRYPOINT_IDCT:
 909       format_config = find_format_config(dec, idct_format_config, num_idct_format_configs);
 910       break;
 911
 912    case PIPE_VIDEO_ENTRYPOINT_MC:
 913       format_config = find_format_config(dec, mc_format_config, num_mc_format_configs);
 914       break;
 915
 916    default:
 917       assert(0);
 918       return NULL;
 919    }
 920
 921    if (!format_config)
 922       return NULL;
 923
 924    if (!init_zscan(dec, format_config))
 925       goto error_zscan;
 926
 927    if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 928       if (!init_idct(dec, format_config))
 929          goto error_sources;
 930    } else {
 931       if (!init_mc_source_widthout_idct(dec, format_config))
 932          goto error_sources;
 933    }
 934
 935    if (!vl_mc_init(&dec->mc_y, dec->base.context, dec->base.width, dec->base.height,
 936                    MACROBLOCK_HEIGHT, format_config->mc_scale,
 937                    mc_vert_shader_callback, mc_frag_shader_callback, dec))
 938       goto error_mc_y;
 939
 940    // TODO
 941    if (!vl_mc_init(&dec->mc_c, dec->base.context, dec->base.width, dec->base.height,
 942                    BLOCK_HEIGHT, format_config->mc_scale,
 943                    mc_vert_shader_callback, mc_frag_shader_callback, dec))
 944       goto error_mc_c;
 945
 946    if (!init_pipe_state(dec))
 947       goto error_pipe_state;
 948
 949    return &dec->base;
 950
 951 error_pipe_state:
 952    vl_mc_cleanup(&dec->mc_c);
 953
 954 error_mc_c:
 955    vl_mc_cleanup(&dec->mc_y);
 956
 957 error_mc_y:
 958    if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 959       vl_idct_cleanup(&dec->idct_y);
 960       vl_idct_cleanup(&dec->idct_c);
 961       dec->idct_source->destroy(dec->idct_source);
 962    }
 963    dec->mc_source->destroy(dec->mc_source);
 964
 965 error_sources:
 966    vl_zscan_cleanup(&dec->zscan_y);
 967    vl_zscan_cleanup(&dec->zscan_c);
 968
 969 error_zscan:
 970    FREE(dec);
 971    return NULL;
 972 }