src/gallium/drivers/panfrost/pan_resource.c

   1 /*
   2  * Copyright (C) 2008 VMware, Inc.
   3  * Copyright (C) 2014 Broadcom
   4  * Copyright (C) 2018-2019 Alyssa Rosenzweig
   5  * Copyright (C) 2019 Collabora, Ltd.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the next
  15  * paragraph) shall be included in all copies or substantial portions of the
  16  * Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  21  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  22  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  23  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  24  * SOFTWARE.
  25  *
  26  * Authors (Collabora):
  27  *   Tomeu Vizoso <tomeu.vizoso@collabora.com>
  28  *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  29  *
  30  */
  31
  32 #include <xf86drm.h>
  33 #include <fcntl.h>
  34 #include "drm-uapi/drm_fourcc.h"
  35
  36 #include "state_tracker/winsys_handle.h"
  37 #include "util/format/u_format.h"
  38 #include "util/u_memory.h"
  39 #include "util/u_surface.h"
  40 #include "util/u_transfer.h"
  41 #include "util/u_transfer_helper.h"
  42 #include "util/u_gen_mipmap.h"
  43
  44 #include "pan_bo.h"
  45 #include "pan_context.h"
  46 #include "pan_screen.h"
  47 #include "pan_resource.h"
  48 #include "pan_util.h"
  49 #include "pan_tiling.h"
  50 #include "panfrost-quirks.h"
  51
  52 void
  53 panfrost_resource_reset_damage(struct panfrost_resource *pres)
  54 {
  55         /* We set the damage extent to the full resource size but keep the
  56          * damage box empty so that the FB content is reloaded by default.
  57          */
  58         memset(&pres->damage, 0, sizeof(pres->damage));
  59         pres->damage.extent.maxx = pres->base.width0;
  60         pres->damage.extent.maxy = pres->base.height0;
  61 }
  62
  63 static struct pipe_resource *
  64 panfrost_resource_from_handle(struct pipe_screen *pscreen,
  65                               const struct pipe_resource *templat,
  66                               struct winsys_handle *whandle,
  67                               unsigned usage)
  68 {
  69         struct panfrost_screen *screen = pan_screen(pscreen);
  70         struct panfrost_resource *rsc;
  71         struct pipe_resource *prsc;
  72
  73         assert(whandle->type == WINSYS_HANDLE_TYPE_FD);
  74
  75         rsc = rzalloc(pscreen, struct panfrost_resource);
  76         if (!rsc)
  77                 return NULL;
  78
  79         prsc = &rsc->base;
  80
  81         *prsc = *templat;
  82
  83         pipe_reference_init(&prsc->reference, 1);
  84         prsc->screen = pscreen;
  85
  86         rsc->bo = panfrost_bo_import(screen, whandle->handle);
  87         rsc->internal_format = templat->format;
  88         rsc->layout = MALI_TEXTURE_LINEAR;
  89         rsc->slices[0].stride = whandle->stride;
  90         rsc->slices[0].offset = whandle->offset;
  91         rsc->slices[0].initialized = true;
  92         panfrost_resource_reset_damage(rsc);
  93
  94         if (screen->ro) {
  95                 rsc->scanout =
  96                         renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
  97                 /* failure is expected in some cases.. */
  98         }
  99
 100         return prsc;
 101 }
 102
 103 static bool
 104 panfrost_resource_get_handle(struct pipe_screen *pscreen,
 105                              struct pipe_context *ctx,
 106                              struct pipe_resource *pt,
 107                              struct winsys_handle *handle,
 108                              unsigned usage)
 109 {
 110         struct panfrost_screen *screen = pan_screen(pscreen);
 111         struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
 112         struct renderonly_scanout *scanout = rsrc->scanout;
 113
 114         handle->modifier = DRM_FORMAT_MOD_INVALID;
 115
 116         if (handle->type == WINSYS_HANDLE_TYPE_SHARED) {
 117                 return false;
 118         } else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
 119                 if (renderonly_get_handle(scanout, handle))
 120                         return true;
 121
 122                 handle->handle = rsrc->bo->gem_handle;
 123                 handle->stride = rsrc->slices[0].stride;
 124                 handle->offset = rsrc->slices[0].offset;
 125                 return TRUE;
 126         } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
 127                 if (scanout) {
 128                         struct drm_prime_handle args = {
 129                                 .handle = scanout->handle,
 130                                 .flags = DRM_CLOEXEC,
 131                         };
 132
 133                         int ret = drmIoctl(screen->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 134                         if (ret == -1)
 135                                 return false;
 136
 137                         handle->stride = scanout->stride;
 138                         handle->handle = args.fd;
 139
 140                         return true;
 141                 } else {
 142                         int fd = panfrost_bo_export(rsrc->bo);
 143
 144                         if (fd < 0)
 145                                 return false;
 146
 147                         handle->handle = fd;
 148                         handle->stride = rsrc->slices[0].stride;
 149                         handle->offset = rsrc->slices[0].offset;
 150                         return true;
 151                 }
 152         }
 153
 154         return false;
 155 }
 156
 157 static void
 158 panfrost_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 159 {
 160         //DBG("TODO %s\n", __func__);
 161 }
 162
 163 static struct pipe_surface *
 164 panfrost_create_surface(struct pipe_context *pipe,
 165                         struct pipe_resource *pt,
 166                         const struct pipe_surface *surf_tmpl)
 167 {
 168         struct pipe_surface *ps = NULL;
 169
 170         ps = rzalloc(pipe, struct pipe_surface);
 171
 172         if (ps) {
 173                 pipe_reference_init(&ps->reference, 1);
 174                 pipe_resource_reference(&ps->texture, pt);
 175                 ps->context = pipe;
 176                 ps->format = surf_tmpl->format;
 177
 178                 if (pt->target != PIPE_BUFFER) {
 179                         assert(surf_tmpl->u.tex.level <= pt->last_level);
 180                         ps->width = u_minify(pt->width0, surf_tmpl->u.tex.level);
 181                         ps->height = u_minify(pt->height0, surf_tmpl->u.tex.level);
 182                         ps->u.tex.level = surf_tmpl->u.tex.level;
 183                         ps->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
 184                         ps->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
 185                 } else {
 186                         /* setting width as number of elements should get us correct renderbuffer width */
 187                         ps->width = surf_tmpl->u.buf.last_element - surf_tmpl->u.buf.first_element + 1;
 188                         ps->height = pt->height0;
 189                         ps->u.buf.first_element = surf_tmpl->u.buf.first_element;
 190                         ps->u.buf.last_element = surf_tmpl->u.buf.last_element;
 191                         assert(ps->u.buf.first_element <= ps->u.buf.last_element);
 192                         assert(ps->u.buf.last_element < ps->width);
 193                 }
 194         }
 195
 196         return ps;
 197 }
 198
 199 static void
 200 panfrost_surface_destroy(struct pipe_context *pipe,
 201                          struct pipe_surface *surf)
 202 {
 203         assert(surf->texture);
 204         pipe_resource_reference(&surf->texture, NULL);
 205         ralloc_free(surf);
 206 }
 207
 208 static struct pipe_resource *
 209 panfrost_create_scanout_res(struct pipe_screen *screen,
 210                             const struct pipe_resource *template)
 211 {
 212         struct panfrost_screen *pscreen = pan_screen(screen);
 213         struct pipe_resource scanout_templat = *template;
 214         struct renderonly_scanout *scanout;
 215         struct winsys_handle handle;
 216         struct pipe_resource *res;
 217
 218         scanout = renderonly_scanout_for_resource(&scanout_templat,
 219                         pscreen->ro, &handle);
 220         if (!scanout)
 221                 return NULL;
 222
 223         assert(handle.type == WINSYS_HANDLE_TYPE_FD);
 224         /* TODO: handle modifiers? */
 225         res = screen->resource_from_handle(screen, template, &handle,
 226                                            PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE);
 227         close(handle.handle);
 228         if (!res)
 229                 return NULL;
 230
 231         struct panfrost_resource *pres = pan_resource(res);
 232
 233         pres->scanout = scanout;
 234
 235         return res;
 236 }
 237
 238 /* Setup the mip tree given a particular layout, possibly with checksumming */
 239
 240 static void
 241 panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size)
 242 {
 243         struct pipe_resource *res = &pres->base;
 244         unsigned width = res->width0;
 245         unsigned height = res->height0;
 246         unsigned depth = res->depth0;
 247         unsigned bytes_per_pixel = util_format_get_blocksize(res->format);
 248
 249         assert(depth > 0);
 250
 251         /* Tiled operates blockwise; linear is packed. Also, anything
 252          * we render to has to be tile-aligned. Maybe not strictly
 253          * necessary, but we're not *that* pressed for memory and it
 254          * makes code a lot simpler */
 255
 256         bool renderable = res->bind &
 257                           (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL);
 258         bool afbc = pres->layout == MALI_TEXTURE_AFBC;
 259         bool tiled = pres->layout == MALI_TEXTURE_TILED;
 260         bool should_align = renderable || tiled;
 261
 262         /* We don't know how to specify a 2D stride for 3D textures */
 263
 264         bool can_align_stride =
 265                 res->target != PIPE_TEXTURE_3D;
 266
 267         should_align &= can_align_stride;
 268
 269         unsigned offset = 0;
 270         unsigned size_2d = 0;
 271
 272         for (unsigned l = 0; l <= res->last_level; ++l) {
 273                 struct panfrost_slice *slice = &pres->slices[l];
 274
 275                 unsigned effective_width = width;
 276                 unsigned effective_height = height;
 277                 unsigned effective_depth = depth;
 278
 279                 if (should_align) {
 280                         effective_width = ALIGN_POT(effective_width, 16);
 281                         effective_height = ALIGN_POT(effective_height, 16);
 282
 283                         /* We don't need to align depth */
 284                 }
 285
 286                 /* Align levels to cache-line as a performance improvement for
 287                  * linear/tiled and as a requirement for AFBC */
 288
 289                 offset = ALIGN_POT(offset, 64);
 290
 291                 slice->offset = offset;
 292
 293                 /* Compute the would-be stride */
 294                 unsigned stride = bytes_per_pixel * effective_width;
 295
 296                 if (util_format_is_compressed(res->format))
 297                         stride /= 4;
 298
 299                 /* ..but cache-line align it for performance */
 300                 if (can_align_stride && pres->layout == MALI_TEXTURE_LINEAR)
 301                         stride = ALIGN_POT(stride, 64);
 302
 303                 slice->stride = stride;
 304
 305                 unsigned slice_one_size = slice->stride * effective_height;
 306                 unsigned slice_full_size = slice_one_size * effective_depth;
 307
 308                 slice->size0 = slice_one_size;
 309
 310                 /* Report 2D size for 3D texturing */
 311
 312                 if (l == 0)
 313                         size_2d = slice_one_size;
 314
 315                 /* Compute AFBC sizes if necessary */
 316                 if (afbc) {
 317                         slice->header_size =
 318                                 panfrost_afbc_header_size(width, height);
 319
 320                         offset += slice->header_size;
 321                 }
 322
 323                 offset += slice_full_size;
 324
 325                 /* Add a checksum region if necessary */
 326                 if (pres->checksummed) {
 327                         slice->checksum_offset = offset;
 328
 329                         unsigned size = panfrost_compute_checksum_size(
 330                                                 slice, width, height);
 331
 332                         offset += size;
 333                 }
 334
 335                 width = u_minify(width, 1);
 336                 height = u_minify(height, 1);
 337                 depth = u_minify(depth, 1);
 338         }
 339
 340         assert(res->array_size);
 341
 342         if (res->target != PIPE_TEXTURE_3D) {
 343                 /* Arrays and cubemaps have the entire miptree duplicated */
 344
 345                 pres->cubemap_stride = ALIGN_POT(offset, 64);
 346                 *bo_size = ALIGN_POT(pres->cubemap_stride * res->array_size, 4096);
 347         } else {
 348                 /* 3D strides across the 2D layers */
 349                 assert(res->array_size == 1);
 350
 351                 pres->cubemap_stride = size_2d;
 352                 *bo_size = ALIGN_POT(offset, 4096);
 353         }
 354 }
 355
 356 static void
 357 panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_resource *pres)
 358 {
 359         struct pipe_resource *res = &pres->base;
 360
 361         /* Based on the usage, figure out what storing will be used. There are
 362          * various tradeoffs:
 363          *
 364          * Linear: the basic format, bad for memory bandwidth, bad for cache
 365          * use. Zero-copy, though. Renderable.
 366          *
 367          * Tiled: Not compressed, but cache-optimized. Expensive to write into
 368          * (due to software tiling), but cheap to sample from. Ideal for most
 369          * textures.
 370          *
 371          * AFBC: Compressed and renderable (so always desirable for non-scanout
 372          * rendertargets). Cheap to sample from. The format is black box, so we
 373          * can't read/write from software.
 374          *
 375          * Tiling textures is almost always faster, unless we only use it once.
 376          * Only a few types of resources can be tiled, ensure the bind is only
 377          * (a combination of) one of the following */
 378
 379         const unsigned valid_binding =
 380                 PIPE_BIND_DEPTH_STENCIL |
 381                 PIPE_BIND_RENDER_TARGET |
 382                 PIPE_BIND_BLENDABLE |
 383                 PIPE_BIND_SAMPLER_VIEW |
 384                 PIPE_BIND_DISPLAY_TARGET;
 385
 386         unsigned bpp = util_format_get_blocksizebits(res->format);
 387         bool is_2d = (res->target == PIPE_TEXTURE_2D);
 388         bool is_sane_bpp = bpp == 8 || bpp == 16 || bpp == 32 || bpp == 64 || bpp == 128;
 389         bool should_tile = (res->usage != PIPE_USAGE_STREAM);
 390         bool must_tile = (res->bind & PIPE_BIND_DEPTH_STENCIL) && (screen->quirks & MIDGARD_SFBD);
 391         bool can_tile = is_2d && is_sane_bpp && ((res->bind & ~valid_binding) == 0);
 392
 393         /* FBOs we would like to checksum, if at all possible */
 394         bool can_checksum = !(res->bind & ~valid_binding);
 395         bool should_checksum = res->bind & PIPE_BIND_RENDER_TARGET;
 396
 397         pres->checksummed = can_checksum && should_checksum;
 398
 399         /* Set the layout appropriately */
 400         assert(!(must_tile && !can_tile)); /* must_tile => can_tile */
 401         pres->layout = ((can_tile && should_tile) || must_tile) ? MALI_TEXTURE_TILED : MALI_TEXTURE_LINEAR;
 402
 403         size_t bo_size;
 404
 405         panfrost_setup_slices(pres, &bo_size);
 406
 407         /* We create a BO immediately but don't bother mapping, since we don't
 408          * care to map e.g. FBOs which the CPU probably won't touch */
 409         pres->bo = panfrost_bo_create(screen, bo_size, PAN_BO_DELAY_MMAP);
 410 }
 411
 412 void
 413 panfrost_resource_set_damage_region(struct pipe_screen *screen,
 414                                     struct pipe_resource *res,
 415                                     unsigned int nrects,
 416                                     const struct pipe_box *rects)
 417 {
 418         struct panfrost_resource *pres = pan_resource(res);
 419         struct pipe_box *damage_rect = &pres->damage.biggest_rect;
 420         struct pipe_scissor_state *damage_extent = &pres->damage.extent;
 421         unsigned int i;
 422
 423         if (!nrects) {
 424                 panfrost_resource_reset_damage(pres);
 425                 return;
 426         }
 427
 428         /* We keep track of 2 different things here:
 429          * 1 the damage extent: the quad including all damage regions. Will be
 430          *   used restrict the rendering area
 431          * 2 the biggest damage rectangle: when there are more than one damage
 432          *   rect we keep the biggest one and will generate 4 wallpaper quads
 433          *   out of it (see panfrost_draw_wallpaper() for more details). We
 434          *   might want to do something smarter at some point.
 435          *
 436          *                _________________________________
 437          *                |                               |
 438          *                |    _________________________  |
 439          *                |   | rect1|         _________| |
 440          *                |   |______|_____   | rect 3: | |
 441          *                |   |    | rect2 |  | biggest | |
 442          *                |   |    |_______|  |  rect   | |
 443          *                |   |_______________|_________| |
 444          *                |        damage extent          |
 445          *                |_______________________________|
 446          *                            resource
 447          */
 448         memset(&pres->damage, 0, sizeof(pres->damage));
 449         damage_extent->minx = 0xffff;
 450         damage_extent->miny = 0xffff;
 451         for (i = 0; i < nrects; i++) {
 452                 int x = rects[i].x, w = rects[i].width, h = rects[i].height;
 453                 int y = res->height0 - (rects[i].y + h);
 454
 455                 /* Clamp x,y,w,h to prevent negative values. */
 456                 if (x < 0) {
 457                         h += x;
 458                         x = 0;
 459                 }
 460                 if (y < 0) {
 461                         w += y;
 462                         y = 0;
 463                 }
 464                 w = MAX2(w, 0);
 465                 h = MAX2(h, 0);
 466
 467                 if (damage_rect->width * damage_rect->height < w * h)
 468                        u_box_2d(x, y, w, h, damage_rect);
 469
 470                 damage_extent->minx = MIN2(damage_extent->minx, x);
 471                 damage_extent->miny = MIN2(damage_extent->miny, y);
 472                 damage_extent->maxx = MAX2(damage_extent->maxx,
 473                                            MIN2(x + w, res->width0));
 474                 damage_extent->maxy = MAX2(damage_extent->maxy,
 475                                            MIN2(y + h, res->height0));
 476         }
 477 }
 478
 479 static struct pipe_resource *
 480 panfrost_resource_create(struct pipe_screen *screen,
 481                          const struct pipe_resource *template)
 482 {
 483         /* Make sure we're familiar */
 484         switch (template->target) {
 485         case PIPE_BUFFER:
 486         case PIPE_TEXTURE_1D:
 487         case PIPE_TEXTURE_2D:
 488         case PIPE_TEXTURE_3D:
 489         case PIPE_TEXTURE_CUBE:
 490         case PIPE_TEXTURE_RECT:
 491         case PIPE_TEXTURE_2D_ARRAY:
 492                 break;
 493         default:
 494                 DBG("Unknown texture target %d\n", template->target);
 495                 assert(0);
 496         }
 497
 498         if (template->bind &
 499             (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED))
 500                 return panfrost_create_scanout_res(screen, template);
 501
 502         struct panfrost_resource *so = rzalloc(screen, struct panfrost_resource);
 503         struct panfrost_screen *pscreen = (struct panfrost_screen *) screen;
 504
 505         so->base = *template;
 506         so->base.screen = screen;
 507         so->internal_format = template->format;
 508
 509         pipe_reference_init(&so->base.reference, 1);
 510
 511         util_range_init(&so->valid_buffer_range);
 512
 513         panfrost_resource_create_bo(pscreen, so);
 514         panfrost_resource_reset_damage(so);
 515
 516         if (template->bind & PIPE_BIND_INDEX_BUFFER)
 517                 so->index_cache = rzalloc(so, struct panfrost_minmax_cache);
 518
 519         return (struct pipe_resource *)so;
 520 }
 521
 522 static void
 523 panfrost_resource_destroy(struct pipe_screen *screen,
 524                           struct pipe_resource *pt)
 525 {
 526         struct panfrost_screen *pscreen = pan_screen(screen);
 527         struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
 528
 529         if (rsrc->scanout)
 530                 renderonly_scanout_destroy(rsrc->scanout, pscreen->ro);
 531
 532         if (rsrc->bo)
 533                 panfrost_bo_unreference(rsrc->bo);
 534
 535         util_range_destroy(&rsrc->valid_buffer_range);
 536         ralloc_free(rsrc);
 537 }
 538
 539
 540 static void *
 541 panfrost_transfer_map(struct pipe_context *pctx,
 542                       struct pipe_resource *resource,
 543                       unsigned level,
 544                       unsigned usage,  /* a combination of PIPE_TRANSFER_x */
 545                       const struct pipe_box *box,
 546                       struct pipe_transfer **out_transfer)
 547 {
 548         struct panfrost_context *ctx = pan_context(pctx);
 549         int bytes_per_pixel = util_format_get_blocksize(resource->format);
 550         struct panfrost_resource *rsrc = pan_resource(resource);
 551         struct panfrost_bo *bo = rsrc->bo;
 552
 553         struct panfrost_gtransfer *transfer = rzalloc(pctx, struct panfrost_gtransfer);
 554         transfer->base.level = level;
 555         transfer->base.usage = usage;
 556         transfer->base.box = *box;
 557
 558         pipe_resource_reference(&transfer->base.resource, resource);
 559
 560         *out_transfer = &transfer->base;
 561
 562         /* If we haven't already mmaped, now's the time */
 563         panfrost_bo_mmap(bo);
 564
 565         if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 566                 /* If the BO is used by one of the pending batches or if it's
 567                  * not ready yet (still accessed by one of the already flushed
 568                  * batches), we try to allocate a new one to avoid waiting.
 569                  */
 570                 if (panfrost_pending_batches_access_bo(ctx, bo) ||
 571                     !panfrost_bo_wait(bo, 0, PAN_BO_ACCESS_RW)) {
 572                         struct panfrost_screen *screen = pan_screen(pctx->screen);
 573                         /* We want the BO to be MMAPed. */
 574                         uint32_t flags = bo->flags & ~PAN_BO_DELAY_MMAP;
 575                         struct panfrost_bo *newbo = NULL;
 576
 577                         /* When the BO has been imported/exported, we can't
 578                          * replace it by another one, otherwise the
 579                          * importer/exporter wouldn't see the change we're
 580                          * doing to it.
 581                          */
 582                         if (!(bo->flags & (PAN_BO_IMPORTED | PAN_BO_EXPORTED)))
 583                                 newbo = panfrost_bo_create(screen, bo->size,
 584                                                            flags);
 585
 586                         if (newbo) {
 587                                 panfrost_bo_unreference(bo);
 588                                 rsrc->bo = newbo;
 589                                 bo = newbo;
 590                         } else {
 591                                 uint32_t access = PAN_BO_ACCESS_RW;
 592
 593                                 /* Allocation failed or was impossible, let's
 594                                  * fall back on a flush+wait.
 595                                  */
 596                                 panfrost_flush_batches_accessing_bo(ctx, bo,
 597                                                                     access);
 598                                 panfrost_bo_wait(bo, INT64_MAX, access);
 599                         }
 600                 }
 601         } else if ((usage & PIPE_TRANSFER_WRITE)
 602                    && resource->target == PIPE_BUFFER
 603                    && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) {
 604                 /* No flush for writes to uninitialized */
 605         } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
 606                 if (usage & PIPE_TRANSFER_WRITE) {
 607                         panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_RW);
 608                         panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_RW);
 609                 } else if (usage & PIPE_TRANSFER_READ) {
 610                         panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_WRITE);
 611                         panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
 612                 }
 613         }
 614
 615         if (rsrc->layout != MALI_TEXTURE_LINEAR) {
 616                 /* Non-linear resources need to be indirectly mapped */
 617
 618                 if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
 619                         return NULL;
 620
 621                 transfer->base.stride = box->width * bytes_per_pixel;
 622                 transfer->base.layer_stride = transfer->base.stride * box->height;
 623                 transfer->map = rzalloc_size(transfer, transfer->base.layer_stride * box->depth);
 624                 assert(box->depth == 1);
 625
 626                 if ((usage & PIPE_TRANSFER_READ) && rsrc->slices[level].initialized) {
 627                         if (rsrc->layout == MALI_TEXTURE_AFBC) {
 628                                 DBG("Unimplemented: reads from AFBC");
 629                         } else if (rsrc->layout == MALI_TEXTURE_TILED) {
 630                                 panfrost_load_tiled_image(
 631                                         transfer->map,
 632                                         bo->cpu + rsrc->slices[level].offset,
 633                                         box->x, box->y, box->width, box->height,
 634                                         transfer->base.stride,
 635                                         rsrc->slices[level].stride,
 636                                         resource->format);
 637                         }
 638                 }
 639
 640                 return transfer->map;
 641         } else {
 642                 /* Direct, persistent writes create holes in time for
 643                  * caching... I don't know if this is actually possible but we
 644                  * should still get it right */
 645
 646                 unsigned dpw = PIPE_TRANSFER_MAP_DIRECTLY | PIPE_TRANSFER_WRITE | PIPE_TRANSFER_PERSISTENT;
 647
 648                 if ((usage & dpw) == dpw && rsrc->index_cache)
 649                         return NULL;
 650
 651                 transfer->base.stride = rsrc->slices[level].stride;
 652                 transfer->base.layer_stride = panfrost_get_layer_stride(
 653                                 rsrc->slices, rsrc->base.target == PIPE_TEXTURE_3D,
 654                                 rsrc->cubemap_stride, level);
 655
 656                 /* By mapping direct-write, we're implicitly already
 657                  * initialized (maybe), so be conservative */
 658
 659                 if ((usage & PIPE_TRANSFER_WRITE) && (usage & PIPE_TRANSFER_MAP_DIRECTLY)) {
 660                         rsrc->slices[level].initialized = true;
 661                         panfrost_minmax_cache_invalidate(rsrc->index_cache, &transfer->base);
 662                 }
 663
 664                 return bo->cpu
 665                        + rsrc->slices[level].offset
 666                        + transfer->base.box.z * transfer->base.layer_stride
 667                        + transfer->base.box.y * rsrc->slices[level].stride
 668                        + transfer->base.box.x * bytes_per_pixel;
 669         }
 670 }
 671
 672 static void
 673 panfrost_transfer_unmap(struct pipe_context *pctx,
 674                         struct pipe_transfer *transfer)
 675 {
 676         /* Gallium expects writeback here, so we tile */
 677
 678         struct panfrost_gtransfer *trans = pan_transfer(transfer);
 679         struct panfrost_resource *prsrc = (struct panfrost_resource *) transfer->resource;
 680
 681         /* Mark whatever we wrote as written */
 682         if (transfer->usage & PIPE_TRANSFER_WRITE)
 683                 prsrc->slices[transfer->level].initialized = true;
 684
 685         if (trans->map) {
 686                 struct panfrost_bo *bo = prsrc->bo;
 687
 688                 if (transfer->usage & PIPE_TRANSFER_WRITE) {
 689                         if (prsrc->layout == MALI_TEXTURE_AFBC) {
 690                                 DBG("Unimplemented: writes to AFBC\n");
 691                         } else if (prsrc->layout == MALI_TEXTURE_TILED) {
 692                                 assert(transfer->box.depth == 1);
 693
 694                                 panfrost_store_tiled_image(
 695                                         bo->cpu + prsrc->slices[transfer->level].offset,
 696                                         trans->map,
 697                                         transfer->box.x, transfer->box.y,
 698                                         transfer->box.width, transfer->box.height,
 699                                         prsrc->slices[transfer->level].stride,
 700                                         transfer->stride,
 701                                         prsrc->base.format);
 702                         }
 703                 }
 704         }
 705
 706
 707         util_range_add(&prsrc->base, &prsrc->valid_buffer_range,
 708                        transfer->box.x,
 709                        transfer->box.x + transfer->box.width);
 710
 711         panfrost_minmax_cache_invalidate(prsrc->index_cache, transfer);
 712
 713         /* Derefence the resource */
 714         pipe_resource_reference(&transfer->resource, NULL);
 715
 716         /* Transfer itself is RALLOCed at the moment */
 717         ralloc_free(transfer);
 718 }
 719
 720 static void
 721 panfrost_transfer_flush_region(struct pipe_context *pctx,
 722                                struct pipe_transfer *transfer,
 723                                const struct pipe_box *box)
 724 {
 725         struct panfrost_resource *rsc = pan_resource(transfer->resource);
 726
 727         if (transfer->resource->target == PIPE_BUFFER) {
 728                 util_range_add(&rsc->base, &rsc->valid_buffer_range,
 729                                transfer->box.x + box->x,
 730                                transfer->box.x + box->x + box->width);
 731         } else {
 732                 unsigned level = transfer->level;
 733                 rsc->slices[level].initialized = true;
 734         }
 735 }
 736
 737 static void
 738 panfrost_invalidate_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 739 {
 740         //DBG("TODO %s\n", __func__);
 741 }
 742
 743 static enum pipe_format
 744 panfrost_resource_get_internal_format(struct pipe_resource *rsrc) {
 745         struct panfrost_resource *prsrc = (struct panfrost_resource *) rsrc;
 746         return prsrc->internal_format;
 747 }
 748
 749 static bool
 750 panfrost_generate_mipmap(
 751         struct pipe_context *pctx,
 752         struct pipe_resource *prsrc,
 753         enum pipe_format format,
 754         unsigned base_level,
 755         unsigned last_level,
 756         unsigned first_layer,
 757         unsigned last_layer)
 758 {
 759         struct panfrost_context *ctx = pan_context(pctx);
 760         struct panfrost_resource *rsrc = pan_resource(prsrc);
 761
 762         /* Generating a mipmap invalidates the written levels, so make that
 763          * explicit so we don't try to wallpaper them back and end up with
 764          * u_blitter recursion */
 765
 766         assert(rsrc->bo);
 767         for (unsigned l = base_level + 1; l <= last_level; ++l)
 768                 rsrc->slices[l].initialized = false;
 769
 770         /* Beyond that, we just delegate the hard stuff. We're careful to
 771          * include flushes on both ends to make sure the data is really valid.
 772          * We could be doing a lot better perf-wise, especially once we have
 773          * reorder-type optimizations in place. But for now prioritize
 774          * correctness. */
 775
 776         panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_RW);
 777         panfrost_bo_wait(rsrc->bo, INT64_MAX, PAN_BO_ACCESS_RW);
 778
 779         /* We've flushed the original buffer if needed, now trigger a blit */
 780
 781         bool blit_res = util_gen_mipmap(
 782                                 pctx, prsrc, format,
 783                                 base_level, last_level,
 784                                 first_layer, last_layer,
 785                                 PIPE_TEX_FILTER_LINEAR);
 786
 787         /* If the blit was successful, flush once more. If it wasn't, well, let
 788          * the state tracker deal with it. */
 789
 790         if (blit_res) {
 791                 panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_WRITE);
 792                 panfrost_bo_wait(rsrc->bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
 793         }
 794
 795         return blit_res;
 796 }
 797
 798 /* Computes the address to a texture at a particular slice */
 799
 800 mali_ptr
 801 panfrost_get_texture_address(
 802         struct panfrost_resource *rsrc,
 803         unsigned level, unsigned face)
 804 {
 805         bool is_3d = rsrc->base.target == PIPE_TEXTURE_3D;
 806         return rsrc->bo->gpu + panfrost_texture_offset(rsrc->slices, is_3d, rsrc->cubemap_stride, level, face);
 807 }
 808
 809 /* Given a resource that has already been allocated, hint that it should use a
 810  * given layout. These are suggestions, not commands; it is perfectly legal to
 811  * stub out this function, but there will be performance implications. */
 812
 813 void
 814 panfrost_resource_hint_layout(
 815                 struct panfrost_screen *screen,
 816                 struct panfrost_resource *rsrc,
 817                 enum mali_texture_layout layout,
 818                 signed weight)
 819 {
 820         /* Nothing to do, although a sophisticated implementation might store
 821          * the hint */
 822
 823         if (rsrc->layout == layout)
 824                 return;
 825
 826         /* We don't use the weight yet, but we should check that it's positive
 827          * (semantically meaning that we should choose the given `layout`) */
 828
 829         if (weight <= 0)
 830                 return;
 831
 832         /* Check if the preferred layout is legal for this buffer */
 833
 834         if (layout == MALI_TEXTURE_AFBC) {
 835                 bool can_afbc = panfrost_format_supports_afbc(rsrc->base.format);
 836                 bool is_scanout = rsrc->base.bind &
 837                         (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED);
 838
 839                 if (!can_afbc || is_scanout)
 840                         return;
 841         }
 842
 843         /* Simple heuristic so far: if the resource is uninitialized, switch to
 844          * the hinted layout. If it is initialized, keep the original layout.
 845          * This misses some cases where it would be beneficial to switch and
 846          * blit. */
 847
 848         bool is_initialized = false;
 849
 850         for (unsigned i = 0; i < MAX_MIP_LEVELS; ++i)
 851                 is_initialized |= rsrc->slices[i].initialized;
 852
 853         if (is_initialized)
 854                 return;
 855
 856         /* We're uninitialized, so do a layout switch. Reinitialize slices. */
 857
 858         size_t new_size;
 859         rsrc->layout = layout;
 860         panfrost_setup_slices(rsrc, &new_size);
 861
 862         /* If we grew in size, reallocate the BO */
 863         if (new_size > rsrc->bo->size) {
 864                 panfrost_bo_unreference(rsrc->bo);
 865                 rsrc->bo = panfrost_bo_create(screen, new_size, PAN_BO_DELAY_MMAP);
 866         }
 867
 868         /* TODO: If there are textures bound, regenerate their descriptors */
 869 }
 870
 871 static void
 872 panfrost_resource_set_stencil(struct pipe_resource *prsrc,
 873                               struct pipe_resource *stencil)
 874 {
 875         pan_resource(prsrc)->separate_stencil = pan_resource(stencil);
 876 }
 877
 878 static struct pipe_resource *
 879 panfrost_resource_get_stencil(struct pipe_resource *prsrc)
 880 {
 881         return &pan_resource(prsrc)->separate_stencil->base;
 882 }
 883
 884 static const struct u_transfer_vtbl transfer_vtbl = {
 885         .resource_create          = panfrost_resource_create,
 886         .resource_destroy         = panfrost_resource_destroy,
 887         .transfer_map             = panfrost_transfer_map,
 888         .transfer_unmap           = panfrost_transfer_unmap,
 889         .transfer_flush_region    = panfrost_transfer_flush_region,
 890         .get_internal_format      = panfrost_resource_get_internal_format,
 891         .set_stencil              = panfrost_resource_set_stencil,
 892         .get_stencil              = panfrost_resource_get_stencil,
 893 };
 894
 895 void
 896 panfrost_resource_screen_init(struct panfrost_screen *pscreen)
 897 {
 898         //pscreen->base.resource_create_with_modifiers =
 899         //        panfrost_resource_create_with_modifiers;
 900         pscreen->base.resource_create = u_transfer_helper_resource_create;
 901         pscreen->base.resource_destroy = u_transfer_helper_resource_destroy;
 902         pscreen->base.resource_from_handle = panfrost_resource_from_handle;
 903         pscreen->base.resource_get_handle = panfrost_resource_get_handle;
 904         pscreen->base.transfer_helper = u_transfer_helper_create(&transfer_vtbl,
 905                                         true, false,
 906                                         true, true);
 907 }
 908
 909 void
 910 panfrost_resource_context_init(struct pipe_context *pctx)
 911 {
 912         pctx->transfer_map = u_transfer_helper_transfer_map;
 913         pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
 914         pctx->create_surface = panfrost_create_surface;
 915         pctx->surface_destroy = panfrost_surface_destroy;
 916         pctx->resource_copy_region = util_resource_copy_region;
 917         pctx->blit = panfrost_blit;
 918         pctx->generate_mipmap = panfrost_generate_mipmap;
 919         pctx->flush_resource = panfrost_flush_resource;
 920         pctx->invalidate_resource = panfrost_invalidate_resource;
 921         pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
 922         pctx->buffer_subdata = u_default_buffer_subdata;
 923         pctx->texture_subdata = u_default_texture_subdata;
 924 }