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 #include "brw_blorp.h"
  30
  31 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
  32
  33 /* Make sure one doesn't end up shrinking base level zero unnecessarily.
  34  * Determining the base level dimension by shifting higher level dimension
  35  * ends up in off-by-one value in case base level has NPOT size (for example,
  36  * 293 != 146 << 1).
  37  * Choose the original base level dimension when shifted dimensions agree.
  38  * Otherwise assume real resize is intended and use the new shifted value.
  39  */
  40 static unsigned
  41 get_base_dim(unsigned old_base_dim, unsigned new_level_dim, unsigned level)
  42 {
  43    const unsigned old_level_dim = old_base_dim >> level;
  44    const unsigned new_base_dim = new_level_dim << level;
  45
  46    return old_level_dim == new_level_dim ? old_base_dim : new_base_dim;
  47 }
  48
  49 /* Work back from the specified level of the image to the baselevel and create a
  50  * miptree of that size.
  51  */
  52 struct intel_mipmap_tree *
  53 intel_miptree_create_for_teximage(struct brw_context *brw,
  54                                   struct intel_texture_object *intelObj,
  55                                   struct intel_texture_image *intelImage,
  56                                   enum intel_miptree_create_flags flags)
  57 {
  58    GLuint lastLevel;
  59    int width, height, depth;
  60    unsigned old_width = 0, old_height = 0, old_depth = 0;
  61    const struct intel_mipmap_tree *old_mt = intelObj->mt;
  62    const unsigned level = intelImage->base.Base.Level;
  63
  64    intel_get_image_dims(&intelImage->base.Base, &width, &height, &depth);
  65
  66    if (old_mt) {
  67       old_width = old_mt->surf.logical_level0_px.width;
  68       old_height = old_mt->surf.logical_level0_px.height;
  69       old_depth = old_mt->surf.dim == ISL_SURF_DIM_3D ?
  70                      old_mt->surf.logical_level0_px.depth :
  71                      old_mt->surf.logical_level0_px.array_len;
  72    }
  73
  74    DBG("%s\n", __func__);
  75
  76    /* Figure out image dimensions at start level. */
  77    switch(intelObj->base.Target) {
  78    case GL_TEXTURE_2D_MULTISAMPLE:
  79    case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
  80    case GL_TEXTURE_RECTANGLE:
  81    case GL_TEXTURE_EXTERNAL_OES:
  82       assert(level == 0);
  83       break;
  84    case GL_TEXTURE_3D:
  85       depth = old_mt ? get_base_dim(old_depth, depth, level) :
  86                        depth << level;
  87       /* Fall through */
  88    case GL_TEXTURE_2D:
  89    case GL_TEXTURE_2D_ARRAY:
  90    case GL_TEXTURE_CUBE_MAP:
  91    case GL_TEXTURE_CUBE_MAP_ARRAY:
  92       height = old_mt ? get_base_dim(old_height, height, level) :
  93                         height << level;
  94       /* Fall through */
  95    case GL_TEXTURE_1D:
  96    case GL_TEXTURE_1D_ARRAY:
  97       width = old_mt ? get_base_dim(old_width, width, level) :
  98                        width << level;
  99       break;
 100    default:
 101       unreachable("Unexpected target");
 102    }
 103
 104    /* Guess a reasonable value for lastLevel.  This is probably going
 105     * to be wrong fairly often and might mean that we have to look at
 106     * resizable buffers, or require that buffers implement lazy
 107     * pagetable arrangements.
 108     */
 109    if ((intelObj->base.Sampler.MinFilter == GL_NEAREST ||
 110         intelObj->base.Sampler.MinFilter == GL_LINEAR) &&
 111        intelImage->base.Base.Level == 0 &&
 112        !intelObj->base.GenerateMipmap) {
 113       lastLevel = 0;
 114    } else {
 115       lastLevel = _mesa_get_tex_max_num_levels(intelObj->base.Target,
 116                                                width, height, depth) - 1;
 117    }
 118
 119    return intel_miptree_create(brw,
 120                                intelObj->base.Target,
 121                                intelImage->base.Base.TexFormat,
 122                                0,
 123                                lastLevel,
 124                                width,
 125                                height,
 126                                depth,
 127                                MAX2(intelImage->base.Base.NumSamples, 1),
 128                                flags);
 129 }
 130
 131 static bool
 132 intel_texsubimage_blorp(struct brw_context *brw, GLuint dims,
 133                         struct gl_texture_image *tex_image,
 134                         unsigned x, unsigned y, unsigned z,
 135                         unsigned width, unsigned height, unsigned depth,
 136                         GLenum format, GLenum type, const void *pixels,
 137                         const struct gl_pixelstore_attrib *packing)
 138 {
 139    struct intel_texture_image *intel_image = intel_texture_image(tex_image);
 140    const unsigned mt_level = tex_image->Level + tex_image->TexObject->MinLevel;
 141    const unsigned mt_z = tex_image->TexObject->MinLayer + tex_image->Face + z;
 142
 143    /* The blorp path can't understand crazy format hackery */
 144    if (_mesa_base_tex_format(&brw->ctx, tex_image->InternalFormat) !=
 145        _mesa_get_format_base_format(tex_image->TexFormat))
 146       return false;
 147
 148    return brw_blorp_upload_miptree(brw, intel_image->mt, tex_image->TexFormat,
 149                                    mt_level, x, y, mt_z, width, height, depth,
 150                                    tex_image->TexObject->Target, format, type,
 151                                    pixels, packing);
 152 }
 153
 154 /**
 155  * \brief A fast path for glTexImage and glTexSubImage.
 156  *
 157  * This fast path is taken when the texture format is BGRA, RGBA,
 158  * A or L and when the texture memory is X- or Y-tiled.  It uploads
 159  * the texture data by mapping the texture memory without a GTT fence, thus
 160  * acquiring a tiled view of the memory, and then copying sucessive
 161  * spans within each tile.
 162  *
 163  * This is a performance win over the conventional texture upload path because
 164  * it avoids the performance penalty of writing through the write-combine
 165  * buffer. In the conventional texture upload path,
 166  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
 167  * fence, thus acquiring a linear view of the memory, then each row in the
 168  * image is memcpy'd. In this fast path, we replace each row's copy with
 169  * a sequence of copies over each linear span in tile.
 170  *
 171  * One use case is Google Chrome's paint rectangles.  Chrome (as
 172  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
 173  * Each page's content is initially uploaded with glTexImage2D and damaged
 174  * regions are updated with glTexSubImage2D. On some workloads, the
 175  * performance gain of this fastpath on Sandybridge is over 5x.
 176  */
 177 static bool
 178 intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 179                                GLuint dims,
 180                                struct gl_texture_image *texImage,
 181                                GLint xoffset, GLint yoffset, GLint zoffset,
 182                                GLsizei width, GLsizei height, GLsizei depth,
 183                                GLenum format, GLenum type,
 184                                const GLvoid *pixels,
 185                                const struct gl_pixelstore_attrib *packing)
 186 {
 187    struct brw_context *brw = brw_context(ctx);
 188    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 189    struct intel_texture_image *image = intel_texture_image(texImage);
 190    int src_pitch;
 191
 192    /* The miptree's buffer. */
 193    struct brw_bo *bo;
 194
 195    uint32_t cpp;
 196    mem_copy_fn mem_copy = NULL;
 197
 198    /* This fastpath is restricted to specific texture types:
 199     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 200     * more types.
 201     *
 202     * FINISHME: The restrictions below on packing alignment and packing row
 203     * length are likely unneeded now because we calculate the source stride
 204     * with _mesa_image_row_stride. However, before removing the restrictions
 205     * we need tests.
 206     */
 207    if (!devinfo->has_llc ||
 208        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 209        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 210          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 211        pixels == NULL ||
 212        _mesa_is_bufferobj(packing->BufferObj) ||
 213        packing->Alignment > 4 ||
 214        packing->SkipPixels > 0 ||
 215        packing->SkipRows > 0 ||
 216        (packing->RowLength != 0 && packing->RowLength != width) ||
 217        packing->SwapBytes ||
 218        packing->LsbFirst ||
 219        packing->Invert)
 220       return false;
 221
 222    /* Only a simple blit, no scale, bias or other mapping. */
 223    if (ctx->_ImageTransferState)
 224       return false;
 225
 226    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 227       return false;
 228
 229    /* If this is a nontrivial texture view, let another path handle it instead. */
 230    if (texImage->TexObject->MinLayer)
 231       return false;
 232
 233    if (!image->mt ||
 234        (image->mt->surf.tiling != ISL_TILING_X &&
 235         image->mt->surf.tiling != ISL_TILING_Y0)) {
 236       /* The algorithm is written only for X- or Y-tiled memory. */
 237       return false;
 238    }
 239
 240    /* linear_to_tiled() assumes that if the object is swizzled, it is using
 241     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 242     * true on gen5 and above.
 243     *
 244     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 245     * parts of the memory aren't swizzled at all. Userspace just can't handle
 246     * that.
 247     */
 248    if (devinfo->gen < 5 && brw->has_swizzling)
 249       return false;
 250
 251    int level = texImage->Level + texImage->TexObject->MinLevel;
 252
 253    /* Since we are going to write raw data to the miptree, we need to resolve
 254     * any pending fast color clears before we start.
 255     */
 256    assert(image->mt->surf.logical_level0_px.depth == 1);
 257    assert(image->mt->surf.logical_level0_px.array_len == 1);
 258
 259    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 260
 261    bo = image->mt->bo;
 262
 263    if (brw_batch_references(&brw->batch, bo)) {
 264       perf_debug("Flushing before mapping a referenced bo.\n");
 265       intel_batchbuffer_flush(brw);
 266    }
 267
 268    void *map = brw_bo_map(brw, bo, MAP_WRITE | MAP_RAW);
 269    if (map == NULL) {
 270       DBG("%s: failed to map bo\n", __func__);
 271       return false;
 272    }
 273
 274    src_pitch = _mesa_image_row_stride(packing, width, format, type);
 275
 276    /* We postponed printing this message until having committed to executing
 277     * the function.
 278     */
 279    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 280        "mesa_format=0x%x tiling=%d "
 281        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d) ",
 282        __func__, texImage->Level, xoffset, yoffset, width, height,
 283        format, type, texImage->TexFormat, image->mt->surf.tiling,
 284        packing->Alignment, packing->RowLength, packing->SkipPixels,
 285        packing->SkipRows);
 286
 287    /* Adjust x and y offset based on miplevel */
 288    unsigned level_x, level_y;
 289    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 290    xoffset += level_x;
 291    yoffset += level_y;
 292
 293    linear_to_tiled(
 294       xoffset * cpp, (xoffset + width) * cpp,
 295       yoffset, yoffset + height,
 296       map,
 297       pixels - (ptrdiff_t) yoffset * src_pitch - (ptrdiff_t) xoffset * cpp,
 298       image->mt->surf.row_pitch, src_pitch,
 299       brw->has_swizzling,
 300       image->mt->surf.tiling,
 301       mem_copy
 302    );
 303
 304    brw_bo_unmap(bo);
 305    return true;
 306 }
 307
 308
 309 static void
 310 intel_upload_tex(struct gl_context * ctx,
 311                  GLuint dims,
 312                  struct gl_texture_image *texImage,
 313                  GLint xoffset, GLint yoffset, GLint zoffset,
 314                  GLsizei width, GLsizei height, GLsizei depth,
 315                  GLenum format, GLenum type,
 316                  const GLvoid * pixels,
 317                  const struct gl_pixelstore_attrib *packing)
 318 {
 319    struct brw_context *brw = brw_context(ctx);
 320    struct intel_mipmap_tree *mt = intel_texture_image(texImage)->mt;
 321    bool ok;
 322
 323    /* Check that there is actually data to store. */
 324    if (pixels == NULL && !_mesa_is_bufferobj(packing->BufferObj))
 325       return;
 326
 327    bool tex_busy = mt && brw_bo_busy(mt->bo);
 328
 329    if (mt && mt->format == MESA_FORMAT_S_UINT8)
 330       mt->r8stencil_needs_update = true;
 331
 332    if (_mesa_is_bufferobj(packing->BufferObj) || tex_busy ||
 333        mt->aux_usage == ISL_AUX_USAGE_CCS_E) {
 334       ok = intel_texsubimage_blorp(brw, dims, texImage,
 335                                    xoffset, yoffset, zoffset,
 336                                    width, height, depth, format, type,
 337                                    pixels, packing);
 338       if (ok)
 339          return;
 340    }
 341
 342    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 343                                        xoffset, yoffset, zoffset,
 344                                        width, height, depth,
 345                                        format, type, pixels, packing);
 346    if (ok)
 347      return;
 348
 349    _mesa_store_texsubimage(ctx, dims, texImage,
 350                            xoffset, yoffset, zoffset,
 351                            width, height, depth,
 352                            format, type, pixels, packing);
 353 }
 354
 355
 356 static void
 357 intelTexImage(struct gl_context * ctx,
 358               GLuint dims,
 359               struct gl_texture_image *texImage,
 360               GLenum format, GLenum type, const void *pixels,
 361               const struct gl_pixelstore_attrib *unpack)
 362 {
 363    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 364        __func__, _mesa_get_format_name(texImage->TexFormat),
 365        _mesa_enum_to_string(texImage->TexObject->Target),
 366        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 367        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 368
 369    /* Allocate storage for texture data. */
 370    if (!ctx->Driver.AllocTextureImageBuffer(ctx, texImage)) {
 371       _mesa_error(ctx, GL_OUT_OF_MEMORY, "glTexImage%uD", dims);
 372       return;
 373    }
 374
 375    assert(intel_texture_image(texImage)->mt);
 376
 377    intel_upload_tex(ctx, dims, texImage, 0, 0, 0,
 378                     texImage->Width, texImage->Height, texImage->Depth,
 379                     format, type, pixels, unpack);
 380 }
 381
 382
 383 static void
 384 intelTexSubImage(struct gl_context * ctx,
 385                  GLuint dims,
 386                  struct gl_texture_image *texImage,
 387                  GLint xoffset, GLint yoffset, GLint zoffset,
 388                  GLsizei width, GLsizei height, GLsizei depth,
 389                  GLenum format, GLenum type,
 390                  const GLvoid * pixels,
 391                  const struct gl_pixelstore_attrib *packing)
 392 {
 393    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 394        __func__, _mesa_get_format_name(texImage->TexFormat),
 395        _mesa_enum_to_string(texImage->TexObject->Target),
 396        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 397        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 398
 399    intel_upload_tex(ctx, dims, texImage, xoffset, yoffset, zoffset,
 400                     width, height, depth, format, type, pixels, packing);
 401 }
 402
 403
 404 static void
 405 intel_set_texture_image_mt(struct brw_context *brw,
 406                            struct gl_texture_image *image,
 407                            GLenum internal_format,
 408                            struct intel_mipmap_tree *mt)
 409
 410 {
 411    struct gl_texture_object *texobj = image->TexObject;
 412    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 413    struct intel_texture_image *intel_image = intel_texture_image(image);
 414
 415    _mesa_init_teximage_fields(&brw->ctx, image,
 416                               mt->surf.logical_level0_px.width,
 417                               mt->surf.logical_level0_px.height, 1,
 418                               0, internal_format, mt->format);
 419
 420    brw->ctx.Driver.FreeTextureImageBuffer(&brw->ctx, image);
 421
 422    intel_texobj->needs_validate = true;
 423    intel_image->base.RowStride = mt->surf.row_pitch / mt->cpp;
 424    assert(mt->surf.row_pitch % mt->cpp == 0);
 425
 426    intel_miptree_reference(&intel_image->mt, mt);
 427
 428    /* Immediately validate the image to the object. */
 429    intel_miptree_reference(&intel_texobj->mt, mt);
 430 }
 431
 432
 433 void
 434 intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
 435                    GLint texture_format,
 436                    __DRIdrawable *dPriv)
 437 {
 438    struct gl_framebuffer *fb = dPriv->driverPrivate;
 439    struct brw_context *brw = pDRICtx->driverPrivate;
 440    struct gl_context *ctx = &brw->ctx;
 441    struct intel_renderbuffer *rb;
 442    struct gl_texture_object *texObj;
 443    struct gl_texture_image *texImage;
 444    GLenum internal_format = 0;
 445
 446    texObj = _mesa_get_current_tex_object(ctx, target);
 447
 448    if (!texObj)
 449       return;
 450
 451    if (dPriv->lastStamp != dPriv->dri2.stamp ||
 452        !pDRICtx->driScreenPriv->dri2.useInvalidate)
 453       intel_update_renderbuffers(pDRICtx, dPriv);
 454
 455    rb = intel_get_renderbuffer(fb, BUFFER_FRONT_LEFT);
 456    /* If the miptree isn't set, then intel_update_renderbuffers was unable
 457     * to get the BO for the drawable from the window system.
 458     */
 459    if (!rb || !rb->mt)
 460       return;
 461
 462    if (rb->mt->cpp == 4) {
 463       if (texture_format == __DRI_TEXTURE_FORMAT_RGB)
 464          internal_format = GL_RGB;
 465       else
 466          internal_format = GL_RGBA;
 467    } else if (rb->mt->cpp == 2) {
 468       /* This is 565 */
 469       internal_format = GL_RGB;
 470    }
 471
 472    intel_miptree_prepare_external(brw, rb->mt);
 473
 474    _mesa_lock_texture(&brw->ctx, texObj);
 475    texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
 476    intel_set_texture_image_mt(brw, texImage, internal_format, rb->mt);
 477    _mesa_unlock_texture(&brw->ctx, texObj);
 478 }
 479
 480 static GLboolean
 481 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
 482                                   struct gl_renderbuffer *rb,
 483                                   struct gl_texture_image *image)
 484 {
 485    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 486    struct intel_texture_image *intel_image = intel_texture_image(image);
 487    struct gl_texture_object *texobj = image->TexObject;
 488    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 489
 490    /* We can only handle RB allocated with AllocRenderbufferStorage, or
 491     * window-system renderbuffers.
 492     */
 493    assert(!rb->TexImage);
 494
 495    if (!irb->mt)
 496       return false;
 497
 498    _mesa_lock_texture(ctx, texobj);
 499    _mesa_init_teximage_fields(ctx, image,
 500                               rb->Width, rb->Height, 1,
 501                               0, rb->InternalFormat, rb->Format);
 502    image->NumSamples = rb->NumSamples;
 503
 504    intel_miptree_reference(&intel_image->mt, irb->mt);
 505
 506    /* Immediately validate the image to the object. */
 507    intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
 508
 509    intel_texobj->needs_validate = true;
 510    _mesa_unlock_texture(ctx, texobj);
 511
 512    return true;
 513 }
 514
 515 void
 516 intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 517 {
 518    /* The old interface didn't have the format argument, so copy our
 519     * implementation's behavior at the time.
 520     */
 521    intelSetTexBuffer2(pDRICtx, target, __DRI_TEXTURE_FORMAT_RGBA, dPriv);
 522 }
 523
 524 static void
 525 intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
 526                               struct gl_texture_object *texObj,
 527                               struct gl_texture_image *texImage,
 528                               GLeglImageOES image_handle)
 529 {
 530    struct brw_context *brw = brw_context(ctx);
 531    struct intel_mipmap_tree *mt;
 532    __DRIscreen *dri_screen = brw->screen->driScrnPriv;
 533    __DRIimage *image;
 534
 535    image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
 536                                                   dri_screen->loaderPrivate);
 537    if (image == NULL)
 538       return;
 539
 540    /* We support external textures only for EGLImages created with
 541     * EGL_EXT_image_dma_buf_import. We may lift that restriction in the future.
 542     */
 543    if (target == GL_TEXTURE_EXTERNAL_OES && !image->dma_buf_imported) {
 544       _mesa_error(ctx, GL_INVALID_OPERATION,
 545             "glEGLImageTargetTexture2DOES(external target is enabled only "
 546                "for images created with EGL_EXT_image_dma_buf_import");
 547       return;
 548    }
 549
 550    /* Disallow depth/stencil textures: we don't have a way to pass the
 551     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
 552     */
 553    if (image->has_depthstencil) {
 554       _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
 555       return;
 556    }
 557
 558    mt = intel_miptree_create_for_dri_image(brw, image, target, image->format,
 559                                            false);
 560    if (mt == NULL)
 561       return;
 562
 563    struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
 564    intel_texobj->planar_format = image->planar_format;
 565
 566    const GLenum internal_format =
 567       image->internal_format != 0 ?
 568       image->internal_format : _mesa_get_format_base_format(mt->format);
 569    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 570    intel_miptree_release(&mt);
 571 }
 572
 573 static bool
 574 intel_gettexsubimage_blorp(struct brw_context *brw,
 575                            struct gl_texture_image *tex_image,
 576                            unsigned x, unsigned y, unsigned z,
 577                            unsigned width, unsigned height, unsigned depth,
 578                            GLenum format, GLenum type, const void *pixels,
 579                            const struct gl_pixelstore_attrib *packing)
 580 {
 581    struct intel_texture_image *intel_image = intel_texture_image(tex_image);
 582    const unsigned mt_level = tex_image->Level + tex_image->TexObject->MinLevel;
 583    const unsigned mt_z = tex_image->TexObject->MinLayer + tex_image->Face + z;
 584
 585    /* The blorp path can't understand crazy format hackery */
 586    if (_mesa_base_tex_format(&brw->ctx, tex_image->InternalFormat) !=
 587        _mesa_get_format_base_format(tex_image->TexFormat))
 588       return false;
 589
 590    return brw_blorp_download_miptree(brw, intel_image->mt,
 591                                      tex_image->TexFormat, SWIZZLE_XYZW,
 592                                      mt_level, x, y, mt_z,
 593                                      width, height, depth,
 594                                      tex_image->TexObject->Target,
 595                                      format, type, false, pixels, packing);
 596 }
 597
 598 /**
 599  * \brief A fast path for glGetTexImage.
 600  *
 601  * \see intel_readpixels_tiled_memcpy()
 602  */
 603 static bool
 604 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
 605                                   struct gl_texture_image *texImage,
 606                                   GLint xoffset, GLint yoffset,
 607                                   GLsizei width, GLsizei height,
 608                                   GLenum format, GLenum type,
 609                                   GLvoid *pixels,
 610                                   const struct gl_pixelstore_attrib *packing)
 611 {
 612    struct brw_context *brw = brw_context(ctx);
 613    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 614    struct intel_texture_image *image = intel_texture_image(texImage);
 615    int dst_pitch;
 616
 617    /* The miptree's buffer. */
 618    struct brw_bo *bo;
 619
 620    uint32_t cpp;
 621    mem_copy_fn mem_copy = NULL;
 622
 623    /* This fastpath is restricted to specific texture types:
 624     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 625     * more types.
 626     *
 627     * FINISHME: The restrictions below on packing alignment and packing row
 628     * length are likely unneeded now because we calculate the destination stride
 629     * with _mesa_image_row_stride. However, before removing the restrictions
 630     * we need tests.
 631     */
 632    if (!devinfo->has_llc ||
 633        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 634        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 635          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 636        pixels == NULL ||
 637        _mesa_is_bufferobj(packing->BufferObj) ||
 638        packing->Alignment > 4 ||
 639        packing->SkipPixels > 0 ||
 640        packing->SkipRows > 0 ||
 641        (packing->RowLength != 0 && packing->RowLength != width) ||
 642        packing->SwapBytes ||
 643        packing->LsbFirst ||
 644        packing->Invert)
 645       return false;
 646
 647    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
 648     * function doesn't set the last channel to 1. Note this checks BaseFormat
 649     * rather than TexFormat in case the RGBX format is being simulated with an
 650     * RGBA format.
 651     */
 652    if (texImage->_BaseFormat == GL_RGB)
 653       return false;
 654
 655    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 656       return false;
 657
 658    /* If this is a nontrivial texture view, let another path handle it instead. */
 659    if (texImage->TexObject->MinLayer)
 660       return false;
 661
 662    if (!image->mt ||
 663        (image->mt->surf.tiling != ISL_TILING_X &&
 664         image->mt->surf.tiling != ISL_TILING_Y0)) {
 665       /* The algorithm is written only for X- or Y-tiled memory. */
 666       return false;
 667    }
 668
 669    /* tiled_to_linear() assumes that if the object is swizzled, it is using
 670     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 671     * true on gen5 and above.
 672     *
 673     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 674     * parts of the memory aren't swizzled at all. Userspace just can't handle
 675     * that.
 676     */
 677    if (devinfo->gen < 5 && brw->has_swizzling)
 678       return false;
 679
 680    int level = texImage->Level + texImage->TexObject->MinLevel;
 681
 682    /* Since we are going to write raw data to the miptree, we need to resolve
 683     * any pending fast color clears before we start.
 684     */
 685    assert(image->mt->surf.logical_level0_px.depth == 1);
 686    assert(image->mt->surf.logical_level0_px.array_len == 1);
 687
 688    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 689
 690    bo = image->mt->bo;
 691
 692    if (brw_batch_references(&brw->batch, bo)) {
 693       perf_debug("Flushing before mapping a referenced bo.\n");
 694       intel_batchbuffer_flush(brw);
 695    }
 696
 697    void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
 698    if (map == NULL) {
 699       DBG("%s: failed to map bo\n", __func__);
 700       return false;
 701    }
 702
 703    dst_pitch = _mesa_image_row_stride(packing, width, format, type);
 704
 705    DBG("%s: level=%d x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 706        "mesa_format=0x%x tiling=%d "
 707        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
 708        __func__, texImage->Level, xoffset, yoffset, width, height,
 709        format, type, texImage->TexFormat, image->mt->surf.tiling,
 710        packing->Alignment, packing->RowLength, packing->SkipPixels,
 711        packing->SkipRows);
 712
 713    /* Adjust x and y offset based on miplevel */
 714    unsigned level_x, level_y;
 715    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 716    xoffset += level_x;
 717    yoffset += level_y;
 718
 719    tiled_to_linear(
 720       xoffset * cpp, (xoffset + width) * cpp,
 721       yoffset, yoffset + height,
 722       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
 723       map,
 724       dst_pitch, image->mt->surf.row_pitch,
 725       brw->has_swizzling,
 726       image->mt->surf.tiling,
 727       mem_copy
 728    );
 729
 730    brw_bo_unmap(bo);
 731    return true;
 732 }
 733
 734 static void
 735 intel_get_tex_sub_image(struct gl_context *ctx,
 736                         GLint xoffset, GLint yoffset, GLint zoffset,
 737                         GLsizei width, GLsizei height, GLint depth,
 738                         GLenum format, GLenum type, GLvoid *pixels,
 739                         struct gl_texture_image *texImage)
 740 {
 741    struct brw_context *brw = brw_context(ctx);
 742    bool ok;
 743
 744    DBG("%s\n", __func__);
 745
 746    if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
 747       if (intel_gettexsubimage_blorp(brw, texImage,
 748                                      xoffset, yoffset, zoffset,
 749                                      width, height, depth, format, type,
 750                                      pixels, &ctx->Pack)) {
 751          /* Flush to guarantee coherency between the render cache and other
 752           * caches the PBO could potentially be bound to after this point.
 753           * See the related comment in intelReadPixels() for a more detailed
 754           * explanation.
 755           */
 756          brw_emit_mi_flush(brw);
 757          return;
 758       }
 759
 760       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
 761    }
 762
 763    ok = intel_gettexsubimage_tiled_memcpy(ctx, texImage, xoffset, yoffset,
 764                                           width, height,
 765                                           format, type, pixels, &ctx->Pack);
 766
 767    if(ok)
 768       return;
 769
 770    _mesa_meta_GetTexSubImage(ctx, xoffset, yoffset, zoffset,
 771                              width, height, depth,
 772                              format, type, pixels, texImage);
 773
 774    DBG("%s - DONE\n", __func__);
 775 }
 776
 777 static void
 778 flush_astc_denorms(struct gl_context *ctx, GLuint dims,
 779                    struct gl_texture_image *texImage,
 780                    GLint xoffset, GLint yoffset, GLint zoffset,
 781                    GLsizei width, GLsizei height, GLsizei depth)
 782 {
 783    struct compressed_pixelstore store;
 784    _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
 785                                        width, height, depth,
 786                                        &ctx->Unpack, &store);
 787
 788    for (int slice = 0; slice < store.CopySlices; slice++) {
 789
 790       /* Map dest texture buffer */
 791       GLubyte *dstMap;
 792       GLint dstRowStride;
 793       ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
 794                                   xoffset, yoffset, width, height,
 795                                   GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
 796                                   &dstMap, &dstRowStride);
 797       if (!dstMap)
 798          continue;
 799
 800       for (int i = 0; i < store.CopyRowsPerSlice; i++) {
 801
 802          /* An ASTC block is stored in little endian mode. The byte that
 803           * contains bits 0..7 is stored at the lower address in memory.
 804           */
 805          struct astc_void_extent {
 806             uint16_t header : 12;
 807             uint16_t dontcare[3];
 808             uint16_t R;
 809             uint16_t G;
 810             uint16_t B;
 811             uint16_t A;
 812          } *blocks = (struct astc_void_extent*) dstMap;
 813
 814          /* Iterate over every copied block in the row */
 815          for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
 816
 817             /* Check if the header matches that of an LDR void-extent block */
 818             if (blocks[j].header == 0xDFC) {
 819
 820                /* Flush UNORM16 values that would be denormalized */
 821                if (blocks[j].A < 4) blocks[j].A = 0;
 822                if (blocks[j].B < 4) blocks[j].B = 0;
 823                if (blocks[j].G < 4) blocks[j].G = 0;
 824                if (blocks[j].R < 4) blocks[j].R = 0;
 825             }
 826          }
 827
 828          dstMap += dstRowStride;
 829       }
 830
 831       ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
 832    }
 833 }
 834
 835
 836 static void
 837 intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
 838                         struct gl_texture_image *texImage,
 839                         GLint xoffset, GLint yoffset, GLint zoffset,
 840                         GLsizei width, GLsizei height, GLsizei depth,
 841                         GLenum format,
 842                         GLsizei imageSize, const GLvoid *data)
 843 {
 844    /* Upload the compressed data blocks */
 845    _mesa_store_compressed_texsubimage(ctx, dims, texImage,
 846                                       xoffset, yoffset, zoffset,
 847                                       width, height, depth,
 848                                       format, imageSize, data);
 849
 850    /* Fix up copied ASTC blocks if necessary */
 851    GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
 852                                                         texImage->TexFormat);
 853    bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
 854                         !_mesa_is_srgb_format(gl_format);
 855    struct brw_context *brw = (struct brw_context*) ctx;
 856    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 857    if (devinfo->gen == 9 && is_linear_astc)
 858       flush_astc_denorms(ctx, dims, texImage,
 859                          xoffset, yoffset, zoffset,
 860                          width, height, depth);
 861 }
 862
 863 void
 864 intelInitTextureImageFuncs(struct dd_function_table *functions)
 865 {
 866    functions->TexImage = intelTexImage;
 867    functions->TexSubImage = intelTexSubImage;
 868    functions->CompressedTexSubImage = intelCompressedTexSubImage;
 869    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
 870    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
 871    functions->GetTexSubImage = intel_get_tex_sub_image;
 872 }