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 "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_format.h"
  33 #include "util/u_math.h"
  34 #include "util/u_memory.h"
  35 #include "util/u_resource.h"
  36
  37 #include "svga_format.h"
  38 #include "svga_screen.h"
  39 #include "svga_context.h"
  40 #include "svga_resource_texture.h"
  41 #include "svga_resource_buffer.h"
  42 #include "svga_sampler_view.h"
  43 #include "svga_winsys.h"
  44 #include "svga_debug.h"
  45
  46
  47 /* XXX: This isn't a real hardware flag, but just a hack for kernel to
  48  * know about primary surfaces. Find a better way to accomplish this.
  49  */
  50 #define SVGA3D_SURFACE_HINT_SCANOUT (1 << 9)
  51
  52
  53 static INLINE void
  54 svga_transfer_dma_band(struct svga_context *svga,
  55                        struct svga_transfer *st,
  56                        SVGA3dTransferType transfer,
  57                        unsigned y, unsigned h, unsigned srcy,
  58                        SVGA3dSurfaceDMAFlags flags)
  59 {
  60    struct svga_texture *texture = svga_texture(st->base.resource);
  61    SVGA3dCopyBox box;
  62    enum pipe_error ret;
  63
  64    box.x = st->base.box.x;
  65    box.y = y;
  66    box.z = st->base.box.z;
  67    box.w = st->base.box.width;
  68    box.h = h;
  69    box.d = 1;
  70    box.srcx = 0;
  71    box.srcy = srcy;
  72    box.srcz = 0;
  73
  74    if (st->base.resource->target == PIPE_TEXTURE_CUBE) {
  75       st->face = st->base.box.z;
  76       box.z = 0;
  77    }
  78    else
  79       st->face = 0;
  80
  81    SVGA_DBG(DEBUG_DMA, "dma %s sid %p, face %u, (%u, %u, %u) - (%u, %u, %u), %ubpp\n",
  82                 transfer == SVGA3D_WRITE_HOST_VRAM ? "to" : "from",
  83                 texture->handle,
  84                 st->face,
  85                 st->base.box.x,
  86                 y,
  87                 box.z,
  88                 st->base.box.x + st->base.box.width,
  89                 y + h,
  90                 box.z + 1,
  91                 util_format_get_blocksize(texture->b.b.format) * 8 /
  92                 (util_format_get_blockwidth(texture->b.b.format)*util_format_get_blockheight(texture->b.b.format)));
  93
  94    ret = SVGA3D_SurfaceDMA(svga->swc, st, transfer, &box, 1, flags);
  95    if(ret != PIPE_OK) {
  96       svga_context_flush(svga, NULL);
  97       ret = SVGA3D_SurfaceDMA(svga->swc, st, transfer, &box, 1, flags);
  98       assert(ret == PIPE_OK);
  99    }
 100 }
 101
 102
 103 static INLINE void
 104 svga_transfer_dma(struct svga_context *svga,
 105                   struct svga_transfer *st,
 106                   SVGA3dTransferType transfer,
 107                   SVGA3dSurfaceDMAFlags flags)
 108 {
 109    struct svga_texture *texture = svga_texture(st->base.resource);
 110    struct svga_screen *screen = svga_screen(texture->b.b.screen);
 111    struct svga_winsys_screen *sws = screen->sws;
 112    struct pipe_fence_handle *fence = NULL;
 113
 114    if (transfer == SVGA3D_READ_HOST_VRAM) {
 115       SVGA_DBG(DEBUG_PERF, "%s: readback transfer\n", __FUNCTION__);
 116    }
 117
 118    /* Ensure any pending operations on host surfaces are queued on the command
 119     * buffer first.
 120     */
 121    svga_surfaces_flush( svga );
 122
 123    if(!st->swbuf) {
 124       /* Do the DMA transfer in a single go */
 125
 126       svga_transfer_dma_band(svga, st, transfer,
 127                              st->base.box.y, st->base.box.height, 0,
 128                              flags);
 129
 130       if(transfer == SVGA3D_READ_HOST_VRAM) {
 131          svga_context_flush(svga, &fence);
 132          sws->fence_finish(sws, fence, 0);
 133          sws->fence_reference(sws, &fence, NULL);
 134       }
 135    }
 136    else {
 137       int y, h, srcy;
 138       unsigned blockheight = util_format_get_blockheight(st->base.resource->format);
 139       h = st->hw_nblocksy * blockheight;
 140       srcy = 0;
 141       for(y = 0; y < st->base.box.height; y += h) {
 142          unsigned offset, length;
 143          void *hw, *sw;
 144
 145          if (y + h > st->base.box.height)
 146             h = st->base.box.height - y;
 147
 148          /* Transfer band must be aligned to pixel block boundaries */
 149          assert(y % blockheight == 0);
 150          assert(h % blockheight == 0);
 151
 152          offset = y * st->base.stride / blockheight;
 153          length = h * st->base.stride / blockheight;
 154
 155          sw = (uint8_t *)st->swbuf + offset;
 156
 157          if (transfer == SVGA3D_WRITE_HOST_VRAM) {
 158             unsigned usage = PIPE_TRANSFER_WRITE;
 159
 160             /* Wait for the previous DMAs to complete */
 161             /* TODO: keep one DMA (at half the size) in the background */
 162             if (y) {
 163                svga_context_flush(svga, NULL);
 164                usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
 165             }
 166
 167             hw = sws->buffer_map(sws, st->hwbuf, usage);
 168             assert(hw);
 169             if (hw) {
 170                memcpy(hw, sw, length);
 171                sws->buffer_unmap(sws, st->hwbuf);
 172             }
 173          }
 174
 175          svga_transfer_dma_band(svga, st, transfer, y, h, srcy, flags);
 176
 177          /*
 178           * Prevent the texture contents to be discarded on the next band
 179           * upload.
 180           */
 181
 182          flags.discard = FALSE;
 183
 184          if(transfer == SVGA3D_READ_HOST_VRAM) {
 185             svga_context_flush(svga, &fence);
 186             sws->fence_finish(sws, fence, 0);
 187
 188             hw = sws->buffer_map(sws, st->hwbuf, PIPE_TRANSFER_READ);
 189             assert(hw);
 190             if(hw) {
 191                memcpy(sw, hw, length);
 192                sws->buffer_unmap(sws, st->hwbuf);
 193             }
 194          }
 195       }
 196    }
 197 }
 198
 199
 200 static boolean
 201 svga_texture_get_handle(struct pipe_screen *screen,
 202                                struct pipe_resource *texture,
 203                                struct winsys_handle *whandle)
 204 {
 205    struct svga_winsys_screen *sws = svga_winsys_screen(texture->screen);
 206    unsigned stride;
 207
 208    assert(svga_texture(texture)->key.cachable == 0);
 209    svga_texture(texture)->key.cachable = 0;
 210    stride = util_format_get_nblocksx(texture->format, texture->width0) *
 211             util_format_get_blocksize(texture->format);
 212    return sws->surface_get_handle(sws, svga_texture(texture)->handle, stride, whandle);
 213 }
 214
 215
 216 static void
 217 svga_texture_destroy(struct pipe_screen *screen,
 218                      struct pipe_resource *pt)
 219 {
 220    struct svga_screen *ss = svga_screen(screen);
 221    struct svga_texture *tex = svga_texture(pt);
 222
 223    ss->texture_timestamp++;
 224
 225    svga_sampler_view_reference(&tex->cached_view, NULL);
 226
 227    /*
 228      DBG("%s deleting %p\n", __FUNCTION__, (void *) tex);
 229    */
 230    SVGA_DBG(DEBUG_DMA, "unref sid %p (texture)\n", tex->handle);
 231    svga_screen_surface_destroy(ss, &tex->key, &tex->handle);
 232
 233    ss->total_resource_bytes -= tex->size;
 234
 235    FREE(tex);
 236 }
 237
 238
 239 /* XXX: Still implementing this as if it was a screen function, but
 240  * can now modify it to queue transfers on the context.
 241  */
 242 static void *
 243 svga_texture_transfer_map(struct pipe_context *pipe,
 244                           struct pipe_resource *texture,
 245                           unsigned level,
 246                           unsigned usage,
 247                           const struct pipe_box *box,
 248                           struct pipe_transfer **ptransfer)
 249 {
 250    struct svga_context *svga = svga_context(pipe);
 251    struct svga_screen *ss = svga_screen(pipe->screen);
 252    struct svga_winsys_screen *sws = ss->sws;
 253    struct svga_transfer *st;
 254    unsigned nblocksx = util_format_get_nblocksx(texture->format, box->width);
 255    unsigned nblocksy = util_format_get_nblocksy(texture->format, box->height);
 256
 257    /* We can't map texture storage directly */
 258    if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
 259       return NULL;
 260
 261    assert(box->depth == 1);
 262    st = CALLOC_STRUCT(svga_transfer);
 263    if (!st)
 264       return NULL;
 265
 266    st->base.resource = texture;
 267    st->base.level = level;
 268    st->base.usage = usage;
 269    st->base.box = *box;
 270    st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
 271    st->base.layer_stride = 0;
 272
 273    st->hw_nblocksy = nblocksy;
 274
 275    st->hwbuf = svga_winsys_buffer_create(svga,
 276                                          1,
 277                                          0,
 278                                          st->hw_nblocksy*st->base.stride);
 279    while(!st->hwbuf && (st->hw_nblocksy /= 2)) {
 280       st->hwbuf = svga_winsys_buffer_create(svga,
 281                                             1,
 282                                             0,
 283                                             st->hw_nblocksy*st->base.stride);
 284    }
 285
 286    if(!st->hwbuf)
 287       goto no_hwbuf;
 288
 289    if(st->hw_nblocksy < nblocksy) {
 290       /* We couldn't allocate a hardware buffer big enough for the transfer,
 291        * so allocate regular malloc memory instead */
 292       if (0) {
 293          debug_printf("%s: failed to allocate %u KB of DMA, "
 294                       "splitting into %u x %u KB DMA transfers\n",
 295                       __FUNCTION__,
 296                       (nblocksy*st->base.stride + 1023)/1024,
 297                       (nblocksy + st->hw_nblocksy - 1)/st->hw_nblocksy,
 298                       (st->hw_nblocksy*st->base.stride + 1023)/1024);
 299       }
 300
 301       st->swbuf = MALLOC(nblocksy*st->base.stride);
 302       if(!st->swbuf)
 303          goto no_swbuf;
 304    }
 305
 306    if (usage & PIPE_TRANSFER_READ) {
 307       SVGA3dSurfaceDMAFlags flags;
 308       memset(&flags, 0, sizeof flags);
 309       svga_transfer_dma(svga, st, SVGA3D_READ_HOST_VRAM, flags);
 310    }
 311
 312    if (st->swbuf) {
 313       *ptransfer = &st->base;
 314       return st->swbuf;
 315    } else {
 316       /* The wait for read transfers already happened when svga_transfer_dma
 317        * was called. */
 318       void *map = sws->buffer_map(sws, st->hwbuf, usage);
 319       if (!map)
 320          goto fail;
 321
 322       *ptransfer = &st->base;
 323       return map;
 324    }
 325
 326 fail:
 327    FREE(st->swbuf);
 328 no_swbuf:
 329    sws->buffer_destroy(sws, st->hwbuf);
 330 no_hwbuf:
 331    FREE(st);
 332    return NULL;
 333 }
 334
 335
 336 /* XXX: Still implementing this as if it was a screen function, but
 337  * can now modify it to queue transfers on the context.
 338  */
 339 static void
 340 svga_texture_transfer_unmap(struct pipe_context *pipe,
 341                             struct pipe_transfer *transfer)
 342 {
 343    struct svga_context *svga = svga_context(pipe);
 344    struct svga_screen *ss = svga_screen(pipe->screen);
 345    struct svga_winsys_screen *sws = ss->sws;
 346    struct svga_transfer *st = svga_transfer(transfer);
 347    struct svga_texture *tex = svga_texture(transfer->resource);
 348
 349    if(!st->swbuf)
 350       sws->buffer_unmap(sws, st->hwbuf);
 351
 352    if (st->base.usage & PIPE_TRANSFER_WRITE) {
 353       SVGA3dSurfaceDMAFlags flags;
 354
 355       memset(&flags, 0, sizeof flags);
 356       if (transfer->usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
 357          flags.discard = TRUE;
 358       }
 359       if (transfer->usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
 360          flags.unsynchronized = TRUE;
 361       }
 362
 363       svga_transfer_dma(svga, st, SVGA3D_WRITE_HOST_VRAM, flags);
 364       ss->texture_timestamp++;
 365       tex->view_age[transfer->level] = ++(tex->age);
 366       if (transfer->resource->target == PIPE_TEXTURE_CUBE)
 367          tex->defined[transfer->box.z][transfer->level] = TRUE;
 368       else
 369          tex->defined[0][transfer->level] = TRUE;
 370    }
 371
 372    FREE(st->swbuf);
 373    sws->buffer_destroy(sws, st->hwbuf);
 374    FREE(st);
 375 }
 376
 377
 378 struct u_resource_vtbl svga_texture_vtbl =
 379 {
 380    svga_texture_get_handle,           /* get_handle */
 381    svga_texture_destroy,              /* resource_destroy */
 382    svga_texture_transfer_map,         /* transfer_map */
 383    u_default_transfer_flush_region,   /* transfer_flush_region */
 384    svga_texture_transfer_unmap,       /* transfer_unmap */
 385    u_default_transfer_inline_write    /* transfer_inline_write */
 386 };
 387
 388
 389 struct pipe_resource *
 390 svga_texture_create(struct pipe_screen *screen,
 391                     const struct pipe_resource *template)
 392 {
 393    struct svga_screen *svgascreen = svga_screen(screen);
 394    struct svga_texture *tex = CALLOC_STRUCT(svga_texture);
 395
 396    if (!tex)
 397       goto error1;
 398
 399    tex->b.b = *template;
 400    tex->b.vtbl = &svga_texture_vtbl;
 401    pipe_reference_init(&tex->b.b.reference, 1);
 402    tex->b.b.screen = screen;
 403
 404    assert(template->last_level < SVGA_MAX_TEXTURE_LEVELS);
 405    if(template->last_level >= SVGA_MAX_TEXTURE_LEVELS)
 406       goto error2;
 407
 408    tex->key.flags = 0;
 409    tex->key.size.width = template->width0;
 410    tex->key.size.height = template->height0;
 411    tex->key.size.depth = template->depth0;
 412
 413    if(template->target == PIPE_TEXTURE_CUBE) {
 414       tex->key.flags |= SVGA3D_SURFACE_CUBEMAP;
 415       tex->key.numFaces = 6;
 416    }
 417    else {
 418       tex->key.numFaces = 1;
 419    }
 420
 421    if (template->target == PIPE_TEXTURE_3D) {
 422       tex->key.flags |= SVGA3D_SURFACE_HINT_VOLUME;
 423    }
 424
 425    tex->key.cachable = 1;
 426
 427    if (template->bind & PIPE_BIND_SAMPLER_VIEW)
 428       tex->key.flags |= SVGA3D_SURFACE_HINT_TEXTURE;
 429
 430    if (template->bind & PIPE_BIND_DISPLAY_TARGET) {
 431       tex->key.cachable = 0;
 432    }
 433
 434    if (template->bind & PIPE_BIND_SHARED) {
 435       tex->key.cachable = 0;
 436    }
 437
 438    if (template->bind & (PIPE_BIND_SCANOUT |
 439                          PIPE_BIND_CURSOR)) {
 440       tex->key.flags |= SVGA3D_SURFACE_HINT_SCANOUT;
 441       tex->key.cachable = 0;
 442    }
 443
 444    /*
 445     * Note: Previously we never passed the
 446     * SVGA3D_SURFACE_HINT_RENDERTARGET hint. Mesa cannot
 447     * know beforehand whether a texture will be used as a rendertarget or not
 448     * and it always requests PIPE_BIND_RENDER_TARGET, therefore
 449     * passing the SVGA3D_SURFACE_HINT_RENDERTARGET here defeats its purpose.
 450     *
 451     * However, this was changed since other state trackers
 452     * (XA for example) uses it accurately and certain device versions
 453     * relies on it in certain situations to render correctly.
 454     */
 455    if((template->bind & PIPE_BIND_RENDER_TARGET) &&
 456       !util_format_is_s3tc(template->format))
 457       tex->key.flags |= SVGA3D_SURFACE_HINT_RENDERTARGET;
 458
 459    if(template->bind & PIPE_BIND_DEPTH_STENCIL)
 460       tex->key.flags |= SVGA3D_SURFACE_HINT_DEPTHSTENCIL;
 461
 462    tex->key.numMipLevels = template->last_level + 1;
 463
 464    tex->key.format = svga_translate_format(svgascreen, template->format, template->bind);
 465    if(tex->key.format == SVGA3D_FORMAT_INVALID)
 466       goto error2;
 467
 468    SVGA_DBG(DEBUG_DMA, "surface_create for texture\n", tex->handle);
 469    tex->handle = svga_screen_surface_create(svgascreen, &tex->key);
 470    if (tex->handle)
 471       SVGA_DBG(DEBUG_DMA, "  --> got sid %p (texture)\n", tex->handle);
 472
 473    debug_reference(&tex->b.b.reference,
 474                    (debug_reference_descriptor)debug_describe_resource, 0);
 475
 476    tex->size = util_resource_size(template);
 477    svgascreen->total_resource_bytes += tex->size;
 478
 479    return &tex->b.b;
 480
 481 error2:
 482    FREE(tex);
 483 error1:
 484    return NULL;
 485 }
 486
 487
 488 struct pipe_resource *
 489 svga_texture_from_handle(struct pipe_screen *screen,
 490                          const struct pipe_resource *template,
 491                          struct winsys_handle *whandle)
 492 {
 493    struct svga_winsys_screen *sws = svga_winsys_screen(screen);
 494    struct svga_winsys_surface *srf;
 495    struct svga_texture *tex;
 496    enum SVGA3dSurfaceFormat format = 0;
 497    assert(screen);
 498
 499    /* Only supports one type */
 500    if ((template->target != PIPE_TEXTURE_2D &&
 501        template->target != PIPE_TEXTURE_RECT) ||
 502        template->last_level != 0 ||
 503        template->depth0 != 1) {
 504       return NULL;
 505    }
 506
 507    srf = sws->surface_from_handle(sws, whandle, &format);
 508
 509    if (!srf)
 510       return NULL;
 511
 512    if (svga_translate_format(svga_screen(screen), template->format, template->bind) != format) {
 513       unsigned f1 = svga_translate_format(svga_screen(screen), template->format, template->bind);
 514       unsigned f2 = format;
 515
 516       /* It's okay for XRGB and ARGB or depth with/out stencil to get mixed up */
 517       if ( !( (f1 == SVGA3D_X8R8G8B8 && f2 == SVGA3D_A8R8G8B8) ||
 518               (f1 == SVGA3D_A8R8G8B8 && f2 == SVGA3D_X8R8G8B8) ||
 519               (f1 == SVGA3D_Z_D24X8 && f2 == SVGA3D_Z_D24S8) ||
 520               (f1 == SVGA3D_Z_DF24 && f2 == SVGA3D_Z_D24S8_INT) ) ) {
 521          debug_printf("%s wrong format %u != %u\n", __FUNCTION__, f1, f2);
 522          return NULL;
 523       }
 524    }
 525
 526    tex = CALLOC_STRUCT(svga_texture);
 527    if (!tex)
 528       return NULL;
 529
 530    tex->b.b = *template;
 531    tex->b.vtbl = &svga_texture_vtbl;
 532    pipe_reference_init(&tex->b.b.reference, 1);
 533    tex->b.b.screen = screen;
 534
 535    SVGA_DBG(DEBUG_DMA, "wrap surface sid %p\n", srf);
 536
 537    tex->key.cachable = 0;
 538    tex->handle = srf;
 539
 540    return &tex->b.b;
 541 }