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