src/gallium/drivers/svga/svga_resource_buffer.c

   1 /**********************************************************
   2  * Copyright 2008-2009 VMware, Inc.  All rights reserved.
   3  *
   4  * Permission is hereby granted, free of charge, to any person
   5  * obtaining a copy of this software and associated documentation
   6  * files (the "Software"), to deal in the Software without
   7  * restriction, including without limitation the rights to use, copy,
   8  * modify, merge, publish, distribute, sublicense, and/or sell copies
   9  * of the Software, and to permit persons to whom the Software is
  10  * furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be
  13  * included in all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  *
  24  **********************************************************/
  25
  26 #include "svga_cmd.h"
  27
  28 #include "pipe/p_state.h"
  29 #include "pipe/p_defines.h"
  30 #include "util/u_inlines.h"
  31 #include "os/os_thread.h"
  32 #include "util/u_math.h"
  33 #include "util/u_memory.h"
  34 #include "util/u_resource.h"
  35
  36 #include "svga_context.h"
  37 #include "svga_screen.h"
  38 #include "svga_resource_buffer.h"
  39 #include "svga_resource_buffer_upload.h"
  40 #include "svga_winsys.h"
  41 #include "svga_debug.h"
  42
  43
  44 /**
  45  * Vertex and index buffers need hardware backing.  Constant buffers
  46  * do not.  No other types of buffers currently supported.
  47  */
  48 static inline boolean
  49 svga_buffer_needs_hw_storage(unsigned usage)
  50 {
  51    return (usage & (PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER |
  52                     PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_STREAM_OUTPUT)) != 0;
  53 }
  54
  55
  56 /**
  57  * Create a buffer transfer.
  58  *
  59  * Unlike texture DMAs (which are written immediately to the command buffer and
  60  * therefore inherently serialized with other context operations), for buffers
  61  * we try to coalesce multiple range mappings (i.e, multiple calls to this
  62  * function) into a single DMA command, for better efficiency in command
  63  * processing.  This means we need to exercise extra care here to ensure that
  64  * the end result is exactly the same as if one DMA was used for every mapped
  65  * range.
  66  */
  67 static void *
  68 svga_buffer_transfer_map(struct pipe_context *pipe,
  69                          struct pipe_resource *resource,
  70                          unsigned level,
  71                          unsigned usage,
  72                          const struct pipe_box *box,
  73                          struct pipe_transfer **ptransfer)
  74 {
  75    struct svga_context *svga = svga_context(pipe);
  76    struct svga_screen *ss = svga_screen(pipe->screen);
  77    struct svga_buffer *sbuf = svga_buffer(resource);
  78    struct pipe_transfer *transfer;
  79    uint8_t *map = NULL;
  80    int64_t begin = svga_get_time(svga);
  81
  82    SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_BUFFERTRANSFERMAP);
  83
  84    assert(box->y == 0);
  85    assert(box->z == 0);
  86    assert(box->height == 1);
  87    assert(box->depth == 1);
  88
  89    transfer = MALLOC_STRUCT(pipe_transfer);
  90    if (!transfer) {
  91       goto done;
  92    }
  93
  94    transfer->resource = resource;
  95    transfer->level = level;
  96    transfer->usage = usage;
  97    transfer->box = *box;
  98    transfer->stride = 0;
  99    transfer->layer_stride = 0;
 100
 101    if ((usage & PIPE_TRANSFER_READ) && sbuf->dirty) {
 102       enum pipe_error ret;
 103
 104       /* Host-side buffers can only be dirtied with vgpu10 features
 105        * (streamout and buffer copy).
 106        */
 107       assert(svga_have_vgpu10(svga));
 108
 109       if (!sbuf->user) {
 110          (void) svga_buffer_handle(svga, resource);
 111       }
 112
 113       if (sbuf->dma.pending > 0) {
 114          svga_buffer_upload_flush(svga, sbuf);
 115          svga_context_finish(svga);
 116       }
 117
 118       assert(sbuf->handle);
 119
 120       ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, sbuf->handle, 0);
 121       if (ret != PIPE_OK) {
 122          svga_context_flush(svga, NULL);
 123          ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, sbuf->handle, 0);
 124          assert(ret == PIPE_OK);
 125       }
 126
 127       svga->hud.num_readbacks++;
 128
 129       svga_context_finish(svga);
 130
 131       sbuf->dirty = FALSE;
 132    }
 133
 134    if (usage & PIPE_TRANSFER_WRITE) {
 135       if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 136          /*
 137           * Flush any pending primitives, finish writing any pending DMA
 138           * commands, and tell the host to discard the buffer contents on
 139           * the next DMA operation.
 140           */
 141
 142          svga_hwtnl_flush_buffer(svga, resource);
 143
 144          if (sbuf->dma.pending) {
 145             svga_buffer_upload_flush(svga, sbuf);
 146
 147             /*
 148              * Instead of flushing the context command buffer, simply discard
 149              * the current hwbuf, and start a new one.
 150              * With GB objects, the map operation takes care of this
 151              * if passed the PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE flag,
 152              * and the old backing store is busy.
 153              */
 154
 155             if (!svga_have_gb_objects(svga))
 156                svga_buffer_destroy_hw_storage(ss, sbuf);
 157          }
 158
 159          sbuf->map.num_ranges = 0;
 160          sbuf->dma.flags.discard = TRUE;
 161       }
 162
 163       if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
 164          if (!sbuf->map.num_ranges) {
 165             /*
 166              * No pending ranges to upload so far, so we can tell the host to
 167              * not synchronize on the next DMA command.
 168              */
 169
 170             sbuf->dma.flags.unsynchronized = TRUE;
 171          }
 172       } else {
 173          /*
 174           * Synchronizing, so flush any pending primitives, finish writing any
 175           * pending DMA command, and ensure the next DMA will be done in order.
 176           */
 177
 178          svga_hwtnl_flush_buffer(svga, resource);
 179
 180          if (sbuf->dma.pending) {
 181             svga_buffer_upload_flush(svga, sbuf);
 182
 183             if (svga_buffer_has_hw_storage(sbuf)) {
 184                /*
 185                 * We have a pending DMA upload from a hardware buffer, therefore
 186                 * we need to ensure that the host finishes processing that DMA
 187                 * command before the state tracker can start overwriting the
 188                 * hardware buffer.
 189                 *
 190                 * XXX: This could be avoided by tying the hardware buffer to
 191                 * the transfer (just as done with textures), which would allow
 192                 * overlapping DMAs commands to be queued on the same context
 193                 * buffer. However, due to the likelihood of software vertex
 194                 * processing, it is more convenient to hold on to the hardware
 195                 * buffer, allowing to quickly access the contents from the CPU
 196                 * without having to do a DMA download from the host.
 197                 */
 198
 199                if (usage & PIPE_TRANSFER_DONTBLOCK) {
 200                   /*
 201                    * Flushing the command buffer here will most likely cause
 202                    * the map of the hwbuf below to block, so preemptively
 203                    * return NULL here if DONTBLOCK is set to prevent unnecessary
 204                    * command buffer flushes.
 205                    */
 206
 207                   FREE(transfer);
 208                   goto done;
 209                }
 210
 211                svga_context_flush(svga, NULL);
 212             }
 213          }
 214
 215          sbuf->dma.flags.unsynchronized = FALSE;
 216       }
 217    }
 218
 219    if (!sbuf->swbuf && !svga_buffer_has_hw_storage(sbuf)) {
 220       if (svga_buffer_create_hw_storage(ss, sbuf) != PIPE_OK) {
 221          /*
 222           * We can't create a hardware buffer big enough, so create a malloc
 223           * buffer instead.
 224           */
 225          if (0) {
 226             debug_printf("%s: failed to allocate %u KB of DMA, "
 227                          "splitting DMA transfers\n",
 228                          __FUNCTION__,
 229                          (sbuf->b.b.width0 + 1023)/1024);
 230          }
 231
 232          sbuf->swbuf = align_malloc(sbuf->b.b.width0, 16);
 233          if (!sbuf->swbuf) {
 234             FREE(transfer);
 235             goto done;
 236          }
 237       }
 238    }
 239
 240    if (sbuf->swbuf) {
 241       /* User/malloc buffer */
 242       map = sbuf->swbuf;
 243    }
 244    else if (svga_buffer_has_hw_storage(sbuf)) {
 245       boolean retry;
 246
 247       map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
 248       if (map == NULL && retry) {
 249          /*
 250           * At this point, svga_buffer_get_transfer() has already
 251           * hit the DISCARD_WHOLE_RESOURCE path and flushed HWTNL
 252           * for this buffer.
 253           */
 254          svga_context_flush(svga, NULL);
 255          map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
 256       }
 257    }
 258    else {
 259       map = NULL;
 260    }
 261
 262    if (map) {
 263       ++sbuf->map.count;
 264       map += transfer->box.x;
 265       *ptransfer = transfer;
 266    } else {
 267       FREE(transfer);
 268    }
 269
 270    svga->hud.map_buffer_time += (svga_get_time(svga) - begin);
 271
 272 done:
 273    SVGA_STATS_TIME_POP(svga_sws(svga));
 274    return map;
 275 }
 276
 277
 278 static void
 279 svga_buffer_transfer_flush_region( struct pipe_context *pipe,
 280                                    struct pipe_transfer *transfer,
 281                                    const struct pipe_box *box)
 282 {
 283    struct svga_screen *ss = svga_screen(pipe->screen);
 284    struct svga_buffer *sbuf = svga_buffer(transfer->resource);
 285
 286    unsigned offset = transfer->box.x + box->x;
 287    unsigned length = box->width;
 288
 289    assert(transfer->usage & PIPE_TRANSFER_WRITE);
 290    assert(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT);
 291
 292    pipe_mutex_lock(ss->swc_mutex);
 293    svga_buffer_add_range(sbuf, offset, offset + length);
 294    pipe_mutex_unlock(ss->swc_mutex);
 295 }
 296
 297
 298 static void
 299 svga_buffer_transfer_unmap( struct pipe_context *pipe,
 300                             struct pipe_transfer *transfer )
 301 {
 302    struct svga_screen *ss = svga_screen(pipe->screen);
 303    struct svga_context *svga = svga_context(pipe);
 304    struct svga_buffer *sbuf = svga_buffer(transfer->resource);
 305
 306    SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_BUFFERTRANSFERUNMAP);
 307
 308    pipe_mutex_lock(ss->swc_mutex);
 309
 310    assert(sbuf->map.count);
 311    if (sbuf->map.count) {
 312       --sbuf->map.count;
 313    }
 314
 315    if (svga_buffer_has_hw_storage(sbuf)) {
 316       svga_buffer_hw_storage_unmap(svga, sbuf);
 317    }
 318
 319    if (transfer->usage & PIPE_TRANSFER_WRITE) {
 320       if (!(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
 321          /*
 322           * Mapped range not flushed explicitly, so flush the whole buffer,
 323           * and tell the host to discard the contents when processing the DMA
 324           * command.
 325           */
 326
 327          SVGA_DBG(DEBUG_DMA, "flushing the whole buffer\n");
 328
 329          sbuf->dma.flags.discard = TRUE;
 330
 331          svga_buffer_add_range(sbuf, 0, sbuf->b.b.width0);
 332       }
 333    }
 334
 335    pipe_mutex_unlock(ss->swc_mutex);
 336    FREE(transfer);
 337    SVGA_STATS_TIME_POP(svga_sws(svga));
 338 }
 339
 340
 341 static void
 342 svga_buffer_destroy( struct pipe_screen *screen,
 343                      struct pipe_resource *buf )
 344 {
 345    struct svga_screen *ss = svga_screen(screen);
 346    struct svga_buffer *sbuf = svga_buffer( buf );
 347
 348    assert(!p_atomic_read(&buf->reference.count));
 349
 350    assert(!sbuf->dma.pending);
 351
 352    if (sbuf->handle)
 353       svga_buffer_destroy_host_surface(ss, sbuf);
 354
 355    if (sbuf->uploaded.buffer)
 356       pipe_resource_reference(&sbuf->uploaded.buffer, NULL);
 357
 358    if (sbuf->hwbuf)
 359       svga_buffer_destroy_hw_storage(ss, sbuf);
 360
 361    if (sbuf->swbuf && !sbuf->user)
 362       align_free(sbuf->swbuf);
 363
 364    ss->hud.total_resource_bytes -= sbuf->size;
 365    assert(ss->hud.num_resources > 0);
 366    if (ss->hud.num_resources > 0)
 367       ss->hud.num_resources--;
 368
 369    FREE(sbuf);
 370 }
 371
 372
 373 struct u_resource_vtbl svga_buffer_vtbl =
 374 {
 375    u_default_resource_get_handle,      /* get_handle */
 376    svga_buffer_destroy,              /* resource_destroy */
 377    svga_buffer_transfer_map,         /* transfer_map */
 378    svga_buffer_transfer_flush_region,  /* transfer_flush_region */
 379    svga_buffer_transfer_unmap,       /* transfer_unmap */
 380 };
 381
 382
 383
 384 struct pipe_resource *
 385 svga_buffer_create(struct pipe_screen *screen,
 386                    const struct pipe_resource *template)
 387 {
 388    struct svga_screen *ss = svga_screen(screen);
 389    struct svga_buffer *sbuf;
 390
 391    SVGA_STATS_TIME_PUSH(ss->sws, SVGA_STATS_TIME_CREATEBUFFER);
 392
 393    sbuf = CALLOC_STRUCT(svga_buffer);
 394    if (!sbuf)
 395       goto error1;
 396
 397    sbuf->b.b = *template;
 398    sbuf->b.vtbl = &svga_buffer_vtbl;
 399    pipe_reference_init(&sbuf->b.b.reference, 1);
 400    sbuf->b.b.screen = screen;
 401    sbuf->bind_flags = template->bind;
 402
 403    if (template->bind & PIPE_BIND_CONSTANT_BUFFER) {
 404       /* Constant buffers can only have the PIPE_BIND_CONSTANT_BUFFER
 405        * flag set.
 406        */
 407       if (ss->sws->have_vgpu10) {
 408          sbuf->bind_flags = PIPE_BIND_CONSTANT_BUFFER;
 409
 410          /* Constant buffer size needs to be in multiples of 16. */
 411          sbuf->b.b.width0 = align(sbuf->b.b.width0, 16);
 412       }
 413    }
 414
 415    if (svga_buffer_needs_hw_storage(template->bind)) {
 416
 417       /* If the buffer will be used for vertex/index/stream data, set all
 418        * the flags so that the buffer will be accepted for all those uses.
 419        * Note that the PIPE_BIND_ flags we get from the state tracker are
 420        * just a hint about how the buffer may be used.  And OpenGL buffer
 421        * object may be used for many different things.
 422        */
 423       if (!(template->bind & PIPE_BIND_CONSTANT_BUFFER)) {
 424          /* Not a constant buffer.  The buffer may be used for vertex data,
 425           * indexes or stream-out.
 426           */
 427          sbuf->bind_flags |= (PIPE_BIND_VERTEX_BUFFER |
 428                               PIPE_BIND_INDEX_BUFFER);
 429          if (ss->sws->have_vgpu10)
 430             sbuf->bind_flags |= PIPE_BIND_STREAM_OUTPUT;
 431       }
 432
 433       if (svga_buffer_create_host_surface(ss, sbuf) != PIPE_OK)
 434          goto error2;
 435    }
 436    else {
 437       sbuf->swbuf = align_malloc(sbuf->b.b.width0, 64);
 438       if (!sbuf->swbuf)
 439          goto error2;
 440    }
 441
 442    debug_reference(&sbuf->b.b.reference,
 443                    (debug_reference_descriptor)debug_describe_resource, 0);
 444
 445    sbuf->size = util_resource_size(&sbuf->b.b);
 446    ss->hud.total_resource_bytes += sbuf->size;
 447
 448    ss->hud.num_resources++;
 449    SVGA_STATS_TIME_POP(ss->sws);
 450
 451    return &sbuf->b.b;
 452
 453 error2:
 454    FREE(sbuf);
 455 error1:
 456    SVGA_STATS_TIME_POP(ss->sws);
 457    return NULL;
 458 }
 459
 460
 461 struct pipe_resource *
 462 svga_user_buffer_create(struct pipe_screen *screen,
 463                         void *ptr,
 464                         unsigned bytes,
 465                         unsigned bind)
 466 {
 467    struct svga_buffer *sbuf;
 468    struct svga_screen *ss = svga_screen(screen);
 469
 470    sbuf = CALLOC_STRUCT(svga_buffer);
 471    if (!sbuf)
 472       goto no_sbuf;
 473
 474    pipe_reference_init(&sbuf->b.b.reference, 1);
 475    sbuf->b.vtbl = &svga_buffer_vtbl;
 476    sbuf->b.b.screen = screen;
 477    sbuf->b.b.format = PIPE_FORMAT_R8_UNORM; /* ?? */
 478    sbuf->b.b.usage = PIPE_USAGE_IMMUTABLE;
 479    sbuf->b.b.bind = bind;
 480    sbuf->b.b.width0 = bytes;
 481    sbuf->b.b.height0 = 1;
 482    sbuf->b.b.depth0 = 1;
 483    sbuf->b.b.array_size = 1;
 484
 485    sbuf->bind_flags = bind;
 486    sbuf->swbuf = ptr;
 487    sbuf->user = TRUE;
 488
 489    debug_reference(&sbuf->b.b.reference,
 490                    (debug_reference_descriptor)debug_describe_resource, 0);
 491
 492    ss->hud.num_resources++;
 493
 494    return &sbuf->b.b;
 495
 496 no_sbuf:
 497    return NULL;
 498 }
 499
 500
 501