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       ok = intel_texsubimage_blorp(brw, dims, texImage,
 334                                    xoffset, yoffset, zoffset,
 335                                    width, height, depth, format, type,
 336                                    pixels, packing);
 337       if (ok)
 338          return;
 339    }
 340
 341    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 342                                        xoffset, yoffset, zoffset,
 343                                        width, height, depth,
 344                                        format, type, pixels, packing);
 345    if (ok)
 346      return;
 347
 348    _mesa_store_texsubimage(ctx, dims, texImage,
 349                            xoffset, yoffset, zoffset,
 350                            width, height, depth,
 351                            format, type, pixels, packing);
 352 }
 353
 354
 355 static void
 356 intelTexImage(struct gl_context * ctx,
 357               GLuint dims,
 358               struct gl_texture_image *texImage,
 359               GLenum format, GLenum type, const void *pixels,
 360               const struct gl_pixelstore_attrib *unpack)
 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    /* Allocate storage for texture data. */
 369    if (!ctx->Driver.AllocTextureImageBuffer(ctx, texImage)) {
 370       _mesa_error(ctx, GL_OUT_OF_MEMORY, "glTexImage%uD", dims);
 371       return;
 372    }
 373
 374    assert(intel_texture_image(texImage)->mt);
 375
 376    intel_upload_tex(ctx, dims, texImage, 0, 0, 0,
 377                     texImage->Width, texImage->Height, texImage->Depth,
 378                     format, type, pixels, unpack);
 379 }
 380
 381
 382 static void
 383 intelTexSubImage(struct gl_context * ctx,
 384                  GLuint dims,
 385                  struct gl_texture_image *texImage,
 386                  GLint xoffset, GLint yoffset, GLint zoffset,
 387                  GLsizei width, GLsizei height, GLsizei depth,
 388                  GLenum format, GLenum type,
 389                  const GLvoid * pixels,
 390                  const struct gl_pixelstore_attrib *packing)
 391 {
 392    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 393        __func__, _mesa_get_format_name(texImage->TexFormat),
 394        _mesa_enum_to_string(texImage->TexObject->Target),
 395        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 396        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 397
 398    intel_upload_tex(ctx, dims, texImage, xoffset, yoffset, zoffset,
 399                     width, height, depth, format, type, pixels, packing);
 400 }
 401
 402
 403 static void
 404 intel_set_texture_image_mt(struct brw_context *brw,
 405                            struct gl_texture_image *image,
 406                            GLenum internal_format,
 407                            struct intel_mipmap_tree *mt)
 408
 409 {
 410    struct gl_texture_object *texobj = image->TexObject;
 411    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 412    struct intel_texture_image *intel_image = intel_texture_image(image);
 413
 414    _mesa_init_teximage_fields(&brw->ctx, image,
 415                               mt->surf.logical_level0_px.width,
 416                               mt->surf.logical_level0_px.height, 1,
 417                               0, internal_format, mt->format);
 418
 419    brw->ctx.Driver.FreeTextureImageBuffer(&brw->ctx, image);
 420
 421    intel_texobj->needs_validate = true;
 422    intel_image->base.RowStride = mt->surf.row_pitch / mt->cpp;
 423    assert(mt->surf.row_pitch % mt->cpp == 0);
 424
 425    intel_miptree_reference(&intel_image->mt, mt);
 426
 427    /* Immediately validate the image to the object. */
 428    intel_miptree_reference(&intel_texobj->mt, mt);
 429 }
 430
 431
 432 void
 433 intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
 434                    GLint texture_format,
 435                    __DRIdrawable *dPriv)
 436 {
 437    struct gl_framebuffer *fb = dPriv->driverPrivate;
 438    struct brw_context *brw = pDRICtx->driverPrivate;
 439    struct gl_context *ctx = &brw->ctx;
 440    struct intel_renderbuffer *rb;
 441    struct gl_texture_object *texObj;
 442    struct gl_texture_image *texImage;
 443    GLenum internal_format = 0;
 444
 445    texObj = _mesa_get_current_tex_object(ctx, target);
 446
 447    if (!texObj)
 448       return;
 449
 450    if (dPriv->lastStamp != dPriv->dri2.stamp ||
 451        !pDRICtx->driScreenPriv->dri2.useInvalidate)
 452       intel_update_renderbuffers(pDRICtx, dPriv);
 453
 454    rb = intel_get_renderbuffer(fb, BUFFER_FRONT_LEFT);
 455    /* If the miptree isn't set, then intel_update_renderbuffers was unable
 456     * to get the BO for the drawable from the window system.
 457     */
 458    if (!rb || !rb->mt)
 459       return;
 460
 461    if (rb->mt->cpp == 4) {
 462       if (texture_format == __DRI_TEXTURE_FORMAT_RGB)
 463          internal_format = GL_RGB;
 464       else
 465          internal_format = GL_RGBA;
 466    } else if (rb->mt->cpp == 2) {
 467       /* This is 565 */
 468       internal_format = GL_RGB;
 469    }
 470
 471    intel_miptree_prepare_external(brw, rb->mt);
 472
 473    _mesa_lock_texture(&brw->ctx, texObj);
 474    texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
 475    intel_set_texture_image_mt(brw, texImage, internal_format, rb->mt);
 476    _mesa_unlock_texture(&brw->ctx, texObj);
 477 }
 478
 479 static GLboolean
 480 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
 481                                   struct gl_renderbuffer *rb,
 482                                   struct gl_texture_image *image)
 483 {
 484    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 485    struct intel_texture_image *intel_image = intel_texture_image(image);
 486    struct gl_texture_object *texobj = image->TexObject;
 487    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 488
 489    /* We can only handle RB allocated with AllocRenderbufferStorage, or
 490     * window-system renderbuffers.
 491     */
 492    assert(!rb->TexImage);
 493
 494    if (!irb->mt)
 495       return false;
 496
 497    _mesa_lock_texture(ctx, texobj);
 498    _mesa_init_teximage_fields(ctx, image,
 499                               rb->Width, rb->Height, 1,
 500                               0, rb->InternalFormat, rb->Format);
 501    image->NumSamples = rb->NumSamples;
 502
 503    intel_miptree_reference(&intel_image->mt, irb->mt);
 504
 505    /* Immediately validate the image to the object. */
 506    intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
 507
 508    intel_texobj->needs_validate = true;
 509    _mesa_unlock_texture(ctx, texobj);
 510
 511    return true;
 512 }
 513
 514 void
 515 intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 516 {
 517    /* The old interface didn't have the format argument, so copy our
 518     * implementation's behavior at the time.
 519     */
 520    intelSetTexBuffer2(pDRICtx, target, __DRI_TEXTURE_FORMAT_RGBA, dPriv);
 521 }
 522
 523 static void
 524 intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
 525                               struct gl_texture_object *texObj,
 526                               struct gl_texture_image *texImage,
 527                               GLeglImageOES image_handle)
 528 {
 529    struct brw_context *brw = brw_context(ctx);
 530    struct intel_mipmap_tree *mt;
 531    __DRIscreen *dri_screen = brw->screen->driScrnPriv;
 532    __DRIimage *image;
 533
 534    image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
 535                                                   dri_screen->loaderPrivate);
 536    if (image == NULL)
 537       return;
 538
 539    /* We support external textures only for EGLImages created with
 540     * EGL_EXT_image_dma_buf_import. We may lift that restriction in the future.
 541     */
 542    if (target == GL_TEXTURE_EXTERNAL_OES && !image->dma_buf_imported) {
 543       _mesa_error(ctx, GL_INVALID_OPERATION,
 544             "glEGLImageTargetTexture2DOES(external target is enabled only "
 545                "for images created with EGL_EXT_image_dma_buf_import");
 546       return;
 547    }
 548
 549    /* Disallow depth/stencil textures: we don't have a way to pass the
 550     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
 551     */
 552    if (image->has_depthstencil) {
 553       _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
 554       return;
 555    }
 556
 557    mt = intel_miptree_create_for_dri_image(brw, image, target, image->format,
 558                                            false);
 559    if (mt == NULL)
 560       return;
 561
 562    struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
 563    intel_texobj->planar_format = image->planar_format;
 564
 565    const GLenum internal_format =
 566       image->internal_format != 0 ?
 567       image->internal_format : _mesa_get_format_base_format(mt->format);
 568    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 569    intel_miptree_release(&mt);
 570 }
 571
 572 /**
 573  * \brief A fast path for glGetTexImage.
 574  *
 575  * \see intel_readpixels_tiled_memcpy()
 576  */
 577 static bool
 578 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
 579                                   struct gl_texture_image *texImage,
 580                                   GLint xoffset, GLint yoffset,
 581                                   GLsizei width, GLsizei height,
 582                                   GLenum format, GLenum type,
 583                                   GLvoid *pixels,
 584                                   const struct gl_pixelstore_attrib *packing)
 585 {
 586    struct brw_context *brw = brw_context(ctx);
 587    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 588    struct intel_texture_image *image = intel_texture_image(texImage);
 589    int dst_pitch;
 590
 591    /* The miptree's buffer. */
 592    struct brw_bo *bo;
 593
 594    uint32_t cpp;
 595    mem_copy_fn mem_copy = NULL;
 596
 597    /* This fastpath is restricted to specific texture types:
 598     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 599     * more types.
 600     *
 601     * FINISHME: The restrictions below on packing alignment and packing row
 602     * length are likely unneeded now because we calculate the destination stride
 603     * with _mesa_image_row_stride. However, before removing the restrictions
 604     * we need tests.
 605     */
 606    if (!devinfo->has_llc ||
 607        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 608        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 609          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 610        pixels == NULL ||
 611        _mesa_is_bufferobj(packing->BufferObj) ||
 612        packing->Alignment > 4 ||
 613        packing->SkipPixels > 0 ||
 614        packing->SkipRows > 0 ||
 615        (packing->RowLength != 0 && packing->RowLength != width) ||
 616        packing->SwapBytes ||
 617        packing->LsbFirst ||
 618        packing->Invert)
 619       return false;
 620
 621    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
 622     * function doesn't set the last channel to 1. Note this checks BaseFormat
 623     * rather than TexFormat in case the RGBX format is being simulated with an
 624     * RGBA format.
 625     */
 626    if (texImage->_BaseFormat == GL_RGB)
 627       return false;
 628
 629    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 630       return false;
 631
 632    /* If this is a nontrivial texture view, let another path handle it instead. */
 633    if (texImage->TexObject->MinLayer)
 634       return false;
 635
 636    if (!image->mt ||
 637        (image->mt->surf.tiling != ISL_TILING_X &&
 638         image->mt->surf.tiling != ISL_TILING_Y0)) {
 639       /* The algorithm is written only for X- or Y-tiled memory. */
 640       return false;
 641    }
 642
 643    /* tiled_to_linear() assumes that if the object is swizzled, it is using
 644     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 645     * true on gen5 and above.
 646     *
 647     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 648     * parts of the memory aren't swizzled at all. Userspace just can't handle
 649     * that.
 650     */
 651    if (devinfo->gen < 5 && brw->has_swizzling)
 652       return false;
 653
 654    int level = texImage->Level + texImage->TexObject->MinLevel;
 655
 656    /* Since we are going to write raw data to the miptree, we need to resolve
 657     * any pending fast color clears before we start.
 658     */
 659    assert(image->mt->surf.logical_level0_px.depth == 1);
 660    assert(image->mt->surf.logical_level0_px.array_len == 1);
 661
 662    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 663
 664    bo = image->mt->bo;
 665
 666    if (brw_batch_references(&brw->batch, bo)) {
 667       perf_debug("Flushing before mapping a referenced bo.\n");
 668       intel_batchbuffer_flush(brw);
 669    }
 670
 671    void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
 672    if (map == NULL) {
 673       DBG("%s: failed to map bo\n", __func__);
 674       return false;
 675    }
 676
 677    dst_pitch = _mesa_image_row_stride(packing, width, format, type);
 678
 679    DBG("%s: level=%d x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 680        "mesa_format=0x%x tiling=%d "
 681        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
 682        __func__, texImage->Level, xoffset, yoffset, width, height,
 683        format, type, texImage->TexFormat, image->mt->surf.tiling,
 684        packing->Alignment, packing->RowLength, packing->SkipPixels,
 685        packing->SkipRows);
 686
 687    /* Adjust x and y offset based on miplevel */
 688    unsigned level_x, level_y;
 689    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 690    xoffset += level_x;
 691    yoffset += level_y;
 692
 693    tiled_to_linear(
 694       xoffset * cpp, (xoffset + width) * cpp,
 695       yoffset, yoffset + height,
 696       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
 697       map,
 698       dst_pitch, image->mt->surf.row_pitch,
 699       brw->has_swizzling,
 700       image->mt->surf.tiling,
 701       mem_copy
 702    );
 703
 704    brw_bo_unmap(bo);
 705    return true;
 706 }
 707
 708 static void
 709 intel_get_tex_sub_image(struct gl_context *ctx,
 710                         GLint xoffset, GLint yoffset, GLint zoffset,
 711                         GLsizei width, GLsizei height, GLint depth,
 712                         GLenum format, GLenum type, GLvoid *pixels,
 713                         struct gl_texture_image *texImage)
 714 {
 715    struct brw_context *brw = brw_context(ctx);
 716    bool ok;
 717
 718    DBG("%s\n", __func__);
 719
 720    if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
 721       if (_mesa_meta_pbo_GetTexSubImage(ctx, 3, texImage,
 722                                         xoffset, yoffset, zoffset,
 723                                         width, height, depth, format, type,
 724                                         pixels, &ctx->Pack)) {
 725          /* Flush to guarantee coherency between the render cache and other
 726           * caches the PBO could potentially be bound to after this point.
 727           * See the related comment in intelReadPixels() for a more detailed
 728           * explanation.
 729           */
 730          brw_emit_mi_flush(brw);
 731          return;
 732       }
 733
 734       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
 735    }
 736
 737    ok = intel_gettexsubimage_tiled_memcpy(ctx, texImage, xoffset, yoffset,
 738                                           width, height,
 739                                           format, type, pixels, &ctx->Pack);
 740
 741    if(ok)
 742       return;
 743
 744    _mesa_meta_GetTexSubImage(ctx, xoffset, yoffset, zoffset,
 745                              width, height, depth,
 746                              format, type, pixels, texImage);
 747
 748    DBG("%s - DONE\n", __func__);
 749 }
 750
 751 static void
 752 flush_astc_denorms(struct gl_context *ctx, GLuint dims,
 753                    struct gl_texture_image *texImage,
 754                    GLint xoffset, GLint yoffset, GLint zoffset,
 755                    GLsizei width, GLsizei height, GLsizei depth)
 756 {
 757    struct compressed_pixelstore store;
 758    _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
 759                                        width, height, depth,
 760                                        &ctx->Unpack, &store);
 761
 762    for (int slice = 0; slice < store.CopySlices; slice++) {
 763
 764       /* Map dest texture buffer */
 765       GLubyte *dstMap;
 766       GLint dstRowStride;
 767       ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
 768                                   xoffset, yoffset, width, height,
 769                                   GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
 770                                   &dstMap, &dstRowStride);
 771       if (!dstMap)
 772          continue;
 773
 774       for (int i = 0; i < store.CopyRowsPerSlice; i++) {
 775
 776          /* An ASTC block is stored in little endian mode. The byte that
 777           * contains bits 0..7 is stored at the lower address in memory.
 778           */
 779          struct astc_void_extent {
 780             uint16_t header : 12;
 781             uint16_t dontcare[3];
 782             uint16_t R;
 783             uint16_t G;
 784             uint16_t B;
 785             uint16_t A;
 786          } *blocks = (struct astc_void_extent*) dstMap;
 787
 788          /* Iterate over every copied block in the row */
 789          for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
 790
 791             /* Check if the header matches that of an LDR void-extent block */
 792             if (blocks[j].header == 0xDFC) {
 793
 794                /* Flush UNORM16 values that would be denormalized */
 795                if (blocks[j].A < 4) blocks[j].A = 0;
 796                if (blocks[j].B < 4) blocks[j].B = 0;
 797                if (blocks[j].G < 4) blocks[j].G = 0;
 798                if (blocks[j].R < 4) blocks[j].R = 0;
 799             }
 800          }
 801
 802          dstMap += dstRowStride;
 803       }
 804
 805       ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
 806    }
 807 }
 808
 809
 810 static void
 811 intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
 812                         struct gl_texture_image *texImage,
 813                         GLint xoffset, GLint yoffset, GLint zoffset,
 814                         GLsizei width, GLsizei height, GLsizei depth,
 815                         GLenum format,
 816                         GLsizei imageSize, const GLvoid *data)
 817 {
 818    /* Upload the compressed data blocks */
 819    _mesa_store_compressed_texsubimage(ctx, dims, texImage,
 820                                       xoffset, yoffset, zoffset,
 821                                       width, height, depth,
 822                                       format, imageSize, data);
 823
 824    /* Fix up copied ASTC blocks if necessary */
 825    GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
 826                                                         texImage->TexFormat);
 827    bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
 828                         !_mesa_is_srgb_format(gl_format);
 829    struct brw_context *brw = (struct brw_context*) ctx;
 830    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 831    if (devinfo->gen == 9 && is_linear_astc)
 832       flush_astc_denorms(ctx, dims, texImage,
 833                          xoffset, yoffset, zoffset,
 834                          width, height, depth);
 835 }
 836
 837 void
 838 intelInitTextureImageFuncs(struct dd_function_table *functions)
 839 {
 840    functions->TexImage = intelTexImage;
 841    functions->TexSubImage = intelTexSubImage;
 842    functions->CompressedTexSubImage = intelCompressedTexSubImage;
 843    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
 844    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
 845    functions->GetTexSubImage = intel_get_tex_sub_image;
 846 }