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
  51 static void
  52 panfrost_resource_reset_damage(struct panfrost_resource *pres)
  53 {
  54         /* We set the damage extent to the full resource size but keep the
  55          * damage box empty so that the FB content is reloaded by default.
  56          */
  57         memset(&pres->damage, 0, sizeof(pres->damage));
  58         pres->damage.extent.maxx = pres->base.width0;
  59         pres->damage.extent.maxy = pres->base.height0;
  60 }
  61
  62 static struct pipe_resource *
  63 panfrost_resource_from_handle(struct pipe_screen *pscreen,
  64                               const struct pipe_resource *templat,
  65                               struct winsys_handle *whandle,
  66                               unsigned usage)
  67 {
  68         struct panfrost_screen *screen = pan_screen(pscreen);
  69         struct panfrost_resource *rsc;
  70         struct pipe_resource *prsc;
  71
  72         assert(whandle->type == WINSYS_HANDLE_TYPE_FD);
  73
  74         rsc = rzalloc(pscreen, struct panfrost_resource);
  75         if (!rsc)
  76                 return NULL;
  77
  78         prsc = &rsc->base;
  79
  80         *prsc = *templat;
  81
  82         pipe_reference_init(&prsc->reference, 1);
  83         prsc->screen = pscreen;
  84
  85         rsc->bo = panfrost_bo_import(screen, whandle->handle);
  86         rsc->slices[0].stride = whandle->stride;
  87         rsc->slices[0].offset = whandle->offset;
  88         rsc->slices[0].initialized = true;
  89         panfrost_resource_reset_damage(rsc);
  90
  91         if (screen->ro) {
  92                 rsc->scanout =
  93                         renderonly_create_gpu_import_for_resource(prsc, screen->ro, NULL);
  94                 /* failure is expected in some cases.. */
  95         }
  96
  97         return prsc;
  98 }
  99
 100 static bool
 101 panfrost_resource_get_handle(struct pipe_screen *pscreen,
 102                              struct pipe_context *ctx,
 103                              struct pipe_resource *pt,
 104                              struct winsys_handle *handle,
 105                              unsigned usage)
 106 {
 107         struct panfrost_screen *screen = pan_screen(pscreen);
 108         struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
 109         struct renderonly_scanout *scanout = rsrc->scanout;
 110
 111         handle->modifier = DRM_FORMAT_MOD_INVALID;
 112
 113         if (handle->type == WINSYS_HANDLE_TYPE_SHARED) {
 114                 return false;
 115         } else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
 116                 if (renderonly_get_handle(scanout, handle))
 117                         return true;
 118
 119                 handle->handle = rsrc->bo->gem_handle;
 120                 handle->stride = rsrc->slices[0].stride;
 121                 handle->offset = rsrc->slices[0].offset;
 122                 return TRUE;
 123         } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
 124                 if (scanout) {
 125                         struct drm_prime_handle args = {
 126                                 .handle = scanout->handle,
 127                                 .flags = DRM_CLOEXEC,
 128                         };
 129
 130                         int ret = drmIoctl(screen->ro->kms_fd, DRM_IOCTL_PRIME_HANDLE_TO_FD, &args);
 131                         if (ret == -1)
 132                                 return false;
 133
 134                         handle->stride = scanout->stride;
 135                         handle->handle = args.fd;
 136
 137                         return true;
 138                 } else {
 139                         int fd = panfrost_bo_export(rsrc->bo);
 140
 141                         if (fd < 0)
 142                                 return false;
 143
 144                         handle->handle = fd;
 145                         handle->stride = rsrc->slices[0].stride;
 146                         handle->offset = rsrc->slices[0].offset;
 147                         return true;
 148                 }
 149         }
 150
 151         return false;
 152 }
 153
 154 static void
 155 panfrost_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 156 {
 157         //DBG("TODO %s\n", __func__);
 158 }
 159
 160 static struct pipe_surface *
 161 panfrost_create_surface(struct pipe_context *pipe,
 162                         struct pipe_resource *pt,
 163                         const struct pipe_surface *surf_tmpl)
 164 {
 165         struct pipe_surface *ps = NULL;
 166
 167         ps = rzalloc(pipe, struct pipe_surface);
 168
 169         if (ps) {
 170                 pipe_reference_init(&ps->reference, 1);
 171                 pipe_resource_reference(&ps->texture, pt);
 172                 ps->context = pipe;
 173                 ps->format = surf_tmpl->format;
 174
 175                 if (pt->target != PIPE_BUFFER) {
 176                         assert(surf_tmpl->u.tex.level <= pt->last_level);
 177                         ps->width = u_minify(pt->width0, surf_tmpl->u.tex.level);
 178                         ps->height = u_minify(pt->height0, surf_tmpl->u.tex.level);
 179                         ps->u.tex.level = surf_tmpl->u.tex.level;
 180                         ps->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
 181                         ps->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
 182                 } else {
 183                         /* setting width as number of elements should get us correct renderbuffer width */
 184                         ps->width = surf_tmpl->u.buf.last_element - surf_tmpl->u.buf.first_element + 1;
 185                         ps->height = pt->height0;
 186                         ps->u.buf.first_element = surf_tmpl->u.buf.first_element;
 187                         ps->u.buf.last_element = surf_tmpl->u.buf.last_element;
 188                         assert(ps->u.buf.first_element <= ps->u.buf.last_element);
 189                         assert(ps->u.buf.last_element < ps->width);
 190                 }
 191         }
 192
 193         return ps;
 194 }
 195
 196 static void
 197 panfrost_surface_destroy(struct pipe_context *pipe,
 198                          struct pipe_surface *surf)
 199 {
 200         assert(surf->texture);
 201         pipe_resource_reference(&surf->texture, NULL);
 202         ralloc_free(surf);
 203 }
 204
 205 static struct pipe_resource *
 206 panfrost_create_scanout_res(struct pipe_screen *screen,
 207                             const struct pipe_resource *template)
 208 {
 209         struct panfrost_screen *pscreen = pan_screen(screen);
 210         struct pipe_resource scanout_templat = *template;
 211         struct renderonly_scanout *scanout;
 212         struct winsys_handle handle;
 213         struct pipe_resource *res;
 214
 215         scanout = renderonly_scanout_for_resource(&scanout_templat,
 216                         pscreen->ro, &handle);
 217         if (!scanout)
 218                 return NULL;
 219
 220         assert(handle.type == WINSYS_HANDLE_TYPE_FD);
 221         /* TODO: handle modifiers? */
 222         res = screen->resource_from_handle(screen, template, &handle,
 223                                            PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE);
 224         close(handle.handle);
 225         if (!res)
 226                 return NULL;
 227
 228         struct panfrost_resource *pres = pan_resource(res);
 229
 230         pres->scanout = scanout;
 231
 232         return res;
 233 }
 234
 235 /* Computes sizes for checksumming, which is 8 bytes per 16x16 tile */
 236
 237 #define CHECKSUM_TILE_WIDTH 16
 238 #define CHECKSUM_TILE_HEIGHT 16
 239 #define CHECKSUM_BYTES_PER_TILE 8
 240
 241 static unsigned
 242 panfrost_compute_checksum_sizes(
 243         struct panfrost_slice *slice,
 244         unsigned width,
 245         unsigned height)
 246 {
 247         unsigned aligned_width = ALIGN_POT(width, CHECKSUM_TILE_WIDTH);
 248         unsigned aligned_height = ALIGN_POT(height, CHECKSUM_TILE_HEIGHT);
 249
 250         unsigned tile_count_x = aligned_width / CHECKSUM_TILE_WIDTH;
 251         unsigned tile_count_y = aligned_height / CHECKSUM_TILE_HEIGHT;
 252
 253         slice->checksum_stride = tile_count_x * CHECKSUM_BYTES_PER_TILE;
 254
 255         return slice->checksum_stride * tile_count_y;
 256 }
 257
 258 /* Setup the mip tree given a particular layout, possibly with checksumming */
 259
 260 static void
 261 panfrost_setup_slices(struct panfrost_resource *pres, size_t *bo_size)
 262 {
 263         struct pipe_resource *res = &pres->base;
 264         unsigned width = res->width0;
 265         unsigned height = res->height0;
 266         unsigned depth = res->depth0;
 267         unsigned bytes_per_pixel = util_format_get_blocksize(res->format);
 268
 269         assert(depth > 0);
 270
 271         /* Tiled operates blockwise; linear is packed. Also, anything
 272          * we render to has to be tile-aligned. Maybe not strictly
 273          * necessary, but we're not *that* pressed for memory and it
 274          * makes code a lot simpler */
 275
 276         bool renderable = res->bind &
 277                           (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL);
 278         bool afbc = pres->layout == PAN_AFBC;
 279         bool tiled = pres->layout == PAN_TILED;
 280         bool should_align = renderable || tiled;
 281
 282         /* We don't know how to specify a 2D stride for 3D textures */
 283
 284         bool can_align_stride =
 285                 res->target != PIPE_TEXTURE_3D;
 286
 287         should_align &= can_align_stride;
 288
 289         unsigned offset = 0;
 290         unsigned size_2d = 0;
 291
 292         for (unsigned l = 0; l <= res->last_level; ++l) {
 293                 struct panfrost_slice *slice = &pres->slices[l];
 294
 295                 unsigned effective_width = width;
 296                 unsigned effective_height = height;
 297                 unsigned effective_depth = depth;
 298
 299                 if (should_align) {
 300                         effective_width = ALIGN_POT(effective_width, 16);
 301                         effective_height = ALIGN_POT(effective_height, 16);
 302
 303                         /* We don't need to align depth */
 304                 }
 305
 306                 /* Align levels to cache-line as a performance improvement for
 307                  * linear/tiled and as a requirement for AFBC */
 308
 309                 offset = ALIGN_POT(offset, 64);
 310
 311                 slice->offset = offset;
 312
 313                 /* Compute the would-be stride */
 314                 unsigned stride = bytes_per_pixel * effective_width;
 315
 316                 /* ..but cache-line align it for performance */
 317                 if (can_align_stride && pres->layout == PAN_LINEAR)
 318                         stride = ALIGN_POT(stride, 64);
 319
 320                 slice->stride = stride;
 321
 322                 unsigned slice_one_size = slice->stride * effective_height;
 323                 unsigned slice_full_size = slice_one_size * effective_depth;
 324
 325                 /* Report 2D size for 3D texturing */
 326
 327                 if (l == 0)
 328                         size_2d = slice_one_size;
 329
 330                 /* Compute AFBC sizes if necessary */
 331                 if (afbc) {
 332                         slice->header_size =
 333                                 panfrost_afbc_header_size(width, height);
 334
 335                         offset += slice->header_size;
 336                 }
 337
 338                 offset += slice_full_size;
 339
 340                 /* Add a checksum region if necessary */
 341                 if (pres->checksummed) {
 342                         slice->checksum_offset = offset;
 343
 344                         unsigned size = panfrost_compute_checksum_sizes(
 345                                                 slice, width, height);
 346
 347                         offset += size;
 348                 }
 349
 350                 width = u_minify(width, 1);
 351                 height = u_minify(height, 1);
 352                 depth = u_minify(depth, 1);
 353         }
 354
 355         assert(res->array_size);
 356
 357         if (res->target != PIPE_TEXTURE_3D) {
 358                 /* Arrays and cubemaps have the entire miptree duplicated */
 359
 360                 pres->cubemap_stride = ALIGN_POT(offset, 64);
 361                 *bo_size = ALIGN_POT(pres->cubemap_stride * res->array_size, 4096);
 362         } else {
 363                 /* 3D strides across the 2D layers */
 364                 assert(res->array_size == 1);
 365
 366                 pres->cubemap_stride = size_2d;
 367                 *bo_size = ALIGN_POT(offset, 4096);
 368         }
 369 }
 370
 371 static void
 372 panfrost_resource_create_bo(struct panfrost_screen *screen, struct panfrost_resource *pres)
 373 {
 374         struct pipe_resource *res = &pres->base;
 375
 376         /* Based on the usage, figure out what storing will be used. There are
 377          * various tradeoffs:
 378          *
 379          * Linear: the basic format, bad for memory bandwidth, bad for cache
 380          * use. Zero-copy, though. Renderable.
 381          *
 382          * Tiled: Not compressed, but cache-optimized. Expensive to write into
 383          * (due to software tiling), but cheap to sample from. Ideal for most
 384          * textures.
 385          *
 386          * AFBC: Compressed and renderable (so always desirable for non-scanout
 387          * rendertargets). Cheap to sample from. The format is black box, so we
 388          * can't read/write from software.
 389          */
 390
 391         /* Tiling textures is almost always faster, unless we only use it once */
 392
 393         bool is_texture = (res->bind & PIPE_BIND_SAMPLER_VIEW);
 394         bool is_2d = res->depth0 == 1 && res->array_size == 1;
 395         bool is_streaming = (res->usage != PIPE_USAGE_STREAM);
 396
 397         bool should_tile = is_streaming && is_texture && is_2d;
 398
 399         /* Depth/stencil can't be tiled, only linear or AFBC */
 400         should_tile &= !(res->bind & PIPE_BIND_DEPTH_STENCIL);
 401
 402         /* FBOs we would like to checksum, if at all possible */
 403         bool can_checksum = !(res->bind & (PIPE_BIND_SCANOUT | PIPE_BIND_SHARED));
 404         bool should_checksum = res->bind & PIPE_BIND_RENDER_TARGET;
 405
 406         pres->checksummed = can_checksum && should_checksum;
 407
 408         /* Set the layout appropriately */
 409         pres->layout = should_tile ? PAN_TILED : PAN_LINEAR;
 410
 411         size_t bo_size;
 412
 413         panfrost_setup_slices(pres, &bo_size);
 414
 415         /* We create a BO immediately but don't bother mapping, since we don't
 416          * care to map e.g. FBOs which the CPU probably won't touch */
 417         pres->bo = panfrost_bo_create(screen, bo_size, PAN_BO_DELAY_MMAP);
 418 }
 419
 420 void
 421 panfrost_resource_set_damage_region(struct pipe_screen *screen,
 422                                     struct pipe_resource *res,
 423                                     unsigned int nrects,
 424                                     const struct pipe_box *rects)
 425 {
 426         struct panfrost_resource *pres = pan_resource(res);
 427         struct pipe_box *damage_rect = &pres->damage.biggest_rect;
 428         struct pipe_scissor_state *damage_extent = &pres->damage.extent;
 429         unsigned int i;
 430
 431         if (!nrects) {
 432                 panfrost_resource_reset_damage(pres);
 433                 return;
 434         }
 435
 436         /* We keep track of 2 different things here:
 437          * 1 the damage extent: the quad including all damage regions. Will be
 438          *   used restrict the rendering area
 439          * 2 the biggest damage rectangle: when there are more than one damage
 440          *   rect we keep the biggest one and will generate 4 wallpaper quads
 441          *   out of it (see panfrost_draw_wallpaper() for more details). We
 442          *   might want to do something smarter at some point.
 443          *
 444          *                _________________________________
 445          *                |                               |
 446          *                |    _________________________  |
 447          *                |   | rect1|         _________| |
 448          *                |   |______|_____   | rect 3: | |
 449          *                |   |    | rect2 |  | biggest | |
 450          *                |   |    |_______|  |  rect   | |
 451          *                |   |_______________|_________| |
 452          *                |        damage extent          |
 453          *                |_______________________________|
 454          *                            resource
 455          */
 456         memset(&pres->damage, 0, sizeof(pres->damage));
 457         damage_extent->minx = 0xffff;
 458         damage_extent->miny = 0xffff;
 459         for (i = 0; i < nrects; i++) {
 460                 int x = rects[i].x, w = rects[i].width, h = rects[i].height;
 461                 int y = res->height0 - (rects[i].y + h);
 462
 463                 /* Clamp x,y,w,h to prevent negative values. */
 464                 if (x < 0) {
 465                         h += x;
 466                         x = 0;
 467                 }
 468                 if (y < 0) {
 469                         w += y;
 470                         y = 0;
 471                 }
 472                 w = MAX2(w, 0);
 473                 h = MAX2(h, 0);
 474
 475                 if (damage_rect->width * damage_rect->height < w * h)
 476                        u_box_2d(x, y, w, h, damage_rect);
 477
 478                 damage_extent->minx = MIN2(damage_extent->minx, x);
 479                 damage_extent->miny = MIN2(damage_extent->miny, y);
 480                 damage_extent->maxx = MAX2(damage_extent->maxx,
 481                                            MIN2(x + w, res->width0));
 482                 damage_extent->maxy = MAX2(damage_extent->maxy,
 483                                            MIN2(y + h, res->height0));
 484         }
 485 }
 486
 487 static struct pipe_resource *
 488 panfrost_resource_create(struct pipe_screen *screen,
 489                          const struct pipe_resource *template)
 490 {
 491         /* Make sure we're familiar */
 492         switch (template->target) {
 493         case PIPE_BUFFER:
 494         case PIPE_TEXTURE_1D:
 495         case PIPE_TEXTURE_2D:
 496         case PIPE_TEXTURE_3D:
 497         case PIPE_TEXTURE_CUBE:
 498         case PIPE_TEXTURE_RECT:
 499         case PIPE_TEXTURE_2D_ARRAY:
 500                 break;
 501         default:
 502                 DBG("Unknown texture target %d\n", template->target);
 503                 assert(0);
 504         }
 505
 506         if (template->bind &
 507             (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED))
 508                 return panfrost_create_scanout_res(screen, template);
 509
 510         struct panfrost_resource *so = rzalloc(screen, struct panfrost_resource);
 511         struct panfrost_screen *pscreen = (struct panfrost_screen *) screen;
 512
 513         so->base = *template;
 514         so->base.screen = screen;
 515
 516         pipe_reference_init(&so->base.reference, 1);
 517
 518         util_range_init(&so->valid_buffer_range);
 519
 520         panfrost_resource_create_bo(pscreen, so);
 521         panfrost_resource_reset_damage(so);
 522
 523         return (struct pipe_resource *)so;
 524 }
 525
 526 static void
 527 panfrost_resource_destroy(struct pipe_screen *screen,
 528                           struct pipe_resource *pt)
 529 {
 530         struct panfrost_screen *pscreen = pan_screen(screen);
 531         struct panfrost_resource *rsrc = (struct panfrost_resource *) pt;
 532
 533         if (rsrc->scanout)
 534                 renderonly_scanout_destroy(rsrc->scanout, pscreen->ro);
 535
 536         if (rsrc->bo)
 537                 panfrost_bo_unreference(rsrc->bo);
 538
 539         util_range_destroy(&rsrc->valid_buffer_range);
 540         ralloc_free(rsrc);
 541 }
 542
 543 static void *
 544 panfrost_transfer_map(struct pipe_context *pctx,
 545                       struct pipe_resource *resource,
 546                       unsigned level,
 547                       unsigned usage,  /* a combination of PIPE_TRANSFER_x */
 548                       const struct pipe_box *box,
 549                       struct pipe_transfer **out_transfer)
 550 {
 551         int bytes_per_pixel = util_format_get_blocksize(resource->format);
 552         struct panfrost_resource *rsrc = pan_resource(resource);
 553         struct panfrost_bo *bo = rsrc->bo;
 554
 555         struct panfrost_gtransfer *transfer = rzalloc(pctx, struct panfrost_gtransfer);
 556         transfer->base.level = level;
 557         transfer->base.usage = usage;
 558         transfer->base.box = *box;
 559
 560         pipe_resource_reference(&transfer->base.resource, resource);
 561
 562         *out_transfer = &transfer->base;
 563
 564         /* If we haven't already mmaped, now's the time */
 565         panfrost_bo_mmap(bo);
 566
 567         /* Check if we're bound for rendering and this is a read pixels. If so,
 568          * we need to flush */
 569
 570         struct panfrost_context *ctx = pan_context(pctx);
 571         struct pipe_framebuffer_state *fb = &ctx->pipe_framebuffer;
 572
 573         bool is_bound = false;
 574
 575         for (unsigned c = 0; c < fb->nr_cbufs; ++c) {
 576                 /* If cbufs is NULL, we're definitely not bound here */
 577
 578                 if (fb->cbufs[c])
 579                         is_bound |= fb->cbufs[c]->texture == resource;
 580         }
 581
 582         if (is_bound && (usage & PIPE_TRANSFER_READ))
 583                  assert(level == 0);
 584
 585         /* TODO: Respect usage flags */
 586
 587         if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 588                 /* If the BO is used by one of the pending batches or if it's
 589                  * not ready yet (still accessed by one of the already flushed
 590                  * batches), we try to allocate a new one to avoid waiting.
 591                  */
 592                 if (panfrost_pending_batches_access_bo(ctx, bo) ||
 593                     !panfrost_bo_wait(bo, 0, PAN_BO_ACCESS_RW)) {
 594                         struct panfrost_screen *screen = pan_screen(pctx->screen);
 595                         /* We want the BO to be MMAPed. */
 596                         uint32_t flags = bo->flags & ~PAN_BO_DELAY_MMAP;
 597                         struct panfrost_bo *newbo = NULL;
 598
 599                         /* When the BO has been imported/exported, we can't
 600                          * replace it by another one, otherwise the
 601                          * importer/exporter wouldn't see the change we're
 602                          * doing to it.
 603                          */
 604                         if (!(bo->flags & (PAN_BO_IMPORTED | PAN_BO_EXPORTED)))
 605                                 newbo = panfrost_bo_create(screen, bo->size,
 606                                                            flags);
 607
 608                         if (newbo) {
 609                                 panfrost_bo_unreference(bo);
 610                                 rsrc->bo = newbo;
 611                                 bo = newbo;
 612                         } else {
 613                                 uint32_t access = PAN_BO_ACCESS_RW;
 614
 615                                 /* Allocation failed or was impossible, let's
 616                                  * fall back on a flush+wait.
 617                                  */
 618                                 panfrost_flush_batches_accessing_bo(ctx, bo,
 619                                                                     access);
 620                                 panfrost_bo_wait(bo, INT64_MAX, access);
 621                         }
 622                 }
 623         } else if ((usage & PIPE_TRANSFER_WRITE)
 624                    && resource->target == PIPE_BUFFER
 625                    && !util_ranges_intersect(&rsrc->valid_buffer_range, box->x, box->x + box->width)) {
 626                 /* No flush for writes to uninitialized */
 627         } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
 628                 if (usage & PIPE_TRANSFER_WRITE) {
 629                         panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_RW);
 630                         panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_RW);
 631                 } else if (usage & PIPE_TRANSFER_READ) {
 632                         panfrost_flush_batches_accessing_bo(ctx, bo, PAN_BO_ACCESS_WRITE);
 633                         panfrost_bo_wait(bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
 634                 } else {
 635                         /* Why are you even mapping?! */
 636                 }
 637         }
 638
 639         if (rsrc->layout != PAN_LINEAR) {
 640                 /* Non-linear resources need to be indirectly mapped */
 641
 642                 if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
 643                         return NULL;
 644
 645                 transfer->base.stride = box->width * bytes_per_pixel;
 646                 transfer->base.layer_stride = transfer->base.stride * box->height;
 647                 transfer->map = rzalloc_size(transfer, transfer->base.layer_stride * box->depth);
 648                 assert(box->depth == 1);
 649
 650                 if ((usage & PIPE_TRANSFER_READ) && rsrc->slices[level].initialized) {
 651                         if (rsrc->layout == PAN_AFBC) {
 652                                 DBG("Unimplemented: reads from AFBC");
 653                         } else if (rsrc->layout == PAN_TILED) {
 654                                 panfrost_load_tiled_image(
 655                                         transfer->map,
 656                                         bo->cpu + rsrc->slices[level].offset,
 657                                         box,
 658                                         transfer->base.stride,
 659                                         rsrc->slices[level].stride,
 660                                         util_format_get_blocksize(resource->format));
 661                         }
 662                 }
 663
 664                 return transfer->map;
 665         } else {
 666                 transfer->base.stride = rsrc->slices[level].stride;
 667                 transfer->base.layer_stride = rsrc->cubemap_stride;
 668
 669                 /* By mapping direct-write, we're implicitly already
 670                  * initialized (maybe), so be conservative */
 671
 672                 if ((usage & PIPE_TRANSFER_WRITE) && (usage & PIPE_TRANSFER_MAP_DIRECTLY))
 673                         rsrc->slices[level].initialized = true;
 674
 675                 return bo->cpu
 676                        + rsrc->slices[level].offset
 677                        + transfer->base.box.z * rsrc->cubemap_stride
 678                        + transfer->base.box.y * rsrc->slices[level].stride
 679                        + transfer->base.box.x * bytes_per_pixel;
 680         }
 681 }
 682
 683 static void
 684 panfrost_transfer_unmap(struct pipe_context *pctx,
 685                         struct pipe_transfer *transfer)
 686 {
 687         /* Gallium expects writeback here, so we tile */
 688
 689         struct panfrost_gtransfer *trans = pan_transfer(transfer);
 690         struct panfrost_resource *prsrc = (struct panfrost_resource *) transfer->resource;
 691
 692         /* Mark whatever we wrote as written */
 693         if (transfer->usage & PIPE_TRANSFER_WRITE)
 694                 prsrc->slices[transfer->level].initialized = true;
 695
 696         if (trans->map) {
 697                 struct panfrost_bo *bo = prsrc->bo;
 698
 699                 if (transfer->usage & PIPE_TRANSFER_WRITE) {
 700                         if (prsrc->layout == PAN_AFBC) {
 701                                 DBG("Unimplemented: writes to AFBC\n");
 702                         } else if (prsrc->layout == PAN_TILED) {
 703                                 assert(transfer->box.depth == 1);
 704
 705                                 panfrost_store_tiled_image(
 706                                         bo->cpu + prsrc->slices[transfer->level].offset,
 707                                         trans->map,
 708                                         &transfer->box,
 709                                         prsrc->slices[transfer->level].stride,
 710                                         transfer->stride,
 711                                         util_format_get_blocksize(prsrc->base.format));
 712                         }
 713                 }
 714         }
 715
 716
 717         util_range_add(&prsrc->base, &prsrc->valid_buffer_range,
 718                        transfer->box.x,
 719                        transfer->box.x + transfer->box.width);
 720
 721         /* Derefence the resource */
 722         pipe_resource_reference(&transfer->resource, NULL);
 723
 724         /* Transfer itself is RALLOCed at the moment */
 725         ralloc_free(transfer);
 726 }
 727
 728 static void
 729 panfrost_transfer_flush_region(struct pipe_context *pctx,
 730                                struct pipe_transfer *transfer,
 731                                const struct pipe_box *box)
 732 {
 733         struct panfrost_resource *rsc = pan_resource(transfer->resource);
 734
 735         if (transfer->resource->target == PIPE_BUFFER) {
 736                 util_range_add(&rsc->base, &rsc->valid_buffer_range,
 737                                transfer->box.x + box->x,
 738                                transfer->box.x + box->x + box->width);
 739         } else {
 740                 unsigned level = transfer->level;
 741                 rsc->slices[level].initialized = true;
 742         }
 743 }
 744
 745 static void
 746 panfrost_invalidate_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 747 {
 748         //DBG("TODO %s\n", __func__);
 749 }
 750
 751 static enum pipe_format
 752 panfrost_resource_get_internal_format(struct pipe_resource *prsrc) {
 753         return prsrc->format;
 754 }
 755
 756 static bool
 757 panfrost_generate_mipmap(
 758         struct pipe_context *pctx,
 759         struct pipe_resource *prsrc,
 760         enum pipe_format format,
 761         unsigned base_level,
 762         unsigned last_level,
 763         unsigned first_layer,
 764         unsigned last_layer)
 765 {
 766         struct panfrost_context *ctx = pan_context(pctx);
 767         struct panfrost_resource *rsrc = pan_resource(prsrc);
 768
 769         /* Generating a mipmap invalidates the written levels, so make that
 770          * explicit so we don't try to wallpaper them back and end up with
 771          * u_blitter recursion */
 772
 773         assert(rsrc->bo);
 774         for (unsigned l = base_level + 1; l <= last_level; ++l)
 775                 rsrc->slices[l].initialized = false;
 776
 777         /* Beyond that, we just delegate the hard stuff. We're careful to
 778          * include flushes on both ends to make sure the data is really valid.
 779          * We could be doing a lot better perf-wise, especially once we have
 780          * reorder-type optimizations in place. But for now prioritize
 781          * correctness. */
 782
 783         panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_RW);
 784         panfrost_bo_wait(rsrc->bo, INT64_MAX, PAN_BO_ACCESS_RW);
 785
 786         /* We've flushed the original buffer if needed, now trigger a blit */
 787
 788         bool blit_res = util_gen_mipmap(
 789                                 pctx, prsrc, format,
 790                                 base_level, last_level,
 791                                 first_layer, last_layer,
 792                                 PIPE_TEX_FILTER_LINEAR);
 793
 794         /* If the blit was successful, flush once more. If it wasn't, well, let
 795          * the state tracker deal with it. */
 796
 797         if (blit_res) {
 798                 panfrost_flush_batches_accessing_bo(ctx, rsrc->bo, PAN_BO_ACCESS_WRITE);
 799                 panfrost_bo_wait(rsrc->bo, INT64_MAX, PAN_BO_ACCESS_WRITE);
 800         }
 801
 802         return blit_res;
 803 }
 804
 805 /* Computes the address to a texture at a particular slice */
 806
 807 mali_ptr
 808 panfrost_get_texture_address(
 809         struct panfrost_resource *rsrc,
 810         unsigned level, unsigned face)
 811 {
 812         unsigned level_offset = rsrc->slices[level].offset;
 813         unsigned face_offset = face * rsrc->cubemap_stride;
 814
 815         return rsrc->bo->gpu + level_offset + face_offset;
 816 }
 817
 818 /* Given a resource that has already been allocated, hint that it should use a
 819  * given layout. These are suggestions, not commands; it is perfectly legal to
 820  * stub out this function, but there will be performance implications. */
 821
 822 void
 823 panfrost_resource_hint_layout(
 824                 struct panfrost_screen *screen,
 825                 struct panfrost_resource *rsrc,
 826                 enum panfrost_memory_layout layout,
 827                 signed weight)
 828 {
 829         /* Nothing to do, although a sophisticated implementation might store
 830          * the hint */
 831
 832         if (rsrc->layout == layout)
 833                 return;
 834
 835         /* We don't use the weight yet, but we should check that it's positive
 836          * (semantically meaning that we should choose the given `layout`) */
 837
 838         if (weight <= 0)
 839                 return;
 840
 841         /* Check if the preferred layout is legal for this buffer */
 842
 843         if (layout == PAN_AFBC) {
 844                 bool can_afbc = panfrost_format_supports_afbc(rsrc->base.format);
 845                 bool is_scanout = rsrc->base.bind &
 846                         (PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT | PIPE_BIND_SHARED);
 847
 848                 if (!can_afbc || is_scanout)
 849                         return;
 850         }
 851
 852         /* Simple heuristic so far: if the resource is uninitialized, switch to
 853          * the hinted layout. If it is initialized, keep the original layout.
 854          * This misses some cases where it would be beneficial to switch and
 855          * blit. */
 856
 857         bool is_initialized = false;
 858
 859         for (unsigned i = 0; i < MAX_MIP_LEVELS; ++i)
 860                 is_initialized |= rsrc->slices[i].initialized;
 861
 862         if (is_initialized)
 863                 return;
 864
 865         /* We're uninitialized, so do a layout switch. Reinitialize slices. */
 866
 867         size_t new_size;
 868         rsrc->layout = layout;
 869         panfrost_setup_slices(rsrc, &new_size);
 870
 871         /* If we grew in size, reallocate the BO */
 872         if (new_size > rsrc->bo->size) {
 873                 panfrost_bo_unreference(rsrc->bo);
 874                 rsrc->bo = panfrost_bo_create(screen, new_size, PAN_BO_DELAY_MMAP);
 875         }
 876 }
 877
 878 static void
 879 panfrost_resource_set_stencil(struct pipe_resource *prsrc,
 880                               struct pipe_resource *stencil)
 881 {
 882         pan_resource(prsrc)->separate_stencil = pan_resource(stencil);
 883 }
 884
 885 static struct pipe_resource *
 886 panfrost_resource_get_stencil(struct pipe_resource *prsrc)
 887 {
 888         return &pan_resource(prsrc)->separate_stencil->base;
 889 }
 890
 891 static const struct u_transfer_vtbl transfer_vtbl = {
 892         .resource_create          = panfrost_resource_create,
 893         .resource_destroy         = panfrost_resource_destroy,
 894         .transfer_map             = panfrost_transfer_map,
 895         .transfer_unmap           = panfrost_transfer_unmap,
 896         .transfer_flush_region    = panfrost_transfer_flush_region,
 897         .get_internal_format      = panfrost_resource_get_internal_format,
 898         .set_stencil              = panfrost_resource_set_stencil,
 899         .get_stencil              = panfrost_resource_get_stencil,
 900 };
 901
 902 void
 903 panfrost_resource_screen_init(struct panfrost_screen *pscreen)
 904 {
 905         //pscreen->base.resource_create_with_modifiers =
 906         //        panfrost_resource_create_with_modifiers;
 907         pscreen->base.resource_create = u_transfer_helper_resource_create;
 908         pscreen->base.resource_destroy = u_transfer_helper_resource_destroy;
 909         pscreen->base.resource_from_handle = panfrost_resource_from_handle;
 910         pscreen->base.resource_get_handle = panfrost_resource_get_handle;
 911         pscreen->base.transfer_helper = u_transfer_helper_create(&transfer_vtbl,
 912                                         true, false,
 913                                         true, true);
 914 }
 915
 916 void
 917 panfrost_resource_context_init(struct pipe_context *pctx)
 918 {
 919         pctx->transfer_map = u_transfer_helper_transfer_map;
 920         pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
 921         pctx->create_surface = panfrost_create_surface;
 922         pctx->surface_destroy = panfrost_surface_destroy;
 923         pctx->resource_copy_region = util_resource_copy_region;
 924         pctx->blit = panfrost_blit;
 925         pctx->generate_mipmap = panfrost_generate_mipmap;
 926         pctx->flush_resource = panfrost_flush_resource;
 927         pctx->invalidate_resource = panfrost_invalidate_resource;
 928         pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
 929         pctx->buffer_subdata = u_default_buffer_subdata;
 930         pctx->texture_subdata = u_default_texture_subdata;
 931 }