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(i == 1 ? &dec->idct_c : &dec->idct_y, &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    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
 201 static bool
 202 init_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
 222 error_mc_cr:
 223    vl_mc_cleanup_buffer(&buf->mc[1]);
 224
 225 error_mc_cb:
 226    vl_mc_cleanup_buffer(&buf->mc[0]);
 227
 228 error_mc_y:
 229    return false;
 230 }
 231
 232 static void
 233 cleanup_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
 243 static void
 244 vl_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
 266 static void
 267 vl_mpeg12_buffer_begin_frame(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->base.context);
 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->base.context->get_transfer
 297       (
 298          dec->base.context, tex,
 299          0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
 300          &rect
 301       );
 302
 303       buf->texels[i] = dec->base.context->transfer_map(dec->base.context, 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
 319       for (i = 0; i < VL_MAX_PLANES; ++i)
 320          vl_zscan_set_layout(&buf->zscan[i], dec->zscan_linear);
 321    }
 322 }
 323
 324 static void
 325 vl_mpeg12_buffer_set_quant_matrix(struct pipe_video_decode_buffer *buffer,
 326                                   const uint8_t intra_matrix[64],
 327                                   const uint8_t non_intra_matrix[64])
 328 {
 329    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 330    unsigned i;
 331
 332    for (i = 0; i < VL_MAX_PLANES; ++i)
 333       vl_zscan_upload_quant(&buf->zscan[i], intra_matrix, non_intra_matrix);
 334 }
 335
 336 static struct pipe_ycbcr_block *
 337 vl_mpeg12_buffer_get_ycbcr_stream(struct pipe_video_decode_buffer *buffer, int component)
 338 {
 339    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 340
 341    assert(buf);
 342
 343    return vl_vb_get_ycbcr_stream(&buf->vertex_stream, component);
 344 }
 345
 346 static short *
 347 vl_mpeg12_buffer_get_ycbcr_buffer(struct pipe_video_decode_buffer *buffer, int component)
 348 {
 349    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 350
 351    assert(buf);
 352    assert(component < VL_MAX_PLANES);
 353
 354    return buf->texels[component];
 355 }
 356
 357 static unsigned
 358 vl_mpeg12_buffer_get_mv_stream_stride(struct pipe_video_decode_buffer *buffer)
 359 {
 360    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 361
 362    assert(buf);
 363
 364    return vl_vb_get_mv_stream_stride(&buf->vertex_stream);
 365 }
 366
 367 static struct pipe_motionvector *
 368 vl_mpeg12_buffer_get_mv_stream(struct pipe_video_decode_buffer *buffer, int ref_frame)
 369 {
 370    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 371
 372    assert(buf);
 373
 374    return vl_vb_get_mv_stream(&buf->vertex_stream, ref_frame);
 375 }
 376
 377 static void
 378 vl_mpeg12_buffer_decode_bitstream(struct pipe_video_decode_buffer *buffer,
 379                                   unsigned num_bytes, const void *data,
 380                                   struct pipe_mpeg12_picture_desc *picture,
 381                                   unsigned num_ycbcr_blocks[3])
 382 {
 383    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 384    struct vl_mpeg12_decoder *dec;
 385    unsigned i;
 386
 387    assert(buf);
 388
 389    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 390    assert(dec);
 391
 392    for (i = 0; i < VL_MAX_PLANES; ++i)
 393       vl_zscan_set_layout(&buf->zscan[i], picture->alternate_scan ? dec->zscan_alternate : dec->zscan_normal);
 394
 395    vl_mpg12_bs_decode(&buf->bs, num_bytes, data, picture, num_ycbcr_blocks);
 396 }
 397
 398 static void
 399 vl_mpeg12_buffer_end_frame(struct pipe_video_decode_buffer *buffer)
 400 {
 401    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer;
 402    struct vl_mpeg12_decoder *dec;
 403    unsigned i;
 404
 405    assert(buf);
 406
 407    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 408    assert(dec);
 409
 410    vl_vb_unmap(&buf->vertex_stream, dec->base.context);
 411
 412    for (i = 0; i < VL_MAX_PLANES; ++i) {
 413       dec->base.context->transfer_unmap(dec->base.context, buf->tex_transfer[i]);
 414       dec->base.context->transfer_destroy(dec->base.context, buf->tex_transfer[i]);
 415    }
 416 }
 417
 418 static void
 419 vl_mpeg12_destroy(struct pipe_video_decoder *decoder)
 420 {
 421    struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
 422
 423    assert(decoder);
 424
 425    /* Asserted in softpipe_delete_fs_state() for some reason */
 426    dec->base.context->bind_vs_state(dec->base.context, NULL);
 427    dec->base.context->bind_fs_state(dec->base.context, NULL);
 428
 429    dec->base.context->delete_depth_stencil_alpha_state(dec->base.context, dec->dsa);
 430    dec->base.context->delete_sampler_state(dec->base.context, dec->sampler_ycbcr);
 431
 432    vl_mc_cleanup(&dec->mc_y);
 433    vl_mc_cleanup(&dec->mc_c);
 434    dec->mc_source->destroy(dec->mc_source);
 435
 436    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 437       vl_idct_cleanup(&dec->idct_y);
 438       vl_idct_cleanup(&dec->idct_c);
 439       dec->idct_source->destroy(dec->idct_source);
 440    }
 441
 442    vl_zscan_cleanup(&dec->zscan_y);
 443    vl_zscan_cleanup(&dec->zscan_c);
 444
 445    dec->base.context->delete_vertex_elements_state(dec->base.context, dec->ves_ycbcr);
 446    dec->base.context->delete_vertex_elements_state(dec->base.context, dec->ves_mv);
 447
 448    pipe_resource_reference(&dec->quads.buffer, NULL);
 449    pipe_resource_reference(&dec->pos.buffer, NULL);
 450
 451    pipe_sampler_view_reference(&dec->zscan_linear, NULL);
 452    pipe_sampler_view_reference(&dec->zscan_normal, NULL);
 453    pipe_sampler_view_reference(&dec->zscan_alternate, NULL);
 454
 455    FREE(dec);
 456 }
 457
 458 static struct pipe_video_decode_buffer *
 459 vl_mpeg12_create_buffer(struct pipe_video_decoder *decoder)
 460 {
 461    struct vl_mpeg12_decoder *dec = (struct vl_mpeg12_decoder*)decoder;
 462    struct vl_mpeg12_buffer *buffer;
 463
 464    assert(dec);
 465
 466    buffer = CALLOC_STRUCT(vl_mpeg12_buffer);
 467    if (buffer == NULL)
 468       return NULL;
 469
 470    buffer->base.decoder = decoder;
 471    buffer->base.destroy = vl_mpeg12_buffer_destroy;
 472    buffer->base.begin_frame = vl_mpeg12_buffer_begin_frame;
 473    buffer->base.set_quant_matrix = vl_mpeg12_buffer_set_quant_matrix;
 474    buffer->base.get_ycbcr_stream = vl_mpeg12_buffer_get_ycbcr_stream;
 475    buffer->base.get_ycbcr_buffer = vl_mpeg12_buffer_get_ycbcr_buffer;
 476    buffer->base.get_mv_stream_stride = vl_mpeg12_buffer_get_mv_stream_stride;
 477    buffer->base.get_mv_stream = vl_mpeg12_buffer_get_mv_stream;
 478    buffer->base.decode_bitstream = vl_mpeg12_buffer_decode_bitstream;
 479    buffer->base.end_frame = vl_mpeg12_buffer_end_frame;
 480
 481    if (!vl_vb_init(&buffer->vertex_stream, dec->base.context,
 482                    dec->base.width / MACROBLOCK_WIDTH,
 483                    dec->base.height / MACROBLOCK_HEIGHT))
 484       goto error_vertex_buffer;
 485
 486    if (!init_mc_buffer(buffer))
 487       goto error_mc;
 488
 489    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 490       if (!init_idct_buffer(buffer))
 491          goto error_idct;
 492
 493    if (!init_zscan_buffer(buffer))
 494       goto error_zscan;
 495
 496    if (dec->base.entrypoint == PIPE_VIDEO_ENTRYPOINT_BITSTREAM)
 497       vl_mpg12_bs_init(&buffer->bs,
 498                        dec->base.width / MACROBLOCK_WIDTH,
 499                        dec->base.height / MACROBLOCK_HEIGHT);
 500
 501    return &buffer->base;
 502
 503 error_zscan:
 504    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 505       cleanup_idct_buffer(buffer);
 506
 507 error_idct:
 508    cleanup_mc_buffer(buffer);
 509
 510 error_mc:
 511    vl_vb_cleanup(&buffer->vertex_stream);
 512
 513 error_vertex_buffer:
 514    FREE(buffer);
 515    return NULL;
 516 }
 517
 518 static void
 519 vl_mpeg12_decoder_flush_buffer(struct pipe_video_decode_buffer *buffer,
 520                                unsigned num_ycbcr_blocks[3],
 521                                struct pipe_video_buffer *refs[2],
 522                                struct pipe_video_buffer *dst)
 523 {
 524    struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer *)buffer;
 525    struct vl_mpeg12_decoder *dec;
 526
 527    struct pipe_sampler_view **sv[VL_MAX_REF_FRAMES], **mc_source_sv;
 528    struct pipe_surface **surfaces;
 529
 530    struct pipe_vertex_buffer vb[3];
 531
 532    unsigned i, j, component;
 533    unsigned nr_components;
 534
 535    assert(buf);
 536
 537    dec = (struct vl_mpeg12_decoder *)buf->base.decoder;
 538    assert(dec);
 539
 540    for (i = 0; i < 2; ++i)
 541       sv[i] = refs[i] ? refs[i]->get_sampler_view_planes(refs[i]) : NULL;
 542
 543    vb[0] = dec->quads;
 544    vb[1] = dec->pos;
 545
 546    surfaces = dst->get_surfaces(dst);
 547
 548    dec->base.context->bind_vertex_elements_state(dec->base.context, dec->ves_mv);
 549    for (i = 0; i < VL_MAX_PLANES; ++i) {
 550       if (!surfaces[i]) continue;
 551
 552       vl_mc_set_surface(&buf->mc[i], surfaces[i]);
 553
 554       for (j = 0; j < VL_MAX_REF_FRAMES; ++j) {
 555          if (!sv[j]) continue;
 556
 557          vb[2] = vl_vb_get_mv(&buf->vertex_stream, j);;
 558          dec->base.context->set_vertex_buffers(dec->base.context, 3, vb);
 559
 560          vl_mc_render_ref(&buf->mc[i], sv[j][i]);
 561       }
 562    }
 563
 564    vb[2] = dec->block_num;
 565
 566    dec->base.context->bind_vertex_elements_state(dec->base.context, dec->ves_ycbcr);
 567    for (i = 0; i < VL_MAX_PLANES; ++i) {
 568       if (!num_ycbcr_blocks[i]) continue;
 569
 570       vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, i);
 571       dec->base.context->set_vertex_buffers(dec->base.context, 3, vb);
 572
 573       vl_zscan_render(&buf->zscan[i] , num_ycbcr_blocks[i]);
 574
 575       if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 576          vl_idct_flush(i == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[i], num_ycbcr_blocks[i]);
 577    }
 578
 579    mc_source_sv = dec->mc_source->get_sampler_view_planes(dec->mc_source);
 580    for (i = 0, component = 0; i < VL_MAX_PLANES; ++i) {
 581       if (!surfaces[i]) continue;
 582
 583       nr_components = util_format_get_nr_components(surfaces[i]->texture->format);
 584       for (j = 0; j < nr_components; ++j, ++component) {
 585          if (!num_ycbcr_blocks[i]) continue;
 586
 587          vb[1] = vl_vb_get_ycbcr(&buf->vertex_stream, component);
 588          dec->base.context->set_vertex_buffers(dec->base.context, 3, vb);
 589
 590          if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT)
 591             vl_idct_prepare_stage2(component == 0 ? &dec->idct_y : &dec->idct_c, &buf->idct[component]);
 592          else {
 593             dec->base.context->set_fragment_sampler_views(dec->base.context, 1, &mc_source_sv[component]);
 594             dec->base.context->bind_fragment_sampler_states(dec->base.context, 1, &dec->sampler_ycbcr);
 595          }
 596          vl_mc_render_ycbcr(&buf->mc[i], j, num_ycbcr_blocks[component]);
 597       }
 598    }
 599 }
 600
 601 static bool
 602 init_pipe_state(struct vl_mpeg12_decoder *dec)
 603 {
 604    struct pipe_depth_stencil_alpha_state dsa;
 605    struct pipe_sampler_state sampler;
 606    unsigned i;
 607
 608    assert(dec);
 609
 610    memset(&dsa, 0, sizeof dsa);
 611    dsa.depth.enabled = 0;
 612    dsa.depth.writemask = 0;
 613    dsa.depth.func = PIPE_FUNC_ALWAYS;
 614    for (i = 0; i < 2; ++i) {
 615       dsa.stencil[i].enabled = 0;
 616       dsa.stencil[i].func = PIPE_FUNC_ALWAYS;
 617       dsa.stencil[i].fail_op = PIPE_STENCIL_OP_KEEP;
 618       dsa.stencil[i].zpass_op = PIPE_STENCIL_OP_KEEP;
 619       dsa.stencil[i].zfail_op = PIPE_STENCIL_OP_KEEP;
 620       dsa.stencil[i].valuemask = 0;
 621       dsa.stencil[i].writemask = 0;
 622    }
 623    dsa.alpha.enabled = 0;
 624    dsa.alpha.func = PIPE_FUNC_ALWAYS;
 625    dsa.alpha.ref_value = 0;
 626    dec->dsa = dec->base.context->create_depth_stencil_alpha_state(dec->base.context, &dsa);
 627    dec->base.context->bind_depth_stencil_alpha_state(dec->base.context, dec->dsa);
 628
 629    memset(&sampler, 0, sizeof(sampler));
 630    sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 631    sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 632    sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_BORDER;
 633    sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
 634    sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 635    sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
 636    sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
 637    sampler.compare_func = PIPE_FUNC_ALWAYS;
 638    sampler.normalized_coords = 1;
 639    dec->sampler_ycbcr = dec->base.context->create_sampler_state(dec->base.context, &sampler);
 640    if (!dec->sampler_ycbcr)
 641       return false;
 642
 643    return true;
 644 }
 645
 646 static const struct format_config*
 647 find_format_config(struct vl_mpeg12_decoder *dec, const struct format_config configs[], unsigned num_configs)
 648 {
 649    struct pipe_screen *screen;
 650    unsigned i;
 651
 652    assert(dec);
 653
 654    screen = dec->base.context->screen;
 655
 656    for (i = 0; i < num_configs; ++i) {
 657       if (!screen->is_format_supported(screen, configs[i].zscan_source_format, PIPE_TEXTURE_2D,
 658                                        1, PIPE_BIND_SAMPLER_VIEW))
 659          continue;
 660
 661       if (configs[i].idct_source_format != PIPE_FORMAT_NONE) {
 662          if (!screen->is_format_supported(screen, configs[i].idct_source_format, PIPE_TEXTURE_2D,
 663                                           1, PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 664             continue;
 665
 666          if (!screen->is_format_supported(screen, configs[i].mc_source_format, PIPE_TEXTURE_3D,
 667                                           1, PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 668             continue;
 669       } else {
 670          if (!screen->is_format_supported(screen, configs[i].mc_source_format, PIPE_TEXTURE_2D,
 671                                           1, PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET))
 672             continue;
 673       }
 674       return &configs[i];
 675    }
 676
 677    return NULL;
 678 }
 679
 680 static bool
 681 init_zscan(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 682 {
 683    unsigned num_channels;
 684
 685    assert(dec);
 686
 687    dec->zscan_source_format = format_config->zscan_source_format;
 688    dec->zscan_linear = vl_zscan_layout(dec->base.context, vl_zscan_linear, dec->blocks_per_line);
 689    dec->zscan_normal = vl_zscan_layout(dec->base.context, vl_zscan_normal, dec->blocks_per_line);
 690    dec->zscan_alternate = vl_zscan_layout(dec->base.context, vl_zscan_alternate, dec->blocks_per_line);
 691
 692    num_channels = dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT ? 4 : 1;
 693
 694    if (!vl_zscan_init(&dec->zscan_y, dec->base.context, dec->base.width, dec->base.height,
 695                       dec->blocks_per_line, dec->num_blocks, num_channels))
 696       return false;
 697
 698    if (!vl_zscan_init(&dec->zscan_c, dec->base.context, dec->chroma_width, dec->chroma_height,
 699                       dec->blocks_per_line, dec->num_blocks, num_channels))
 700       return false;
 701
 702    return true;
 703 }
 704
 705 static bool
 706 init_idct(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 707 {
 708    unsigned nr_of_idct_render_targets, max_inst;
 709    enum pipe_format formats[3];
 710
 711    struct pipe_sampler_view *matrix = NULL;
 712
 713    nr_of_idct_render_targets = dec->base.context->screen->get_param
 714    (
 715       dec->base.context->screen, PIPE_CAP_MAX_RENDER_TARGETS
 716    );
 717
 718    max_inst = dec->base.context->screen->get_shader_param
 719    (
 720       dec->base.context->screen, PIPE_SHADER_FRAGMENT, PIPE_SHADER_CAP_MAX_INSTRUCTIONS
 721    );
 722
 723    // Just assume we need 32 inst per render target, not 100% true, but should work in most cases
 724    if (nr_of_idct_render_targets >= 4 && max_inst >= 32*4)
 725       // more than 4 render targets usually doesn't makes any seens
 726       nr_of_idct_render_targets = 4;
 727    else
 728       nr_of_idct_render_targets = 1;
 729
 730    formats[0] = formats[1] = formats[2] = format_config->idct_source_format;
 731    dec->idct_source = vl_video_buffer_create_ex
 732    (
 733       dec->base.context, dec->base.width / 4, dec->base.height, 1,
 734       dec->base.chroma_format, formats, PIPE_USAGE_STATIC
 735    );
 736
 737    if (!dec->idct_source)
 738       goto error_idct_source;
 739
 740    formats[0] = formats[1] = formats[2] = format_config->mc_source_format;
 741    dec->mc_source = vl_video_buffer_create_ex
 742    (
 743       dec->base.context, dec->base.width / nr_of_idct_render_targets,
 744       dec->base.height / 4, nr_of_idct_render_targets,
 745       dec->base.chroma_format, formats, PIPE_USAGE_STATIC
 746    );
 747
 748    if (!dec->mc_source)
 749       goto error_mc_source;
 750
 751    if (!(matrix = vl_idct_upload_matrix(dec->base.context, format_config->idct_scale)))
 752       goto error_matrix;
 753
 754    if (!vl_idct_init(&dec->idct_y, dec->base.context, dec->base.width, dec->base.height,
 755                      nr_of_idct_render_targets, matrix, matrix))
 756       goto error_y;
 757
 758    if(!vl_idct_init(&dec->idct_c, dec->base.context, dec->chroma_width, dec->chroma_height,
 759                     nr_of_idct_render_targets, matrix, matrix))
 760       goto error_c;
 761
 762    pipe_sampler_view_reference(&matrix, NULL);
 763
 764    return true;
 765
 766 error_c:
 767    vl_idct_cleanup(&dec->idct_y);
 768
 769 error_y:
 770    pipe_sampler_view_reference(&matrix, NULL);
 771
 772 error_matrix:
 773    dec->mc_source->destroy(dec->mc_source);
 774
 775 error_mc_source:
 776    dec->idct_source->destroy(dec->idct_source);
 777
 778 error_idct_source:
 779    return false;
 780 }
 781
 782 static bool
 783 init_mc_source_widthout_idct(struct vl_mpeg12_decoder *dec, const struct format_config* format_config)
 784 {
 785    enum pipe_format formats[3];
 786
 787    formats[0] = formats[1] = formats[2] = format_config->mc_source_format;
 788    dec->mc_source = vl_video_buffer_create_ex
 789    (
 790       dec->base.context, dec->base.width, dec->base.height, 1,
 791       dec->base.chroma_format, formats, PIPE_USAGE_STATIC
 792    );
 793
 794    return dec->mc_source != NULL;
 795 }
 796
 797 static void
 798 mc_vert_shader_callback(void *priv, struct vl_mc *mc,
 799                         struct ureg_program *shader,
 800                         unsigned first_output,
 801                         struct ureg_dst tex)
 802 {
 803    struct vl_mpeg12_decoder *dec = priv;
 804    struct ureg_dst o_vtex;
 805
 806    assert(priv && mc);
 807    assert(shader);
 808
 809    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 810       struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c;
 811       vl_idct_stage2_vert_shader(idct, shader, first_output, tex);
 812    } else {
 813       o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, first_output);
 814       ureg_MOV(shader, ureg_writemask(o_vtex, TGSI_WRITEMASK_XY), ureg_src(tex));
 815    }
 816 }
 817
 818 static void
 819 mc_frag_shader_callback(void *priv, struct vl_mc *mc,
 820                         struct ureg_program *shader,
 821                         unsigned first_input,
 822                         struct ureg_dst dst)
 823 {
 824    struct vl_mpeg12_decoder *dec = priv;
 825    struct ureg_src src, sampler;
 826
 827    assert(priv && mc);
 828    assert(shader);
 829
 830    if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 831       struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c;
 832       vl_idct_stage2_frag_shader(idct, shader, first_input, dst);
 833    } else {
 834       src = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, first_input, TGSI_INTERPOLATE_LINEAR);
 835       sampler = ureg_DECL_sampler(shader, 0);
 836       ureg_TEX(shader, dst, TGSI_TEXTURE_2D, src, sampler);
 837    }
 838 }
 839
 840 struct pipe_video_decoder *
 841 vl_create_mpeg12_decoder(struct pipe_context *context,
 842                          enum pipe_video_profile profile,
 843                          enum pipe_video_entrypoint entrypoint,
 844                          enum pipe_video_chroma_format chroma_format,
 845                          unsigned width, unsigned height)
 846 {
 847    const unsigned block_size_pixels = BLOCK_WIDTH * BLOCK_HEIGHT;
 848    const struct format_config *format_config;
 849    struct vl_mpeg12_decoder *dec;
 850
 851    assert(u_reduce_video_profile(profile) == PIPE_VIDEO_CODEC_MPEG12);
 852
 853    dec = CALLOC_STRUCT(vl_mpeg12_decoder);
 854
 855    if (!dec)
 856       return NULL;
 857
 858    dec->base.context = context;
 859    dec->base.profile = profile;
 860    dec->base.entrypoint = entrypoint;
 861    dec->base.chroma_format = chroma_format;
 862    dec->base.width = width;
 863    dec->base.height = height;
 864
 865    dec->base.destroy = vl_mpeg12_destroy;
 866    dec->base.create_buffer = vl_mpeg12_create_buffer;
 867    dec->base.flush_buffer = vl_mpeg12_decoder_flush_buffer;
 868
 869    dec->blocks_per_line = MAX2(util_next_power_of_two(dec->base.width) / block_size_pixels, 4);
 870    dec->num_blocks = (dec->base.width * dec->base.height) / block_size_pixels;
 871
 872    dec->quads = vl_vb_upload_quads(dec->base.context);
 873    dec->pos = vl_vb_upload_pos(
 874       dec->base.context,
 875       dec->base.width / MACROBLOCK_WIDTH,
 876       dec->base.height / MACROBLOCK_HEIGHT
 877    );
 878    dec->block_num = vl_vb_upload_block_num(dec->base.context, dec->num_blocks);
 879
 880    dec->ves_ycbcr = vl_vb_get_ves_ycbcr(dec->base.context);
 881    dec->ves_mv = vl_vb_get_ves_mv(dec->base.context);
 882
 883    /* TODO: Implement 422, 444 */
 884    assert(dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420);
 885
 886    if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420) {
 887       dec->chroma_width = dec->base.width / 2;
 888       dec->chroma_height = dec->base.height / 2;
 889    } else if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_422) {
 890       dec->chroma_width = dec->base.width;
 891       dec->chroma_height = dec->base.height / 2;
 892    } else {
 893       dec->chroma_width = dec->base.width;
 894       dec->chroma_height = dec->base.height;
 895    }
 896
 897    switch (entrypoint) {
 898    case PIPE_VIDEO_ENTRYPOINT_BITSTREAM:
 899       format_config = find_format_config(dec, bitstream_format_config, num_bitstream_format_configs);
 900       break;
 901
 902    case PIPE_VIDEO_ENTRYPOINT_IDCT:
 903       format_config = find_format_config(dec, idct_format_config, num_idct_format_configs);
 904       break;
 905
 906    case PIPE_VIDEO_ENTRYPOINT_MC:
 907       format_config = find_format_config(dec, mc_format_config, num_mc_format_configs);
 908       break;
 909
 910    default:
 911       assert(0);
 912       return NULL;
 913    }
 914
 915    if (!format_config)
 916       return NULL;
 917
 918    if (!init_zscan(dec, format_config))
 919       goto error_zscan;
 920
 921    if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 922       if (!init_idct(dec, format_config))
 923          goto error_sources;
 924    } else {
 925       if (!init_mc_source_widthout_idct(dec, format_config))
 926          goto error_sources;
 927    }
 928
 929    if (!vl_mc_init(&dec->mc_y, dec->base.context, dec->base.width, dec->base.height,
 930                    MACROBLOCK_HEIGHT, format_config->mc_scale,
 931                    mc_vert_shader_callback, mc_frag_shader_callback, dec))
 932       goto error_mc_y;
 933
 934    // TODO
 935    if (!vl_mc_init(&dec->mc_c, dec->base.context, dec->base.width, dec->base.height,
 936                    BLOCK_HEIGHT, format_config->mc_scale,
 937                    mc_vert_shader_callback, mc_frag_shader_callback, dec))
 938       goto error_mc_c;
 939
 940    if (!init_pipe_state(dec))
 941       goto error_pipe_state;
 942
 943    return &dec->base;
 944
 945 error_pipe_state:
 946    vl_mc_cleanup(&dec->mc_c);
 947
 948 error_mc_c:
 949    vl_mc_cleanup(&dec->mc_y);
 950
 951 error_mc_y:
 952    if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) {
 953       vl_idct_cleanup(&dec->idct_y);
 954       vl_idct_cleanup(&dec->idct_c);
 955       dec->idct_source->destroy(dec->idct_source);
 956    }
 957    dec->mc_source->destroy(dec->mc_source);
 958
 959 error_sources:
 960    vl_zscan_cleanup(&dec->zscan_y);
 961    vl_zscan_cleanup(&dec->zscan_c);
 962
 963 error_zscan:
 964    FREE(dec);
 965    return NULL;
 966 }