src/gallium/drivers/svga/svga_resource_texture.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 "svga3d_reg.h"
  27 #include "svga3d_surfacedefs.h"
  28
  29 #include "pipe/p_state.h"
  30 #include "pipe/p_defines.h"
  31 #include "os/os_thread.h"
  32 #include "util/u_format.h"
  33 #include "util/u_inlines.h"
  34 #include "util/u_math.h"
  35 #include "util/u_memory.h"
  36 #include "util/u_resource.h"
  37 #include "util/u_upload_mgr.h"
  38
  39 #include "svga_cmd.h"
  40 #include "svga_format.h"
  41 #include "svga_screen.h"
  42 #include "svga_context.h"
  43 #include "svga_resource_texture.h"
  44 #include "svga_resource_buffer.h"
  45 #include "svga_sampler_view.h"
  46 #include "svga_winsys.h"
  47 #include "svga_debug.h"
  48
  49
  50 static void
  51 svga_transfer_dma_band(struct svga_context *svga,
  52                        struct svga_transfer *st,
  53                        SVGA3dTransferType transfer,
  54                        unsigned x, unsigned y, unsigned z,
  55                        unsigned w, unsigned h, unsigned d,
  56                        unsigned srcx, unsigned srcy, unsigned srcz,
  57                        SVGA3dSurfaceDMAFlags flags)
  58 {
  59    struct svga_texture *texture = svga_texture(st->base.resource);
  60    SVGA3dCopyBox box;
  61    enum pipe_error ret;
  62
  63    assert(!st->use_direct_map);
  64
  65    box.x = x;
  66    box.y = y;
  67    box.z = z;
  68    box.w = w;
  69    box.h = h;
  70    box.d = d;
  71    box.srcx = srcx;
  72    box.srcy = srcy;
  73    box.srcz = srcz;
  74
  75    SVGA_DBG(DEBUG_DMA, "dma %s sid %p, face %u, (%u, %u, %u) - "
  76             "(%u, %u, %u), %ubpp\n",
  77             transfer == SVGA3D_WRITE_HOST_VRAM ? "to" : "from",
  78             texture->handle,
  79             st->slice,
  80             x,
  81             y,
  82             z,
  83             x + w,
  84             y + h,
  85             z + 1,
  86             util_format_get_blocksize(texture->b.b.format) * 8 /
  87             (util_format_get_blockwidth(texture->b.b.format)
  88              * util_format_get_blockheight(texture->b.b.format)));
  89
  90    ret = SVGA3D_SurfaceDMA(svga->swc, st, transfer, &box, 1, flags);
  91    if (ret != PIPE_OK) {
  92       svga_context_flush(svga, NULL);
  93       ret = SVGA3D_SurfaceDMA(svga->swc, st, transfer, &box, 1, flags);
  94       assert(ret == PIPE_OK);
  95    }
  96 }
  97
  98
  99 static void
 100 svga_transfer_dma(struct svga_context *svga,
 101                   struct svga_transfer *st,
 102                   SVGA3dTransferType transfer,
 103                   SVGA3dSurfaceDMAFlags flags)
 104 {
 105    struct svga_texture *texture = svga_texture(st->base.resource);
 106    struct svga_screen *screen = svga_screen(texture->b.b.screen);
 107    struct svga_winsys_screen *sws = screen->sws;
 108    struct pipe_fence_handle *fence = NULL;
 109
 110    assert(!st->use_direct_map);
 111
 112    if (transfer == SVGA3D_READ_HOST_VRAM) {
 113       SVGA_DBG(DEBUG_PERF, "%s: readback transfer\n", __FUNCTION__);
 114    }
 115
 116    /* Ensure any pending operations on host surfaces are queued on the command
 117     * buffer first.
 118     */
 119    svga_surfaces_flush( svga );
 120
 121    if (!st->swbuf) {
 122       /* Do the DMA transfer in a single go */
 123       svga_transfer_dma_band(svga, st, transfer,
 124                              st->base.box.x, st->base.box.y, st->base.box.z,
 125                              st->base.box.width, st->base.box.height, st->base.box.depth,
 126                              0, 0, 0,
 127                              flags);
 128
 129       if (transfer == SVGA3D_READ_HOST_VRAM) {
 130          svga_context_flush(svga, &fence);
 131          sws->fence_finish(sws, fence, 0);
 132          sws->fence_reference(sws, &fence, NULL);
 133       }
 134    }
 135    else {
 136       int y, h, srcy;
 137       unsigned blockheight =
 138          util_format_get_blockheight(st->base.resource->format);
 139
 140       h = st->hw_nblocksy * blockheight;
 141       srcy = 0;
 142
 143       for (y = 0; y < st->base.box.height; y += h) {
 144          unsigned offset, length;
 145          void *hw, *sw;
 146
 147          if (y + h > st->base.box.height)
 148             h = st->base.box.height - y;
 149
 150          /* Transfer band must be aligned to pixel block boundaries */
 151          assert(y % blockheight == 0);
 152          assert(h % blockheight == 0);
 153
 154          offset = y * st->base.stride / blockheight;
 155          length = h * st->base.stride / blockheight;
 156
 157          sw = (uint8_t *) st->swbuf + offset;
 158
 159          if (transfer == SVGA3D_WRITE_HOST_VRAM) {
 160             unsigned usage = PIPE_TRANSFER_WRITE;
 161
 162             /* Wait for the previous DMAs to complete */
 163             /* TODO: keep one DMA (at half the size) in the background */
 164             if (y) {
 165                svga_context_flush(svga, NULL);
 166                usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
 167             }
 168
 169             hw = sws->buffer_map(sws, st->hwbuf, usage);
 170             assert(hw);
 171             if (hw) {
 172                memcpy(hw, sw, length);
 173                sws->buffer_unmap(sws, st->hwbuf);
 174             }
 175          }
 176
 177          svga_transfer_dma_band(svga, st, transfer,
 178                                 st->base.box.x, y, st->base.box.z,
 179                                 st->base.box.width, h, st->base.box.depth,
 180                                 0, srcy, 0, flags);
 181
 182          /*
 183           * Prevent the texture contents to be discarded on the next band
 184           * upload.
 185           */
 186          flags.discard = FALSE;
 187
 188          if (transfer == SVGA3D_READ_HOST_VRAM) {
 189             svga_context_flush(svga, &fence);
 190             sws->fence_finish(sws, fence, 0);
 191
 192             hw = sws->buffer_map(sws, st->hwbuf, PIPE_TRANSFER_READ);
 193             assert(hw);
 194             if (hw) {
 195                memcpy(sw, hw, length);
 196                sws->buffer_unmap(sws, st->hwbuf);
 197             }
 198          }
 199       }
 200    }
 201 }
 202
 203
 204
 205 static boolean
 206 svga_texture_get_handle(struct pipe_screen *screen,
 207                         struct pipe_resource *texture,
 208                         struct winsys_handle *whandle)
 209 {
 210    struct svga_winsys_screen *sws = svga_winsys_screen(texture->screen);
 211    unsigned stride;
 212
 213    assert(svga_texture(texture)->key.cachable == 0);
 214    svga_texture(texture)->key.cachable = 0;
 215
 216    stride = util_format_get_nblocksx(texture->format, texture->width0) *
 217             util_format_get_blocksize(texture->format);
 218
 219    return sws->surface_get_handle(sws, svga_texture(texture)->handle,
 220                                   stride, whandle);
 221 }
 222
 223
 224 static void
 225 svga_texture_destroy(struct pipe_screen *screen,
 226                      struct pipe_resource *pt)
 227 {
 228    struct svga_screen *ss = svga_screen(screen);
 229    struct svga_texture *tex = svga_texture(pt);
 230
 231    ss->texture_timestamp++;
 232
 233    svga_sampler_view_reference(&tex->cached_view, NULL);
 234
 235    /*
 236      DBG("%s deleting %p\n", __FUNCTION__, (void *) tex);
 237    */
 238    SVGA_DBG(DEBUG_DMA, "unref sid %p (texture)\n", tex->handle);
 239    svga_screen_surface_destroy(ss, &tex->key, &tex->handle);
 240
 241    /* Destroy the backed surface handle if exists */
 242    if (tex->backed_handle)
 243       svga_screen_surface_destroy(ss, &tex->backed_key, &tex->backed_handle);
 244
 245    ss->hud.total_resource_bytes -= tex->size;
 246
 247    FREE(tex->defined);
 248    FREE(tex->rendered_to);
 249    FREE(tex->dirty);
 250    FREE(tex);
 251
 252    assert(ss->hud.num_resources > 0);
 253    if (ss->hud.num_resources > 0)
 254       ss->hud.num_resources--;
 255 }
 256
 257
 258 /**
 259  * Determine if the resource was rendered to
 260  */
 261 static inline boolean
 262 was_tex_rendered_to(struct pipe_resource *resource,
 263                     const struct pipe_transfer *transfer)
 264 {
 265    unsigned face;
 266
 267    if (resource->target == PIPE_TEXTURE_CUBE) {
 268       assert(transfer->box.depth == 1);
 269       face = transfer->box.z;
 270    }
 271    else {
 272       face = 0;
 273    }
 274
 275    return svga_was_texture_rendered_to(svga_texture(resource),
 276                                        face, transfer->level);
 277 }
 278
 279
 280 /**
 281  * Determine if we need to read back a texture image before mapping it.
 282  */
 283 static inline boolean
 284 need_tex_readback(struct pipe_transfer *transfer)
 285 {
 286    if (transfer->usage & PIPE_TRANSFER_READ)
 287       return TRUE;
 288
 289    if ((transfer->usage & PIPE_TRANSFER_WRITE) &&
 290        ((transfer->usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) == 0)) {
 291       return was_tex_rendered_to(transfer->resource, transfer);
 292    }
 293
 294    return FALSE;
 295 }
 296
 297
 298 static enum pipe_error
 299 readback_image_vgpu9(struct svga_context *svga,
 300                    struct svga_winsys_surface *surf,
 301                    unsigned slice,
 302                    unsigned level)
 303 {
 304    enum pipe_error ret;
 305
 306    ret = SVGA3D_ReadbackGBImage(svga->swc, surf, slice, level);
 307    if (ret != PIPE_OK) {
 308       svga_context_flush(svga, NULL);
 309       ret = SVGA3D_ReadbackGBImage(svga->swc, surf, slice, level);
 310    }
 311    return ret;
 312 }
 313
 314
 315 static enum pipe_error
 316 readback_image_vgpu10(struct svga_context *svga,
 317                     struct svga_winsys_surface *surf,
 318                     unsigned slice,
 319                     unsigned level,
 320                     unsigned numMipLevels)
 321 {
 322    enum pipe_error ret;
 323    unsigned subResource;
 324
 325    subResource = slice * numMipLevels + level;
 326    ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, surf, subResource);
 327    if (ret != PIPE_OK) {
 328       svga_context_flush(svga, NULL);
 329       ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, surf, subResource);
 330    }
 331    return ret;
 332 }
 333
 334
 335 /**
 336  * Use DMA for the transfer request
 337  */
 338 static void *
 339 svga_texture_transfer_map_dma(struct svga_context *svga,
 340                               struct svga_transfer *st)
 341 {
 342    struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 343    struct pipe_resource *texture = st->base.resource;
 344    unsigned nblocksx, nblocksy;
 345    unsigned d;
 346    unsigned usage = st->base.usage;
 347
 348    /* we'll put the data into a tightly packed buffer */
 349    nblocksx = util_format_get_nblocksx(texture->format, st->base.box.width);
 350    nblocksy = util_format_get_nblocksy(texture->format, st->base.box.height);
 351    d = st->base.box.depth;
 352
 353    st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
 354    st->base.layer_stride = st->base.stride * nblocksy;
 355    st->hw_nblocksy = nblocksy;
 356
 357    st->hwbuf = svga_winsys_buffer_create(svga, 1, 0,
 358                                          st->hw_nblocksy * st->base.stride * d);
 359
 360    while (!st->hwbuf && (st->hw_nblocksy /= 2)) {
 361       st->hwbuf =
 362          svga_winsys_buffer_create(svga, 1, 0,
 363                                    st->hw_nblocksy * st->base.stride * d);
 364    }
 365
 366    if (!st->hwbuf)
 367       return NULL;
 368
 369    if (st->hw_nblocksy < nblocksy) {
 370       /* We couldn't allocate a hardware buffer big enough for the transfer,
 371        * so allocate regular malloc memory instead
 372        */
 373       if (0) {
 374          debug_printf("%s: failed to allocate %u KB of DMA, "
 375                       "splitting into %u x %u KB DMA transfers\n",
 376                       __FUNCTION__,
 377                       (nblocksy * st->base.stride + 1023) / 1024,
 378                       (nblocksy + st->hw_nblocksy - 1) / st->hw_nblocksy,
 379                       (st->hw_nblocksy * st->base.stride + 1023) / 1024);
 380       }
 381
 382       st->swbuf = MALLOC(nblocksy * st->base.stride * d);
 383       if (!st->swbuf) {
 384          sws->buffer_destroy(sws, st->hwbuf);
 385          return NULL;
 386       }
 387    }
 388
 389    if (usage & PIPE_TRANSFER_READ) {
 390       SVGA3dSurfaceDMAFlags flags;
 391       memset(&flags, 0, sizeof flags);
 392       svga_transfer_dma(svga, st, SVGA3D_READ_HOST_VRAM, flags);
 393    }
 394
 395    if (st->swbuf) {
 396       return st->swbuf;
 397    }
 398    else {
 399       return sws->buffer_map(sws, st->hwbuf, usage);
 400    }
 401 }
 402
 403
 404 /**
 405  * Use direct map for the transfer request
 406  */
 407 static void *
 408 svga_texture_transfer_map_direct(struct svga_context *svga,
 409                                  struct svga_transfer *st)
 410 {
 411    struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 412    struct pipe_transfer *transfer = &st->base;
 413    struct pipe_resource *texture = transfer->resource;
 414    struct svga_texture *tex = svga_texture(texture);
 415    struct svga_winsys_surface *surf = tex->handle;
 416    unsigned level = st->base.level;
 417    unsigned w, h, nblocksx, nblocksy;
 418    unsigned usage = st->base.usage;
 419
 420    if (need_tex_readback(transfer)) {
 421       enum pipe_error ret;
 422
 423       svga_surfaces_flush(svga);
 424
 425       if (svga_have_vgpu10(svga)) {
 426          ret = readback_image_vgpu10(svga, surf, st->slice, level,
 427                                      tex->b.b.last_level + 1);
 428       } else {
 429          ret = readback_image_vgpu9(svga, surf, st->slice, level);
 430       }
 431
 432       svga->hud.num_readbacks++;
 433       SVGA_STATS_COUNT_INC(sws, SVGA_STATS_COUNT_TEXREADBACK);
 434
 435       assert(ret == PIPE_OK);
 436       (void) ret;
 437
 438       svga_context_flush(svga, NULL);
 439       /*
 440        * Note: if PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE were specified
 441        * we could potentially clear the flag for all faces/layers/mips.
 442        */
 443       svga_clear_texture_rendered_to(tex, st->slice, level);
 444    }
 445    else {
 446       assert(usage & PIPE_TRANSFER_WRITE);
 447       if ((usage & PIPE_TRANSFER_UNSYNCHRONIZED) == 0) {
 448          if (svga_is_texture_dirty(tex, st->slice, level)) {
 449             /*
 450              * do a surface flush if the subresource has been modified
 451              * in this command buffer.
 452              */
 453             svga_surfaces_flush(svga);
 454             if (!sws->surface_is_flushed(sws, surf)) {
 455                svga->hud.surface_write_flushes++;
 456                SVGA_STATS_COUNT_INC(sws, SVGA_STATS_COUNT_SURFACEWRITEFLUSH);
 457                svga_context_flush(svga, NULL);
 458             }
 459          }
 460       }
 461    }
 462
 463    /* we'll directly access the guest-backed surface */
 464    w = u_minify(texture->width0, level);
 465    h = u_minify(texture->height0, level);
 466    nblocksx = util_format_get_nblocksx(texture->format, w);
 467    nblocksy = util_format_get_nblocksy(texture->format, h);
 468    st->hw_nblocksy = nblocksy;
 469    st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
 470    st->base.layer_stride = st->base.stride * nblocksy;
 471
 472    /*
 473     * Begin mapping code
 474     */
 475    {
 476       SVGA3dSize baseLevelSize;
 477       uint8_t *map;
 478       boolean retry;
 479       unsigned offset, mip_width, mip_height;
 480
 481       map = svga->swc->surface_map(svga->swc, surf, usage, &retry);
 482       if (map == NULL && retry) {
 483          /*
 484           * At this point, the svga_surfaces_flush() should already have
 485           * called in svga_texture_get_transfer().
 486           */
 487          svga->hud.surface_write_flushes++;
 488          svga_context_flush(svga, NULL);
 489          map = svga->swc->surface_map(svga->swc, surf, usage, &retry);
 490       }
 491
 492       /*
 493        * Make sure we return NULL if the map fails
 494        */
 495       if (!map) {
 496          return NULL;
 497       }
 498
 499       /**
 500        * Compute the offset to the specific texture slice in the buffer.
 501        */
 502       baseLevelSize.width = tex->b.b.width0;
 503       baseLevelSize.height = tex->b.b.height0;
 504       baseLevelSize.depth = tex->b.b.depth0;
 505
 506       if ((tex->b.b.target == PIPE_TEXTURE_1D_ARRAY) ||
 507           (tex->b.b.target == PIPE_TEXTURE_2D_ARRAY)) {
 508          st->base.layer_stride =
 509             svga3dsurface_get_image_offset(tex->key.format, baseLevelSize,
 510                                            tex->b.b.last_level + 1, 1, 0);
 511       }
 512
 513       offset = svga3dsurface_get_image_offset(tex->key.format, baseLevelSize,
 514                                               tex->b.b.last_level + 1, /* numMips */
 515                                               st->slice, level);
 516       if (level > 0) {
 517          assert(offset > 0);
 518       }
 519
 520       mip_width = u_minify(tex->b.b.width0, level);
 521       mip_height = u_minify(tex->b.b.height0, level);
 522
 523       offset += svga3dsurface_get_pixel_offset(tex->key.format,
 524                                                mip_width, mip_height,
 525                                                st->base.box.x,
 526                                                st->base.box.y,
 527                                                st->base.box.z);
 528       return (void *) (map + offset);
 529    }
 530 }
 531
 532
 533 /**
 534  * Request a transfer map to the texture resource
 535  */
 536 static void *
 537 svga_texture_transfer_map(struct pipe_context *pipe,
 538                           struct pipe_resource *texture,
 539                           unsigned level,
 540                           unsigned usage,
 541                           const struct pipe_box *box,
 542                           struct pipe_transfer **ptransfer)
 543 {
 544    struct svga_context *svga = svga_context(pipe);
 545    struct svga_winsys_screen *sws = svga_screen(pipe->screen)->sws;
 546    struct svga_texture *tex = svga_texture(texture);
 547    struct svga_transfer *st;
 548    struct svga_winsys_surface *surf = tex->handle;
 549    boolean use_direct_map = svga_have_gb_objects(svga) &&
 550                             !svga_have_gb_dma(svga);
 551    void *map = NULL;
 552    int64_t begin = svga_get_time(svga);
 553
 554    SVGA_STATS_TIME_PUSH(sws, SVGA_STATS_TIME_TEXTRANSFERMAP);
 555
 556    if (!surf)
 557       goto done;
 558
 559    /* We can't map texture storage directly unless we have GB objects */
 560    if (usage & PIPE_TRANSFER_MAP_DIRECTLY) {
 561       if (svga_have_gb_objects(svga))
 562          use_direct_map = TRUE;
 563       else
 564          goto done;
 565    }
 566
 567    st = CALLOC_STRUCT(svga_transfer);
 568    if (!st)
 569       goto done;
 570
 571    st->base.level = level;
 572    st->base.usage = usage;
 573    st->base.box = *box;
 574
 575    switch (tex->b.b.target) {
 576    case PIPE_TEXTURE_CUBE:
 577       st->slice = st->base.box.z;
 578       st->base.box.z = 0;   /* so we don't apply double offsets below */
 579       break;
 580    case PIPE_TEXTURE_2D_ARRAY:
 581    case PIPE_TEXTURE_1D_ARRAY:
 582       st->slice = st->base.box.z;
 583       st->base.box.z = 0;   /* so we don't apply double offsets below */
 584
 585       /* Force direct map for transfering multiple slices */
 586       if (st->base.box.depth > 1)
 587          use_direct_map = svga_have_gb_objects(svga);
 588
 589       break;
 590    default:
 591       st->slice = 0;
 592       break;
 593    }
 594
 595    st->use_direct_map = use_direct_map;
 596    pipe_resource_reference(&st->base.resource, texture);
 597
 598    /* If this is the first time mapping to the surface in this
 599     * command buffer, clear the dirty masks of this surface.
 600     */
 601    if (sws->surface_is_flushed(sws, surf)) {
 602       svga_clear_texture_dirty(tex);
 603    }
 604
 605    if (!use_direct_map) {
 606       /* upload to the DMA buffer */
 607       map = svga_texture_transfer_map_dma(svga, st);
 608    }
 609    else {
 610       boolean can_use_upload = tex->can_use_upload &&
 611                                !(st->base.usage & PIPE_TRANSFER_READ);
 612       boolean was_rendered_to = was_tex_rendered_to(texture, &st->base);
 613
 614       /* If the texture was already rendered to and upload buffer
 615        * is supported, then we will use upload buffer to
 616        * avoid the need to read back the texture content; otherwise,
 617        * we'll first try to map directly to the GB surface, if it is blocked,
 618        * then we'll try the upload buffer.
 619        */
 620       if (was_rendered_to && can_use_upload) {
 621          map = svga_texture_transfer_map_upload(svga, st);
 622       }
 623       else {
 624          unsigned orig_usage = st->base.usage;
 625
 626          /* First try directly map to the GB surface */
 627          if (can_use_upload)
 628             st->base.usage |= PIPE_TRANSFER_DONTBLOCK;
 629          map = svga_texture_transfer_map_direct(svga, st);
 630          st->base.usage = orig_usage;
 631
 632          if (!map && can_use_upload) {
 633             /* if direct map with DONTBLOCK fails, then try upload to the
 634              * texture upload buffer.
 635              */
 636             map = svga_texture_transfer_map_upload(svga, st);
 637          }
 638       }
 639
 640       /* If upload fails, then try direct map again without forcing it
 641        * to DONTBLOCK.
 642        */
 643       if (!map) {
 644          map = svga_texture_transfer_map_direct(svga, st);
 645       }
 646    }
 647
 648    if (!map) {
 649       FREE(st);
 650    }
 651    else {
 652       *ptransfer = &st->base;
 653       svga->hud.num_textures_mapped++;
 654       if (usage & PIPE_TRANSFER_WRITE) {
 655          /* record texture upload for HUD */
 656          svga->hud.num_bytes_uploaded +=
 657             st->base.layer_stride * st->base.box.depth;
 658
 659          /* mark this texture level as dirty */
 660          svga_set_texture_dirty(tex, st->slice, level);
 661       }
 662    }
 663
 664 done:
 665    svga->hud.map_buffer_time += (svga_get_time(svga) - begin);
 666    SVGA_STATS_TIME_POP(sws);
 667    (void) sws;
 668
 669    return map;
 670 }
 671
 672 /**
 673  * Unmap a GB texture surface.
 674  */
 675 static void
 676 svga_texture_surface_unmap(struct svga_context *svga,
 677                            struct pipe_transfer *transfer)
 678 {
 679    struct svga_winsys_surface *surf = svga_texture(transfer->resource)->handle;
 680    struct svga_winsys_context *swc = svga->swc;
 681    boolean rebind;
 682
 683    assert(surf);
 684
 685    swc->surface_unmap(swc, surf, &rebind);
 686    if (rebind) {
 687       enum pipe_error ret;
 688       ret = SVGA3D_BindGBSurface(swc, surf);
 689       if (ret != PIPE_OK) {
 690          /* flush and retry */
 691          svga_context_flush(svga, NULL);
 692          ret = SVGA3D_BindGBSurface(swc, surf);
 693          assert(ret == PIPE_OK);
 694       }
 695    }
 696 }
 697
 698
 699 static enum pipe_error
 700 update_image_vgpu9(struct svga_context *svga,
 701                    struct svga_winsys_surface *surf,
 702                    const SVGA3dBox *box,
 703                    unsigned slice,
 704                    unsigned level)
 705 {
 706    enum pipe_error ret;
 707
 708    ret = SVGA3D_UpdateGBImage(svga->swc, surf, box, slice, level);
 709    if (ret != PIPE_OK) {
 710       svga_context_flush(svga, NULL);
 711       ret = SVGA3D_UpdateGBImage(svga->swc, surf, box, slice, level);
 712    }
 713    return ret;
 714 }
 715
 716
 717 static enum pipe_error
 718 update_image_vgpu10(struct svga_context *svga,
 719                     struct svga_winsys_surface *surf,
 720                     const SVGA3dBox *box,
 721                     unsigned slice,
 722                     unsigned level,
 723                     unsigned numMipLevels)
 724 {
 725    enum pipe_error ret;
 726    unsigned subResource;
 727
 728    subResource = slice * numMipLevels + level;
 729    ret = SVGA3D_vgpu10_UpdateSubResource(svga->swc, surf, box, subResource);
 730    if (ret != PIPE_OK) {
 731       svga_context_flush(svga, NULL);
 732       ret = SVGA3D_vgpu10_UpdateSubResource(svga->swc, surf, box, subResource);
 733    }
 734    return ret;
 735 }
 736
 737
 738 /**
 739  * unmap DMA transfer request
 740  */
 741 static void
 742 svga_texture_transfer_unmap_dma(struct svga_context *svga,
 743                                 struct svga_transfer *st)
 744 {
 745    struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
 746
 747    if (st->hwbuf)
 748       sws->buffer_unmap(sws, st->hwbuf);
 749
 750    if (st->base.usage & PIPE_TRANSFER_WRITE) {
 751       /* Use DMA to transfer texture data */
 752       SVGA3dSurfaceDMAFlags flags;
 753
 754       memset(&flags, 0, sizeof flags);
 755       if (st->base.usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 756          flags.discard = TRUE;
 757       }
 758       if (st->base.usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
 759          flags.unsynchronized = TRUE;
 760       }
 761
 762       svga_transfer_dma(svga, st, SVGA3D_WRITE_HOST_VRAM, flags);
 763    }
 764
 765    FREE(st->swbuf);
 766    sws->buffer_destroy(sws, st->hwbuf);
 767 }
 768
 769
 770 /**
 771  * unmap direct map transfer request
 772  */
 773 static void
 774 svga_texture_transfer_unmap_direct(struct svga_context *svga,
 775                                    struct svga_transfer *st)
 776 {
 777    struct pipe_transfer *transfer = &st->base;
 778    struct svga_texture *tex = svga_texture(transfer->resource);
 779
 780    svga_texture_surface_unmap(svga, transfer);
 781
 782    /* Now send an update command to update the content in the backend. */
 783    if (st->base.usage & PIPE_TRANSFER_WRITE) {
 784       struct svga_winsys_surface *surf = tex->handle;
 785       SVGA3dBox box;
 786       enum pipe_error ret;
 787       unsigned nlayers = 1;
 788
 789       assert(svga_have_gb_objects(svga));
 790
 791       /* update the effected region */
 792       box.x = transfer->box.x;
 793       box.y = transfer->box.y;
 794       box.w = transfer->box.width;
 795       box.h = transfer->box.height;
 796       box.d = transfer->box.depth;
 797
 798       switch (tex->b.b.target) {
 799       case PIPE_TEXTURE_CUBE:
 800          box.z = 0;
 801          break;
 802       case PIPE_TEXTURE_2D_ARRAY:
 803          nlayers = box.d;
 804          box.z = 0;
 805          box.d = 1;
 806          break;
 807       case PIPE_TEXTURE_1D_ARRAY:
 808          nlayers = box.d;
 809          box.y = box.z = 0;
 810          box.d = 1;
 811          break;
 812       default:
 813          box.z = transfer->box.z;
 814          break;
 815       }
 816
 817       if (0)
 818          debug_printf("%s %d, %d, %d  %d x %d x %d\n",
 819                       __FUNCTION__,
 820                       box.x, box.y, box.z,
 821                       box.w, box.h, box.d);
 822
 823       if (svga_have_vgpu10(svga)) {
 824          unsigned i;
 825          for (i = 0; i < nlayers; i++) {
 826             ret = update_image_vgpu10(svga, surf, &box,
 827                                       st->slice + i, transfer->level,
 828                                       tex->b.b.last_level + 1);
 829             assert(ret == PIPE_OK);
 830          }
 831       } else {
 832          assert(nlayers == 1);
 833          ret = update_image_vgpu9(svga, surf, &box, st->slice, transfer->level);
 834          assert(ret == PIPE_OK);
 835       }
 836       (void) ret;
 837    }
 838 }
 839
 840 static void
 841 svga_texture_transfer_unmap(struct pipe_context *pipe,
 842                             struct pipe_transfer *transfer)
 843 {
 844    struct svga_context *svga = svga_context(pipe);
 845    struct svga_screen *ss = svga_screen(pipe->screen);
 846    struct svga_winsys_screen *sws = ss->sws;
 847    struct svga_transfer *st = svga_transfer(transfer);
 848    struct svga_texture *tex = svga_texture(transfer->resource);
 849
 850    SVGA_STATS_TIME_PUSH(sws, SVGA_STATS_TIME_TEXTRANSFERUNMAP);
 851
 852    if (!st->use_direct_map) {
 853       svga_texture_transfer_unmap_dma(svga, st);
 854    }
 855    else if (st->upload.buf) {
 856       svga_texture_transfer_unmap_upload(svga, st);
 857    }
 858    else {
 859       svga_texture_transfer_unmap_direct(svga, st);
 860    }
 861
 862    if (st->base.usage & PIPE_TRANSFER_WRITE) {
 863       svga->hud.num_resource_updates++;
 864
 865       /* Mark the texture level as dirty */
 866       ss->texture_timestamp++;
 867       svga_age_texture_view(tex, transfer->level);
 868       if (transfer->resource->target == PIPE_TEXTURE_CUBE)
 869          svga_define_texture_level(tex, st->slice, transfer->level);
 870       else
 871          svga_define_texture_level(tex, 0, transfer->level);
 872    }
 873
 874    pipe_resource_reference(&st->base.resource, NULL);
 875    FREE(st);
 876    SVGA_STATS_TIME_POP(sws);
 877    (void) sws;
 878 }
 879
 880
 881 /**
 882  * Does format store depth values?
 883  */
 884 static inline boolean
 885 format_has_depth(enum pipe_format format)
 886 {
 887    const struct util_format_description *desc = util_format_description(format);
 888    return util_format_has_depth(desc);
 889 }
 890
 891
 892 struct u_resource_vtbl svga_texture_vtbl =
 893 {
 894    svga_texture_get_handle,           /* get_handle */
 895    svga_texture_destroy,              /* resource_destroy */
 896    svga_texture_transfer_map,         /* transfer_map */
 897    u_default_transfer_flush_region,   /* transfer_flush_region */
 898    svga_texture_transfer_unmap,       /* transfer_unmap */
 899 };
 900
 901
 902 struct pipe_resource *
 903 svga_texture_create(struct pipe_screen *screen,
 904                     const struct pipe_resource *template)
 905 {
 906    struct svga_screen *svgascreen = svga_screen(screen);
 907    struct svga_texture *tex;
 908    unsigned bindings = template->bind;
 909
 910    SVGA_STATS_TIME_PUSH(svgascreen->sws,
 911                         SVGA_STATS_TIME_CREATETEXTURE);
 912
 913    assert(template->last_level < SVGA_MAX_TEXTURE_LEVELS);
 914    if (template->last_level >= SVGA_MAX_TEXTURE_LEVELS) {
 915       goto fail_notex;
 916    }
 917
 918    tex = CALLOC_STRUCT(svga_texture);
 919    if (!tex) {
 920       goto fail_notex;
 921    }
 922
 923    tex->defined = CALLOC(template->depth0 * template->array_size,
 924                          sizeof(tex->defined[0]));
 925    if (!tex->defined) {
 926       FREE(tex);
 927       goto fail_notex;
 928    }
 929
 930    tex->rendered_to = CALLOC(template->depth0 * template->array_size,
 931                              sizeof(tex->rendered_to[0]));
 932    if (!tex->rendered_to) {
 933       goto fail;
 934    }
 935
 936    tex->dirty = CALLOC(template->depth0 * template->array_size,
 937                              sizeof(tex->dirty[0]));
 938    if (!tex->dirty) {
 939       goto fail;
 940    }
 941
 942    tex->b.b = *template;
 943    tex->b.vtbl = &svga_texture_vtbl;
 944    pipe_reference_init(&tex->b.b.reference, 1);
 945    tex->b.b.screen = screen;
 946
 947    tex->key.flags = 0;
 948    tex->key.size.width = template->width0;
 949    tex->key.size.height = template->height0;
 950    tex->key.size.depth = template->depth0;
 951    tex->key.arraySize = 1;
 952    tex->key.numFaces = 1;
 953
 954    /* single sample texture can be treated as non-multisamples texture */
 955    tex->key.sampleCount = template->nr_samples > 1 ? template->nr_samples : 0;
 956
 957    if (template->nr_samples > 1) {
 958       tex->key.flags |= SVGA3D_SURFACE_MASKABLE_ANTIALIAS;
 959    }
 960
 961    if (svgascreen->sws->have_vgpu10) {
 962       switch (template->target) {
 963       case PIPE_TEXTURE_1D:
 964          tex->key.flags |= SVGA3D_SURFACE_1D;
 965          break;
 966       case PIPE_TEXTURE_1D_ARRAY:
 967          tex->key.flags |= SVGA3D_SURFACE_1D;
 968          /* fall-through */
 969       case PIPE_TEXTURE_2D_ARRAY:
 970          tex->key.flags |= SVGA3D_SURFACE_ARRAY;
 971          tex->key.arraySize = template->array_size;
 972          break;
 973       case PIPE_TEXTURE_3D:
 974          tex->key.flags |= SVGA3D_SURFACE_VOLUME;
 975          break;
 976       case PIPE_TEXTURE_CUBE:
 977          tex->key.flags |= (SVGA3D_SURFACE_CUBEMAP | SVGA3D_SURFACE_ARRAY);
 978          tex->key.numFaces = 6;
 979          break;
 980       default:
 981          break;
 982       }
 983    }
 984    else {
 985       switch (template->target) {
 986       case PIPE_TEXTURE_3D:
 987          tex->key.flags |= SVGA3D_SURFACE_VOLUME;
 988          break;
 989       case PIPE_TEXTURE_CUBE:
 990          tex->key.flags |= SVGA3D_SURFACE_CUBEMAP;
 991          tex->key.numFaces = 6;
 992          break;
 993       default:
 994          break;
 995       }
 996    }
 997
 998    tex->key.cachable = 1;
 999
1000    if ((bindings & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL)) &&
1001        !(bindings & PIPE_BIND_SAMPLER_VIEW)) {
1002       /* Also check if the format can be sampled from */
1003       if (screen->is_format_supported(screen, template->format,
1004                                       template->target,
1005                                       template->nr_samples,
1006                                       PIPE_BIND_SAMPLER_VIEW)) {
1007          bindings |= PIPE_BIND_SAMPLER_VIEW;
1008       }
1009    }
1010
1011    if (bindings & PIPE_BIND_SAMPLER_VIEW) {
1012       tex->key.flags |= SVGA3D_SURFACE_HINT_TEXTURE;
1013       tex->key.flags |= SVGA3D_SURFACE_BIND_SHADER_RESOURCE;
1014
1015       if (!(bindings & PIPE_BIND_RENDER_TARGET)) {
1016          /* Also check if the format is color renderable */
1017          if (screen->is_format_supported(screen, template->format,
1018                                          template->target,
1019                                          template->nr_samples,
1020                                          PIPE_BIND_RENDER_TARGET)) {
1021             bindings |= PIPE_BIND_RENDER_TARGET;
1022          }
1023       }
1024
1025       if (!(bindings & PIPE_BIND_DEPTH_STENCIL)) {
1026          /* Also check if the format is depth/stencil renderable */
1027          if (screen->is_format_supported(screen, template->format,
1028                                          template->target,
1029                                          template->nr_samples,
1030                                          PIPE_BIND_DEPTH_STENCIL)) {
1031             bindings |= PIPE_BIND_DEPTH_STENCIL;
1032          }
1033       }
1034    }
1035
1036    if (bindings & PIPE_BIND_DISPLAY_TARGET) {
1037       tex->key.cachable = 0;
1038    }
1039
1040    if (bindings & PIPE_BIND_SHARED) {
1041       tex->key.cachable = 0;
1042    }
1043
1044    if (bindings & (PIPE_BIND_SCANOUT | PIPE_BIND_CURSOR)) {
1045       tex->key.scanout = 1;
1046       tex->key.cachable = 0;
1047    }
1048
1049    /*
1050     * Note: Previously we never passed the
1051     * SVGA3D_SURFACE_HINT_RENDERTARGET hint. Mesa cannot
1052     * know beforehand whether a texture will be used as a rendertarget or not
1053     * and it always requests PIPE_BIND_RENDER_TARGET, therefore
1054     * passing the SVGA3D_SURFACE_HINT_RENDERTARGET here defeats its purpose.
1055     *
1056     * However, this was changed since other state trackers
1057     * (XA for example) uses it accurately and certain device versions
1058     * relies on it in certain situations to render correctly.
1059     */
1060    if ((bindings & PIPE_BIND_RENDER_TARGET) &&
1061        !util_format_is_s3tc(template->format)) {
1062       tex->key.flags |= SVGA3D_SURFACE_HINT_RENDERTARGET;
1063       tex->key.flags |= SVGA3D_SURFACE_BIND_RENDER_TARGET;
1064    }
1065
1066    if (bindings & PIPE_BIND_DEPTH_STENCIL) {
1067       tex->key.flags |= SVGA3D_SURFACE_HINT_DEPTHSTENCIL;
1068       tex->key.flags |= SVGA3D_SURFACE_BIND_DEPTH_STENCIL;
1069    }
1070
1071    tex->key.numMipLevels = template->last_level + 1;
1072
1073    tex->key.format = svga_translate_format(svgascreen, template->format,
1074                                            bindings);
1075    if (tex->key.format == SVGA3D_FORMAT_INVALID) {
1076       goto fail;
1077    }
1078
1079    /* Use typeless formats for sRGB and depth resources.  Typeless
1080     * formats can be reinterpreted as other formats.  For example,
1081     * SVGA3D_R8G8B8A8_UNORM_TYPELESS can be interpreted as
1082     * SVGA3D_R8G8B8A8_UNORM_SRGB or SVGA3D_R8G8B8A8_UNORM.
1083     */
1084    if (svgascreen->sws->have_vgpu10 &&
1085        (util_format_is_srgb(template->format) ||
1086         format_has_depth(template->format))) {
1087       SVGA3dSurfaceFormat typeless = svga_typeless_format(tex->key.format);
1088       if (0) {
1089          debug_printf("Convert resource type %s -> %s (bind 0x%x)\n",
1090                       svga_format_name(tex->key.format),
1091                       svga_format_name(typeless),
1092                       bindings);
1093       }
1094
1095       if (svga_format_is_uncompressed_snorm(tex->key.format)) {
1096          /* We can't normally render to snorm surfaces, but once we
1097           * substitute a typeless format, we can if the rendertarget view
1098           * is unorm.  This can happen with GL_ARB_copy_image.
1099           */
1100          tex->key.flags |= SVGA3D_SURFACE_HINT_RENDERTARGET;
1101          tex->key.flags |= SVGA3D_SURFACE_BIND_RENDER_TARGET;
1102       }
1103
1104       tex->key.format = typeless;
1105    }
1106
1107    SVGA_DBG(DEBUG_DMA, "surface_create for texture\n", tex->handle);
1108    tex->handle = svga_screen_surface_create(svgascreen, bindings,
1109                                             tex->b.b.usage,
1110                                             &tex->validated, &tex->key);
1111    if (!tex->handle) {
1112       goto fail;
1113    }
1114
1115    SVGA_DBG(DEBUG_DMA, "  --> got sid %p (texture)\n", tex->handle);
1116
1117    debug_reference(&tex->b.b.reference,
1118                    (debug_reference_descriptor)debug_describe_resource, 0);
1119
1120    tex->size = util_resource_size(template);
1121
1122    /* Determine if texture upload buffer can be used to upload this texture */
1123    tex->can_use_upload = svga_texture_transfer_map_can_upload(svgascreen,
1124                                                               &tex->b.b);
1125
1126    /* Initialize the backing resource cache */
1127    tex->backed_handle = NULL;
1128
1129    svgascreen->hud.total_resource_bytes += tex->size;
1130    svgascreen->hud.num_resources++;
1131
1132    SVGA_STATS_TIME_POP(svgascreen->sws);
1133
1134    return &tex->b.b;
1135
1136 fail:
1137    if (tex->dirty)
1138       FREE(tex->dirty);
1139    if (tex->rendered_to)
1140       FREE(tex->rendered_to);
1141    if (tex->defined)
1142       FREE(tex->defined);
1143    FREE(tex);
1144 fail_notex:
1145    SVGA_STATS_TIME_POP(svgascreen->sws);
1146    return NULL;
1147 }
1148
1149
1150 struct pipe_resource *
1151 svga_texture_from_handle(struct pipe_screen *screen,
1152                          const struct pipe_resource *template,
1153                          struct winsys_handle *whandle)
1154 {
1155    struct svga_winsys_screen *sws = svga_winsys_screen(screen);
1156    struct svga_screen *ss = svga_screen(screen);
1157    struct svga_winsys_surface *srf;
1158    struct svga_texture *tex;
1159    enum SVGA3dSurfaceFormat format = 0;
1160    assert(screen);
1161
1162    /* Only supports one type */
1163    if ((template->target != PIPE_TEXTURE_2D &&
1164        template->target != PIPE_TEXTURE_RECT) ||
1165        template->last_level != 0 ||
1166        template->depth0 != 1) {
1167       return NULL;
1168    }
1169
1170    srf = sws->surface_from_handle(sws, whandle, &format);
1171
1172    if (!srf)
1173       return NULL;
1174
1175    if (!svga_format_is_shareable(ss, template->format, format,
1176                                  template->bind, true))
1177       goto out_unref;
1178
1179    tex = CALLOC_STRUCT(svga_texture);
1180    if (!tex)
1181       goto out_unref;
1182
1183    tex->defined = CALLOC(template->depth0 * template->array_size,
1184                          sizeof(tex->defined[0]));
1185    if (!tex->defined)
1186       goto out_no_defined;
1187
1188    tex->b.b = *template;
1189    tex->b.vtbl = &svga_texture_vtbl;
1190    pipe_reference_init(&tex->b.b.reference, 1);
1191    tex->b.b.screen = screen;
1192
1193    SVGA_DBG(DEBUG_DMA, "wrap surface sid %p\n", srf);
1194
1195    tex->key.cachable = 0;
1196    tex->key.format = format;
1197    tex->handle = srf;
1198
1199    tex->rendered_to = CALLOC(1, sizeof(tex->rendered_to[0]));
1200    if (!tex->rendered_to)
1201       goto out_no_rendered_to;
1202
1203    tex->dirty = CALLOC(1, sizeof(tex->dirty[0]));
1204    if (!tex->dirty)
1205       goto out_no_dirty;
1206
1207    tex->imported = TRUE;
1208
1209    ss->hud.num_resources++;
1210
1211    return &tex->b.b;
1212
1213 out_no_dirty:
1214    FREE(tex->rendered_to);
1215 out_no_rendered_to:
1216    FREE(tex->defined);
1217 out_no_defined:
1218    FREE(tex);
1219 out_unref:
1220    sws->surface_reference(sws, &srf, NULL);
1221    return NULL;
1222 }
1223
1224 boolean
1225 svga_texture_generate_mipmap(struct pipe_context *pipe,
1226                              struct pipe_resource *pt,
1227                              enum pipe_format format,
1228                              unsigned base_level,
1229                              unsigned last_level,
1230                              unsigned first_layer,
1231                              unsigned last_layer)
1232 {
1233    struct pipe_sampler_view templ, *psv;
1234    struct svga_pipe_sampler_view *sv;
1235    struct svga_context *svga = svga_context(pipe);
1236    struct svga_texture *tex = svga_texture(pt);
1237    enum pipe_error ret;
1238
1239    assert(svga_have_vgpu10(svga));
1240
1241    /* Only support 2D texture for now */
1242    if (pt->target != PIPE_TEXTURE_2D)
1243       return FALSE;
1244
1245    /* Fallback to the mipmap generation utility for those formats that
1246     * do not support hw generate mipmap
1247     */
1248    if (!svga_format_support_gen_mips(format))
1249       return FALSE;
1250
1251    /* Make sure the texture surface was created with
1252     * SVGA3D_SURFACE_BIND_RENDER_TARGET
1253     */
1254    if (!tex->handle || !(tex->key.flags & SVGA3D_SURFACE_BIND_RENDER_TARGET))
1255       return FALSE;
1256
1257    templ.format = format;
1258    templ.u.tex.first_layer = first_layer;
1259    templ.u.tex.last_layer = last_layer;
1260    templ.u.tex.first_level = base_level;
1261    templ.u.tex.last_level = last_level;
1262
1263    psv = pipe->create_sampler_view(pipe, pt, &templ);
1264    if (psv == NULL)
1265       return FALSE;
1266
1267    sv = svga_pipe_sampler_view(psv);
1268    ret = svga_validate_pipe_sampler_view(svga, sv);
1269    if (ret != PIPE_OK) {
1270       svga_context_flush(svga, NULL);
1271       ret = svga_validate_pipe_sampler_view(svga, sv);
1272       assert(ret == PIPE_OK);
1273    }
1274
1275    ret = SVGA3D_vgpu10_GenMips(svga->swc, sv->id, tex->handle);
1276    if (ret != PIPE_OK) {
1277       svga_context_flush(svga, NULL);
1278       ret = SVGA3D_vgpu10_GenMips(svga->swc, sv->id, tex->handle);
1279    }
1280    pipe_sampler_view_reference(&psv, NULL);
1281
1282    svga->hud.num_generate_mipmap++;
1283
1284    return TRUE;
1285 }
1286
1287
1288 /* texture upload buffer default size in bytes */
1289 #define TEX_UPLOAD_DEFAULT_SIZE (1024 * 1024)
1290
1291 /**
1292  * Create a texture upload buffer
1293  */
1294 boolean
1295 svga_texture_transfer_map_upload_create(struct svga_context *svga)
1296 {
1297    svga->tex_upload = u_upload_create(&svga->pipe, TEX_UPLOAD_DEFAULT_SIZE,
1298                                       0, PIPE_USAGE_STAGING);
1299    return svga->tex_upload != NULL;
1300 }
1301
1302
1303 /**
1304  * Destroy the texture upload buffer
1305  */
1306 void
1307 svga_texture_transfer_map_upload_destroy(struct svga_context *svga)
1308 {
1309    u_upload_destroy(svga->tex_upload);
1310 }
1311
1312
1313 /**
1314  * Returns true if this transfer map request can use the upload buffer.
1315  */
1316 boolean
1317 svga_texture_transfer_map_can_upload(const struct svga_screen *svgascreen,
1318                                      const struct pipe_resource *texture)
1319 {
1320    if (svgascreen->sws->have_transfer_from_buffer_cmd == FALSE)
1321       return FALSE;
1322
1323    /* TransferFromBuffer command is not well supported with multi-samples surface */
1324    if (texture->nr_samples > 1)
1325       return FALSE;
1326
1327    if (util_format_is_compressed(texture->format)) {
1328       /* XXX Need to take a closer look to see why texture upload
1329        * with 3D texture with compressed format fails
1330        */
1331       if (texture->target == PIPE_TEXTURE_3D)
1332           return FALSE;
1333    }
1334    else if (texture->format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
1335       return FALSE;
1336    }
1337
1338    return TRUE;
1339 }
1340
1341
1342 /**
1343  * Use upload buffer for the transfer map request.
1344  */
1345 void *
1346 svga_texture_transfer_map_upload(struct svga_context *svga,
1347                                  struct svga_transfer *st)
1348 {
1349    struct pipe_resource *texture = st->base.resource;
1350    struct pipe_resource *tex_buffer = NULL;
1351    void *tex_map;
1352    unsigned nblocksx, nblocksy;
1353    unsigned offset;
1354    unsigned upload_size;
1355
1356    assert(svga->tex_upload);
1357
1358    st->upload.box.x = st->base.box.x;
1359    st->upload.box.y = st->base.box.y;
1360    st->upload.box.z = st->base.box.z;
1361    st->upload.box.w = st->base.box.width;
1362    st->upload.box.h = st->base.box.height;
1363    st->upload.box.d = st->base.box.depth;
1364    st->upload.nlayers = 1;
1365
1366    switch (texture->target) {
1367    case PIPE_TEXTURE_CUBE:
1368       st->upload.box.z = 0;
1369       break;
1370    case PIPE_TEXTURE_2D_ARRAY:
1371       st->upload.nlayers = st->base.box.depth;
1372       st->upload.box.z = 0;
1373       st->upload.box.d = 1;
1374       break;
1375    case PIPE_TEXTURE_1D_ARRAY:
1376       st->upload.nlayers = st->base.box.depth;
1377       st->upload.box.y = st->upload.box.z = 0;
1378       st->upload.box.d = 1;
1379       break;
1380    default:
1381       break;
1382    }
1383
1384    nblocksx = util_format_get_nblocksx(texture->format, st->base.box.width);
1385    nblocksy = util_format_get_nblocksy(texture->format, st->base.box.height);
1386
1387    st->base.stride = nblocksx * util_format_get_blocksize(texture->format);
1388    st->base.layer_stride = st->base.stride * nblocksy;
1389
1390    /* In order to use the TransferFromBuffer command to update the
1391     * texture content from the buffer, the layer stride for a multi-layers
1392     * surface needs to be in multiples of 16 bytes.
1393     */
1394    if (st->upload.nlayers > 1 && st->base.layer_stride & 15)
1395       return NULL;
1396
1397    upload_size = st->base.layer_stride * st->base.box.depth;
1398    upload_size = align(upload_size, 16);
1399
1400 #ifdef DEBUG
1401    if (util_format_is_compressed(texture->format)) {
1402       struct svga_texture *tex = svga_texture(texture);
1403       unsigned blockw, blockh, bytesPerBlock;
1404
1405       svga_format_size(tex->key.format, &blockw, &blockh, &bytesPerBlock);
1406
1407       /* dest box must start on block boundary */
1408       assert((st->base.box.x % blockw) == 0);
1409       assert((st->base.box.y % blockh) == 0);
1410    }
1411 #endif
1412
1413    /* If the upload size exceeds the default buffer size, the
1414     * upload buffer manager code will try to allocate a new buffer
1415     * with the new buffer size.
1416     */
1417    u_upload_alloc(svga->tex_upload, 0, upload_size, 16,
1418                   &offset, &tex_buffer, &tex_map);
1419
1420    if (!tex_map) {
1421       return NULL;
1422    }
1423
1424    st->upload.buf = tex_buffer;
1425    st->upload.map = tex_map;
1426    st->upload.offset = offset;
1427
1428    return tex_map;
1429 }
1430
1431
1432 /**
1433  * Unmap upload map transfer request
1434  */
1435 void
1436 svga_texture_transfer_unmap_upload(struct svga_context *svga,
1437                                    struct svga_transfer *st)
1438 {
1439    struct svga_winsys_surface *srcsurf;
1440    struct svga_winsys_surface *dstsurf;
1441    struct pipe_resource *texture = st->base.resource;
1442    struct svga_texture *tex = svga_texture(texture);
1443    enum pipe_error ret;
1444    unsigned subResource;
1445    unsigned numMipLevels;
1446    unsigned i, layer;
1447    unsigned offset = st->upload.offset;
1448
1449    assert(svga->tex_upload);
1450    assert(st->upload.buf);
1451
1452    /* unmap the texture upload buffer */
1453    u_upload_unmap(svga->tex_upload);
1454
1455    srcsurf = svga_buffer_handle(svga, st->upload.buf);
1456    dstsurf = svga_texture(texture)->handle;
1457    assert(dstsurf);
1458
1459    numMipLevels = texture->last_level + 1;
1460
1461    for (i = 0, layer = st->slice; i < st->upload.nlayers; i++, layer++) {
1462       subResource = layer * numMipLevels + st->base.level;
1463
1464       /* send a transferFromBuffer command to update the host texture surface */
1465       assert((offset & 15) == 0);
1466
1467       ret = SVGA3D_vgpu10_TransferFromBuffer(svga->swc, srcsurf,
1468                                              offset,
1469                                              st->base.stride,
1470                                              st->base.layer_stride,
1471                                              dstsurf, subResource,
1472                                              &st->upload.box);
1473       if (ret != PIPE_OK) {
1474          svga_context_flush(svga, NULL);
1475          ret = SVGA3D_vgpu10_TransferFromBuffer(svga->swc, srcsurf,
1476                                                 offset,
1477                                                 st->base.stride,
1478                                                 st->base.layer_stride,
1479                                                 dstsurf, subResource,
1480                                                 &st->upload.box);
1481          assert(ret == PIPE_OK);
1482       }
1483       offset += st->base.layer_stride;
1484
1485       /* Set rendered-to flag */
1486       svga_set_texture_rendered_to(tex, layer, st->base.level);
1487    }
1488
1489    pipe_resource_reference(&st->upload.buf, NULL);
1490 }