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  * \param for_glTexImage Was this called from glTexImage or glTexSubImage?
 135  *
 136  * This fast path is taken when the texture format is BGRA, RGBA,
 137  * A or L and when the texture memory is X- or Y-tiled.  It uploads
 138  * the texture data by mapping the texture memory without a GTT fence, thus
 139  * acquiring a tiled view of the memory, and then copying sucessive
 140  * spans within each tile.
 141  *
 142  * This is a performance win over the conventional texture upload path because
 143  * it avoids the performance penalty of writing through the write-combine
 144  * buffer. In the conventional texture upload path,
 145  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
 146  * fence, thus acquiring a linear view of the memory, then each row in the
 147  * image is memcpy'd. In this fast path, we replace each row's copy with
 148  * a sequence of copies over each linear span in tile.
 149  *
 150  * One use case is Google Chrome's paint rectangles.  Chrome (as
 151  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
 152  * Each page's content is initially uploaded with glTexImage2D and damaged
 153  * regions are updated with glTexSubImage2D. On some workloads, the
 154  * performance gain of this fastpath on Sandybridge is over 5x.
 155  */
 156 static bool
 157 intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 158                                GLuint dims,
 159                                struct gl_texture_image *texImage,
 160                                GLint xoffset, GLint yoffset, GLint zoffset,
 161                                GLsizei width, GLsizei height, GLsizei depth,
 162                                GLenum format, GLenum type,
 163                                const GLvoid *pixels,
 164                                const struct gl_pixelstore_attrib *packing,
 165                                bool for_glTexImage)
 166 {
 167    struct brw_context *brw = brw_context(ctx);
 168    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 169    struct intel_texture_image *image = intel_texture_image(texImage);
 170    int src_pitch;
 171
 172    /* The miptree's buffer. */
 173    struct brw_bo *bo;
 174
 175    uint32_t cpp;
 176    mem_copy_fn mem_copy = NULL;
 177
 178    /* This fastpath is restricted to specific texture types:
 179     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 180     * more types.
 181     *
 182     * FINISHME: The restrictions below on packing alignment and packing row
 183     * length are likely unneeded now because we calculate the source stride
 184     * with _mesa_image_row_stride. However, before removing the restrictions
 185     * we need tests.
 186     */
 187    if (!devinfo->has_llc ||
 188        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 189        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 190          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 191        pixels == NULL ||
 192        _mesa_is_bufferobj(packing->BufferObj) ||
 193        packing->Alignment > 4 ||
 194        packing->SkipPixels > 0 ||
 195        packing->SkipRows > 0 ||
 196        (packing->RowLength != 0 && packing->RowLength != width) ||
 197        packing->SwapBytes ||
 198        packing->LsbFirst ||
 199        packing->Invert)
 200       return false;
 201
 202    /* Only a simple blit, no scale, bias or other mapping. */
 203    if (ctx->_ImageTransferState)
 204       return false;
 205
 206    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 207       return false;
 208
 209    /* If this is a nontrivial texture view, let another path handle it instead. */
 210    if (texImage->TexObject->MinLayer)
 211       return false;
 212
 213    if (for_glTexImage)
 214       ctx->Driver.AllocTextureImageBuffer(ctx, texImage);
 215
 216    if (!image->mt ||
 217        (image->mt->surf.tiling != ISL_TILING_X &&
 218         image->mt->surf.tiling != ISL_TILING_Y0)) {
 219       /* The algorithm is written only for X- or Y-tiled memory. */
 220       return false;
 221    }
 222
 223    /* linear_to_tiled() assumes that if the object is swizzled, it is using
 224     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 225     * true on gen5 and above.
 226     *
 227     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 228     * parts of the memory aren't swizzled at all. Userspace just can't handle
 229     * that.
 230     */
 231    if (devinfo->gen < 5 && brw->has_swizzling)
 232       return false;
 233
 234    int level = texImage->Level + texImage->TexObject->MinLevel;
 235
 236    /* Since we are going to write raw data to the miptree, we need to resolve
 237     * any pending fast color clears before we start.
 238     */
 239    assert(image->mt->surf.logical_level0_px.depth == 1);
 240    assert(image->mt->surf.logical_level0_px.array_len == 1);
 241
 242    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 243
 244    bo = image->mt->bo;
 245
 246    if (brw_batch_references(&brw->batch, bo)) {
 247       perf_debug("Flushing before mapping a referenced bo.\n");
 248       intel_batchbuffer_flush(brw);
 249    }
 250
 251    void *map = brw_bo_map(brw, bo, MAP_WRITE | MAP_RAW);
 252    if (map == NULL) {
 253       DBG("%s: failed to map bo\n", __func__);
 254       return false;
 255    }
 256
 257    src_pitch = _mesa_image_row_stride(packing, width, format, type);
 258
 259    /* We postponed printing this message until having committed to executing
 260     * the function.
 261     */
 262    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 263        "mesa_format=0x%x tiling=%d "
 264        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d) "
 265        "for_glTexImage=%d\n",
 266        __func__, texImage->Level, xoffset, yoffset, width, height,
 267        format, type, texImage->TexFormat, image->mt->surf.tiling,
 268        packing->Alignment, packing->RowLength, packing->SkipPixels,
 269        packing->SkipRows, for_glTexImage);
 270
 271    /* Adjust x and y offset based on miplevel */
 272    unsigned level_x, level_y;
 273    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 274    xoffset += level_x;
 275    yoffset += level_y;
 276
 277    linear_to_tiled(
 278       xoffset * cpp, (xoffset + width) * cpp,
 279       yoffset, yoffset + height,
 280       map,
 281       pixels - (ptrdiff_t) yoffset * src_pitch - (ptrdiff_t) xoffset * cpp,
 282       image->mt->surf.row_pitch, src_pitch,
 283       brw->has_swizzling,
 284       image->mt->surf.tiling,
 285       mem_copy
 286    );
 287
 288    brw_bo_unmap(bo);
 289    return true;
 290 }
 291
 292
 293 static void
 294 intelTexImage(struct gl_context * ctx,
 295               GLuint dims,
 296               struct gl_texture_image *texImage,
 297               GLenum format, GLenum type, const void *pixels,
 298               const struct gl_pixelstore_attrib *unpack)
 299 {
 300    struct intel_texture_image *intelImage = intel_texture_image(texImage);
 301    bool ok;
 302
 303    bool tex_busy = intelImage->mt && brw_bo_busy(intelImage->mt->bo);
 304
 305    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 306        __func__, _mesa_get_format_name(texImage->TexFormat),
 307        _mesa_enum_to_string(texImage->TexObject->Target),
 308        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 309        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 310
 311    /* Allocate storage for texture data. */
 312    if (!ctx->Driver.AllocTextureImageBuffer(ctx, texImage)) {
 313       _mesa_error(ctx, GL_OUT_OF_MEMORY, "glTexImage%uD", dims);
 314       return;
 315    }
 316
 317    assert(intelImage->mt);
 318
 319    if (intelImage->mt->format == MESA_FORMAT_S_UINT8)
 320       intelImage->mt->r8stencil_needs_update = true;
 321
 322    ok = _mesa_meta_pbo_TexSubImage(ctx, dims, texImage, 0, 0, 0,
 323                                    texImage->Width, texImage->Height,
 324                                    texImage->Depth,
 325                                    format, type, pixels,
 326                                    tex_busy, unpack);
 327    if (ok)
 328       return;
 329
 330    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 331                                        0, 0, 0, /*x,y,z offsets*/
 332                                        texImage->Width,
 333                                        texImage->Height,
 334                                        texImage->Depth,
 335                                        format, type, pixels, unpack,
 336                                        false /*allocate_storage*/);
 337    if (ok)
 338       return;
 339
 340    DBG("%s: upload image %dx%dx%d pixels %p\n",
 341        __func__, texImage->Width, texImage->Height, texImage->Depth,
 342        pixels);
 343
 344    _mesa_store_teximage(ctx, dims, texImage,
 345                         format, type, pixels, unpack);
 346 }
 347
 348
 349 static void
 350 intelTexSubImage(struct gl_context * ctx,
 351                  GLuint dims,
 352                  struct gl_texture_image *texImage,
 353                  GLint xoffset, GLint yoffset, GLint zoffset,
 354                  GLsizei width, GLsizei height, GLsizei depth,
 355                  GLenum format, GLenum type,
 356                  const GLvoid * pixels,
 357                  const struct gl_pixelstore_attrib *packing)
 358 {
 359    struct intel_mipmap_tree *mt = intel_texture_image(texImage)->mt;
 360    bool ok;
 361
 362    bool tex_busy = mt && brw_bo_busy(mt->bo);
 363
 364    if (mt && mt->format == MESA_FORMAT_S_UINT8)
 365       mt->r8stencil_needs_update = true;
 366
 367    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 368        __func__, _mesa_get_format_name(texImage->TexFormat),
 369        _mesa_enum_to_string(texImage->TexObject->Target),
 370        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 371        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 372
 373    ok = _mesa_meta_pbo_TexSubImage(ctx, dims, texImage,
 374                                    xoffset, yoffset, zoffset,
 375                                    width, height, depth, format, type,
 376                                    pixels, tex_busy, packing);
 377    if (ok)
 378       return;
 379
 380    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 381                                        xoffset, yoffset, zoffset,
 382                                        width, height, depth,
 383                                        format, type, pixels, packing,
 384                                        false /*for_glTexImage*/);
 385    if (ok)
 386      return;
 387
 388    _mesa_store_texsubimage(ctx, dims, texImage,
 389                            xoffset, yoffset, zoffset,
 390                            width, height, depth,
 391                            format, type, pixels, packing);
 392 }
 393
 394
 395 static void
 396 intel_set_texture_image_mt(struct brw_context *brw,
 397                            struct gl_texture_image *image,
 398                            GLenum internal_format,
 399                            struct intel_mipmap_tree *mt)
 400
 401 {
 402    struct gl_texture_object *texobj = image->TexObject;
 403    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 404    struct intel_texture_image *intel_image = intel_texture_image(image);
 405
 406    _mesa_init_teximage_fields(&brw->ctx, image,
 407                               mt->surf.logical_level0_px.width,
 408                               mt->surf.logical_level0_px.height, 1,
 409                               0, internal_format, mt->format);
 410
 411    brw->ctx.Driver.FreeTextureImageBuffer(&brw->ctx, image);
 412
 413    intel_texobj->needs_validate = true;
 414    intel_image->base.RowStride = mt->surf.row_pitch / mt->cpp;
 415    assert(mt->surf.row_pitch % mt->cpp == 0);
 416
 417    intel_miptree_reference(&intel_image->mt, mt);
 418
 419    /* Immediately validate the image to the object. */
 420    intel_miptree_reference(&intel_texobj->mt, mt);
 421 }
 422
 423
 424 void
 425 intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
 426                    GLint texture_format,
 427                    __DRIdrawable *dPriv)
 428 {
 429    struct gl_framebuffer *fb = dPriv->driverPrivate;
 430    struct brw_context *brw = pDRICtx->driverPrivate;
 431    struct gl_context *ctx = &brw->ctx;
 432    struct intel_renderbuffer *rb;
 433    struct gl_texture_object *texObj;
 434    struct gl_texture_image *texImage;
 435    mesa_format texFormat = MESA_FORMAT_NONE;
 436    struct intel_mipmap_tree *mt;
 437    GLenum internal_format = 0;
 438
 439    texObj = _mesa_get_current_tex_object(ctx, target);
 440
 441    if (!texObj)
 442       return;
 443
 444    if (dPriv->lastStamp != dPriv->dri2.stamp ||
 445        !pDRICtx->driScreenPriv->dri2.useInvalidate)
 446       intel_update_renderbuffers(pDRICtx, dPriv);
 447
 448    rb = intel_get_renderbuffer(fb, BUFFER_FRONT_LEFT);
 449    /* If the miptree isn't set, then intel_update_renderbuffers was unable
 450     * to get the BO for the drawable from the window system.
 451     */
 452    if (!rb || !rb->mt)
 453       return;
 454
 455    if (rb->mt->cpp == 4) {
 456       if (texture_format == __DRI_TEXTURE_FORMAT_RGB) {
 457          internal_format = GL_RGB;
 458          texFormat = MESA_FORMAT_B8G8R8X8_UNORM;
 459       }
 460       else {
 461          internal_format = GL_RGBA;
 462          texFormat = MESA_FORMAT_B8G8R8A8_UNORM;
 463       }
 464    } else if (rb->mt->cpp == 2) {
 465       internal_format = GL_RGB;
 466       texFormat = MESA_FORMAT_B5G6R5_UNORM;
 467    }
 468
 469    intel_miptree_make_shareable(brw, rb->mt);
 470    mt = intel_miptree_create_for_bo(brw, rb->mt->bo, texFormat, 0,
 471                                     rb->Base.Base.Width,
 472                                     rb->Base.Base.Height,
 473                                     1, rb->mt->surf.row_pitch,
 474                                     MIPTREE_CREATE_DEFAULT);
 475    if (mt == NULL)
 476        return;
 477    mt->target = target;
 478
 479    _mesa_lock_texture(&brw->ctx, texObj);
 480    texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
 481    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 482    intel_miptree_release(&mt);
 483    _mesa_unlock_texture(&brw->ctx, texObj);
 484 }
 485
 486 static GLboolean
 487 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
 488                                   struct gl_renderbuffer *rb,
 489                                   struct gl_texture_image *image)
 490 {
 491    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 492    struct intel_texture_image *intel_image = intel_texture_image(image);
 493    struct gl_texture_object *texobj = image->TexObject;
 494    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 495
 496    /* We can only handle RB allocated with AllocRenderbufferStorage, or
 497     * window-system renderbuffers.
 498     */
 499    assert(!rb->TexImage);
 500
 501    if (!irb->mt)
 502       return false;
 503
 504    _mesa_lock_texture(ctx, texobj);
 505    _mesa_init_teximage_fields(ctx, image,
 506                               rb->Width, rb->Height, 1,
 507                               0, rb->InternalFormat, rb->Format);
 508    image->NumSamples = rb->NumSamples;
 509
 510    intel_miptree_reference(&intel_image->mt, irb->mt);
 511
 512    /* Immediately validate the image to the object. */
 513    intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
 514
 515    intel_texobj->needs_validate = true;
 516    _mesa_unlock_texture(ctx, texobj);
 517
 518    return true;
 519 }
 520
 521 void
 522 intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 523 {
 524    /* The old interface didn't have the format argument, so copy our
 525     * implementation's behavior at the time.
 526     */
 527    intelSetTexBuffer2(pDRICtx, target, __DRI_TEXTURE_FORMAT_RGBA, dPriv);
 528 }
 529
 530 static void
 531 intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
 532                               struct gl_texture_object *texObj,
 533                               struct gl_texture_image *texImage,
 534                               GLeglImageOES image_handle)
 535 {
 536    struct brw_context *brw = brw_context(ctx);
 537    struct intel_mipmap_tree *mt;
 538    __DRIscreen *dri_screen = brw->screen->driScrnPriv;
 539    __DRIimage *image;
 540
 541    image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
 542                                                   dri_screen->loaderPrivate);
 543    if (image == NULL)
 544       return;
 545
 546    /* We support external textures only for EGLImages created with
 547     * EGL_EXT_image_dma_buf_import. We may lift that restriction in the future.
 548     */
 549    if (target == GL_TEXTURE_EXTERNAL_OES && !image->dma_buf_imported) {
 550       _mesa_error(ctx, GL_INVALID_OPERATION,
 551             "glEGLImageTargetTexture2DOES(external target is enabled only "
 552                "for images created with EGL_EXT_image_dma_buf_import");
 553       return;
 554    }
 555
 556    /* Disallow depth/stencil textures: we don't have a way to pass the
 557     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
 558     */
 559    if (image->has_depthstencil) {
 560       _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
 561       return;
 562    }
 563
 564    mt = intel_miptree_create_for_dri_image(brw, image, target,
 565                                            ISL_COLORSPACE_NONE, false);
 566    if (mt == NULL)
 567       return;
 568
 569    struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
 570    intel_texobj->planar_format = image->planar_format;
 571
 572    const GLenum internal_format =
 573       image->internal_format != 0 ?
 574       image->internal_format : _mesa_get_format_base_format(mt->format);
 575    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 576    intel_miptree_release(&mt);
 577 }
 578
 579 /**
 580  * \brief A fast path for glGetTexImage.
 581  *
 582  * \see intel_readpixels_tiled_memcpy()
 583  */
 584 static bool
 585 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
 586                                   struct gl_texture_image *texImage,
 587                                   GLint xoffset, GLint yoffset,
 588                                   GLsizei width, GLsizei height,
 589                                   GLenum format, GLenum type,
 590                                   GLvoid *pixels,
 591                                   const struct gl_pixelstore_attrib *packing)
 592 {
 593    struct brw_context *brw = brw_context(ctx);
 594    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 595    struct intel_texture_image *image = intel_texture_image(texImage);
 596    int dst_pitch;
 597
 598    /* The miptree's buffer. */
 599    struct brw_bo *bo;
 600
 601    uint32_t cpp;
 602    mem_copy_fn mem_copy = NULL;
 603
 604    /* This fastpath is restricted to specific texture types:
 605     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 606     * more types.
 607     *
 608     * FINISHME: The restrictions below on packing alignment and packing row
 609     * length are likely unneeded now because we calculate the destination stride
 610     * with _mesa_image_row_stride. However, before removing the restrictions
 611     * we need tests.
 612     */
 613    if (!devinfo->has_llc ||
 614        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 615        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 616          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 617        pixels == NULL ||
 618        _mesa_is_bufferobj(packing->BufferObj) ||
 619        packing->Alignment > 4 ||
 620        packing->SkipPixels > 0 ||
 621        packing->SkipRows > 0 ||
 622        (packing->RowLength != 0 && packing->RowLength != width) ||
 623        packing->SwapBytes ||
 624        packing->LsbFirst ||
 625        packing->Invert)
 626       return false;
 627
 628    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
 629     * function doesn't set the last channel to 1. Note this checks BaseFormat
 630     * rather than TexFormat in case the RGBX format is being simulated with an
 631     * RGBA format.
 632     */
 633    if (texImage->_BaseFormat == GL_RGB)
 634       return false;
 635
 636    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 637       return false;
 638
 639    /* If this is a nontrivial texture view, let another path handle it instead. */
 640    if (texImage->TexObject->MinLayer)
 641       return false;
 642
 643    if (!image->mt ||
 644        (image->mt->surf.tiling != ISL_TILING_X &&
 645         image->mt->surf.tiling != ISL_TILING_Y0)) {
 646       /* The algorithm is written only for X- or Y-tiled memory. */
 647       return false;
 648    }
 649
 650    /* tiled_to_linear() assumes that if the object is swizzled, it is using
 651     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 652     * true on gen5 and above.
 653     *
 654     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 655     * parts of the memory aren't swizzled at all. Userspace just can't handle
 656     * that.
 657     */
 658    if (devinfo->gen < 5 && brw->has_swizzling)
 659       return false;
 660
 661    int level = texImage->Level + texImage->TexObject->MinLevel;
 662
 663    /* Since we are going to write raw data to the miptree, we need to resolve
 664     * any pending fast color clears before we start.
 665     */
 666    assert(image->mt->surf.logical_level0_px.depth == 1);
 667    assert(image->mt->surf.logical_level0_px.array_len == 1);
 668
 669    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 670
 671    bo = image->mt->bo;
 672
 673    if (brw_batch_references(&brw->batch, bo)) {
 674       perf_debug("Flushing before mapping a referenced bo.\n");
 675       intel_batchbuffer_flush(brw);
 676    }
 677
 678    void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
 679    if (map == NULL) {
 680       DBG("%s: failed to map bo\n", __func__);
 681       return false;
 682    }
 683
 684    dst_pitch = _mesa_image_row_stride(packing, width, format, type);
 685
 686    DBG("%s: level=%d x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 687        "mesa_format=0x%x tiling=%d "
 688        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
 689        __func__, texImage->Level, xoffset, yoffset, width, height,
 690        format, type, texImage->TexFormat, image->mt->surf.tiling,
 691        packing->Alignment, packing->RowLength, packing->SkipPixels,
 692        packing->SkipRows);
 693
 694    /* Adjust x and y offset based on miplevel */
 695    unsigned level_x, level_y;
 696    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 697    xoffset += level_x;
 698    yoffset += level_y;
 699
 700    tiled_to_linear(
 701       xoffset * cpp, (xoffset + width) * cpp,
 702       yoffset, yoffset + height,
 703       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
 704       map,
 705       dst_pitch, image->mt->surf.row_pitch,
 706       brw->has_swizzling,
 707       image->mt->surf.tiling,
 708       mem_copy
 709    );
 710
 711    brw_bo_unmap(bo);
 712    return true;
 713 }
 714
 715 static void
 716 intel_get_tex_sub_image(struct gl_context *ctx,
 717                         GLint xoffset, GLint yoffset, GLint zoffset,
 718                         GLsizei width, GLsizei height, GLint depth,
 719                         GLenum format, GLenum type, GLvoid *pixels,
 720                         struct gl_texture_image *texImage)
 721 {
 722    struct brw_context *brw = brw_context(ctx);
 723    bool ok;
 724
 725    DBG("%s\n", __func__);
 726
 727    if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
 728       if (_mesa_meta_pbo_GetTexSubImage(ctx, 3, texImage,
 729                                         xoffset, yoffset, zoffset,
 730                                         width, height, depth, format, type,
 731                                         pixels, &ctx->Pack)) {
 732          /* Flush to guarantee coherency between the render cache and other
 733           * caches the PBO could potentially be bound to after this point.
 734           * See the related comment in intelReadPixels() for a more detailed
 735           * explanation.
 736           */
 737          brw_emit_mi_flush(brw);
 738          return;
 739       }
 740
 741       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
 742    }
 743
 744    ok = intel_gettexsubimage_tiled_memcpy(ctx, texImage, xoffset, yoffset,
 745                                           width, height,
 746                                           format, type, pixels, &ctx->Pack);
 747
 748    if(ok)
 749       return;
 750
 751    _mesa_meta_GetTexSubImage(ctx, xoffset, yoffset, zoffset,
 752                              width, height, depth,
 753                              format, type, pixels, texImage);
 754
 755    DBG("%s - DONE\n", __func__);
 756 }
 757
 758 static void
 759 flush_astc_denorms(struct gl_context *ctx, GLuint dims,
 760                    struct gl_texture_image *texImage,
 761                    GLint xoffset, GLint yoffset, GLint zoffset,
 762                    GLsizei width, GLsizei height, GLsizei depth)
 763 {
 764    struct compressed_pixelstore store;
 765    _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
 766                                        width, height, depth,
 767                                        &ctx->Unpack, &store);
 768
 769    for (int slice = 0; slice < store.CopySlices; slice++) {
 770
 771       /* Map dest texture buffer */
 772       GLubyte *dstMap;
 773       GLint dstRowStride;
 774       ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
 775                                   xoffset, yoffset, width, height,
 776                                   GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
 777                                   &dstMap, &dstRowStride);
 778       if (!dstMap)
 779          continue;
 780
 781       for (int i = 0; i < store.CopyRowsPerSlice; i++) {
 782
 783          /* An ASTC block is stored in little endian mode. The byte that
 784           * contains bits 0..7 is stored at the lower address in memory.
 785           */
 786          struct astc_void_extent {
 787             uint16_t header : 12;
 788             uint16_t dontcare[3];
 789             uint16_t R;
 790             uint16_t G;
 791             uint16_t B;
 792             uint16_t A;
 793          } *blocks = (struct astc_void_extent*) dstMap;
 794
 795          /* Iterate over every copied block in the row */
 796          for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
 797
 798             /* Check if the header matches that of an LDR void-extent block */
 799             if (blocks[j].header == 0xDFC) {
 800
 801                /* Flush UNORM16 values that would be denormalized */
 802                if (blocks[j].A < 4) blocks[j].A = 0;
 803                if (blocks[j].B < 4) blocks[j].B = 0;
 804                if (blocks[j].G < 4) blocks[j].G = 0;
 805                if (blocks[j].R < 4) blocks[j].R = 0;
 806             }
 807          }
 808
 809          dstMap += dstRowStride;
 810       }
 811
 812       ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
 813    }
 814 }
 815
 816
 817 static void
 818 intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
 819                         struct gl_texture_image *texImage,
 820                         GLint xoffset, GLint yoffset, GLint zoffset,
 821                         GLsizei width, GLsizei height, GLsizei depth,
 822                         GLenum format,
 823                         GLsizei imageSize, const GLvoid *data)
 824 {
 825    /* Upload the compressed data blocks */
 826    _mesa_store_compressed_texsubimage(ctx, dims, texImage,
 827                                       xoffset, yoffset, zoffset,
 828                                       width, height, depth,
 829                                       format, imageSize, data);
 830
 831    /* Fix up copied ASTC blocks if necessary */
 832    GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
 833                                                         texImage->TexFormat);
 834    bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
 835                         !_mesa_is_srgb_format(gl_format);
 836    struct brw_context *brw = (struct brw_context*) ctx;
 837    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 838    if (devinfo->gen == 9 && is_linear_astc)
 839       flush_astc_denorms(ctx, dims, texImage,
 840                          xoffset, yoffset, zoffset,
 841                          width, height, depth);
 842 }
 843
 844 void
 845 intelInitTextureImageFuncs(struct dd_function_table *functions)
 846 {
 847    functions->TexImage = intelTexImage;
 848    functions->TexSubImage = intelTexSubImage;
 849    functions->CompressedTexSubImage = intelCompressedTexSubImage;
 850    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
 851    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
 852    functions->GetTexSubImage = intel_get_tex_sub_image;
 853 }