src/gallium/drivers/radeon/r600_buffer_common.c

   1 /*
   2  * Copyright 2013 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "radeonsi/si_pipe.h"
  25 #include "r600_cs.h"
  26 #include "util/u_memory.h"
  27 #include "util/u_upload_mgr.h"
  28 #include <inttypes.h>
  29 #include <stdio.h>
  30
  31 bool si_rings_is_buffer_referenced(struct r600_common_context *ctx,
  32                                    struct pb_buffer *buf,
  33                                    enum radeon_bo_usage usage)
  34 {
  35         if (ctx->ws->cs_is_buffer_referenced(ctx->gfx_cs, buf, usage)) {
  36                 return true;
  37         }
  38         if (radeon_emitted(ctx->dma_cs, 0) &&
  39             ctx->ws->cs_is_buffer_referenced(ctx->dma_cs, buf, usage)) {
  40                 return true;
  41         }
  42         return false;
  43 }
  44
  45 void *si_buffer_map_sync_with_rings(struct r600_common_context *ctx,
  46                                     struct r600_resource *resource,
  47                                     unsigned usage)
  48 {
  49         enum radeon_bo_usage rusage = RADEON_USAGE_READWRITE;
  50         bool busy = false;
  51
  52         assert(!(resource->flags & RADEON_FLAG_SPARSE));
  53
  54         if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
  55                 return ctx->ws->buffer_map(resource->buf, NULL, usage);
  56         }
  57
  58         if (!(usage & PIPE_TRANSFER_WRITE)) {
  59                 /* have to wait for the last write */
  60                 rusage = RADEON_USAGE_WRITE;
  61         }
  62
  63         if (radeon_emitted(ctx->gfx_cs, ctx->initial_gfx_cs_size) &&
  64             ctx->ws->cs_is_buffer_referenced(ctx->gfx_cs,
  65                                              resource->buf, rusage)) {
  66                 if (usage & PIPE_TRANSFER_DONTBLOCK) {
  67                         si_flush_gfx_cs(ctx, PIPE_FLUSH_ASYNC, NULL);
  68                         return NULL;
  69                 } else {
  70                         si_flush_gfx_cs(ctx, 0, NULL);
  71                         busy = true;
  72                 }
  73         }
  74         if (radeon_emitted(ctx->dma_cs, 0) &&
  75             ctx->ws->cs_is_buffer_referenced(ctx->dma_cs,
  76                                              resource->buf, rusage)) {
  77                 if (usage & PIPE_TRANSFER_DONTBLOCK) {
  78                         si_flush_dma_cs(ctx, PIPE_FLUSH_ASYNC, NULL);
  79                         return NULL;
  80                 } else {
  81                         si_flush_dma_cs(ctx, 0, NULL);
  82                         busy = true;
  83                 }
  84         }
  85
  86         if (busy || !ctx->ws->buffer_wait(resource->buf, 0, rusage)) {
  87                 if (usage & PIPE_TRANSFER_DONTBLOCK) {
  88                         return NULL;
  89                 } else {
  90                         /* We will be wait for the GPU. Wait for any offloaded
  91                          * CS flush to complete to avoid busy-waiting in the winsys. */
  92                         ctx->ws->cs_sync_flush(ctx->gfx_cs);
  93                         if (ctx->dma_cs)
  94                                 ctx->ws->cs_sync_flush(ctx->dma_cs);
  95                 }
  96         }
  97
  98         /* Setting the CS to NULL will prevent doing checks we have done already. */
  99         return ctx->ws->buffer_map(resource->buf, NULL, usage);
 100 }
 101
 102 void si_init_resource_fields(struct si_screen *sscreen,
 103                              struct r600_resource *res,
 104                              uint64_t size, unsigned alignment)
 105 {
 106         struct r600_texture *rtex = (struct r600_texture*)res;
 107
 108         res->bo_size = size;
 109         res->bo_alignment = alignment;
 110         res->flags = 0;
 111         res->texture_handle_allocated = false;
 112         res->image_handle_allocated = false;
 113
 114         switch (res->b.b.usage) {
 115         case PIPE_USAGE_STREAM:
 116                 res->flags = RADEON_FLAG_GTT_WC;
 117                 /* fall through */
 118         case PIPE_USAGE_STAGING:
 119                 /* Transfers are likely to occur more often with these
 120                  * resources. */
 121                 res->domains = RADEON_DOMAIN_GTT;
 122                 break;
 123         case PIPE_USAGE_DYNAMIC:
 124                 /* Older kernels didn't always flush the HDP cache before
 125                  * CS execution
 126                  */
 127                 if (sscreen->info.drm_major == 2 &&
 128                     sscreen->info.drm_minor < 40) {
 129                         res->domains = RADEON_DOMAIN_GTT;
 130                         res->flags |= RADEON_FLAG_GTT_WC;
 131                         break;
 132                 }
 133                 /* fall through */
 134         case PIPE_USAGE_DEFAULT:
 135         case PIPE_USAGE_IMMUTABLE:
 136         default:
 137                 /* Not listing GTT here improves performance in some
 138                  * apps. */
 139                 res->domains = RADEON_DOMAIN_VRAM;
 140                 res->flags |= RADEON_FLAG_GTT_WC;
 141                 break;
 142         }
 143
 144         if (res->b.b.target == PIPE_BUFFER &&
 145             res->b.b.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT |
 146                               PIPE_RESOURCE_FLAG_MAP_COHERENT)) {
 147                 /* Use GTT for all persistent mappings with older
 148                  * kernels, because they didn't always flush the HDP
 149                  * cache before CS execution.
 150                  *
 151                  * Write-combined CPU mappings are fine, the kernel
 152                  * ensures all CPU writes finish before the GPU
 153                  * executes a command stream.
 154                  */
 155                 if (sscreen->info.drm_major == 2 &&
 156                     sscreen->info.drm_minor < 40)
 157                         res->domains = RADEON_DOMAIN_GTT;
 158         }
 159
 160         /* Tiled textures are unmappable. Always put them in VRAM. */
 161         if ((res->b.b.target != PIPE_BUFFER && !rtex->surface.is_linear) ||
 162             res->b.b.flags & R600_RESOURCE_FLAG_UNMAPPABLE) {
 163                 res->domains = RADEON_DOMAIN_VRAM;
 164                 res->flags |= RADEON_FLAG_NO_CPU_ACCESS |
 165                          RADEON_FLAG_GTT_WC;
 166         }
 167
 168         /* Displayable and shareable surfaces are not suballocated. */
 169         if (res->b.b.bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT))
 170                 res->flags |= RADEON_FLAG_NO_SUBALLOC; /* shareable */
 171         else
 172                 res->flags |= RADEON_FLAG_NO_INTERPROCESS_SHARING;
 173
 174         if (sscreen->debug_flags & DBG(NO_WC))
 175                 res->flags &= ~RADEON_FLAG_GTT_WC;
 176
 177         if (res->b.b.flags & R600_RESOURCE_FLAG_READ_ONLY)
 178                 res->flags |= RADEON_FLAG_READ_ONLY;
 179
 180         if (res->b.b.flags & R600_RESOURCE_FLAG_32BIT)
 181                 res->flags |= RADEON_FLAG_32BIT;
 182
 183         /* Set expected VRAM and GART usage for the buffer. */
 184         res->vram_usage = 0;
 185         res->gart_usage = 0;
 186         res->max_forced_staging_uploads = 0;
 187         res->b.max_forced_staging_uploads = 0;
 188
 189         if (res->domains & RADEON_DOMAIN_VRAM) {
 190                 res->vram_usage = size;
 191
 192                 res->max_forced_staging_uploads =
 193                 res->b.max_forced_staging_uploads =
 194                         sscreen->info.has_dedicated_vram &&
 195                         size >= sscreen->info.vram_vis_size / 4 ? 1 : 0;
 196         } else if (res->domains & RADEON_DOMAIN_GTT) {
 197                 res->gart_usage = size;
 198         }
 199 }
 200
 201 bool si_alloc_resource(struct si_screen *sscreen,
 202                        struct r600_resource *res)
 203 {
 204         struct pb_buffer *old_buf, *new_buf;
 205
 206         /* Allocate a new resource. */
 207         new_buf = sscreen->ws->buffer_create(sscreen->ws, res->bo_size,
 208                                              res->bo_alignment,
 209                                              res->domains, res->flags);
 210         if (!new_buf) {
 211                 return false;
 212         }
 213
 214         /* Replace the pointer such that if res->buf wasn't NULL, it won't be
 215          * NULL. This should prevent crashes with multiple contexts using
 216          * the same buffer where one of the contexts invalidates it while
 217          * the others are using it. */
 218         old_buf = res->buf;
 219         res->buf = new_buf; /* should be atomic */
 220         res->gpu_address = sscreen->ws->buffer_get_virtual_address(res->buf);
 221
 222         if (res->flags & RADEON_FLAG_32BIT) {
 223                 uint64_t start = res->gpu_address;
 224                 uint64_t last = start + res->bo_size - 1;
 225                 (void)start;
 226                 (void)last;
 227
 228                 assert((start >> 32) == sscreen->info.address32_hi);
 229                 assert((last >> 32) == sscreen->info.address32_hi);
 230         }
 231
 232         pb_reference(&old_buf, NULL);
 233
 234         util_range_set_empty(&res->valid_buffer_range);
 235         res->TC_L2_dirty = false;
 236
 237         /* Print debug information. */
 238         if (sscreen->debug_flags & DBG(VM) && res->b.b.target == PIPE_BUFFER) {
 239                 fprintf(stderr, "VM start=0x%"PRIX64"  end=0x%"PRIX64" | Buffer %"PRIu64" bytes\n",
 240                         res->gpu_address, res->gpu_address + res->buf->size,
 241                         res->buf->size);
 242         }
 243         return true;
 244 }
 245
 246 static void r600_buffer_destroy(struct pipe_screen *screen,
 247                                 struct pipe_resource *buf)
 248 {
 249         struct r600_resource *rbuffer = r600_resource(buf);
 250
 251         threaded_resource_deinit(buf);
 252         util_range_destroy(&rbuffer->valid_buffer_range);
 253         pb_reference(&rbuffer->buf, NULL);
 254         FREE(rbuffer);
 255 }
 256
 257 /* Reallocate the buffer a update all resource bindings where the buffer is
 258  * bound.
 259  *
 260  * This is used to avoid CPU-GPU synchronizations, because it makes the buffer
 261  * idle by discarding its contents.
 262  */
 263 static bool
 264 r600_invalidate_buffer(struct r600_common_context *rctx,
 265                        struct r600_resource *rbuffer)
 266 {
 267         /* Shared buffers can't be reallocated. */
 268         if (rbuffer->b.is_shared)
 269                 return false;
 270
 271         /* Sparse buffers can't be reallocated. */
 272         if (rbuffer->flags & RADEON_FLAG_SPARSE)
 273                 return false;
 274
 275         /* In AMD_pinned_memory, the user pointer association only gets
 276          * broken when the buffer is explicitly re-allocated.
 277          */
 278         if (rbuffer->b.is_user_ptr)
 279                 return false;
 280
 281         /* Check if mapping this buffer would cause waiting for the GPU. */
 282         if (si_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
 283             !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) {
 284                 uint64_t old_va = rbuffer->gpu_address;
 285
 286                 /* Reallocate the buffer in the same pipe_resource. */
 287                 si_alloc_resource(rctx->screen, rbuffer);
 288                 si_rebind_buffer(&rctx->b, &rbuffer->b.b, old_va);
 289         } else {
 290                 util_range_set_empty(&rbuffer->valid_buffer_range);
 291         }
 292
 293         return true;
 294 }
 295
 296 /* Replace the storage of dst with src. */
 297 void si_replace_buffer_storage(struct pipe_context *ctx,
 298                                  struct pipe_resource *dst,
 299                                  struct pipe_resource *src)
 300 {
 301         struct r600_resource *rdst = r600_resource(dst);
 302         struct r600_resource *rsrc = r600_resource(src);
 303         uint64_t old_gpu_address = rdst->gpu_address;
 304
 305         pb_reference(&rdst->buf, rsrc->buf);
 306         rdst->gpu_address = rsrc->gpu_address;
 307         rdst->b.b.bind = rsrc->b.b.bind;
 308         rdst->b.max_forced_staging_uploads = rsrc->b.max_forced_staging_uploads;
 309         rdst->max_forced_staging_uploads = rsrc->max_forced_staging_uploads;
 310         rdst->flags = rsrc->flags;
 311
 312         assert(rdst->vram_usage == rsrc->vram_usage);
 313         assert(rdst->gart_usage == rsrc->gart_usage);
 314         assert(rdst->bo_size == rsrc->bo_size);
 315         assert(rdst->bo_alignment == rsrc->bo_alignment);
 316         assert(rdst->domains == rsrc->domains);
 317
 318         si_rebind_buffer(ctx, dst, old_gpu_address);
 319 }
 320
 321 static void si_invalidate_resource(struct pipe_context *ctx,
 322                                    struct pipe_resource *resource)
 323 {
 324         struct r600_common_context *rctx = (struct r600_common_context*)ctx;
 325         struct r600_resource *rbuffer = r600_resource(resource);
 326
 327         /* We currently only do anyting here for buffers */
 328         if (resource->target == PIPE_BUFFER)
 329                 (void)r600_invalidate_buffer(rctx, rbuffer);
 330 }
 331
 332 static void *r600_buffer_get_transfer(struct pipe_context *ctx,
 333                                       struct pipe_resource *resource,
 334                                       unsigned usage,
 335                                       const struct pipe_box *box,
 336                                       struct pipe_transfer **ptransfer,
 337                                       void *data, struct r600_resource *staging,
 338                                       unsigned offset)
 339 {
 340         struct r600_common_context *rctx = (struct r600_common_context*)ctx;
 341         struct r600_transfer *transfer;
 342
 343         if (usage & TC_TRANSFER_MAP_THREADED_UNSYNC)
 344                 transfer = slab_alloc(&rctx->pool_transfers_unsync);
 345         else
 346                 transfer = slab_alloc(&rctx->pool_transfers);
 347
 348         transfer->b.b.resource = NULL;
 349         pipe_resource_reference(&transfer->b.b.resource, resource);
 350         transfer->b.b.level = 0;
 351         transfer->b.b.usage = usage;
 352         transfer->b.b.box = *box;
 353         transfer->b.b.stride = 0;
 354         transfer->b.b.layer_stride = 0;
 355         transfer->b.staging = NULL;
 356         transfer->offset = offset;
 357         transfer->staging = staging;
 358         *ptransfer = &transfer->b.b;
 359         return data;
 360 }
 361
 362 static void *r600_buffer_transfer_map(struct pipe_context *ctx,
 363                                       struct pipe_resource *resource,
 364                                       unsigned level,
 365                                       unsigned usage,
 366                                       const struct pipe_box *box,
 367                                       struct pipe_transfer **ptransfer)
 368 {
 369         struct r600_common_context *rctx = (struct r600_common_context*)ctx;
 370         struct r600_resource *rbuffer = r600_resource(resource);
 371         uint8_t *data;
 372
 373         assert(box->x + box->width <= resource->width0);
 374
 375         /* From GL_AMD_pinned_memory issues:
 376          *
 377          *     4) Is glMapBuffer on a shared buffer guaranteed to return the
 378          *        same system address which was specified at creation time?
 379          *
 380          *        RESOLVED: NO. The GL implementation might return a different
 381          *        virtual mapping of that memory, although the same physical
 382          *        page will be used.
 383          *
 384          * So don't ever use staging buffers.
 385          */
 386         if (rbuffer->b.is_user_ptr)
 387                 usage |= PIPE_TRANSFER_PERSISTENT;
 388
 389         /* See if the buffer range being mapped has never been initialized,
 390          * in which case it can be mapped unsynchronized. */
 391         if (!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
 392                        TC_TRANSFER_MAP_NO_INFER_UNSYNCHRONIZED)) &&
 393             usage & PIPE_TRANSFER_WRITE &&
 394             !rbuffer->b.is_shared &&
 395             !util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
 396                 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
 397         }
 398
 399         /* If discarding the entire range, discard the whole resource instead. */
 400         if (usage & PIPE_TRANSFER_DISCARD_RANGE &&
 401             box->x == 0 && box->width == resource->width0) {
 402                 usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
 403         }
 404
 405         /* If a buffer in VRAM is too large and the range is discarded, don't
 406          * map it directly. This makes sure that the buffer stays in VRAM.
 407          */
 408         bool force_discard_range = false;
 409         if (usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
 410                      PIPE_TRANSFER_DISCARD_RANGE) &&
 411             !(usage & PIPE_TRANSFER_PERSISTENT) &&
 412             /* Try not to decrement the counter if it's not positive. Still racy,
 413              * but it makes it harder to wrap the counter from INT_MIN to INT_MAX. */
 414             rbuffer->max_forced_staging_uploads > 0 &&
 415             p_atomic_dec_return(&rbuffer->max_forced_staging_uploads) >= 0) {
 416                 usage &= ~(PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
 417                            PIPE_TRANSFER_UNSYNCHRONIZED);
 418                 usage |= PIPE_TRANSFER_DISCARD_RANGE;
 419                 force_discard_range = true;
 420         }
 421
 422         if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
 423             !(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
 424                        TC_TRANSFER_MAP_NO_INVALIDATE))) {
 425                 assert(usage & PIPE_TRANSFER_WRITE);
 426
 427                 if (r600_invalidate_buffer(rctx, rbuffer)) {
 428                         /* At this point, the buffer is always idle. */
 429                         usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
 430                 } else {
 431                         /* Fall back to a temporary buffer. */
 432                         usage |= PIPE_TRANSFER_DISCARD_RANGE;
 433                 }
 434         }
 435
 436         if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
 437             ((!(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
 438                          PIPE_TRANSFER_PERSISTENT))) ||
 439              (rbuffer->flags & RADEON_FLAG_SPARSE))) {
 440                 assert(usage & PIPE_TRANSFER_WRITE);
 441
 442                 /* Check if mapping this buffer would cause waiting for the GPU.
 443                  */
 444                 if (rbuffer->flags & RADEON_FLAG_SPARSE ||
 445                     force_discard_range ||
 446                     si_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
 447                     !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) {
 448                         /* Do a wait-free write-only transfer using a temporary buffer. */
 449                         unsigned offset;
 450                         struct r600_resource *staging = NULL;
 451
 452                         u_upload_alloc(ctx->stream_uploader, 0,
 453                                        box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
 454                                        rctx->screen->info.tcc_cache_line_size,
 455                                        &offset, (struct pipe_resource**)&staging,
 456                                        (void**)&data);
 457
 458                         if (staging) {
 459                                 data += box->x % R600_MAP_BUFFER_ALIGNMENT;
 460                                 return r600_buffer_get_transfer(ctx, resource, usage, box,
 461                                                                 ptransfer, data, staging, offset);
 462                         } else if (rbuffer->flags & RADEON_FLAG_SPARSE) {
 463                                 return NULL;
 464                         }
 465                 } else {
 466                         /* At this point, the buffer is always idle (we checked it above). */
 467                         usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
 468                 }
 469         }
 470         /* Use a staging buffer in cached GTT for reads. */
 471         else if (((usage & PIPE_TRANSFER_READ) &&
 472                   !(usage & PIPE_TRANSFER_PERSISTENT) &&
 473                   (rbuffer->domains & RADEON_DOMAIN_VRAM ||
 474                    rbuffer->flags & RADEON_FLAG_GTT_WC)) ||
 475                  (rbuffer->flags & RADEON_FLAG_SPARSE)) {
 476                 struct r600_resource *staging;
 477
 478                 assert(!(usage & TC_TRANSFER_MAP_THREADED_UNSYNC));
 479                 staging = (struct r600_resource*) pipe_buffer_create(
 480                                 ctx->screen, 0, PIPE_USAGE_STAGING,
 481                                 box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT));
 482                 if (staging) {
 483                         /* Copy the VRAM buffer to the staging buffer. */
 484                         rctx->dma_copy(ctx, &staging->b.b, 0,
 485                                        box->x % R600_MAP_BUFFER_ALIGNMENT,
 486                                        0, 0, resource, 0, box);
 487
 488                         data = si_buffer_map_sync_with_rings(rctx, staging,
 489                                                                usage & ~PIPE_TRANSFER_UNSYNCHRONIZED);
 490                         if (!data) {
 491                                 r600_resource_reference(&staging, NULL);
 492                                 return NULL;
 493                         }
 494                         data += box->x % R600_MAP_BUFFER_ALIGNMENT;
 495
 496                         return r600_buffer_get_transfer(ctx, resource, usage, box,
 497                                                         ptransfer, data, staging, 0);
 498                 } else if (rbuffer->flags & RADEON_FLAG_SPARSE) {
 499                         return NULL;
 500                 }
 501         }
 502
 503         data = si_buffer_map_sync_with_rings(rctx, rbuffer, usage);
 504         if (!data) {
 505                 return NULL;
 506         }
 507         data += box->x;
 508
 509         return r600_buffer_get_transfer(ctx, resource, usage, box,
 510                                         ptransfer, data, NULL, 0);
 511 }
 512
 513 static void r600_buffer_do_flush_region(struct pipe_context *ctx,
 514                                         struct pipe_transfer *transfer,
 515                                         const struct pipe_box *box)
 516 {
 517         struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
 518         struct r600_resource *rbuffer = r600_resource(transfer->resource);
 519
 520         if (rtransfer->staging) {
 521                 struct pipe_resource *dst, *src;
 522                 unsigned soffset;
 523                 struct pipe_box dma_box;
 524
 525                 dst = transfer->resource;
 526                 src = &rtransfer->staging->b.b;
 527                 soffset = rtransfer->offset + box->x % R600_MAP_BUFFER_ALIGNMENT;
 528
 529                 u_box_1d(soffset, box->width, &dma_box);
 530
 531                 /* Copy the staging buffer into the original one. */
 532                 ctx->resource_copy_region(ctx, dst, 0, box->x, 0, 0, src, 0, &dma_box);
 533         }
 534
 535         util_range_add(&rbuffer->valid_buffer_range, box->x,
 536                        box->x + box->width);
 537 }
 538
 539 static void r600_buffer_flush_region(struct pipe_context *ctx,
 540                                      struct pipe_transfer *transfer,
 541                                      const struct pipe_box *rel_box)
 542 {
 543         unsigned required_usage = PIPE_TRANSFER_WRITE |
 544                                   PIPE_TRANSFER_FLUSH_EXPLICIT;
 545
 546         if ((transfer->usage & required_usage) == required_usage) {
 547                 struct pipe_box box;
 548
 549                 u_box_1d(transfer->box.x + rel_box->x, rel_box->width, &box);
 550                 r600_buffer_do_flush_region(ctx, transfer, &box);
 551         }
 552 }
 553
 554 static void r600_buffer_transfer_unmap(struct pipe_context *ctx,
 555                                        struct pipe_transfer *transfer)
 556 {
 557         struct r600_common_context *rctx = (struct r600_common_context*)ctx;
 558         struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
 559
 560         if (transfer->usage & PIPE_TRANSFER_WRITE &&
 561             !(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT))
 562                 r600_buffer_do_flush_region(ctx, transfer, &transfer->box);
 563
 564         r600_resource_reference(&rtransfer->staging, NULL);
 565         assert(rtransfer->b.staging == NULL); /* for threaded context only */
 566         pipe_resource_reference(&transfer->resource, NULL);
 567
 568         /* Don't use pool_transfers_unsync. We are always in the driver
 569          * thread. */
 570         slab_free(&rctx->pool_transfers, transfer);
 571 }
 572
 573 static void si_buffer_subdata(struct pipe_context *ctx,
 574                               struct pipe_resource *buffer,
 575                               unsigned usage, unsigned offset,
 576                               unsigned size, const void *data)
 577 {
 578         struct pipe_transfer *transfer = NULL;
 579         struct pipe_box box;
 580         uint8_t *map = NULL;
 581
 582         u_box_1d(offset, size, &box);
 583         map = r600_buffer_transfer_map(ctx, buffer, 0,
 584                                        PIPE_TRANSFER_WRITE |
 585                                        PIPE_TRANSFER_DISCARD_RANGE |
 586                                        usage,
 587                                        &box, &transfer);
 588         if (!map)
 589                 return;
 590
 591         memcpy(map, data, size);
 592         r600_buffer_transfer_unmap(ctx, transfer);
 593 }
 594
 595 static const struct u_resource_vtbl r600_buffer_vtbl =
 596 {
 597         NULL,                           /* get_handle */
 598         r600_buffer_destroy,            /* resource_destroy */
 599         r600_buffer_transfer_map,       /* transfer_map */
 600         r600_buffer_flush_region,       /* transfer_flush_region */
 601         r600_buffer_transfer_unmap,     /* transfer_unmap */
 602 };
 603
 604 static struct r600_resource *
 605 r600_alloc_buffer_struct(struct pipe_screen *screen,
 606                          const struct pipe_resource *templ)
 607 {
 608         struct r600_resource *rbuffer;
 609
 610         rbuffer = MALLOC_STRUCT(r600_resource);
 611
 612         rbuffer->b.b = *templ;
 613         rbuffer->b.b.next = NULL;
 614         pipe_reference_init(&rbuffer->b.b.reference, 1);
 615         rbuffer->b.b.screen = screen;
 616
 617         rbuffer->b.vtbl = &r600_buffer_vtbl;
 618         threaded_resource_init(&rbuffer->b.b);
 619
 620         rbuffer->buf = NULL;
 621         rbuffer->bind_history = 0;
 622         rbuffer->TC_L2_dirty = false;
 623         util_range_init(&rbuffer->valid_buffer_range);
 624         return rbuffer;
 625 }
 626
 627 static struct pipe_resource *si_buffer_create(struct pipe_screen *screen,
 628                                               const struct pipe_resource *templ,
 629                                               unsigned alignment)
 630 {
 631         struct si_screen *sscreen = (struct si_screen*)screen;
 632         struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
 633
 634         if (templ->flags & PIPE_RESOURCE_FLAG_SPARSE)
 635                 rbuffer->b.b.flags |= R600_RESOURCE_FLAG_UNMAPPABLE;
 636
 637         si_init_resource_fields(sscreen, rbuffer, templ->width0, alignment);
 638
 639         if (templ->flags & PIPE_RESOURCE_FLAG_SPARSE)
 640                 rbuffer->flags |= RADEON_FLAG_SPARSE;
 641
 642         if (!si_alloc_resource(sscreen, rbuffer)) {
 643                 FREE(rbuffer);
 644                 return NULL;
 645         }
 646         return &rbuffer->b.b;
 647 }
 648
 649 struct pipe_resource *si_aligned_buffer_create(struct pipe_screen *screen,
 650                                                unsigned flags,
 651                                                unsigned usage,
 652                                                unsigned size,
 653                                                unsigned alignment)
 654 {
 655         struct pipe_resource buffer;
 656
 657         memset(&buffer, 0, sizeof buffer);
 658         buffer.target = PIPE_BUFFER;
 659         buffer.format = PIPE_FORMAT_R8_UNORM;
 660         buffer.bind = 0;
 661         buffer.usage = usage;
 662         buffer.flags = flags;
 663         buffer.width0 = size;
 664         buffer.height0 = 1;
 665         buffer.depth0 = 1;
 666         buffer.array_size = 1;
 667         return si_buffer_create(screen, &buffer, alignment);
 668 }
 669
 670 static struct pipe_resource *
 671 si_buffer_from_user_memory(struct pipe_screen *screen,
 672                            const struct pipe_resource *templ,
 673                            void *user_memory)
 674 {
 675         struct si_screen *sscreen = (struct si_screen*)screen;
 676         struct radeon_winsys *ws = sscreen->ws;
 677         struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
 678
 679         rbuffer->domains = RADEON_DOMAIN_GTT;
 680         rbuffer->flags = 0;
 681         rbuffer->b.is_user_ptr = true;
 682         util_range_add(&rbuffer->valid_buffer_range, 0, templ->width0);
 683         util_range_add(&rbuffer->b.valid_buffer_range, 0, templ->width0);
 684
 685         /* Convert a user pointer to a buffer. */
 686         rbuffer->buf = ws->buffer_from_ptr(ws, user_memory, templ->width0);
 687         if (!rbuffer->buf) {
 688                 FREE(rbuffer);
 689                 return NULL;
 690         }
 691
 692         rbuffer->gpu_address = ws->buffer_get_virtual_address(rbuffer->buf);
 693         rbuffer->vram_usage = 0;
 694         rbuffer->gart_usage = templ->width0;
 695
 696         return &rbuffer->b.b;
 697 }
 698
 699 static struct pipe_resource *si_resource_create(struct pipe_screen *screen,
 700                                                 const struct pipe_resource *templ)
 701 {
 702         if (templ->target == PIPE_BUFFER) {
 703                 return si_buffer_create(screen, templ, 256);
 704         } else {
 705                 return si_texture_create(screen, templ);
 706         }
 707 }
 708
 709 void si_init_screen_buffer_functions(struct si_screen *sscreen)
 710 {
 711         sscreen->b.resource_create = si_resource_create;
 712         sscreen->b.resource_destroy = u_resource_destroy_vtbl;
 713         sscreen->b.resource_from_user_memory = si_buffer_from_user_memory;
 714 }
 715
 716 void si_init_buffer_functions(struct si_context *sctx)
 717 {
 718         sctx->b.b.invalidate_resource = si_invalidate_resource;
 719         sctx->b.b.transfer_map = u_transfer_map_vtbl;
 720         sctx->b.b.transfer_flush_region = u_transfer_flush_region_vtbl;
 721         sctx->b.b.transfer_unmap = u_transfer_unmap_vtbl;
 722         sctx->b.b.texture_subdata = u_default_texture_subdata;
 723         sctx->b.b.buffer_subdata = si_buffer_subdata;
 724 }