vl_mpeg12_decoder.c revision 4e837f557bf5f5afb286e1f2244ed69c0092c2d6
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 41struct 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 50static 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 57static const unsigned num_bitstream_format_configs = 58 sizeof(bitstream_format_config) / sizeof(struct format_config); 59 60static 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 67static const unsigned num_idct_format_configs = 68 sizeof(idct_format_config) / sizeof(struct format_config); 69 70static 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 75static const unsigned num_mc_format_configs = 76 sizeof(mc_format_config) / sizeof(struct format_config); 77 78static bool 79init_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->pipe, 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 125error_plane: 126 for (; i > 0; --i) 127 vl_zscan_cleanup_buffer(&buffer->zscan[i - 1]); 128 129error_surface: 130error_sampler: 131 buffer->zscan_source->destroy(buffer->zscan_source); 132 133error_source: 134 return false; 135} 136 137static void 138cleanup_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 149static bool 150init_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 178error_plane: 179 for (; i > 0; --i) 180 vl_idct_cleanup_buffer(i == 1 ? &dec->idct_c : &dec->idct_y, &buffer->idct[i - 1]); 181 182error_mc_source_sv: 183error_source_sv: 184 return false; 185} 186 187static void 188cleanup_idct_buffer(struct vl_mpeg12_buffer *buf) 189{ 190 struct vl_mpeg12_decoder *dec; 191 assert(buf); 192 193 dec = (struct vl_mpeg12_decoder*)buf->base.decoder; 194 assert(dec); 195 196 vl_idct_cleanup_buffer(&dec->idct_y, &buf->idct[0]); 197 vl_idct_cleanup_buffer(&dec->idct_c, &buf->idct[1]); 198 vl_idct_cleanup_buffer(&dec->idct_c, &buf->idct[2]); 199} 200 201static bool 202init_mc_buffer(struct vl_mpeg12_buffer *buf) 203{ 204 struct vl_mpeg12_decoder *dec; 205 206 assert(buf); 207 208 dec = (struct vl_mpeg12_decoder*)buf->base.decoder; 209 assert(dec); 210 211 if(!vl_mc_init_buffer(&dec->mc_y, &buf->mc[0])) 212 goto error_mc_y; 213 214 if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[1])) 215 goto error_mc_cb; 216 217 if(!vl_mc_init_buffer(&dec->mc_c, &buf->mc[2])) 218 goto error_mc_cr; 219 220 return true; 221 222error_mc_cr: 223 vl_mc_cleanup_buffer(&buf->mc[1]); 224 225error_mc_cb: 226 vl_mc_cleanup_buffer(&buf->mc[0]); 227 228error_mc_y: 229 return false; 230} 231 232static void 233cleanup_mc_buffer(struct vl_mpeg12_buffer *buf) 234{ 235 unsigned i; 236 237 assert(buf); 238 239 for (i = 0; i < VL_MAX_PLANES; ++i) 240 vl_mc_cleanup_buffer(&buf->mc[i]); 241} 242 243static void 244vl_mpeg12_buffer_destroy(struct pipe_video_decode_buffer *buffer) 245{ 246 struct vl_mpeg12_buffer *buf = (struct vl_mpeg12_buffer*)buffer; 247 struct vl_mpeg12_decoder *dec; 248 249 assert(buf); 250 251 dec = (struct vl_mpeg12_decoder*)buf->base.decoder; 252 assert(dec); 253 254 cleanup_zscan_buffer(buf); 255 256 if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) 257 cleanup_idct_buffer(buf); 258 259 cleanup_mc_buffer(buf); 260 261 vl_vb_cleanup(&buf->vertex_stream); 262 263 FREE(buf); 264} 265 266static void 267vl_mpeg12_buffer_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->pipe); 281 282 sampler_views = buf->zscan_source->get_sampler_view_planes(buf->zscan_source); 283 284 assert(sampler_views); 285 286 for (i = 0; i < VL_MAX_PLANES; ++i) { 287 struct pipe_resource *tex = sampler_views[i]->texture; 288 struct pipe_box rect = 289 { 290 0, 0, 0, 291 tex->width0, 292 tex->height0, 293 1 294 }; 295 296 buf->tex_transfer[i] = dec->pipe->get_transfer 297 ( 298 dec->pipe, tex, 299 0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD, 300 &rect 301 ); 302 303 buf->texels[i] = dec->pipe->transfer_map(dec->pipe, buf->tex_transfer[i]); 304 } 305 306 if (dec->base.entrypoint == PIPE_VIDEO_ENTRYPOINT_BITSTREAM) { 307 struct pipe_ycbcr_block *ycbcr_stream[VL_MAX_PLANES]; 308 struct pipe_motionvector *mv_stream[VL_MAX_REF_FRAMES]; 309 310 for (i = 0; i < VL_MAX_PLANES; ++i) 311 ycbcr_stream[i] = vl_vb_get_ycbcr_stream(&buf->vertex_stream, i); 312 313 for (i = 0; i < VL_MAX_REF_FRAMES; ++i) 314 mv_stream[i] = vl_vb_get_mv_stream(&buf->vertex_stream, i); 315 316 vl_mpg12_bs_set_buffers(&buf->bs, ycbcr_stream, buf->texels, mv_stream); 317 } else { 318 319 for (i = 0; i < VL_MAX_PLANES; ++i) 320 vl_zscan_set_layout(&buf->zscan[i], dec->zscan_linear); 321 } 322} 323 324static void 325vl_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 336static struct pipe_ycbcr_block * 337vl_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 346static short * 347vl_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 357static unsigned 358vl_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 367static struct pipe_motionvector * 368vl_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 377static void 378vl_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 398static void 399vl_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->pipe); 411 412 for (i = 0; i < VL_MAX_PLANES; ++i) { 413 dec->pipe->transfer_unmap(dec->pipe, buf->tex_transfer[i]); 414 dec->pipe->transfer_destroy(dec->pipe, buf->tex_transfer[i]); 415 } 416} 417 418static void 419vl_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->pipe->bind_vs_state(dec->pipe, NULL); 427 dec->pipe->bind_fs_state(dec->pipe, NULL); 428 429 dec->pipe->delete_depth_stencil_alpha_state(dec->pipe, dec->dsa); 430 dec->pipe->delete_sampler_state(dec->pipe, 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->pipe->delete_vertex_elements_state(dec->pipe, dec->ves_ycbcr); 446 dec->pipe->delete_vertex_elements_state(dec->pipe, 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 458static struct pipe_video_decode_buffer * 459vl_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->pipe, 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 503error_zscan: 504 if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) 505 cleanup_idct_buffer(buffer); 506 507error_idct: 508 cleanup_mc_buffer(buffer); 509 510error_mc: 511 vl_vb_cleanup(&buffer->vertex_stream); 512 513error_vertex_buffer: 514 FREE(buffer); 515 return NULL; 516} 517 518static void 519vl_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->pipe->bind_vertex_elements_state(dec->pipe, 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->pipe->set_vertex_buffers(dec->pipe, 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->pipe->bind_vertex_elements_state(dec->pipe, 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->pipe->set_vertex_buffers(dec->pipe, 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->pipe->set_vertex_buffers(dec->pipe, 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->pipe->set_fragment_sampler_views(dec->pipe, 1, &mc_source_sv[component]); 594 dec->pipe->bind_fragment_sampler_states(dec->pipe, 1, &dec->sampler_ycbcr); 595 } 596 vl_mc_render_ycbcr(&buf->mc[i], j, num_ycbcr_blocks[component]); 597 } 598 } 599} 600 601static bool 602init_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->pipe->create_depth_stencil_alpha_state(dec->pipe, &dsa); 627 dec->pipe->bind_depth_stencil_alpha_state(dec->pipe, 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->pipe->create_sampler_state(dec->pipe, &sampler); 640 if (!dec->sampler_ycbcr) 641 return false; 642 643 return true; 644} 645 646static const struct format_config* 647find_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->pipe->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 680static bool 681init_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->pipe, vl_zscan_linear, dec->blocks_per_line); 689 dec->zscan_normal = vl_zscan_layout(dec->pipe, vl_zscan_normal, dec->blocks_per_line); 690 dec->zscan_alternate = vl_zscan_layout(dec->pipe, 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->pipe, 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->pipe, 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 705static bool 706init_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->pipe->screen->get_param(dec->pipe->screen, PIPE_CAP_MAX_RENDER_TARGETS); 714 max_inst = dec->pipe->screen->get_shader_param(dec->pipe->screen, PIPE_SHADER_FRAGMENT, PIPE_SHADER_CAP_MAX_INSTRUCTIONS); 715 716 // Just assume we need 32 inst per render target, not 100% true, but should work in most cases 717 if (nr_of_idct_render_targets >= 4 && max_inst >= 32*4) 718 // more than 4 render targets usually doesn't makes any seens 719 nr_of_idct_render_targets = 4; 720 else 721 nr_of_idct_render_targets = 1; 722 723 formats[0] = formats[1] = formats[2] = format_config->idct_source_format; 724 dec->idct_source = vl_video_buffer_create_ex 725 ( 726 dec->pipe, dec->base.width / 4, dec->base.height, 1, 727 dec->base.chroma_format, formats, PIPE_USAGE_STATIC 728 ); 729 730 if (!dec->idct_source) 731 goto error_idct_source; 732 733 formats[0] = formats[1] = formats[2] = format_config->mc_source_format; 734 dec->mc_source = vl_video_buffer_create_ex 735 ( 736 dec->pipe, dec->base.width / nr_of_idct_render_targets, 737 dec->base.height / 4, nr_of_idct_render_targets, 738 dec->base.chroma_format, formats, PIPE_USAGE_STATIC 739 ); 740 741 if (!dec->mc_source) 742 goto error_mc_source; 743 744 if (!(matrix = vl_idct_upload_matrix(dec->pipe, format_config->idct_scale))) 745 goto error_matrix; 746 747 if (!vl_idct_init(&dec->idct_y, dec->pipe, dec->base.width, dec->base.height, 748 nr_of_idct_render_targets, matrix, matrix)) 749 goto error_y; 750 751 if(!vl_idct_init(&dec->idct_c, dec->pipe, dec->chroma_width, dec->chroma_height, 752 nr_of_idct_render_targets, matrix, matrix)) 753 goto error_c; 754 755 pipe_sampler_view_reference(&matrix, NULL); 756 757 return true; 758 759error_c: 760 vl_idct_cleanup(&dec->idct_y); 761 762error_y: 763 pipe_sampler_view_reference(&matrix, NULL); 764 765error_matrix: 766 dec->mc_source->destroy(dec->mc_source); 767 768error_mc_source: 769 dec->idct_source->destroy(dec->idct_source); 770 771error_idct_source: 772 return false; 773} 774 775static bool 776init_mc_source_widthout_idct(struct vl_mpeg12_decoder *dec, const struct format_config* format_config) 777{ 778 enum pipe_format formats[3]; 779 780 formats[0] = formats[1] = formats[2] = format_config->mc_source_format; 781 dec->mc_source = vl_video_buffer_create_ex 782 ( 783 dec->pipe, dec->base.width, dec->base.height, 1, 784 dec->base.chroma_format, formats, PIPE_USAGE_STATIC 785 ); 786 787 return dec->mc_source != NULL; 788} 789 790static void 791mc_vert_shader_callback(void *priv, struct vl_mc *mc, 792 struct ureg_program *shader, 793 unsigned first_output, 794 struct ureg_dst tex) 795{ 796 struct vl_mpeg12_decoder *dec = priv; 797 struct ureg_dst o_vtex; 798 799 assert(priv && mc); 800 assert(shader); 801 802 if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) { 803 struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c; 804 vl_idct_stage2_vert_shader(idct, shader, first_output, tex); 805 } else { 806 o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, first_output); 807 ureg_MOV(shader, ureg_writemask(o_vtex, TGSI_WRITEMASK_XY), ureg_src(tex)); 808 } 809} 810 811static void 812mc_frag_shader_callback(void *priv, struct vl_mc *mc, 813 struct ureg_program *shader, 814 unsigned first_input, 815 struct ureg_dst dst) 816{ 817 struct vl_mpeg12_decoder *dec = priv; 818 struct ureg_src src, sampler; 819 820 assert(priv && mc); 821 assert(shader); 822 823 if (dec->base.entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) { 824 struct vl_idct *idct = mc == &dec->mc_y ? &dec->idct_y : &dec->idct_c; 825 vl_idct_stage2_frag_shader(idct, shader, first_input, dst); 826 } else { 827 src = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, first_input, TGSI_INTERPOLATE_LINEAR); 828 sampler = ureg_DECL_sampler(shader, 0); 829 ureg_TEX(shader, dst, TGSI_TEXTURE_2D, src, sampler); 830 } 831} 832 833struct pipe_video_decoder * 834vl_create_mpeg12_decoder(struct pipe_video_context *context, 835 struct pipe_context *pipe, 836 enum pipe_video_profile profile, 837 enum pipe_video_entrypoint entrypoint, 838 enum pipe_video_chroma_format chroma_format, 839 unsigned width, unsigned height) 840{ 841 const unsigned block_size_pixels = BLOCK_WIDTH * BLOCK_HEIGHT; 842 const struct format_config *format_config; 843 struct vl_mpeg12_decoder *dec; 844 845 assert(u_reduce_video_profile(profile) == PIPE_VIDEO_CODEC_MPEG12); 846 847 dec = CALLOC_STRUCT(vl_mpeg12_decoder); 848 849 if (!dec) 850 return NULL; 851 852 dec->base.context = context; 853 dec->base.profile = profile; 854 dec->base.entrypoint = entrypoint; 855 dec->base.chroma_format = chroma_format; 856 dec->base.width = width; 857 dec->base.height = height; 858 859 dec->base.destroy = vl_mpeg12_destroy; 860 dec->base.create_buffer = vl_mpeg12_create_buffer; 861 dec->base.flush_buffer = vl_mpeg12_decoder_flush_buffer; 862 863 dec->pipe = pipe; 864 865 dec->blocks_per_line = MAX2(util_next_power_of_two(dec->base.width) / block_size_pixels, 4); 866 dec->num_blocks = (dec->base.width * dec->base.height) / block_size_pixels; 867 868 dec->quads = vl_vb_upload_quads(dec->pipe); 869 dec->pos = vl_vb_upload_pos( 870 dec->pipe, 871 dec->base.width / MACROBLOCK_WIDTH, 872 dec->base.height / MACROBLOCK_HEIGHT 873 ); 874 dec->block_num = vl_vb_upload_block_num(dec->pipe, dec->num_blocks); 875 876 dec->ves_ycbcr = vl_vb_get_ves_ycbcr(dec->pipe); 877 dec->ves_mv = vl_vb_get_ves_mv(dec->pipe); 878 879 /* TODO: Implement 422, 444 */ 880 assert(dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420); 881 882 if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420) { 883 dec->chroma_width = dec->base.width / 2; 884 dec->chroma_height = dec->base.height / 2; 885 } else if (dec->base.chroma_format == PIPE_VIDEO_CHROMA_FORMAT_422) { 886 dec->chroma_width = dec->base.width; 887 dec->chroma_height = dec->base.height / 2; 888 } else { 889 dec->chroma_width = dec->base.width; 890 dec->chroma_height = dec->base.height; 891 } 892 893 switch (entrypoint) { 894 case PIPE_VIDEO_ENTRYPOINT_BITSTREAM: 895 format_config = find_format_config(dec, bitstream_format_config, num_bitstream_format_configs); 896 break; 897 898 case PIPE_VIDEO_ENTRYPOINT_IDCT: 899 format_config = find_format_config(dec, idct_format_config, num_idct_format_configs); 900 break; 901 902 case PIPE_VIDEO_ENTRYPOINT_MC: 903 format_config = find_format_config(dec, mc_format_config, num_mc_format_configs); 904 break; 905 906 default: 907 assert(0); 908 return NULL; 909 } 910 911 if (!format_config) 912 return NULL; 913 914 if (!init_zscan(dec, format_config)) 915 goto error_zscan; 916 917 if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) { 918 if (!init_idct(dec, format_config)) 919 goto error_sources; 920 } else { 921 if (!init_mc_source_widthout_idct(dec, format_config)) 922 goto error_sources; 923 } 924 925 if (!vl_mc_init(&dec->mc_y, dec->pipe, dec->base.width, dec->base.height, MACROBLOCK_HEIGHT, format_config->mc_scale, 926 mc_vert_shader_callback, mc_frag_shader_callback, dec)) 927 goto error_mc_y; 928 929 // TODO 930 if (!vl_mc_init(&dec->mc_c, dec->pipe, dec->base.width, dec->base.height, BLOCK_HEIGHT, format_config->mc_scale, 931 mc_vert_shader_callback, mc_frag_shader_callback, dec)) 932 goto error_mc_c; 933 934 if (!init_pipe_state(dec)) 935 goto error_pipe_state; 936 937 return &dec->base; 938 939error_pipe_state: 940 vl_mc_cleanup(&dec->mc_c); 941 942error_mc_c: 943 vl_mc_cleanup(&dec->mc_y); 944 945error_mc_y: 946 if (entrypoint <= PIPE_VIDEO_ENTRYPOINT_IDCT) { 947 vl_idct_cleanup(&dec->idct_y); 948 vl_idct_cleanup(&dec->idct_c); 949 dec->idct_source->destroy(dec->idct_source); 950 } 951 dec->mc_source->destroy(dec->mc_source); 952 953error_sources: 954 vl_zscan_cleanup(&dec->zscan_y); 955 vl_zscan_cleanup(&dec->zscan_c); 956 957error_zscan: 958 FREE(dec); 959 return NULL; 960} 961