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