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_WRITE) {
 102       /* If we write to the buffer for any reason, free any saved translated
 103        * vertices.
 104        */
 105       pipe_resource_reference(&sbuf->translated_indices.buffer, NULL);
 106    }
 107
 108    if ((usage & PIPE_TRANSFER_READ) && sbuf->dirty) {
 109       enum pipe_error ret;
 110
 111       /* Host-side buffers can only be dirtied with vgpu10 features
 112        * (streamout and buffer copy).
 113        */
 114       assert(svga_have_vgpu10(svga));
 115
 116       if (!sbuf->user) {
 117          (void) svga_buffer_handle(svga, resource);
 118       }
 119
 120       if (sbuf->dma.pending > 0) {
 121          svga_buffer_upload_flush(svga, sbuf);
 122          svga_context_finish(svga);
 123       }
 124
 125       assert(sbuf->handle);
 126
 127       ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, sbuf->handle, 0);
 128       if (ret != PIPE_OK) {
 129          svga_context_flush(svga, NULL);
 130          ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, sbuf->handle, 0);
 131          assert(ret == PIPE_OK);
 132       }
 133
 134       svga->hud.num_readbacks++;
 135
 136       svga_context_finish(svga);
 137
 138       sbuf->dirty = FALSE;
 139    }
 140
 141    if (usage & PIPE_TRANSFER_WRITE) {
 142       if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 143          /*
 144           * Flush any pending primitives, finish writing any pending DMA
 145           * commands, and tell the host to discard the buffer contents on
 146           * the next DMA operation.
 147           */
 148
 149          svga_hwtnl_flush_buffer(svga, resource);
 150
 151          if (sbuf->dma.pending) {
 152             svga_buffer_upload_flush(svga, sbuf);
 153
 154             /*
 155              * Instead of flushing the context command buffer, simply discard
 156              * the current hwbuf, and start a new one.
 157              * With GB objects, the map operation takes care of this
 158              * if passed the PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE flag,
 159              * and the old backing store is busy.
 160              */
 161
 162             if (!svga_have_gb_objects(svga))
 163                svga_buffer_destroy_hw_storage(ss, sbuf);
 164          }
 165
 166          sbuf->map.num_ranges = 0;
 167          sbuf->dma.flags.discard = TRUE;
 168       }
 169
 170       if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
 171          if (!sbuf->map.num_ranges) {
 172             /*
 173              * No pending ranges to upload so far, so we can tell the host to
 174              * not synchronize on the next DMA command.
 175              */
 176
 177             sbuf->dma.flags.unsynchronized = TRUE;
 178          }
 179       } else {
 180          /*
 181           * Synchronizing, so flush any pending primitives, finish writing any
 182           * pending DMA command, and ensure the next DMA will be done in order.
 183           */
 184
 185          svga_hwtnl_flush_buffer(svga, resource);
 186
 187          if (sbuf->dma.pending) {
 188             svga_buffer_upload_flush(svga, sbuf);
 189
 190             if (svga_buffer_has_hw_storage(sbuf)) {
 191                /*
 192                 * We have a pending DMA upload from a hardware buffer, therefore
 193                 * we need to ensure that the host finishes processing that DMA
 194                 * command before the state tracker can start overwriting the
 195                 * hardware buffer.
 196                 *
 197                 * XXX: This could be avoided by tying the hardware buffer to
 198                 * the transfer (just as done with textures), which would allow
 199                 * overlapping DMAs commands to be queued on the same context
 200                 * buffer. However, due to the likelihood of software vertex
 201                 * processing, it is more convenient to hold on to the hardware
 202                 * buffer, allowing to quickly access the contents from the CPU
 203                 * without having to do a DMA download from the host.
 204                 */
 205
 206                if (usage & PIPE_TRANSFER_DONTBLOCK) {
 207                   /*
 208                    * Flushing the command buffer here will most likely cause
 209                    * the map of the hwbuf below to block, so preemptively
 210                    * return NULL here if DONTBLOCK is set to prevent unnecessary
 211                    * command buffer flushes.
 212                    */
 213
 214                   FREE(transfer);
 215                   goto done;
 216                }
 217
 218                svga_context_flush(svga, NULL);
 219             }
 220          }
 221
 222          sbuf->dma.flags.unsynchronized = FALSE;
 223       }
 224    }
 225
 226    if (!sbuf->swbuf && !svga_buffer_has_hw_storage(sbuf)) {
 227       if (svga_buffer_create_hw_storage(ss, sbuf) != PIPE_OK) {
 228          /*
 229           * We can't create a hardware buffer big enough, so create a malloc
 230           * buffer instead.
 231           */
 232          if (0) {
 233             debug_printf("%s: failed to allocate %u KB of DMA, "
 234                          "splitting DMA transfers\n",
 235                          __FUNCTION__,
 236                          (sbuf->b.b.width0 + 1023)/1024);
 237          }
 238
 239          sbuf->swbuf = align_malloc(sbuf->b.b.width0, 16);
 240          if (!sbuf->swbuf) {
 241             FREE(transfer);
 242             goto done;
 243          }
 244       }
 245    }
 246
 247    if (sbuf->swbuf) {
 248       /* User/malloc buffer */
 249       map = sbuf->swbuf;
 250    }
 251    else if (svga_buffer_has_hw_storage(sbuf)) {
 252       boolean retry;
 253
 254       map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
 255       if (map == NULL && retry) {
 256          /*
 257           * At this point, svga_buffer_get_transfer() has already
 258           * hit the DISCARD_WHOLE_RESOURCE path and flushed HWTNL
 259           * for this buffer.
 260           */
 261          svga_context_flush(svga, NULL);
 262          map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
 263       }
 264    }
 265    else {
 266       map = NULL;
 267    }
 268
 269    if (map) {
 270       ++sbuf->map.count;
 271       map += transfer->box.x;
 272       *ptransfer = transfer;
 273    } else {
 274       FREE(transfer);
 275    }
 276
 277    svga->hud.map_buffer_time += (svga_get_time(svga) - begin);
 278
 279 done:
 280    SVGA_STATS_TIME_POP(svga_sws(svga));
 281    return map;
 282 }
 283
 284
 285 static void
 286 svga_buffer_transfer_flush_region( struct pipe_context *pipe,
 287                                    struct pipe_transfer *transfer,
 288                                    const struct pipe_box *box)
 289 {
 290    struct svga_screen *ss = svga_screen(pipe->screen);
 291    struct svga_buffer *sbuf = svga_buffer(transfer->resource);
 292
 293    unsigned offset = transfer->box.x + box->x;
 294    unsigned length = box->width;
 295
 296    assert(transfer->usage & PIPE_TRANSFER_WRITE);
 297    assert(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT);
 298
 299    mtx_lock(&ss->swc_mutex);
 300    svga_buffer_add_range(sbuf, offset, offset + length);
 301    pipe_mutex_unlock(ss->swc_mutex);
 302 }
 303
 304
 305 static void
 306 svga_buffer_transfer_unmap( struct pipe_context *pipe,
 307                             struct pipe_transfer *transfer )
 308 {
 309    struct svga_screen *ss = svga_screen(pipe->screen);
 310    struct svga_context *svga = svga_context(pipe);
 311    struct svga_buffer *sbuf = svga_buffer(transfer->resource);
 312
 313    SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_BUFFERTRANSFERUNMAP);
 314
 315    mtx_lock(&ss->swc_mutex);
 316
 317    assert(sbuf->map.count);
 318    if (sbuf->map.count) {
 319       --sbuf->map.count;
 320    }
 321
 322    if (svga_buffer_has_hw_storage(sbuf)) {
 323       svga_buffer_hw_storage_unmap(svga, sbuf);
 324    }
 325
 326    if (transfer->usage & PIPE_TRANSFER_WRITE) {
 327       if (!(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
 328          /*
 329           * Mapped range not flushed explicitly, so flush the whole buffer,
 330           * and tell the host to discard the contents when processing the DMA
 331           * command.
 332           */
 333
 334          SVGA_DBG(DEBUG_DMA, "flushing the whole buffer\n");
 335
 336          sbuf->dma.flags.discard = TRUE;
 337
 338          svga_buffer_add_range(sbuf, 0, sbuf->b.b.width0);
 339       }
 340    }
 341
 342    pipe_mutex_unlock(ss->swc_mutex);
 343    FREE(transfer);
 344    SVGA_STATS_TIME_POP(svga_sws(svga));
 345 }
 346
 347
 348 static void
 349 svga_buffer_destroy( struct pipe_screen *screen,
 350                      struct pipe_resource *buf )
 351 {
 352    struct svga_screen *ss = svga_screen(screen);
 353    struct svga_buffer *sbuf = svga_buffer( buf );
 354
 355    assert(!p_atomic_read(&buf->reference.count));
 356
 357    assert(!sbuf->dma.pending);
 358
 359    if (sbuf->handle)
 360       svga_buffer_destroy_host_surface(ss, sbuf);
 361
 362    if (sbuf->uploaded.buffer)
 363       pipe_resource_reference(&sbuf->uploaded.buffer, NULL);
 364
 365    if (sbuf->hwbuf)
 366       svga_buffer_destroy_hw_storage(ss, sbuf);
 367
 368    if (sbuf->swbuf && !sbuf->user)
 369       align_free(sbuf->swbuf);
 370
 371    pipe_resource_reference(&sbuf->translated_indices.buffer, NULL);
 372
 373    ss->hud.total_resource_bytes -= sbuf->size;
 374    assert(ss->hud.num_resources > 0);
 375    if (ss->hud.num_resources > 0)
 376       ss->hud.num_resources--;
 377
 378    FREE(sbuf);
 379 }
 380
 381
 382 struct u_resource_vtbl svga_buffer_vtbl =
 383 {
 384    u_default_resource_get_handle,      /* get_handle */
 385    svga_buffer_destroy,              /* resource_destroy */
 386    svga_buffer_transfer_map,         /* transfer_map */
 387    svga_buffer_transfer_flush_region,  /* transfer_flush_region */
 388    svga_buffer_transfer_unmap,       /* transfer_unmap */
 389 };
 390
 391
 392
 393 struct pipe_resource *
 394 svga_buffer_create(struct pipe_screen *screen,
 395                    const struct pipe_resource *template)
 396 {
 397    struct svga_screen *ss = svga_screen(screen);
 398    struct svga_buffer *sbuf;
 399
 400    SVGA_STATS_TIME_PUSH(ss->sws, SVGA_STATS_TIME_CREATEBUFFER);
 401
 402    sbuf = CALLOC_STRUCT(svga_buffer);
 403    if (!sbuf)
 404       goto error1;
 405
 406    sbuf->b.b = *template;
 407    sbuf->b.vtbl = &svga_buffer_vtbl;
 408    pipe_reference_init(&sbuf->b.b.reference, 1);
 409    sbuf->b.b.screen = screen;
 410    sbuf->bind_flags = template->bind;
 411
 412    if (template->bind & PIPE_BIND_CONSTANT_BUFFER) {
 413       /* Constant buffers can only have the PIPE_BIND_CONSTANT_BUFFER
 414        * flag set.
 415        */
 416       if (ss->sws->have_vgpu10) {
 417          sbuf->bind_flags = PIPE_BIND_CONSTANT_BUFFER;
 418
 419          /* Constant buffer size needs to be in multiples of 16. */
 420          sbuf->b.b.width0 = align(sbuf->b.b.width0, 16);
 421       }
 422    }
 423
 424    if (svga_buffer_needs_hw_storage(template->bind)) {
 425
 426       /* If the buffer will be used for vertex/index/stream data, set all
 427        * the flags so that the buffer will be accepted for all those uses.
 428        * Note that the PIPE_BIND_ flags we get from the state tracker are
 429        * just a hint about how the buffer may be used.  And OpenGL buffer
 430        * object may be used for many different things.
 431        */
 432       if (!(template->bind & PIPE_BIND_CONSTANT_BUFFER)) {
 433          /* Not a constant buffer.  The buffer may be used for vertex data,
 434           * indexes or stream-out.
 435           */
 436          sbuf->bind_flags |= (PIPE_BIND_VERTEX_BUFFER |
 437                               PIPE_BIND_INDEX_BUFFER);
 438          if (ss->sws->have_vgpu10)
 439             sbuf->bind_flags |= PIPE_BIND_STREAM_OUTPUT;
 440       }
 441
 442       if (svga_buffer_create_host_surface(ss, sbuf) != PIPE_OK)
 443          goto error2;
 444    }
 445    else {
 446       sbuf->swbuf = align_malloc(sbuf->b.b.width0, 64);
 447       if (!sbuf->swbuf)
 448          goto error2;
 449    }
 450
 451    debug_reference(&sbuf->b.b.reference,
 452                    (debug_reference_descriptor)debug_describe_resource, 0);
 453
 454    sbuf->size = util_resource_size(&sbuf->b.b);
 455    ss->hud.total_resource_bytes += sbuf->size;
 456
 457    ss->hud.num_resources++;
 458    SVGA_STATS_TIME_POP(ss->sws);
 459
 460    return &sbuf->b.b;
 461
 462 error2:
 463    FREE(sbuf);
 464 error1:
 465    SVGA_STATS_TIME_POP(ss->sws);
 466    return NULL;
 467 }
 468
 469
 470 struct pipe_resource *
 471 svga_user_buffer_create(struct pipe_screen *screen,
 472                         void *ptr,
 473                         unsigned bytes,
 474                         unsigned bind)
 475 {
 476    struct svga_buffer *sbuf;
 477    struct svga_screen *ss = svga_screen(screen);
 478
 479    sbuf = CALLOC_STRUCT(svga_buffer);
 480    if (!sbuf)
 481       goto no_sbuf;
 482
 483    pipe_reference_init(&sbuf->b.b.reference, 1);
 484    sbuf->b.vtbl = &svga_buffer_vtbl;
 485    sbuf->b.b.screen = screen;
 486    sbuf->b.b.format = PIPE_FORMAT_R8_UNORM; /* ?? */
 487    sbuf->b.b.usage = PIPE_USAGE_IMMUTABLE;
 488    sbuf->b.b.bind = bind;
 489    sbuf->b.b.width0 = bytes;
 490    sbuf->b.b.height0 = 1;
 491    sbuf->b.b.depth0 = 1;
 492    sbuf->b.b.array_size = 1;
 493
 494    sbuf->bind_flags = bind;
 495    sbuf->swbuf = ptr;
 496    sbuf->user = TRUE;
 497
 498    debug_reference(&sbuf->b.b.reference,
 499                    (debug_reference_descriptor)debug_describe_resource, 0);
 500
 501    ss->hud.num_resources++;
 502
 503    return &sbuf->b.b;
 504
 505 no_sbuf:
 506    return NULL;
 507 }
 508
 509
 510