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 static const enum pipe_format const_zscan_source_formats[] = {
  42    PIPE_FORMAT_R16_SNORM,
  43    PIPE_FORMAT_R16_SSCALED
  44 };
  45
  46 static const unsigned num_zscan_source_formats =
  47    sizeof(const_zscan_source_formats) / sizeof(enum pipe_format);
  48
  49 static const enum pipe_format const_idct_source_formats[] = {
  50    PIPE_FORMAT_R16G16B16A16_SNORM,
  51    PIPE_FORMAT_R16G16B16A16_SSCALED
  52 };
  53
  54 static const unsigned num_idct_source_formats =
  55    sizeof(const_idct_source_formats) / sizeof(enum pipe_format);
  56
  57 static 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
  64 static const unsigned num_idct_intermediate_formats =
  65    sizeof(const_idct_intermediate_formats) / sizeof(enum pipe_format);
  66
  67 static const enum pipe_format const_mc_source_formats[] = {
  68    PIPE_FORMAT_R16_SNORM,
  69    PIPE_FORMAT_R16_SSCALED
  70 };
  71
  72 static const unsigned num_mc_source_formats =
  73    sizeof(const_mc_source_formats) / sizeof(enum pipe_format);
  74
  75 static bool
  76 init_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_views(buffer->zscan_source);
 101    if (!source)
 102       goto error_sampler;
 103
 104    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 105       destination = buffer->idct_source->get_surfaces(buffer->idct_source);
 106    else
 107       destination = buffer->mc_source->get_surfaces(buffer->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
 119 error_plane:
 120    for (; i > 0; --i)
 121       vl_zscan_cleanup_buffer(&buffer->zscan[i - 1]);
 122
 123 error_surface:
 124 error_sampler:
 125    buffer->zscan_source->destroy(buffer->zscan_source);
 126
 127 error_source:
 128    return false;
 129 }
 130
 131 static void
 132 cleanup_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
 143 static bool
 144 init_idct_buffer(struct vl_mpeg12_buffer *buffer)
 145 {
 146    enum pipe_format formats[3];
 147
 148    struct pipe_sampler_view **idct_source_sv, **idct_intermediate_sv;
 149    struct pipe_surface **idct_surfaces;
 150
 151    struct vl_mpeg12_decoder *dec;
 152
 153    unsigned i;
 154
 155    assert(buffer);
 156
 157    dec = (struct vl_mpeg12_decoder*)buffer->base.decoder;
 158
 159    formats[0] = formats[1] = formats[2] = dec->idct_source_format;
 160    buffer->idct_source = vl_video_buffer_init(dec->base.context, dec->pipe,
 161                                               dec->base.width / 4, dec->base.height, 1,
 162                                               dec->base.chroma_format,
 163                                               formats, PIPE_USAGE_STATIC);
 164    if (!buffer->idct_source)
 165       goto error_source;
 166
 167    formats[0] = formats[1] = formats[2] = dec->idct_intermediate_format;
 168    buffer->idct_intermediate = vl_video_buffer_init(dec->base.context, dec->pipe,
 169                                                     dec->base.width / dec->nr_of_idct_render_targets,
 170                                                     dec->base.height / 4, dec->nr_of_idct_render_targets,
 171                                                     dec->base.chroma_format,
 172                                                     formats, PIPE_USAGE_STATIC);
 173
 174    if (!buffer->idct_intermediate)
 175       goto error_intermediate;
 176
 177    idct_source_sv = buffer->idct_source->get_sampler_views(buffer->idct_source);
 178    if (!idct_source_sv)
 179       goto error_source_sv;
 180
 181    idct_intermediate_sv = buffer->idct_intermediate->get_sampler_views(buffer->idct_intermediate);
 182    if (!idct_intermediate_sv)
 183       goto error_intermediate_sv;
 184
 185    idct_surfaces = buffer->mc_source->get_surfaces(buffer->mc_source);
 186    if (!idct_surfaces)
 187       goto error_surfaces;
 188
 189    for (i = 0; i < 3; ++i)
 190       if (!vl_idct_init_buffer(i == 0 ? &dec->idct_y : &dec->idct_c,
 191                                &buffer->idct[i], idct_source_sv[i],
 192                                idct_intermediate_sv[i], idct_surfaces[i]))
 193          goto error_plane;
 194
 195    return true;
 196
 197 error_plane:
 198    for (; i > 0; --i)
 199       vl_idct_cleanup_buffer(i == 1 ? &dec->idct_c : &dec->idct_y, &buffer->idct[i - 1]);
 200
 201 error_surfaces:
 202 error_intermediate_sv:
 203 error_source_sv:
 204    buffer->idct_intermediate->destroy(buffer->idct_intermediate);
 205
 206 error_intermediate:
 207    buffer->idct_source->destroy(buffer->idct_source);
 208
 209 error_source:
 210    return false;
 211 }
 212
 213 static void
 214 cleanup_idct_buffer(struct vl_mpeg12_buffer *buf)
 215 {
 216    struct vl_mpeg12_decoder *dec;
 217    assert(buf);
 218
 219    dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
 220    assert(dec);
 221
 222    vl_idct_cleanup_buffer(&dec->idct_y, &buf->idct[0]);
 223    vl_idct_cleanup_buffer(&dec->idct_c, &buf->idct[1]);
 224    vl_idct_cleanup_buffer(&dec->idct_c, &buf->idct[2]);
 225    buf->idct_source->destroy(buf->idct_source);
 226    buf->idct_intermediate->destroy(buf->idct_intermediate);
 227 }
 228
 229 static bool
 230 init_mc_buffer(struct vl_mpeg12_buffer *buf)
 231 {
 232    struct vl_mpeg12_decoder *dec;
 233    enum pipe_format formats[3];
 234    struct pipe_sampler_view **mc_source_sv;
 235
 236    assert(buf);
 237
 238    dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
 239    assert(dec);
 240
 241    formats[0] = formats[1] = formats[2] =dec->mc_source_format;
 242    buf->mc_source = vl_video_buffer_init(dec->base.context, dec->pipe,
 243                                             dec->base.width, dec->base.height, 1,
 244                                             dec->base.chroma_format,
 245                                             formats, PIPE_USAGE_STATIC);
 246
 247    if (!buf->mc_source)
 248       goto error_mc_source;
 249
 250    mc_source_sv = buf->mc_source->get_sampler_views(buf->mc_source);
 251    if (!mc_source_sv)
 252       goto error_mc_source_sv;
 253
 254    if(!vl_mc_init_buffer(&dec->mc_y, &buf->mc[0], mc_source_sv[0]))
 255       goto error_mc_y;
 256
 257    if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[1], mc_source_sv[1]))
 258       goto error_mc_cb;
 259
 260    if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[2], mc_source_sv[2]))
 261       goto error_mc_cr;
 262
 263    return true;
 264
 265 error_mc_cr:
 266    vl_mc_cleanup_buffer(&buf->mc[1]);
 267
 268 error_mc_cb:
 269    vl_mc_cleanup_buffer(&buf->mc[0]);
 270
 271 error_mc_y:
 272 error_mc_source_sv:
 273    buf->mc_source->destroy(buf->mc_source);
 274
 275 error_mc_source:
 276    return false;
 277 }
 278
 279 static void
 280 cleanup_mc_buffer(struct vl_mpeg12_buffer *buf)
 281 {
 282    unsigned i;
 283
 284    assert(buf);
 285
 286    for (i = 0; i < VL_MAX_PLANES; ++i)
 287       vl_mc_cleanup_buffer(&buf->mc[i]);
 288
 289    buf->mc_source->destroy(buf->mc_source);
 290 }
 291
 292 static void
 293 vl_mpeg12_buffer_destroy(struct pipe_video_decode_buffer *buffer)
 294 {
 295    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 296    struct vl_mpeg12_decoder *dec;
 297
 298    assert(buf);
 299
 300    dec = (struct vl_mpeg12_decoder*)buf->base.decoder;
 301    assert(dec);
 302
 303    cleanup_zscan_buffer(buf);
 304
 305    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 306       cleanup_idct_buffer(buf);
 307
 308    cleanup_mc_buffer(buf);
 309
 310    vl_vb_cleanup(&buf->vertex_stream);
 311
 312    FREE(buf);
 313 }
 314
 315 static void
 316 vl_mpeg12_buffer_map(struct pipe_video_decode_buffer *buffer)
 317 {
 318    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 319    struct vl_mpeg12_decoder *dec;
 320
 321    struct pipe_sampler_view **sampler_views;
 322    unsigned i;
 323
 324    assert(buf);
 325
 326    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 327    assert(dec);
 328
 329    vl_vb_map(&buf->vertex_stream, dec->pipe);
 330
 331    sampler_views = buf->zscan_source->get_sampler_views(buf->zscan_source);
 332
 333    assert(sampler_views);
 334
 335    for (i = 0; i < VL_MAX_PLANES; ++i) {
 336       struct pipe_resource *tex = sampler_views[i]->texture;
 337       struct pipe_box rect =
 338       {
 339          0, 0, 0,
 340          tex->width0,
 341          tex->height0,
 342          1
 343       };
 344
 345       buf->tex_transfer[i] = dec->pipe->get_transfer
 346       (
 347          dec->pipe, tex,
 348          0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
 349          &rect
 350       );
 351
 352       buf->texels[i] = dec->pipe->transfer_map(dec->pipe, buf->tex_transfer[i]);
 353    }
 354 }
 355
 356 static struct pipe_ycbcr_block *
 357 vl_mpeg12_buffer_get_ycbcr_stream(struct pipe_video_decode_buffer *buffer, int component)
 358 {
 359    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 360
 361    assert(buf);
 362
 363    return vl_vb_get_ycbcr_stream(&buf->vertex_stream, component);
 364 }
 365
 366 static short *
 367 vl_mpeg12_buffer_get_ycbcr_buffer(struct pipe_video_decode_buffer *buffer, int component)
 368 {
 369    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 370
 371    assert(buf);
 372    assert(component < VL_MAX_PLANES);
 373
 374    return buf->texels[component];
 375 }
 376
 377 static unsigned
 378 vl_mpeg12_buffer_get_mv_stream_stride(struct pipe_video_decode_buffer *buffer)
 379 {
 380    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 381
 382    assert(buf);
 383
 384    return vl_vb_get_mv_stream_stride(&buf->vertex_stream);
 385 }
 386
 387 static struct pipe_motionvector *
 388 vl_mpeg12_buffer_get_mv_stream(struct pipe_video_decode_buffer *buffer, int ref_frame)
 389 {
 390    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 391
 392    assert(buf);
 393
 394    return vl_vb_get_mv_stream(&buf->vertex_stream, ref_frame);
 395 }
 396
 397 static void
 398 vl_mpeg12_buffer_unmap(struct pipe_video_decode_buffer *buffer)
 399 {
 400    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 401    struct vl_mpeg12_decoder *dec;
 402    unsigned i;
 403
 404    assert(buf);
 405
 406    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 407    assert(dec);
 408
 409    vl_vb_unmap(&buf->vertex_stream, dec->pipe);
 410
 411    for (i = 0; i < VL_MAX_PLANES; ++i) {
 412       dec->pipe->transfer_unmap(dec->pipe, buf->tex_transfer[i]);
 413       dec->pipe->transfer_destroy(dec->pipe, buf->tex_transfer[i]);
 414    }
 415 }
 416
 417 static void
 418 vl_mpeg12_destroy(struct pipe_video_decoder *decoder)
 419 {
 420    struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
 421
 422    assert(decoder);
 423
 424    /* Asserted in softpipe_delete_fs_state() for some reason */
 425    dec->pipe->bind_vs_state(dec->pipe, NULL);
 426    dec->pipe->bind_fs_state(dec->pipe, NULL);
 427
 428    dec->pipe->delete_depth_stencil_alpha_state(dec->pipe, dec->dsa);
 429
 430    vl_mc_cleanup(&dec->mc_y);
 431    vl_mc_cleanup(&dec->mc_c);
 432
 433    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 434       vl_idct_cleanup(&dec->idct_y);
 435       vl_idct_cleanup(&dec->idct_c);
 436    }
 437
 438    vl_zscan_cleanup(&dec->zscan_y);
 439    vl_zscan_cleanup(&dec->zscan_c);
 440
 441    dec->pipe->delete_vertex_elements_state(dec->pipe, dec->ves_ycbcr);
 442    dec->pipe->delete_vertex_elements_state(dec->pipe, dec->ves_mv);
 443
 444    pipe_resource_reference(&dec->quads.buffer, NULL);
 445    pipe_resource_reference(&dec->pos.buffer, NULL);
 446
 447    FREE(dec);
 448 }
 449
 450 static struct pipe_video_decode_buffer *
 451 vl_mpeg12_create_buffer(struct pipe_video_decoder *decoder)
 452 {
 453    struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
 454    struct vl_mpeg12_buffer *buffer;
 455
 456    assert(dec);
 457
 458    buffer = CALLOC_STRUCT(vl_mpeg12_buffer);
 459    if (buffer == NULL)
 460       return NULL;
 461
 462    buffer->base.decoder = decoder;
 463    buffer->base.destroy = vl_mpeg12_buffer_destroy;
 464    buffer->base.map = vl_mpeg12_buffer_map;
 465    buffer->base.get_ycbcr_stream = vl_mpeg12_buffer_get_ycbcr_stream;
 466    buffer->base.get_ycbcr_buffer = vl_mpeg12_buffer_get_ycbcr_buffer;
 467    buffer->base.get_mv_stream_stride = vl_mpeg12_buffer_get_mv_stream_stride;
 468    buffer->base.get_mv_stream = vl_mpeg12_buffer_get_mv_stream;
 469    buffer->base.unmap = vl_mpeg12_buffer_unmap;
 470
 471    if (!vl_vb_init(&buffer->vertex_stream, dec->pipe,
 472                    dec->base.width / MACROBLOCK_WIDTH,
 473                    dec->base.height / MACROBLOCK_HEIGHT))
 474       goto error_vertex_buffer;
 475
 476    if (!init_mc_buffer(buffer))
 477       goto error_mc;
 478
 479    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 480       if (!init_idct_buffer(buffer))
 481          goto error_idct;
 482
 483    if (!init_zscan_buffer(buffer))
 484       goto error_zscan;
 485
 486    return &buffer->base;
 487
 488 error_zscan:
 489    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 490       cleanup_idct_buffer(buffer);
 491
 492 error_idct:
 493    cleanup_mc_buffer(buffer);
 494
 495 error_mc:
 496    vl_vb_cleanup(&buffer->vertex_stream);
 497
 498 error_vertex_buffer:
 499    FREE(buffer);
 500    return NULL;
 501 }
 502
 503 static void
 504 vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
 505                                unsigned num_ycbcr_blocks[3],
 506                                struct pipe_video_buffer *refs[2],
 507                                struct pipe_video_buffer *dst,
 508                                struct pipe_fence_handle **fence)
 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[2];
 514    struct pipe_surface **surfaces;
 515
 516    struct pipe_vertex_buffer vb[3];
 517
 518    unsigned i, j;
 519
 520    assert(buf);
 521
 522    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 523    assert(dec);
 524
 525    for (i = 0; i < 2; ++i)
 526       sv[i] = refs[i] ? refs[i]->get_sampler_views(refs[i]) : NULL;
 527
 528    surfaces = dst->get_surfaces(dst);
 529
 530    vb[0] = dec->quads;
 531    vb[1] = dec->pos;
 532
 533    dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_mv);
 534    for (i = 0; i < VL_MAX_PLANES; ++i) {
 535       vl_mc_set_surface(&buf->mc[i], surfaces[i]);
 536
 537       for (j = 0; j < 2; ++j) {
 538          if (sv[j] == NULL) continue;
 539
 540          vb[2] = vl_vb_get_mv(&buf->vertex_stream, j);;
 541          dec->pipe->set_vertex_buffers(dec->pipe, 3, vb);
 542
 543          vl_mc_render_ref(&buf->mc[i], sv[j][i]);
 544       }
 545    }
 546
 547    dec->pipe->bind_vertex_elements_state(dec->pipe, dec->ves_ycbcr);
 548    for (i = 0; i < VL_MAX_PLANES; ++i) {
 549       if (num_ycbcr_blocks[i] == 0) continue;
 550
 551       vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, i);
 552       dec->pipe->set_vertex_buffers(dec->pipe, 2, vb);
 553
 554       vl_zscan_render(&buf->zscan[i] , num_ycbcr_blocks[i]);
 555
 556       if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 557          vl_idct_flush(i == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[i], num_ycbcr_blocks[i]);
 558
 559       vl_mc_render_ycbcr(&buf->mc[i], num_ycbcr_blocks[i]);
 560    }
 561
 562    dec->pipe->flush(dec->pipe, fence);
 563 }
 564
 565 static bool
 566 init_pipe_state(struct vl_mpeg12_decoder *dec)
 567 {
 568    struct pipe_depth_stencil_alpha_state dsa;
 569    unsigned i;
 570
 571    assert(dec);
 572
 573    memset(&dsa, 0, sizeof dsa);
 574    dsa.depth.enabled = 0;
 575    dsa.depth.writemask = 0;
 576    dsa.depth.func = PIPE_FUNC_ALWAYS;
 577    for (i = 0; i < 2; ++i) {
 578       dsa.stencil[i].enabled = 0;
 579       dsa.stencil[i].func = PIPE_FUNC_ALWAYS;
 580       dsa.stencil[i].fail_op = PIPE_STENCIL_OP_KEEP;
 581       dsa.stencil[i].zpass_op = PIPE_STENCIL_OP_KEEP;
 582       dsa.stencil[i].zfail_op = PIPE_STENCIL_OP_KEEP;
 583       dsa.stencil[i].valuemask = 0;
 584       dsa.stencil[i].writemask = 0;
 585    }
 586    dsa.alpha.enabled = 0;
 587    dsa.alpha.func = PIPE_FUNC_ALWAYS;
 588    dsa.alpha.ref_value = 0;
 589    dec->dsa = dec->pipe->create_depth_stencil_alpha_state(dec->pipe, &dsa);
 590    dec->pipe->bind_depth_stencil_alpha_state(dec->pipe, dec->dsa);
 591
 592    return true;
 593 }
 594
 595 static enum pipe_format
 596 find_first_supported_format(struct vl_mpeg12_decoder *dec,
 597                             const enum pipe_format formats[],
 598                             unsigned num_formats,
 599                             enum pipe_texture_target target)
 600 {
 601    struct pipe_screen *screen;
 602    unsigned i;
 603
 604    assert(dec);
 605
 606    screen = dec->pipe->screen;
 607
 608    for (i = 0; i < num_formats; ++i)
 609       if (screen->is_format_supported(dec->pipe->screen, formats[i], target, 1,
 610                                       PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 611          return formats[i];
 612
 613    return PIPE_FORMAT_NONE;
 614 }
 615
 616 static bool
 617 init_zscan(struct vl_mpeg12_decoder *dec)
 618 {
 619    struct pipe_sampler_view *layout;
 620
 621    unsigned num_channels;
 622
 623    assert(dec);
 624
 625    dec->blocks_per_line = 4;
 626    dec->max_blocks =
 627       (dec->base.width * dec->base.height) /
 628       (BLOCK_WIDTH * BLOCK_HEIGHT);
 629
 630    dec->zscan_source_format = find_first_supported_format(dec, const_zscan_source_formats,
 631                                                           num_zscan_source_formats, PIPE_TEXTURE_2D);
 632
 633    if (dec->zscan_source_format == PIPE_FORMAT_NONE)
 634       return false;
 635
 636    layout = vl_zscan_linear(dec->pipe, dec->blocks_per_line);
 637
 638    num_channels = dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT ? 4 : 1;
 639
 640    if (!vl_zscan_init(&dec->zscan_y, dec->pipe, dec->base.width, dec->base.height,
 641                       dec->blocks_per_line, dec->max_blocks, num_channels))
 642       return false;
 643
 644    vl_zscan_set_layout(&dec->zscan_y, layout);
 645
 646    if (!vl_zscan_init(&dec->zscan_c, dec->pipe, dec->chroma_width, dec->chroma_height,
 647                       dec->blocks_per_line, dec->max_blocks, num_channels))
 648       return false;
 649
 650    vl_zscan_set_layout(&dec->zscan_c, layout);
 651
 652    return true;
 653 }
 654
 655 static bool
 656 init_idct(struct vl_mpeg12_decoder *dec)
 657 {
 658    struct pipe_sampler_view *matrix, *transpose = NULL;
 659    float matrix_scale, transpose_scale;
 660
 661    dec->nr_of_idct_render_targets = dec->pipe->screen->get_param(dec->pipe->screen, PIPE_CAP_MAX_RENDER_TARGETS);
 662
 663    // more than 4 render targets usually doesn't makes any seens
 664    dec->nr_of_idct_render_targets = MIN2(dec->nr_of_idct_render_targets, 4);
 665
 666    dec->idct_source_format = find_first_supported_format(dec, const_idct_source_formats,
 667                                                          num_idct_source_formats, PIPE_TEXTURE_2D);
 668
 669    if (dec->idct_source_format == PIPE_FORMAT_NONE)
 670       return false;
 671
 672    dec->idct_intermediate_format = find_first_supported_format(dec, const_idct_intermediate_formats,
 673                                                                num_idct_intermediate_formats, PIPE_TEXTURE_3D);
 674
 675    if (dec->idct_intermediate_format == PIPE_FORMAT_NONE)
 676       return false;
 677
 678    switch (dec->idct_source_format) {
 679    case PIPE_FORMAT_R16G16B16A16_SSCALED:
 680       matrix_scale = SCALE_FACTOR_SSCALED;
 681       break;
 682
 683    case PIPE_FORMAT_R16G16B16A16_SNORM:
 684       matrix_scale = SCALE_FACTOR_SNORM;
 685       break;
 686
 687    default:
 688       assert(0);
 689       return false;
 690    }
 691
 692    if (dec->idct_intermediate_format == PIPE_FORMAT_R16G16B16A16_FLOAT ||
 693        dec->idct_intermediate_format == PIPE_FORMAT_R32G32B32A32_FLOAT)
 694       transpose_scale = 1.0f;
 695    else
 696       transpose_scale = matrix_scale = sqrt(matrix_scale);
 697
 698    if (dec->mc_source_format == PIPE_FORMAT_R16_SSCALED)
 699       transpose_scale /= SCALE_FACTOR_SSCALED;
 700
 701    if (!(matrix = vl_idct_upload_matrix(dec->pipe, matrix_scale)))
 702       goto error_matrix;
 703
 704    if (matrix_scale != transpose_scale) {
 705       if (!(transpose = vl_idct_upload_matrix(dec->pipe, transpose_scale)))
 706          goto error_transpose;
 707    } else
 708       pipe_sampler_view_reference(&transpose, matrix);
 709
 710    if (!vl_idct_init(&dec->idct_y, dec->pipe, dec->base.width, dec->base.height,
 711                      dec->nr_of_idct_render_targets, matrix, transpose))
 712       goto error_y;
 713
 714    if(!vl_idct_init(&dec->idct_c, dec->pipe, dec->chroma_width, dec->chroma_height,
 715                     dec->nr_of_idct_render_targets, matrix, transpose))
 716       goto error_c;
 717
 718    pipe_sampler_view_reference(&matrix, NULL);
 719    pipe_sampler_view_reference(&transpose, NULL);
 720    return true;
 721
 722 error_c:
 723    vl_idct_cleanup(&dec->idct_y);
 724
 725 error_y:
 726    pipe_sampler_view_reference(&transpose, NULL);
 727
 728 error_transpose:
 729    pipe_sampler_view_reference(&matrix, NULL);
 730
 731 error_matrix:
 732    return false;
 733 }
 734
 735 struct pipe_video_decoder *
 736 vl_create_mpeg12_decoder(struct pipe_video_context *context,
 737                          struct pipe_context *pipe,
 738                          enum pipe_video_profile profile,
 739                          enum pipe_video_entrypoint entrypoint,
 740                          enum pipe_video_chroma_format chroma_format,
 741                          unsigned width, unsigned height)
 742 {
 743    struct vl_mpeg12_decoder *dec;
 744    float mc_scale;
 745
 746    assert(u_reduce_video_profile(profile) == PIPE_VIDEO_CODEC_MPEG12);
 747
 748    dec = CALLOC_STRUCT(vl_mpeg12_decoder);
 749
 750    if (!dec)
 751       return NULL;
 752
 753    dec->base.context = context;
 754    dec->base.profile = profile;
 755    dec->base.entrypoint = entrypoint;
 756    dec->base.chroma_format = chroma_format;
 757    dec->base.width = width;
 758    dec->base.height = height;
 759
 760    dec->base.destroy = vl_mpeg12_destroy;
 761    dec->base.create_buffer = vl_mpeg12_create_buffer;
 762    dec->base.flush_buffer = vl_mpeg12_decoder_flush_buffer;
 763
 764    dec->base.width = align(width, MACROBLOCK_WIDTH);
 765    dec->base.height = align(height, MACROBLOCK_HEIGHT);
 766
 767    dec->pipe = pipe;
 768
 769    dec->quads = vl_vb_upload_quads(dec->pipe);
 770    dec->pos = vl_vb_upload_pos(
 771       dec->pipe,
 772       dec->base.width / MACROBLOCK_WIDTH,
 773       dec->base.height / MACROBLOCK_HEIGHT
 774    );
 775
 776    dec->ves_ycbcr = vl_vb_get_ves_ycbcr(dec->pipe);
 777    dec->ves_mv = vl_vb_get_ves_mv(dec->pipe);
 778
 779    /* TODO: Implement 422, 444 */
 780    assert(dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420);
 781
 782    dec->mc_source_format = find_first_supported_format(dec, const_mc_source_formats,
 783                                                        num_mc_source_formats, PIPE_TEXTURE_3D);
 784
 785    if (dec->mc_source_format == PIPE_FORMAT_NONE)
 786       return NULL;
 787
 788    if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420) {
 789       dec->chroma_width = dec->base.width / 2;
 790       dec->chroma_height = dec->base.height / 2;
 791    } else if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_422) {
 792       dec->chroma_width = dec->base.width;
 793       dec->chroma_height = dec->base.height / 2;
 794    } else {
 795       dec->chroma_width = dec->base.width;
 796       dec->chroma_height = dec->base.height;
 797    }
 798
 799    if (!init_zscan(dec))
 800       goto error_zscan;
 801
 802    if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 803       if (!init_idct(dec))
 804          goto error_idct;
 805       if (dec->mc_source_format == PIPE_FORMAT_R16_SSCALED)
 806          mc_scale = SCALE_FACTOR_SSCALED;
 807       else
 808          mc_scale = 1.0f;
 809    } else {
 810       switch (dec->mc_source_format) {
 811       case PIPE_FORMAT_R16_SNORM:
 812          mc_scale = SCALE_FACTOR_SNORM;
 813          break;
 814
 815       case PIPE_FORMAT_R16_SSCALED:
 816          mc_scale = SCALE_FACTOR_SSCALED;
 817          break;
 818
 819       default:
 820          assert(0);
 821          return NULL;
 822       }
 823    }
 824
 825    if (!vl_mc_init(&dec->mc_y, dec->pipe, dec->base.width, dec->base.height, MACROBLOCK_HEIGHT, mc_scale))
 826       goto error_mc_y;
 827
 828    // TODO
 829    if (!vl_mc_init(&dec->mc_c, dec->pipe, dec->base.width, dec->base.height, BLOCK_HEIGHT, mc_scale))
 830       goto error_mc_c;
 831
 832    if (!init_pipe_state(dec))
 833       goto error_pipe_state;
 834
 835    return &dec->base;
 836
 837 error_pipe_state:
 838    vl_mc_cleanup(&dec->mc_c);
 839
 840 error_mc_c:
 841    vl_mc_cleanup(&dec->mc_y);
 842
 843 error_mc_y:
 844    if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 845       vl_idct_cleanup(&dec->idct_y);
 846       vl_idct_cleanup(&dec->idct_c);
 847    }
 848
 849 error_idct:
 850    vl_zscan_cleanup(&dec->zscan_y);
 851    vl_zscan_cleanup(&dec->zscan_c);
 852
 853 error_zscan:
 854    FREE(dec);
 855    return NULL;
 856 }