1 /* 2 * Copyright 2014 Broadcom 3 * Copyright (C) 2012 Rob Clark <robclark (at) freedesktop.org> 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice (including the next 13 * paragraph) shall be included in all copies or substantial portions of the 14 * Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 22 * IN THE SOFTWARE. 23 */ 24 25 #include "util/u_blit.h" 26 #include "util/u_memory.h" 27 #include "util/u_format.h" 28 #include "util/u_inlines.h" 29 #include "util/u_surface.h" 30 #include "util/u_upload_mgr.h" 31 32 #include "vc4_screen.h" 33 #include "vc4_context.h" 34 #include "vc4_resource.h" 35 #include "vc4_tiling.h" 36 37 static bool miptree_debug = false; 38 39 static bool 40 vc4_resource_bo_alloc(struct vc4_resource *rsc) 41 { 42 struct pipe_resource *prsc = &rsc->base.b; 43 struct pipe_screen *pscreen = prsc->screen; 44 struct vc4_bo *bo; 45 46 if (miptree_debug) { 47 fprintf(stderr, "alloc %p: size %d + offset %d -> %d\n", 48 rsc, 49 rsc->slices[0].size, 50 rsc->slices[0].offset, 51 rsc->slices[0].offset + 52 rsc->slices[0].size + 53 rsc->cube_map_stride * (prsc->array_size - 1)); 54 } 55 56 bo = vc4_bo_alloc(vc4_screen(pscreen), 57 rsc->slices[0].offset + 58 rsc->slices[0].size + 59 rsc->cube_map_stride * (prsc->array_size - 1), 60 "resource"); 61 if (bo) { 62 vc4_bo_unreference(&rsc->bo); 63 rsc->bo = bo; 64 return true; 65 } else { 66 return false; 67 } 68 } 69 70 static void 71 vc4_resource_transfer_unmap(struct pipe_context *pctx, 72 struct pipe_transfer *ptrans) 73 { 74 struct vc4_context *vc4 = vc4_context(pctx); 75 struct vc4_transfer *trans = vc4_transfer(ptrans); 76 77 if (trans->map) { 78 struct vc4_resource *rsc; 79 struct vc4_resource_slice *slice; 80 if (trans->ss_resource) { 81 rsc = vc4_resource(trans->ss_resource); 82 slice = &rsc->slices[0]; 83 } else { 84 rsc = vc4_resource(ptrans->resource); 85 slice = &rsc->slices[ptrans->level]; 86 } 87 88 if (ptrans->usage & PIPE_TRANSFER_WRITE) { 89 vc4_store_tiled_image(rsc->bo->map + slice->offset + 90 ptrans->box.z * rsc->cube_map_stride, 91 slice->stride, 92 trans->map, ptrans->stride, 93 slice->tiling, rsc->cpp, 94 &ptrans->box); 95 } 96 free(trans->map); 97 } 98 99 if (trans->ss_resource && (ptrans->usage & PIPE_TRANSFER_WRITE)) { 100 struct pipe_blit_info blit; 101 memset(&blit, 0, sizeof(blit)); 102 103 blit.src.resource = trans->ss_resource; 104 blit.src.format = trans->ss_resource->format; 105 blit.src.box.width = trans->ss_box.width; 106 blit.src.box.height = trans->ss_box.height; 107 blit.src.box.depth = 1; 108 109 blit.dst.resource = ptrans->resource; 110 blit.dst.format = ptrans->resource->format; 111 blit.dst.level = ptrans->level; 112 blit.dst.box = trans->ss_box; 113 114 blit.mask = util_format_get_mask(ptrans->resource->format); 115 blit.filter = PIPE_TEX_FILTER_NEAREST; 116 117 pctx->blit(pctx, &blit); 118 119 pipe_resource_reference(&trans->ss_resource, NULL); 120 } 121 122 pipe_resource_reference(&ptrans->resource, NULL); 123 slab_free(&vc4->transfer_pool, ptrans); 124 } 125 126 static struct pipe_resource * 127 vc4_get_temp_resource(struct pipe_context *pctx, 128 struct pipe_resource *prsc, 129 const struct pipe_box *box) 130 { 131 struct pipe_resource temp_setup; 132 133 memset(&temp_setup, 0, sizeof(temp_setup)); 134 temp_setup.target = prsc->target; 135 temp_setup.format = prsc->format; 136 temp_setup.width0 = box->width; 137 temp_setup.height0 = box->height; 138 temp_setup.depth0 = 1; 139 temp_setup.array_size = 1; 140 141 return pctx->screen->resource_create(pctx->screen, &temp_setup); 142 } 143 144 static void * 145 vc4_resource_transfer_map(struct pipe_context *pctx, 146 struct pipe_resource *prsc, 147 unsigned level, unsigned usage, 148 const struct pipe_box *box, 149 struct pipe_transfer **pptrans) 150 { 151 struct vc4_context *vc4 = vc4_context(pctx); 152 struct vc4_resource *rsc = vc4_resource(prsc); 153 struct vc4_transfer *trans; 154 struct pipe_transfer *ptrans; 155 enum pipe_format format = prsc->format; 156 char *buf; 157 158 /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is 159 * being mapped. 160 */ 161 if ((usage & PIPE_TRANSFER_DISCARD_RANGE) && 162 !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) && 163 !(prsc->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT) && 164 prsc->last_level == 0 && 165 prsc->width0 == box->width && 166 prsc->height0 == box->height && 167 prsc->depth0 == box->depth && 168 prsc->array_size == 1) { 169 usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE; 170 } 171 172 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) { 173 if (vc4_resource_bo_alloc(rsc)) { 174 /* If it might be bound as one of our vertex buffers, 175 * make sure we re-emit vertex buffer state. 176 */ 177 if (prsc->bind & PIPE_BIND_VERTEX_BUFFER) 178 vc4->dirty |= VC4_DIRTY_VTXBUF; 179 } else { 180 /* If we failed to reallocate, flush users so that we 181 * don't violate any syncing requirements. 182 */ 183 vc4_flush_jobs_reading_resource(vc4, prsc); 184 } 185 } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) { 186 /* If we're writing and the buffer is being used by the CL, we 187 * have to flush the CL first. If we're only reading, we need 188 * to flush if the CL has written our buffer. 189 */ 190 if (usage & PIPE_TRANSFER_WRITE) 191 vc4_flush_jobs_reading_resource(vc4, prsc); 192 else 193 vc4_flush_jobs_writing_resource(vc4, prsc); 194 } 195 196 if (usage & PIPE_TRANSFER_WRITE) { 197 rsc->writes++; 198 rsc->initialized_buffers = ~0; 199 } 200 201 trans = slab_alloc(&vc4->transfer_pool); 202 if (!trans) 203 return NULL; 204 205 /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */ 206 207 /* slab_alloc_st() doesn't zero: */ 208 memset(trans, 0, sizeof(*trans)); 209 ptrans = &trans->base; 210 211 pipe_resource_reference(&ptrans->resource, prsc); 212 ptrans->level = level; 213 ptrans->usage = usage; 214 ptrans->box = *box; 215 216 /* If the resource is multisampled, we need to resolve to single 217 * sample. This seems like it should be handled at a higher layer. 218 */ 219 if (prsc->nr_samples > 1) { 220 trans->ss_resource = vc4_get_temp_resource(pctx, prsc, box); 221 if (!trans->ss_resource) 222 goto fail; 223 assert(!trans->ss_resource->nr_samples); 224 225 /* The ptrans->box gets modified for tile alignment, so save 226 * the original box for unmap time. 227 */ 228 trans->ss_box = *box; 229 230 if (usage & PIPE_TRANSFER_READ) { 231 struct pipe_blit_info blit; 232 memset(&blit, 0, sizeof(blit)); 233 234 blit.src.resource = ptrans->resource; 235 blit.src.format = ptrans->resource->format; 236 blit.src.level = ptrans->level; 237 blit.src.box = trans->ss_box; 238 239 blit.dst.resource = trans->ss_resource; 240 blit.dst.format = trans->ss_resource->format; 241 blit.dst.box.width = trans->ss_box.width; 242 blit.dst.box.height = trans->ss_box.height; 243 blit.dst.box.depth = 1; 244 245 blit.mask = util_format_get_mask(prsc->format); 246 blit.filter = PIPE_TEX_FILTER_NEAREST; 247 248 pctx->blit(pctx, &blit); 249 vc4_flush_jobs_writing_resource(vc4, blit.dst.resource); 250 } 251 252 /* The rest of the mapping process should use our temporary. */ 253 prsc = trans->ss_resource; 254 rsc = vc4_resource(prsc); 255 ptrans->box.x = 0; 256 ptrans->box.y = 0; 257 ptrans->box.z = 0; 258 } 259 260 /* Note that the current kernel implementation is synchronous, so no 261 * need to do syncing stuff here yet. 262 */ 263 264 if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) 265 buf = vc4_bo_map_unsynchronized(rsc->bo); 266 else 267 buf = vc4_bo_map(rsc->bo); 268 if (!buf) { 269 fprintf(stderr, "Failed to map bo\n"); 270 goto fail; 271 } 272 273 *pptrans = ptrans; 274 275 struct vc4_resource_slice *slice = &rsc->slices[level]; 276 if (rsc->tiled) { 277 uint32_t utile_w = vc4_utile_width(rsc->cpp); 278 uint32_t utile_h = vc4_utile_height(rsc->cpp); 279 280 /* No direct mappings of tiled, since we need to manually 281 * tile/untile. 282 */ 283 if (usage & PIPE_TRANSFER_MAP_DIRECTLY) 284 return NULL; 285 286 if (format == PIPE_FORMAT_ETC1_RGB8) { 287 /* ETC1 is arranged as 64-bit blocks, where each block 288 * is 4x4 pixels. Texture tiling operates on the 289 * 64-bit block the way it would an uncompressed 290 * pixels. 291 */ 292 assert(!(ptrans->box.x & 3)); 293 assert(!(ptrans->box.y & 3)); 294 ptrans->box.x >>= 2; 295 ptrans->box.y >>= 2; 296 ptrans->box.width = (ptrans->box.width + 3) >> 2; 297 ptrans->box.height = (ptrans->box.height + 3) >> 2; 298 } 299 300 /* We need to align the box to utile boundaries, since that's 301 * what load/store operates on. This may cause us to need to 302 * read out the original contents in that border area. Right 303 * now we just read out the entire contents, including the 304 * middle area that will just get overwritten. 305 */ 306 uint32_t box_start_x = ptrans->box.x & (utile_w - 1); 307 uint32_t box_start_y = ptrans->box.y & (utile_h - 1); 308 bool needs_load = (usage & PIPE_TRANSFER_READ) != 0; 309 310 if (box_start_x) { 311 ptrans->box.width += box_start_x; 312 ptrans->box.x -= box_start_x; 313 needs_load = true; 314 } 315 if (box_start_y) { 316 ptrans->box.height += box_start_y; 317 ptrans->box.y -= box_start_y; 318 needs_load = true; 319 } 320 if (ptrans->box.width & (utile_w - 1)) { 321 /* We only need to force a load if our border region 322 * we're extending into is actually part of the 323 * texture. 324 */ 325 uint32_t slice_width = u_minify(prsc->width0, level); 326 if (ptrans->box.x + ptrans->box.width != slice_width) 327 needs_load = true; 328 ptrans->box.width = align(ptrans->box.width, utile_w); 329 } 330 if (ptrans->box.height & (utile_h - 1)) { 331 uint32_t slice_height = u_minify(prsc->height0, level); 332 if (ptrans->box.y + ptrans->box.height != slice_height) 333 needs_load = true; 334 ptrans->box.height = align(ptrans->box.height, utile_h); 335 } 336 337 ptrans->stride = ptrans->box.width * rsc->cpp; 338 ptrans->layer_stride = ptrans->stride * ptrans->box.height; 339 340 trans->map = malloc(ptrans->layer_stride * ptrans->box.depth); 341 342 if (needs_load) { 343 vc4_load_tiled_image(trans->map, ptrans->stride, 344 buf + slice->offset + 345 ptrans->box.z * rsc->cube_map_stride, 346 slice->stride, 347 slice->tiling, rsc->cpp, 348 &ptrans->box); 349 } 350 return (trans->map + 351 box_start_x * rsc->cpp + 352 box_start_y * ptrans->stride); 353 } else { 354 ptrans->stride = slice->stride; 355 ptrans->layer_stride = ptrans->stride; 356 357 return buf + slice->offset + 358 ptrans->box.y / util_format_get_blockheight(format) * ptrans->stride + 359 ptrans->box.x / util_format_get_blockwidth(format) * rsc->cpp + 360 ptrans->box.z * rsc->cube_map_stride; 361 } 362 363 364 fail: 365 vc4_resource_transfer_unmap(pctx, ptrans); 366 return NULL; 367 } 368 369 static void 370 vc4_resource_destroy(struct pipe_screen *pscreen, 371 struct pipe_resource *prsc) 372 { 373 struct vc4_resource *rsc = vc4_resource(prsc); 374 pipe_resource_reference(&rsc->shadow_parent, NULL); 375 vc4_bo_unreference(&rsc->bo); 376 free(rsc); 377 } 378 379 static boolean 380 vc4_resource_get_handle(struct pipe_screen *pscreen, 381 struct pipe_resource *prsc, 382 struct winsys_handle *handle) 383 { 384 struct vc4_resource *rsc = vc4_resource(prsc); 385 386 return vc4_screen_bo_get_handle(pscreen, rsc->bo, rsc->slices[0].stride, 387 handle); 388 } 389 390 static const struct u_resource_vtbl vc4_resource_vtbl = { 391 .resource_get_handle = vc4_resource_get_handle, 392 .resource_destroy = vc4_resource_destroy, 393 .transfer_map = vc4_resource_transfer_map, 394 .transfer_flush_region = u_default_transfer_flush_region, 395 .transfer_unmap = vc4_resource_transfer_unmap, 396 }; 397 398 static void 399 vc4_setup_slices(struct vc4_resource *rsc) 400 { 401 struct pipe_resource *prsc = &rsc->base.b; 402 uint32_t width = prsc->width0; 403 uint32_t height = prsc->height0; 404 if (prsc->format == PIPE_FORMAT_ETC1_RGB8) { 405 width = (width + 3) >> 2; 406 height = (height + 3) >> 2; 407 } 408 409 uint32_t pot_width = util_next_power_of_two(width); 410 uint32_t pot_height = util_next_power_of_two(height); 411 uint32_t offset = 0; 412 uint32_t utile_w = vc4_utile_width(rsc->cpp); 413 uint32_t utile_h = vc4_utile_height(rsc->cpp); 414 415 for (int i = prsc->last_level; i >= 0; i--) { 416 struct vc4_resource_slice *slice = &rsc->slices[i]; 417 418 uint32_t level_width, level_height; 419 if (i == 0) { 420 level_width = width; 421 level_height = height; 422 } else { 423 level_width = u_minify(pot_width, i); 424 level_height = u_minify(pot_height, i); 425 } 426 427 if (!rsc->tiled) { 428 slice->tiling = VC4_TILING_FORMAT_LINEAR; 429 if (prsc->nr_samples > 1) { 430 /* MSAA (4x) surfaces are stored as raw tile buffer contents. */ 431 level_width = align(level_width, 32); 432 level_height = align(level_height, 32); 433 } else { 434 level_width = align(level_width, utile_w); 435 } 436 } else { 437 if (vc4_size_is_lt(level_width, level_height, 438 rsc->cpp)) { 439 slice->tiling = VC4_TILING_FORMAT_LT; 440 level_width = align(level_width, utile_w); 441 level_height = align(level_height, utile_h); 442 } else { 443 slice->tiling = VC4_TILING_FORMAT_T; 444 level_width = align(level_width, 445 4 * 2 * utile_w); 446 level_height = align(level_height, 447 4 * 2 * utile_h); 448 } 449 } 450 451 slice->offset = offset; 452 slice->stride = (level_width * rsc->cpp * 453 MAX2(prsc->nr_samples, 1)); 454 slice->size = level_height * slice->stride; 455 456 offset += slice->size; 457 458 if (miptree_debug) { 459 static const char tiling_chars[] = { 460 [VC4_TILING_FORMAT_LINEAR] = 'R', 461 [VC4_TILING_FORMAT_LT] = 'L', 462 [VC4_TILING_FORMAT_T] = 'T' 463 }; 464 fprintf(stderr, 465 "rsc setup %p (format %s: vc4 %d), %dx%d: " 466 "level %d (%c) -> %dx%d, stride %d@0x%08x\n", 467 rsc, 468 util_format_short_name(prsc->format), 469 rsc->vc4_format, 470 prsc->width0, prsc->height0, 471 i, tiling_chars[slice->tiling], 472 level_width, level_height, 473 slice->stride, slice->offset); 474 } 475 } 476 477 /* The texture base pointer that has to point to level 0 doesn't have 478 * intra-page bits, so we have to align it, and thus shift up all the 479 * smaller slices. 480 */ 481 uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) - 482 rsc->slices[0].offset); 483 if (page_align_offset) { 484 for (int i = 0; i <= prsc->last_level; i++) 485 rsc->slices[i].offset += page_align_offset; 486 } 487 488 /* Cube map faces appear as whole miptrees at a page-aligned offset 489 * from the first face's miptree. 490 */ 491 if (prsc->target == PIPE_TEXTURE_CUBE) { 492 rsc->cube_map_stride = align(rsc->slices[0].offset + 493 rsc->slices[0].size, 4096); 494 } 495 } 496 497 static struct vc4_resource * 498 vc4_resource_setup(struct pipe_screen *pscreen, 499 const struct pipe_resource *tmpl) 500 { 501 struct vc4_resource *rsc = CALLOC_STRUCT(vc4_resource); 502 if (!rsc) 503 return NULL; 504 struct pipe_resource *prsc = &rsc->base.b; 505 506 *prsc = *tmpl; 507 508 pipe_reference_init(&prsc->reference, 1); 509 prsc->screen = pscreen; 510 511 rsc->base.vtbl = &vc4_resource_vtbl; 512 if (prsc->nr_samples <= 1) 513 rsc->cpp = util_format_get_blocksize(tmpl->format); 514 else 515 rsc->cpp = sizeof(uint32_t); 516 517 assert(rsc->cpp); 518 519 return rsc; 520 } 521 522 static enum vc4_texture_data_type 523 get_resource_texture_format(struct pipe_resource *prsc) 524 { 525 struct vc4_resource *rsc = vc4_resource(prsc); 526 uint8_t format = vc4_get_tex_format(prsc->format); 527 528 if (!rsc->tiled) { 529 if (prsc->nr_samples > 1) { 530 return ~0; 531 } else { 532 assert(format == VC4_TEXTURE_TYPE_RGBA8888); 533 return VC4_TEXTURE_TYPE_RGBA32R; 534 } 535 } 536 537 return format; 538 } 539 540 struct pipe_resource * 541 vc4_resource_create(struct pipe_screen *pscreen, 542 const struct pipe_resource *tmpl) 543 { 544 struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl); 545 struct pipe_resource *prsc = &rsc->base.b; 546 547 /* We have to make shared be untiled, since we don't have any way to 548 * communicate metadata about tiling currently. 549 */ 550 if (tmpl->target == PIPE_BUFFER || 551 tmpl->nr_samples > 1 || 552 (tmpl->bind & (PIPE_BIND_SCANOUT | 553 PIPE_BIND_LINEAR | 554 PIPE_BIND_SHARED | 555 PIPE_BIND_CURSOR))) { 556 rsc->tiled = false; 557 } else { 558 rsc->tiled = true; 559 } 560 561 if (tmpl->target != PIPE_BUFFER) 562 rsc->vc4_format = get_resource_texture_format(prsc); 563 564 vc4_setup_slices(rsc); 565 if (!vc4_resource_bo_alloc(rsc)) 566 goto fail; 567 568 return prsc; 569 fail: 570 vc4_resource_destroy(pscreen, prsc); 571 return NULL; 572 } 573 574 static struct pipe_resource * 575 vc4_resource_from_handle(struct pipe_screen *pscreen, 576 const struct pipe_resource *tmpl, 577 struct winsys_handle *handle, 578 unsigned usage) 579 { 580 struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl); 581 struct pipe_resource *prsc = &rsc->base.b; 582 struct vc4_resource_slice *slice = &rsc->slices[0]; 583 uint32_t expected_stride = 584 align(prsc->width0, vc4_utile_width(rsc->cpp)) * rsc->cpp; 585 586 if (!rsc) 587 return NULL; 588 589 if (handle->stride != expected_stride) { 590 static bool warned = false; 591 if (!warned) { 592 warned = true; 593 fprintf(stderr, 594 "Attempting to import %dx%d %s with " 595 "unsupported stride %d instead of %d\n", 596 prsc->width0, prsc->height0, 597 util_format_short_name(prsc->format), 598 handle->stride, 599 expected_stride); 600 } 601 goto fail; 602 } 603 604 rsc->tiled = false; 605 rsc->bo = vc4_screen_bo_from_handle(pscreen, handle); 606 if (!rsc->bo) 607 goto fail; 608 609 slice->stride = handle->stride; 610 slice->tiling = VC4_TILING_FORMAT_LINEAR; 611 612 rsc->vc4_format = get_resource_texture_format(prsc); 613 614 if (miptree_debug) { 615 fprintf(stderr, 616 "rsc import %p (format %d), %dx%d: " 617 "level 0 (R) -> stride %d@0x%08x\n", 618 rsc, rsc->vc4_format, 619 prsc->width0, prsc->height0, 620 slice->stride, slice->offset); 621 } 622 623 return prsc; 624 625 fail: 626 vc4_resource_destroy(pscreen, prsc); 627 return NULL; 628 } 629 630 static struct pipe_surface * 631 vc4_create_surface(struct pipe_context *pctx, 632 struct pipe_resource *ptex, 633 const struct pipe_surface *surf_tmpl) 634 { 635 struct vc4_surface *surface = CALLOC_STRUCT(vc4_surface); 636 struct vc4_resource *rsc = vc4_resource(ptex); 637 638 if (!surface) 639 return NULL; 640 641 assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer); 642 643 struct pipe_surface *psurf = &surface->base; 644 unsigned level = surf_tmpl->u.tex.level; 645 646 pipe_reference_init(&psurf->reference, 1); 647 pipe_resource_reference(&psurf->texture, ptex); 648 649 psurf->context = pctx; 650 psurf->format = surf_tmpl->format; 651 psurf->width = u_minify(ptex->width0, level); 652 psurf->height = u_minify(ptex->height0, level); 653 psurf->u.tex.level = level; 654 psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer; 655 psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer; 656 surface->offset = (rsc->slices[level].offset + 657 psurf->u.tex.first_layer * rsc->cube_map_stride); 658 surface->tiling = rsc->slices[level].tiling; 659 660 return &surface->base; 661 } 662 663 static void 664 vc4_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf) 665 { 666 pipe_resource_reference(&psurf->texture, NULL); 667 FREE(psurf); 668 } 669 670 static void 671 vc4_dump_surface_non_msaa(struct pipe_surface *psurf) 672 { 673 struct pipe_resource *prsc = psurf->texture; 674 struct vc4_resource *rsc = vc4_resource(prsc); 675 uint32_t *map = vc4_bo_map(rsc->bo); 676 uint32_t stride = rsc->slices[0].stride / 4; 677 uint32_t width = psurf->width; 678 uint32_t height = psurf->height; 679 uint32_t chunk_w = width / 79; 680 uint32_t chunk_h = height / 40; 681 uint32_t found_colors[10]; 682 uint32_t num_found_colors = 0; 683 684 if (rsc->vc4_format != VC4_TEXTURE_TYPE_RGBA32R) { 685 fprintf(stderr, "%s: Unsupported format %s\n", 686 __func__, util_format_short_name(psurf->format)); 687 return; 688 } 689 690 for (int by = 0; by < height; by += chunk_h) { 691 for (int bx = 0; bx < width; bx += chunk_w) { 692 int all_found_color = -1; /* nothing found */ 693 694 for (int y = by; y < MIN2(height, by + chunk_h); y++) { 695 for (int x = bx; x < MIN2(width, bx + chunk_w); x++) { 696 uint32_t pix = map[y * stride + x]; 697 698 int i; 699 for (i = 0; i < num_found_colors; i++) { 700 if (pix == found_colors[i]) 701 break; 702 } 703 if (i == num_found_colors && 704 num_found_colors < 705 ARRAY_SIZE(found_colors)) { 706 found_colors[num_found_colors++] = pix; 707 } 708 709 if (i < num_found_colors) { 710 if (all_found_color == -1) 711 all_found_color = i; 712 else if (i != all_found_color) 713 all_found_color = ARRAY_SIZE(found_colors); 714 } 715 } 716 } 717 /* If all pixels for this chunk have a consistent 718 * value, then print a character for it. Either a 719 * fixed name (particularly common for piglit tests), 720 * or a runtime-generated number. 721 */ 722 if (all_found_color >= 0 && 723 all_found_color < ARRAY_SIZE(found_colors)) { 724 static const struct { 725 uint32_t val; 726 const char *c; 727 } named_colors[] = { 728 { 0xff000000, "" }, 729 { 0x00000000, "" }, 730 { 0xffff0000, "r" }, 731 { 0xff00ff00, "g" }, 732 { 0xff0000ff, "b" }, 733 { 0xffffffff, "w" }, 734 }; 735 int i; 736 for (i = 0; i < ARRAY_SIZE(named_colors); i++) { 737 if (named_colors[i].val == 738 found_colors[all_found_color]) { 739 fprintf(stderr, "%s", 740 named_colors[i].c); 741 break; 742 } 743 } 744 /* For unnamed colors, print a number and the 745 * numbers will have values printed at the 746 * end. 747 */ 748 if (i == ARRAY_SIZE(named_colors)) { 749 fprintf(stderr, "%c", 750 '0' + all_found_color); 751 } 752 } else { 753 /* If there's no consistent color, print this. 754 */ 755 fprintf(stderr, "."); 756 } 757 } 758 fprintf(stderr, "\n"); 759 } 760 761 for (int i = 0; i < num_found_colors; i++) { 762 fprintf(stderr, "color %d: 0x%08x\n", i, found_colors[i]); 763 } 764 } 765 766 static uint32_t 767 vc4_surface_msaa_get_sample(struct pipe_surface *psurf, 768 uint32_t x, uint32_t y, uint32_t sample) 769 { 770 struct pipe_resource *prsc = psurf->texture; 771 struct vc4_resource *rsc = vc4_resource(prsc); 772 uint32_t tile_w = 32, tile_h = 32; 773 uint32_t tiles_w = DIV_ROUND_UP(psurf->width, 32); 774 775 uint32_t tile_x = x / tile_w; 776 uint32_t tile_y = y / tile_h; 777 uint32_t *tile = (vc4_bo_map(rsc->bo) + 778 VC4_TILE_BUFFER_SIZE * (tile_y * tiles_w + tile_x)); 779 uint32_t subtile_x = x % tile_w; 780 uint32_t subtile_y = y % tile_h; 781 782 uint32_t quad_samples = VC4_MAX_SAMPLES * 4; 783 uint32_t tile_stride = quad_samples * tile_w / 2; 784 785 return *((uint32_t *)tile + 786 (subtile_y >> 1) * tile_stride + 787 (subtile_x >> 1) * quad_samples + 788 ((subtile_y & 1) << 1) + 789 (subtile_x & 1) + 790 sample); 791 } 792 793 static void 794 vc4_dump_surface_msaa_char(struct pipe_surface *psurf, 795 uint32_t start_x, uint32_t start_y, 796 uint32_t w, uint32_t h) 797 { 798 bool all_same_color = true; 799 uint32_t all_pix = 0; 800 801 for (int y = start_y; y < start_y + h; y++) { 802 for (int x = start_x; x < start_x + w; x++) { 803 for (int s = 0; s < VC4_MAX_SAMPLES; s++) { 804 uint32_t pix = vc4_surface_msaa_get_sample(psurf, 805 x, y, 806 s); 807 if (x == start_x && y == start_y) 808 all_pix = pix; 809 else if (all_pix != pix) 810 all_same_color = false; 811 } 812 } 813 } 814 if (all_same_color) { 815 static const struct { 816 uint32_t val; 817 const char *c; 818 } named_colors[] = { 819 { 0xff000000, "" }, 820 { 0x00000000, "" }, 821 { 0xffff0000, "r" }, 822 { 0xff00ff00, "g" }, 823 { 0xff0000ff, "b" }, 824 { 0xffffffff, "w" }, 825 }; 826 int i; 827 for (i = 0; i < ARRAY_SIZE(named_colors); i++) { 828 if (named_colors[i].val == all_pix) { 829 fprintf(stderr, "%s", 830 named_colors[i].c); 831 return; 832 } 833 } 834 fprintf(stderr, "x"); 835 } else { 836 fprintf(stderr, "."); 837 } 838 } 839 840 static void 841 vc4_dump_surface_msaa(struct pipe_surface *psurf) 842 { 843 uint32_t tile_w = 32, tile_h = 32; 844 uint32_t tiles_w = DIV_ROUND_UP(psurf->width, tile_w); 845 uint32_t tiles_h = DIV_ROUND_UP(psurf->height, tile_h); 846 uint32_t char_w = 140, char_h = 60; 847 uint32_t char_w_per_tile = char_w / tiles_w - 1; 848 uint32_t char_h_per_tile = char_h / tiles_h - 1; 849 uint32_t found_colors[10]; 850 uint32_t num_found_colors = 0; 851 852 fprintf(stderr, "Surface: %dx%d (%dx MSAA)\n", 853 psurf->width, psurf->height, psurf->texture->nr_samples); 854 855 for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++) 856 fprintf(stderr, "-"); 857 fprintf(stderr, "\n"); 858 859 for (int ty = 0; ty < psurf->height; ty += tile_h) { 860 for (int y = 0; y < char_h_per_tile; y++) { 861 862 for (int tx = 0; tx < psurf->width; tx += tile_w) { 863 for (int x = 0; x < char_w_per_tile; x++) { 864 uint32_t bx1 = (x * tile_w / 865 char_w_per_tile); 866 uint32_t bx2 = ((x + 1) * tile_w / 867 char_w_per_tile); 868 uint32_t by1 = (y * tile_h / 869 char_h_per_tile); 870 uint32_t by2 = ((y + 1) * tile_h / 871 char_h_per_tile); 872 873 vc4_dump_surface_msaa_char(psurf, 874 tx + bx1, 875 ty + by1, 876 bx2 - bx1, 877 by2 - by1); 878 } 879 fprintf(stderr, "|"); 880 } 881 fprintf(stderr, "\n"); 882 } 883 884 for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++) 885 fprintf(stderr, "-"); 886 fprintf(stderr, "\n"); 887 } 888 889 for (int i = 0; i < num_found_colors; i++) { 890 fprintf(stderr, "color %d: 0x%08x\n", i, found_colors[i]); 891 } 892 } 893 894 /** Debug routine to dump the contents of an 8888 surface to the console */ 895 void 896 vc4_dump_surface(struct pipe_surface *psurf) 897 { 898 if (!psurf) 899 return; 900 901 if (psurf->texture->nr_samples > 1) 902 vc4_dump_surface_msaa(psurf); 903 else 904 vc4_dump_surface_non_msaa(psurf); 905 } 906 907 static void 908 vc4_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource) 909 { 910 /* All calls to flush_resource are followed by a flush of the context, 911 * so there's nothing to do. 912 */ 913 } 914 915 void 916 vc4_update_shadow_baselevel_texture(struct pipe_context *pctx, 917 struct pipe_sampler_view *view) 918 { 919 struct vc4_resource *shadow = vc4_resource(view->texture); 920 struct vc4_resource *orig = vc4_resource(shadow->shadow_parent); 921 assert(orig); 922 923 if (shadow->writes == orig->writes && orig->bo->private) 924 return; 925 926 perf_debug("Updating %dx%d@%d shadow texture due to %s\n", 927 orig->base.b.width0, orig->base.b.height0, 928 view->u.tex.first_level, 929 view->u.tex.first_level ? "base level" : "raster layout"); 930 931 for (int i = 0; i <= shadow->base.b.last_level; i++) { 932 unsigned width = u_minify(shadow->base.b.width0, i); 933 unsigned height = u_minify(shadow->base.b.height0, i); 934 struct pipe_blit_info info = { 935 .dst = { 936 .resource = &shadow->base.b, 937 .level = i, 938 .box = { 939 .x = 0, 940 .y = 0, 941 .z = 0, 942 .width = width, 943 .height = height, 944 .depth = 1, 945 }, 946 .format = shadow->base.b.format, 947 }, 948 .src = { 949 .resource = &orig->base.b, 950 .level = view->u.tex.first_level + i, 951 .box = { 952 .x = 0, 953 .y = 0, 954 .z = 0, 955 .width = width, 956 .height = height, 957 .depth = 1, 958 }, 959 .format = orig->base.b.format, 960 }, 961 .mask = ~0, 962 }; 963 pctx->blit(pctx, &info); 964 } 965 966 shadow->writes = orig->writes; 967 } 968 969 /** 970 * Converts a 4-byte index buffer to 2 bytes. 971 * 972 * Since GLES2 only has support for 1 and 2-byte indices, the hardware doesn't 973 * include 4-byte index support, and we have to shrink it down. 974 * 975 * There's no fallback support for when indices end up being larger than 2^16, 976 * though it will at least assertion fail. Also, if the original index data 977 * was in user memory, it would be nice to not have uploaded it to a VBO 978 * before translating. 979 */ 980 struct pipe_resource * 981 vc4_get_shadow_index_buffer(struct pipe_context *pctx, 982 const struct pipe_index_buffer *ib, 983 uint32_t count, 984 uint32_t *shadow_offset) 985 { 986 struct vc4_context *vc4 = vc4_context(pctx); 987 struct vc4_resource *orig = vc4_resource(ib->buffer); 988 perf_debug("Fallback conversion for %d uint indices\n", count); 989 990 void *data; 991 struct pipe_resource *shadow_rsc = NULL; 992 u_upload_alloc(vc4->uploader, 0, count * 2, 4, 993 shadow_offset, &shadow_rsc, &data); 994 uint16_t *dst = data; 995 996 struct pipe_transfer *src_transfer = NULL; 997 const uint32_t *src; 998 if (ib->user_buffer) { 999 src = ib->user_buffer; 1000 } else { 1001 src = pipe_buffer_map_range(pctx, &orig->base.b, 1002 ib->offset, 1003 count * 4, 1004 PIPE_TRANSFER_READ, &src_transfer); 1005 } 1006 1007 for (int i = 0; i < count; i++) { 1008 uint32_t src_index = src[i]; 1009 assert(src_index <= 0xffff); 1010 dst[i] = src_index; 1011 } 1012 1013 if (src_transfer) 1014 pctx->transfer_unmap(pctx, src_transfer); 1015 1016 return shadow_rsc; 1017 } 1018 1019 void 1020 vc4_resource_screen_init(struct pipe_screen *pscreen) 1021 { 1022 pscreen->resource_create = vc4_resource_create; 1023 pscreen->resource_from_handle = vc4_resource_from_handle; 1024 pscreen->resource_get_handle = u_resource_get_handle_vtbl; 1025 pscreen->resource_destroy = u_resource_destroy_vtbl; 1026 } 1027 1028 void 1029 vc4_resource_context_init(struct pipe_context *pctx) 1030 { 1031 pctx->transfer_map = u_transfer_map_vtbl; 1032 pctx->transfer_flush_region = u_transfer_flush_region_vtbl; 1033 pctx->transfer_unmap = u_transfer_unmap_vtbl; 1034 pctx->buffer_subdata = u_default_buffer_subdata; 1035 pctx->texture_subdata = u_default_texture_subdata; 1036 pctx->create_surface = vc4_create_surface; 1037 pctx->surface_destroy = vc4_surface_destroy; 1038 pctx->resource_copy_region = util_resource_copy_region; 1039 pctx->blit = vc4_blit; 1040 pctx->flush_resource = vc4_flush_resource; 1041 } 1042