src/mesa/drivers/dri/i965/intel_tex_image.c

   1
   2 #include "main/macros.h"
   3 #include "main/mtypes.h"
   4 #include "main/enums.h"
   5 #include "main/bufferobj.h"
   6 #include "main/context.h"
   7 #include "main/formats.h"
   8 #include "main/glformats.h"
   9 #include "main/image.h"
  10 #include "main/pbo.h"
  11 #include "main/renderbuffer.h"
  12 #include "main/texcompress.h"
  13 #include "main/texgetimage.h"
  14 #include "main/texobj.h"
  15 #include "main/teximage.h"
  16 #include "main/texstore.h"
  17
  18 #include "drivers/common/meta.h"
  19
  20 #include "intel_mipmap_tree.h"
  21 #include "intel_buffer_objects.h"
  22 #include "intel_batchbuffer.h"
  23 #include "intel_tex.h"
  24 #include "intel_blit.h"
  25 #include "intel_fbo.h"
  26 #include "intel_image.h"
  27 #include "intel_tiled_memcpy.h"
  28 #include "brw_context.h"
  29
  30 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
  31
  32 /* Make sure one doesn't end up shrinking base level zero unnecessarily.
  33  * Determining the base level dimension by shifting higher level dimension
  34  * ends up in off-by-one value in case base level has NPOT size (for example,
  35  * 293 != 146 << 1).
  36  * Choose the original base level dimension when shifted dimensions agree.
  37  * Otherwise assume real resize is intended and use the new shifted value.
  38  */
  39 static unsigned
  40 get_base_dim(unsigned old_base_dim, unsigned new_level_dim, unsigned level)
  41 {
  42    const unsigned old_level_dim = old_base_dim >> level;
  43    const unsigned new_base_dim = new_level_dim << level;
  44
  45    return old_level_dim == new_level_dim ? old_base_dim : new_base_dim;
  46 }
  47
  48 /* Work back from the specified level of the image to the baselevel and create a
  49  * miptree of that size.
  50  */
  51 struct intel_mipmap_tree *
  52 intel_miptree_create_for_teximage(struct brw_context *brw,
  53                                   struct intel_texture_object *intelObj,
  54                                   struct intel_texture_image *intelImage,
  55                                   enum intel_miptree_create_flags flags)
  56 {
  57    GLuint lastLevel;
  58    int width, height, depth;
  59    unsigned old_width = 0, old_height = 0, old_depth = 0;
  60    const struct intel_mipmap_tree *old_mt = intelObj->mt;
  61    const unsigned level = intelImage->base.Base.Level;
  62
  63    intel_get_image_dims(&intelImage->base.Base, &width, &height, &depth);
  64
  65    if (old_mt) {
  66       old_width = old_mt->surf.logical_level0_px.width;
  67       old_height = old_mt->surf.logical_level0_px.height;
  68       old_depth = old_mt->surf.dim == ISL_SURF_DIM_3D ?
  69                      old_mt->surf.logical_level0_px.depth :
  70                      old_mt->surf.logical_level0_px.array_len;
  71    }
  72
  73    DBG("%s\n", __func__);
  74
  75    /* Figure out image dimensions at start level. */
  76    switch(intelObj->base.Target) {
  77    case GL_TEXTURE_2D_MULTISAMPLE:
  78    case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
  79    case GL_TEXTURE_RECTANGLE:
  80    case GL_TEXTURE_EXTERNAL_OES:
  81       assert(level == 0);
  82       break;
  83    case GL_TEXTURE_3D:
  84       depth = old_mt ? get_base_dim(old_depth, depth, level) :
  85                        depth << level;
  86       /* Fall through */
  87    case GL_TEXTURE_2D:
  88    case GL_TEXTURE_2D_ARRAY:
  89    case GL_TEXTURE_CUBE_MAP:
  90    case GL_TEXTURE_CUBE_MAP_ARRAY:
  91       height = old_mt ? get_base_dim(old_height, height, level) :
  92                         height << level;
  93       /* Fall through */
  94    case GL_TEXTURE_1D:
  95    case GL_TEXTURE_1D_ARRAY:
  96       width = old_mt ? get_base_dim(old_width, width, level) :
  97                        width << level;
  98       break;
  99    default:
 100       unreachable("Unexpected target");
 101    }
 102
 103    /* Guess a reasonable value for lastLevel.  This is probably going
 104     * to be wrong fairly often and might mean that we have to look at
 105     * resizable buffers, or require that buffers implement lazy
 106     * pagetable arrangements.
 107     */
 108    if ((intelObj->base.Sampler.MinFilter == GL_NEAREST ||
 109         intelObj->base.Sampler.MinFilter == GL_LINEAR) &&
 110        intelImage->base.Base.Level == 0 &&
 111        !intelObj->base.GenerateMipmap) {
 112       lastLevel = 0;
 113    } else {
 114       lastLevel = _mesa_get_tex_max_num_levels(intelObj->base.Target,
 115                                                width, height, depth) - 1;
 116    }
 117
 118    return intel_miptree_create(brw,
 119                                intelObj->base.Target,
 120                                intelImage->base.Base.TexFormat,
 121                                0,
 122                                lastLevel,
 123                                width,
 124                                height,
 125                                depth,
 126                                MAX2(intelImage->base.Base.NumSamples, 1),
 127                                flags);
 128 }
 129
 130
 131 /**
 132  * \brief A fast path for glTexImage and glTexSubImage.
 133  *
 134  * This fast path is taken when the texture format is BGRA, RGBA,
 135  * A or L and when the texture memory is X- or Y-tiled.  It uploads
 136  * the texture data by mapping the texture memory without a GTT fence, thus
 137  * acquiring a tiled view of the memory, and then copying sucessive
 138  * spans within each tile.
 139  *
 140  * This is a performance win over the conventional texture upload path because
 141  * it avoids the performance penalty of writing through the write-combine
 142  * buffer. In the conventional texture upload path,
 143  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
 144  * fence, thus acquiring a linear view of the memory, then each row in the
 145  * image is memcpy'd. In this fast path, we replace each row's copy with
 146  * a sequence of copies over each linear span in tile.
 147  *
 148  * One use case is Google Chrome's paint rectangles.  Chrome (as
 149  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
 150  * Each page's content is initially uploaded with glTexImage2D and damaged
 151  * regions are updated with glTexSubImage2D. On some workloads, the
 152  * performance gain of this fastpath on Sandybridge is over 5x.
 153  */
 154 static bool
 155 intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 156                                GLuint dims,
 157                                struct gl_texture_image *texImage,
 158                                GLint xoffset, GLint yoffset, GLint zoffset,
 159                                GLsizei width, GLsizei height, GLsizei depth,
 160                                GLenum format, GLenum type,
 161                                const GLvoid *pixels,
 162                                const struct gl_pixelstore_attrib *packing)
 163 {
 164    struct brw_context *brw = brw_context(ctx);
 165    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 166    struct intel_texture_image *image = intel_texture_image(texImage);
 167    int src_pitch;
 168
 169    /* The miptree's buffer. */
 170    struct brw_bo *bo;
 171
 172    uint32_t cpp;
 173    mem_copy_fn mem_copy = NULL;
 174
 175    /* This fastpath is restricted to specific texture types:
 176     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 177     * more types.
 178     *
 179     * FINISHME: The restrictions below on packing alignment and packing row
 180     * length are likely unneeded now because we calculate the source stride
 181     * with _mesa_image_row_stride. However, before removing the restrictions
 182     * we need tests.
 183     */
 184    if (!devinfo->has_llc ||
 185        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 186        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 187          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 188        pixels == NULL ||
 189        _mesa_is_bufferobj(packing->BufferObj) ||
 190        packing->Alignment > 4 ||
 191        packing->SkipPixels > 0 ||
 192        packing->SkipRows > 0 ||
 193        (packing->RowLength != 0 && packing->RowLength != width) ||
 194        packing->SwapBytes ||
 195        packing->LsbFirst ||
 196        packing->Invert)
 197       return false;
 198
 199    /* Only a simple blit, no scale, bias or other mapping. */
 200    if (ctx->_ImageTransferState)
 201       return false;
 202
 203    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 204       return false;
 205
 206    /* If this is a nontrivial texture view, let another path handle it instead. */
 207    if (texImage->TexObject->MinLayer)
 208       return false;
 209
 210    if (!image->mt ||
 211        (image->mt->surf.tiling != ISL_TILING_X &&
 212         image->mt->surf.tiling != ISL_TILING_Y0)) {
 213       /* The algorithm is written only for X- or Y-tiled memory. */
 214       return false;
 215    }
 216
 217    /* linear_to_tiled() assumes that if the object is swizzled, it is using
 218     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 219     * true on gen5 and above.
 220     *
 221     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 222     * parts of the memory aren't swizzled at all. Userspace just can't handle
 223     * that.
 224     */
 225    if (devinfo->gen < 5 && brw->has_swizzling)
 226       return false;
 227
 228    int level = texImage->Level + texImage->TexObject->MinLevel;
 229
 230    /* Since we are going to write raw data to the miptree, we need to resolve
 231     * any pending fast color clears before we start.
 232     */
 233    assert(image->mt->surf.logical_level0_px.depth == 1);
 234    assert(image->mt->surf.logical_level0_px.array_len == 1);
 235
 236    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 237
 238    bo = image->mt->bo;
 239
 240    if (brw_batch_references(&brw->batch, bo)) {
 241       perf_debug("Flushing before mapping a referenced bo.\n");
 242       intel_batchbuffer_flush(brw);
 243    }
 244
 245    void *map = brw_bo_map(brw, bo, MAP_WRITE | MAP_RAW);
 246    if (map == NULL) {
 247       DBG("%s: failed to map bo\n", __func__);
 248       return false;
 249    }
 250
 251    src_pitch = _mesa_image_row_stride(packing, width, format, type);
 252
 253    /* We postponed printing this message until having committed to executing
 254     * the function.
 255     */
 256    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 257        "mesa_format=0x%x tiling=%d "
 258        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d) ",
 259        __func__, texImage->Level, xoffset, yoffset, width, height,
 260        format, type, texImage->TexFormat, image->mt->surf.tiling,
 261        packing->Alignment, packing->RowLength, packing->SkipPixels,
 262        packing->SkipRows);
 263
 264    /* Adjust x and y offset based on miplevel */
 265    unsigned level_x, level_y;
 266    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 267    xoffset += level_x;
 268    yoffset += level_y;
 269
 270    linear_to_tiled(
 271       xoffset * cpp, (xoffset + width) * cpp,
 272       yoffset, yoffset + height,
 273       map,
 274       pixels - (ptrdiff_t) yoffset * src_pitch - (ptrdiff_t) xoffset * cpp,
 275       image->mt->surf.row_pitch, src_pitch,
 276       brw->has_swizzling,
 277       image->mt->surf.tiling,
 278       mem_copy
 279    );
 280
 281    brw_bo_unmap(bo);
 282    return true;
 283 }
 284
 285
 286 static void
 287 intel_upload_tex(struct gl_context * ctx,
 288                  GLuint dims,
 289                  struct gl_texture_image *texImage,
 290                  GLint xoffset, GLint yoffset, GLint zoffset,
 291                  GLsizei width, GLsizei height, GLsizei depth,
 292                  GLenum format, GLenum type,
 293                  const GLvoid * pixels,
 294                  const struct gl_pixelstore_attrib *packing)
 295 {
 296    struct intel_mipmap_tree *mt = intel_texture_image(texImage)->mt;
 297    bool ok;
 298
 299    bool tex_busy = mt && brw_bo_busy(mt->bo);
 300
 301    if (mt && mt->format == MESA_FORMAT_S_UINT8)
 302       mt->r8stencil_needs_update = true;
 303
 304    ok = _mesa_meta_pbo_TexSubImage(ctx, dims, texImage,
 305                                    xoffset, yoffset, zoffset,
 306                                    width, height, depth, format, type,
 307                                    pixels, tex_busy, packing);
 308    if (ok)
 309       return;
 310
 311    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 312                                        xoffset, yoffset, zoffset,
 313                                        width, height, depth,
 314                                        format, type, pixels, packing);
 315    if (ok)
 316      return;
 317
 318    _mesa_store_texsubimage(ctx, dims, texImage,
 319                            xoffset, yoffset, zoffset,
 320                            width, height, depth,
 321                            format, type, pixels, packing);
 322 }
 323
 324
 325 static void
 326 intelTexImage(struct gl_context * ctx,
 327               GLuint dims,
 328               struct gl_texture_image *texImage,
 329               GLenum format, GLenum type, const void *pixels,
 330               const struct gl_pixelstore_attrib *unpack)
 331 {
 332    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 333        __func__, _mesa_get_format_name(texImage->TexFormat),
 334        _mesa_enum_to_string(texImage->TexObject->Target),
 335        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 336        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 337
 338    /* Allocate storage for texture data. */
 339    if (!ctx->Driver.AllocTextureImageBuffer(ctx, texImage)) {
 340       _mesa_error(ctx, GL_OUT_OF_MEMORY, "glTexImage%uD", dims);
 341       return;
 342    }
 343
 344    assert(intel_texture_image(texImage)->mt);
 345
 346    intel_upload_tex(ctx, dims, texImage, 0, 0, 0,
 347                     texImage->Width, texImage->Height, texImage->Depth,
 348                     format, type, pixels, unpack);
 349 }
 350
 351
 352 static void
 353 intelTexSubImage(struct gl_context * ctx,
 354                  GLuint dims,
 355                  struct gl_texture_image *texImage,
 356                  GLint xoffset, GLint yoffset, GLint zoffset,
 357                  GLsizei width, GLsizei height, GLsizei depth,
 358                  GLenum format, GLenum type,
 359                  const GLvoid * pixels,
 360                  const struct gl_pixelstore_attrib *packing)
 361 {
 362    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 363        __func__, _mesa_get_format_name(texImage->TexFormat),
 364        _mesa_enum_to_string(texImage->TexObject->Target),
 365        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 366        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 367
 368    intel_upload_tex(ctx, dims, texImage, xoffset, yoffset, zoffset,
 369                     width, height, depth, format, type, pixels, packing);
 370 }
 371
 372
 373 static void
 374 intel_set_texture_image_mt(struct brw_context *brw,
 375                            struct gl_texture_image *image,
 376                            GLenum internal_format,
 377                            struct intel_mipmap_tree *mt)
 378
 379 {
 380    struct gl_texture_object *texobj = image->TexObject;
 381    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 382    struct intel_texture_image *intel_image = intel_texture_image(image);
 383
 384    _mesa_init_teximage_fields(&brw->ctx, image,
 385                               mt->surf.logical_level0_px.width,
 386                               mt->surf.logical_level0_px.height, 1,
 387                               0, internal_format, mt->format);
 388
 389    brw->ctx.Driver.FreeTextureImageBuffer(&brw->ctx, image);
 390
 391    intel_texobj->needs_validate = true;
 392    intel_image->base.RowStride = mt->surf.row_pitch / mt->cpp;
 393    assert(mt->surf.row_pitch % mt->cpp == 0);
 394
 395    intel_miptree_reference(&intel_image->mt, mt);
 396
 397    /* Immediately validate the image to the object. */
 398    intel_miptree_reference(&intel_texobj->mt, mt);
 399 }
 400
 401
 402 void
 403 intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
 404                    GLint texture_format,
 405                    __DRIdrawable *dPriv)
 406 {
 407    struct gl_framebuffer *fb = dPriv->driverPrivate;
 408    struct brw_context *brw = pDRICtx->driverPrivate;
 409    struct gl_context *ctx = &brw->ctx;
 410    struct intel_renderbuffer *rb;
 411    struct gl_texture_object *texObj;
 412    struct gl_texture_image *texImage;
 413    GLenum internal_format = 0;
 414
 415    texObj = _mesa_get_current_tex_object(ctx, target);
 416
 417    if (!texObj)
 418       return;
 419
 420    if (dPriv->lastStamp != dPriv->dri2.stamp ||
 421        !pDRICtx->driScreenPriv->dri2.useInvalidate)
 422       intel_update_renderbuffers(pDRICtx, dPriv);
 423
 424    rb = intel_get_renderbuffer(fb, BUFFER_FRONT_LEFT);
 425    /* If the miptree isn't set, then intel_update_renderbuffers was unable
 426     * to get the BO for the drawable from the window system.
 427     */
 428    if (!rb || !rb->mt)
 429       return;
 430
 431    if (rb->mt->cpp == 4) {
 432       if (texture_format == __DRI_TEXTURE_FORMAT_RGB)
 433          internal_format = GL_RGB;
 434       else
 435          internal_format = GL_RGBA;
 436    } else if (rb->mt->cpp == 2) {
 437       /* This is 565 */
 438       internal_format = GL_RGB;
 439    }
 440
 441    intel_miptree_prepare_external(brw, rb->mt);
 442
 443    _mesa_lock_texture(&brw->ctx, texObj);
 444    texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
 445    intel_set_texture_image_mt(brw, texImage, internal_format, rb->mt);
 446    _mesa_unlock_texture(&brw->ctx, texObj);
 447 }
 448
 449 static GLboolean
 450 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
 451                                   struct gl_renderbuffer *rb,
 452                                   struct gl_texture_image *image)
 453 {
 454    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 455    struct intel_texture_image *intel_image = intel_texture_image(image);
 456    struct gl_texture_object *texobj = image->TexObject;
 457    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 458
 459    /* We can only handle RB allocated with AllocRenderbufferStorage, or
 460     * window-system renderbuffers.
 461     */
 462    assert(!rb->TexImage);
 463
 464    if (!irb->mt)
 465       return false;
 466
 467    _mesa_lock_texture(ctx, texobj);
 468    _mesa_init_teximage_fields(ctx, image,
 469                               rb->Width, rb->Height, 1,
 470                               0, rb->InternalFormat, rb->Format);
 471    image->NumSamples = rb->NumSamples;
 472
 473    intel_miptree_reference(&intel_image->mt, irb->mt);
 474
 475    /* Immediately validate the image to the object. */
 476    intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
 477
 478    intel_texobj->needs_validate = true;
 479    _mesa_unlock_texture(ctx, texobj);
 480
 481    return true;
 482 }
 483
 484 void
 485 intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 486 {
 487    /* The old interface didn't have the format argument, so copy our
 488     * implementation's behavior at the time.
 489     */
 490    intelSetTexBuffer2(pDRICtx, target, __DRI_TEXTURE_FORMAT_RGBA, dPriv);
 491 }
 492
 493 static void
 494 intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
 495                               struct gl_texture_object *texObj,
 496                               struct gl_texture_image *texImage,
 497                               GLeglImageOES image_handle)
 498 {
 499    struct brw_context *brw = brw_context(ctx);
 500    struct intel_mipmap_tree *mt;
 501    __DRIscreen *dri_screen = brw->screen->driScrnPriv;
 502    __DRIimage *image;
 503
 504    image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
 505                                                   dri_screen->loaderPrivate);
 506    if (image == NULL)
 507       return;
 508
 509    /* We support external textures only for EGLImages created with
 510     * EGL_EXT_image_dma_buf_import. We may lift that restriction in the future.
 511     */
 512    if (target == GL_TEXTURE_EXTERNAL_OES && !image->dma_buf_imported) {
 513       _mesa_error(ctx, GL_INVALID_OPERATION,
 514             "glEGLImageTargetTexture2DOES(external target is enabled only "
 515                "for images created with EGL_EXT_image_dma_buf_import");
 516       return;
 517    }
 518
 519    /* Disallow depth/stencil textures: we don't have a way to pass the
 520     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
 521     */
 522    if (image->has_depthstencil) {
 523       _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
 524       return;
 525    }
 526
 527    mt = intel_miptree_create_for_dri_image(brw, image, target,
 528                                            ISL_COLORSPACE_NONE, false);
 529    if (mt == NULL)
 530       return;
 531
 532    struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
 533    intel_texobj->planar_format = image->planar_format;
 534
 535    const GLenum internal_format =
 536       image->internal_format != 0 ?
 537       image->internal_format : _mesa_get_format_base_format(mt->format);
 538    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 539    intel_miptree_release(&mt);
 540 }
 541
 542 /**
 543  * \brief A fast path for glGetTexImage.
 544  *
 545  * \see intel_readpixels_tiled_memcpy()
 546  */
 547 static bool
 548 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
 549                                   struct gl_texture_image *texImage,
 550                                   GLint xoffset, GLint yoffset,
 551                                   GLsizei width, GLsizei height,
 552                                   GLenum format, GLenum type,
 553                                   GLvoid *pixels,
 554                                   const struct gl_pixelstore_attrib *packing)
 555 {
 556    struct brw_context *brw = brw_context(ctx);
 557    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 558    struct intel_texture_image *image = intel_texture_image(texImage);
 559    int dst_pitch;
 560
 561    /* The miptree's buffer. */
 562    struct brw_bo *bo;
 563
 564    uint32_t cpp;
 565    mem_copy_fn mem_copy = NULL;
 566
 567    /* This fastpath is restricted to specific texture types:
 568     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 569     * more types.
 570     *
 571     * FINISHME: The restrictions below on packing alignment and packing row
 572     * length are likely unneeded now because we calculate the destination stride
 573     * with _mesa_image_row_stride. However, before removing the restrictions
 574     * we need tests.
 575     */
 576    if (!devinfo->has_llc ||
 577        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 578        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 579          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 580        pixels == NULL ||
 581        _mesa_is_bufferobj(packing->BufferObj) ||
 582        packing->Alignment > 4 ||
 583        packing->SkipPixels > 0 ||
 584        packing->SkipRows > 0 ||
 585        (packing->RowLength != 0 && packing->RowLength != width) ||
 586        packing->SwapBytes ||
 587        packing->LsbFirst ||
 588        packing->Invert)
 589       return false;
 590
 591    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
 592     * function doesn't set the last channel to 1. Note this checks BaseFormat
 593     * rather than TexFormat in case the RGBX format is being simulated with an
 594     * RGBA format.
 595     */
 596    if (texImage->_BaseFormat == GL_RGB)
 597       return false;
 598
 599    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 600       return false;
 601
 602    /* If this is a nontrivial texture view, let another path handle it instead. */
 603    if (texImage->TexObject->MinLayer)
 604       return false;
 605
 606    if (!image->mt ||
 607        (image->mt->surf.tiling != ISL_TILING_X &&
 608         image->mt->surf.tiling != ISL_TILING_Y0)) {
 609       /* The algorithm is written only for X- or Y-tiled memory. */
 610       return false;
 611    }
 612
 613    /* tiled_to_linear() assumes that if the object is swizzled, it is using
 614     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 615     * true on gen5 and above.
 616     *
 617     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 618     * parts of the memory aren't swizzled at all. Userspace just can't handle
 619     * that.
 620     */
 621    if (devinfo->gen < 5 && brw->has_swizzling)
 622       return false;
 623
 624    int level = texImage->Level + texImage->TexObject->MinLevel;
 625
 626    /* Since we are going to write raw data to the miptree, we need to resolve
 627     * any pending fast color clears before we start.
 628     */
 629    assert(image->mt->surf.logical_level0_px.depth == 1);
 630    assert(image->mt->surf.logical_level0_px.array_len == 1);
 631
 632    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 633
 634    bo = image->mt->bo;
 635
 636    if (brw_batch_references(&brw->batch, bo)) {
 637       perf_debug("Flushing before mapping a referenced bo.\n");
 638       intel_batchbuffer_flush(brw);
 639    }
 640
 641    void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
 642    if (map == NULL) {
 643       DBG("%s: failed to map bo\n", __func__);
 644       return false;
 645    }
 646
 647    dst_pitch = _mesa_image_row_stride(packing, width, format, type);
 648
 649    DBG("%s: level=%d x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 650        "mesa_format=0x%x tiling=%d "
 651        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
 652        __func__, texImage->Level, xoffset, yoffset, width, height,
 653        format, type, texImage->TexFormat, image->mt->surf.tiling,
 654        packing->Alignment, packing->RowLength, packing->SkipPixels,
 655        packing->SkipRows);
 656
 657    /* Adjust x and y offset based on miplevel */
 658    unsigned level_x, level_y;
 659    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 660    xoffset += level_x;
 661    yoffset += level_y;
 662
 663    tiled_to_linear(
 664       xoffset * cpp, (xoffset + width) * cpp,
 665       yoffset, yoffset + height,
 666       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
 667       map,
 668       dst_pitch, image->mt->surf.row_pitch,
 669       brw->has_swizzling,
 670       image->mt->surf.tiling,
 671       mem_copy
 672    );
 673
 674    brw_bo_unmap(bo);
 675    return true;
 676 }
 677
 678 static void
 679 intel_get_tex_sub_image(struct gl_context *ctx,
 680                         GLint xoffset, GLint yoffset, GLint zoffset,
 681                         GLsizei width, GLsizei height, GLint depth,
 682                         GLenum format, GLenum type, GLvoid *pixels,
 683                         struct gl_texture_image *texImage)
 684 {
 685    struct brw_context *brw = brw_context(ctx);
 686    bool ok;
 687
 688    DBG("%s\n", __func__);
 689
 690    if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
 691       if (_mesa_meta_pbo_GetTexSubImage(ctx, 3, texImage,
 692                                         xoffset, yoffset, zoffset,
 693                                         width, height, depth, format, type,
 694                                         pixels, &ctx->Pack)) {
 695          /* Flush to guarantee coherency between the render cache and other
 696           * caches the PBO could potentially be bound to after this point.
 697           * See the related comment in intelReadPixels() for a more detailed
 698           * explanation.
 699           */
 700          brw_emit_mi_flush(brw);
 701          return;
 702       }
 703
 704       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
 705    }
 706
 707    ok = intel_gettexsubimage_tiled_memcpy(ctx, texImage, xoffset, yoffset,
 708                                           width, height,
 709                                           format, type, pixels, &ctx->Pack);
 710
 711    if(ok)
 712       return;
 713
 714    _mesa_meta_GetTexSubImage(ctx, xoffset, yoffset, zoffset,
 715                              width, height, depth,
 716                              format, type, pixels, texImage);
 717
 718    DBG("%s - DONE\n", __func__);
 719 }
 720
 721 static void
 722 flush_astc_denorms(struct gl_context *ctx, GLuint dims,
 723                    struct gl_texture_image *texImage,
 724                    GLint xoffset, GLint yoffset, GLint zoffset,
 725                    GLsizei width, GLsizei height, GLsizei depth)
 726 {
 727    struct compressed_pixelstore store;
 728    _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
 729                                        width, height, depth,
 730                                        &ctx->Unpack, &store);
 731
 732    for (int slice = 0; slice < store.CopySlices; slice++) {
 733
 734       /* Map dest texture buffer */
 735       GLubyte *dstMap;
 736       GLint dstRowStride;
 737       ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
 738                                   xoffset, yoffset, width, height,
 739                                   GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
 740                                   &dstMap, &dstRowStride);
 741       if (!dstMap)
 742          continue;
 743
 744       for (int i = 0; i < store.CopyRowsPerSlice; i++) {
 745
 746          /* An ASTC block is stored in little endian mode. The byte that
 747           * contains bits 0..7 is stored at the lower address in memory.
 748           */
 749          struct astc_void_extent {
 750             uint16_t header : 12;
 751             uint16_t dontcare[3];
 752             uint16_t R;
 753             uint16_t G;
 754             uint16_t B;
 755             uint16_t A;
 756          } *blocks = (struct astc_void_extent*) dstMap;
 757
 758          /* Iterate over every copied block in the row */
 759          for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
 760
 761             /* Check if the header matches that of an LDR void-extent block */
 762             if (blocks[j].header == 0xDFC) {
 763
 764                /* Flush UNORM16 values that would be denormalized */
 765                if (blocks[j].A < 4) blocks[j].A = 0;
 766                if (blocks[j].B < 4) blocks[j].B = 0;
 767                if (blocks[j].G < 4) blocks[j].G = 0;
 768                if (blocks[j].R < 4) blocks[j].R = 0;
 769             }
 770          }
 771
 772          dstMap += dstRowStride;
 773       }
 774
 775       ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
 776    }
 777 }
 778
 779
 780 static void
 781 intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
 782                         struct gl_texture_image *texImage,
 783                         GLint xoffset, GLint yoffset, GLint zoffset,
 784                         GLsizei width, GLsizei height, GLsizei depth,
 785                         GLenum format,
 786                         GLsizei imageSize, const GLvoid *data)
 787 {
 788    /* Upload the compressed data blocks */
 789    _mesa_store_compressed_texsubimage(ctx, dims, texImage,
 790                                       xoffset, yoffset, zoffset,
 791                                       width, height, depth,
 792                                       format, imageSize, data);
 793
 794    /* Fix up copied ASTC blocks if necessary */
 795    GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
 796                                                         texImage->TexFormat);
 797    bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
 798                         !_mesa_is_srgb_format(gl_format);
 799    struct brw_context *brw = (struct brw_context*) ctx;
 800    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 801    if (devinfo->gen == 9 && is_linear_astc)
 802       flush_astc_denorms(ctx, dims, texImage,
 803                          xoffset, yoffset, zoffset,
 804                          width, height, depth);
 805 }
 806
 807 void
 808 intelInitTextureImageFuncs(struct dd_function_table *functions)
 809 {
 810    functions->TexImage = intelTexImage;
 811    functions->TexSubImage = intelTexSubImage;
 812    functions->CompressedTexSubImage = intelCompressedTexSubImage;
 813    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
 814    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
 815    functions->GetTexSubImage = intel_get_tex_sub_image;
 816 }