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