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

   1
   2 #include "main/macros.h"
   3 #include "main/mtypes.h"
   4 #include "main/enums.h"
   5 #include "main/bufferobj.h"
   6 #include "main/context.h"
   7 #include "main/formats.h"
   8 #include "main/glformats.h"
   9 #include "main/image.h"
  10 #include "main/pbo.h"
  11 #include "main/renderbuffer.h"
  12 #include "main/texcompress.h"
  13 #include "main/texgetimage.h"
  14 #include "main/texobj.h"
  15 #include "main/teximage.h"
  16 #include "main/texstore.h"
  17
  18 #include "drivers/common/meta.h"
  19
  20 #include "intel_mipmap_tree.h"
  21 #include "intel_buffer_objects.h"
  22 #include "intel_batchbuffer.h"
  23 #include "intel_tex.h"
  24 #include "intel_blit.h"
  25 #include "intel_fbo.h"
  26 #include "intel_image.h"
  27 #include "intel_tiled_memcpy.h"
  28 #include "brw_context.h"
  29
  30 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
  31
  32 /* Make sure one doesn't end up shrinking base level zero unnecessarily.
  33  * Determining the base level dimension by shifting higher level dimension
  34  * ends up in off-by-one value in case base level has NPOT size (for example,
  35  * 293 != 146 << 1).
  36  * Choose the original base level dimension when shifted dimensions agree.
  37  * Otherwise assume real resize is intended and use the new shifted value.
  38  */
  39 static unsigned
  40 get_base_dim(unsigned old_base_dim, unsigned new_level_dim, unsigned level)
  41 {
  42    const unsigned old_level_dim = old_base_dim >> level;
  43    const unsigned new_base_dim = new_level_dim << level;
  44
  45    return old_level_dim == new_level_dim ? old_base_dim : new_base_dim;
  46 }
  47
  48 /* Work back from the specified level of the image to the baselevel and create a
  49  * miptree of that size.
  50  */
  51 struct intel_mipmap_tree *
  52 intel_miptree_create_for_teximage(struct brw_context *brw,
  53                                   struct intel_texture_object *intelObj,
  54                                   struct intel_texture_image *intelImage,
  55                                   enum intel_miptree_create_flags flags)
  56 {
  57    GLuint lastLevel;
  58    int width, height, depth;
  59    unsigned old_width = 0, old_height = 0, old_depth = 0;
  60    const struct intel_mipmap_tree *old_mt = intelObj->mt;
  61    const unsigned level = intelImage->base.Base.Level;
  62
  63    intel_get_image_dims(&intelImage->base.Base, &width, &height, &depth);
  64
  65    if (old_mt) {
  66       old_width = old_mt->surf.logical_level0_px.width;
  67       old_height = old_mt->surf.logical_level0_px.height;
  68       old_depth = old_mt->surf.dim == ISL_SURF_DIM_3D ?
  69                      old_mt->surf.logical_level0_px.depth :
  70                      old_mt->surf.logical_level0_px.array_len;
  71    }
  72
  73    DBG("%s\n", __func__);
  74
  75    /* Figure out image dimensions at start level. */
  76    switch(intelObj->base.Target) {
  77    case GL_TEXTURE_2D_MULTISAMPLE:
  78    case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
  79    case GL_TEXTURE_RECTANGLE:
  80    case GL_TEXTURE_EXTERNAL_OES:
  81       assert(level == 0);
  82       break;
  83    case GL_TEXTURE_3D:
  84       depth = old_mt ? get_base_dim(old_depth, depth, level) :
  85                        depth << level;
  86       /* Fall through */
  87    case GL_TEXTURE_2D:
  88    case GL_TEXTURE_2D_ARRAY:
  89    case GL_TEXTURE_CUBE_MAP:
  90    case GL_TEXTURE_CUBE_MAP_ARRAY:
  91       height = old_mt ? get_base_dim(old_height, height, level) :
  92                         height << level;
  93       /* Fall through */
  94    case GL_TEXTURE_1D:
  95    case GL_TEXTURE_1D_ARRAY:
  96       width = old_mt ? get_base_dim(old_width, width, level) :
  97                        width << level;
  98       break;
  99    default:
 100       unreachable("Unexpected target");
 101    }
 102
 103    /* Guess a reasonable value for lastLevel.  This is probably going
 104     * to be wrong fairly often and might mean that we have to look at
 105     * resizable buffers, or require that buffers implement lazy
 106     * pagetable arrangements.
 107     */
 108    if ((intelObj->base.Sampler.MinFilter == GL_NEAREST ||
 109         intelObj->base.Sampler.MinFilter == GL_LINEAR) &&
 110        intelImage->base.Base.Level == 0 &&
 111        !intelObj->base.GenerateMipmap) {
 112       lastLevel = 0;
 113    } else {
 114       lastLevel = _mesa_get_tex_max_num_levels(intelObj->base.Target,
 115                                                width, height, depth) - 1;
 116    }
 117
 118    return intel_miptree_create(brw,
 119                                intelObj->base.Target,
 120                                intelImage->base.Base.TexFormat,
 121                                0,
 122                                lastLevel,
 123                                width,
 124                                height,
 125                                depth,
 126                                MAX2(intelImage->base.Base.NumSamples, 1),
 127                                flags);
 128 }
 129
 130
 131 /**
 132  * \brief A fast path for glTexImage and glTexSubImage.
 133  *
 134  * This fast path is taken when the texture format is BGRA, RGBA,
 135  * A or L and when the texture memory is X- or Y-tiled.  It uploads
 136  * the texture data by mapping the texture memory without a GTT fence, thus
 137  * acquiring a tiled view of the memory, and then copying sucessive
 138  * spans within each tile.
 139  *
 140  * This is a performance win over the conventional texture upload path because
 141  * it avoids the performance penalty of writing through the write-combine
 142  * buffer. In the conventional texture upload path,
 143  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
 144  * fence, thus acquiring a linear view of the memory, then each row in the
 145  * image is memcpy'd. In this fast path, we replace each row's copy with
 146  * a sequence of copies over each linear span in tile.
 147  *
 148  * One use case is Google Chrome's paint rectangles.  Chrome (as
 149  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
 150  * Each page's content is initially uploaded with glTexImage2D and damaged
 151  * regions are updated with glTexSubImage2D. On some workloads, the
 152  * performance gain of this fastpath on Sandybridge is over 5x.
 153  */
 154 static bool
 155 intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 156                                GLuint dims,
 157                                struct gl_texture_image *texImage,
 158                                GLint xoffset, GLint yoffset, GLint zoffset,
 159                                GLsizei width, GLsizei height, GLsizei depth,
 160                                GLenum format, GLenum type,
 161                                const GLvoid *pixels,
 162                                const struct gl_pixelstore_attrib *packing)
 163 {
 164    struct brw_context *brw = brw_context(ctx);
 165    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 166    struct intel_texture_image *image = intel_texture_image(texImage);
 167    int src_pitch;
 168
 169    /* The miptree's buffer. */
 170    struct brw_bo *bo;
 171
 172    uint32_t cpp;
 173    mem_copy_fn mem_copy = NULL;
 174
 175    /* This fastpath is restricted to specific texture types:
 176     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 177     * more types.
 178     *
 179     * FINISHME: The restrictions below on packing alignment and packing row
 180     * length are likely unneeded now because we calculate the source stride
 181     * with _mesa_image_row_stride. However, before removing the restrictions
 182     * we need tests.
 183     */
 184    if (!devinfo->has_llc ||
 185        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 186        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 187          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 188        pixels == NULL ||
 189        _mesa_is_bufferobj(packing->BufferObj) ||
 190        packing->Alignment > 4 ||
 191        packing->SkipPixels > 0 ||
 192        packing->SkipRows > 0 ||
 193        (packing->RowLength != 0 && packing->RowLength != width) ||
 194        packing->SwapBytes ||
 195        packing->LsbFirst ||
 196        packing->Invert)
 197       return false;
 198
 199    /* Only a simple blit, no scale, bias or other mapping. */
 200    if (ctx->_ImageTransferState)
 201       return false;
 202
 203    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 204       return false;
 205
 206    /* If this is a nontrivial texture view, let another path handle it instead. */
 207    if (texImage->TexObject->MinLayer)
 208       return false;
 209
 210    if (!image->mt ||
 211        (image->mt->surf.tiling != ISL_TILING_X &&
 212         image->mt->surf.tiling != ISL_TILING_Y0)) {
 213       /* The algorithm is written only for X- or Y-tiled memory. */
 214       return false;
 215    }
 216
 217    /* linear_to_tiled() assumes that if the object is swizzled, it is using
 218     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 219     * true on gen5 and above.
 220     *
 221     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 222     * parts of the memory aren't swizzled at all. Userspace just can't handle
 223     * that.
 224     */
 225    if (devinfo->gen < 5 && brw->has_swizzling)
 226       return false;
 227
 228    int level = texImage->Level + texImage->TexObject->MinLevel;
 229
 230    /* Since we are going to write raw data to the miptree, we need to resolve
 231     * any pending fast color clears before we start.
 232     */
 233    assert(image->mt->surf.logical_level0_px.depth == 1);
 234    assert(image->mt->surf.logical_level0_px.array_len == 1);
 235
 236    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 237
 238    bo = image->mt->bo;
 239
 240    if (brw_batch_references(&brw->batch, bo)) {
 241       perf_debug("Flushing before mapping a referenced bo.\n");
 242       intel_batchbuffer_flush(brw);
 243    }
 244
 245    void *map = brw_bo_map(brw, bo, MAP_WRITE | MAP_RAW);
 246    if (map == NULL) {
 247       DBG("%s: failed to map bo\n", __func__);
 248       return false;
 249    }
 250
 251    src_pitch = _mesa_image_row_stride(packing, width, format, type);
 252
 253    /* We postponed printing this message until having committed to executing
 254     * the function.
 255     */
 256    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 257        "mesa_format=0x%x tiling=%d "
 258        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d) ",
 259        __func__, texImage->Level, xoffset, yoffset, width, height,
 260        format, type, texImage->TexFormat, image->mt->surf.tiling,
 261        packing->Alignment, packing->RowLength, packing->SkipPixels,
 262        packing->SkipRows);
 263
 264    /* Adjust x and y offset based on miplevel */
 265    unsigned level_x, level_y;
 266    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 267    xoffset += level_x;
 268    yoffset += level_y;
 269
 270    linear_to_tiled(
 271       xoffset * cpp, (xoffset + width) * cpp,
 272       yoffset, yoffset + height,
 273       map,
 274       pixels - (ptrdiff_t) yoffset * src_pitch - (ptrdiff_t) xoffset * cpp,
 275       image->mt->surf.row_pitch, src_pitch,
 276       brw->has_swizzling,
 277       image->mt->surf.tiling,
 278       mem_copy
 279    );
 280
 281    brw_bo_unmap(bo);
 282    return true;
 283 }
 284
 285
 286 static void
 287 intelTexImage(struct gl_context * ctx,
 288               GLuint dims,
 289               struct gl_texture_image *texImage,
 290               GLenum format, GLenum type, const void *pixels,
 291               const struct gl_pixelstore_attrib *unpack)
 292 {
 293    struct intel_texture_image *intelImage = intel_texture_image(texImage);
 294    bool ok;
 295
 296    bool tex_busy = intelImage->mt && brw_bo_busy(intelImage->mt->bo);
 297
 298    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 299        __func__, _mesa_get_format_name(texImage->TexFormat),
 300        _mesa_enum_to_string(texImage->TexObject->Target),
 301        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 302        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 303
 304    /* Allocate storage for texture data. */
 305    if (!ctx->Driver.AllocTextureImageBuffer(ctx, texImage)) {
 306       _mesa_error(ctx, GL_OUT_OF_MEMORY, "glTexImage%uD", dims);
 307       return;
 308    }
 309
 310    assert(intelImage->mt);
 311
 312    if (intelImage->mt->format == MESA_FORMAT_S_UINT8)
 313       intelImage->mt->r8stencil_needs_update = true;
 314
 315    ok = _mesa_meta_pbo_TexSubImage(ctx, dims, texImage, 0, 0, 0,
 316                                    texImage->Width, texImage->Height,
 317                                    texImage->Depth,
 318                                    format, type, pixels,
 319                                    tex_busy, unpack);
 320    if (ok)
 321       return;
 322
 323    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 324                                        0, 0, 0, /*x,y,z offsets*/
 325                                        texImage->Width,
 326                                        texImage->Height,
 327                                        texImage->Depth,
 328                                        format, type, pixels, unpack);
 329    if (ok)
 330       return;
 331
 332    DBG("%s: upload image %dx%dx%d pixels %p\n",
 333        __func__, texImage->Width, texImage->Height, texImage->Depth,
 334        pixels);
 335
 336    _mesa_store_teximage(ctx, dims, texImage,
 337                         format, type, pixels, unpack);
 338 }
 339
 340
 341 static void
 342 intelTexSubImage(struct gl_context * ctx,
 343                  GLuint dims,
 344                  struct gl_texture_image *texImage,
 345                  GLint xoffset, GLint yoffset, GLint zoffset,
 346                  GLsizei width, GLsizei height, GLsizei depth,
 347                  GLenum format, GLenum type,
 348                  const GLvoid * pixels,
 349                  const struct gl_pixelstore_attrib *packing)
 350 {
 351    struct intel_mipmap_tree *mt = intel_texture_image(texImage)->mt;
 352    bool ok;
 353
 354    bool tex_busy = mt && brw_bo_busy(mt->bo);
 355
 356    if (mt && mt->format == MESA_FORMAT_S_UINT8)
 357       mt->r8stencil_needs_update = true;
 358
 359    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 360        __func__, _mesa_get_format_name(texImage->TexFormat),
 361        _mesa_enum_to_string(texImage->TexObject->Target),
 362        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 363        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 364
 365    ok = _mesa_meta_pbo_TexSubImage(ctx, dims, texImage,
 366                                    xoffset, yoffset, zoffset,
 367                                    width, height, depth, format, type,
 368                                    pixels, tex_busy, packing);
 369    if (ok)
 370       return;
 371
 372    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 373                                        xoffset, yoffset, zoffset,
 374                                        width, height, depth,
 375                                        format, type, pixels, packing);
 376    if (ok)
 377      return;
 378
 379    _mesa_store_texsubimage(ctx, dims, texImage,
 380                            xoffset, yoffset, zoffset,
 381                            width, height, depth,
 382                            format, type, pixels, packing);
 383 }
 384
 385
 386 static void
 387 intel_set_texture_image_mt(struct brw_context *brw,
 388                            struct gl_texture_image *image,
 389                            GLenum internal_format,
 390                            struct intel_mipmap_tree *mt)
 391
 392 {
 393    struct gl_texture_object *texobj = image->TexObject;
 394    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 395    struct intel_texture_image *intel_image = intel_texture_image(image);
 396
 397    _mesa_init_teximage_fields(&brw->ctx, image,
 398                               mt->surf.logical_level0_px.width,
 399                               mt->surf.logical_level0_px.height, 1,
 400                               0, internal_format, mt->format);
 401
 402    brw->ctx.Driver.FreeTextureImageBuffer(&brw->ctx, image);
 403
 404    intel_texobj->needs_validate = true;
 405    intel_image->base.RowStride = mt->surf.row_pitch / mt->cpp;
 406    assert(mt->surf.row_pitch % mt->cpp == 0);
 407
 408    intel_miptree_reference(&intel_image->mt, mt);
 409
 410    /* Immediately validate the image to the object. */
 411    intel_miptree_reference(&intel_texobj->mt, mt);
 412 }
 413
 414
 415 void
 416 intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
 417                    GLint texture_format,
 418                    __DRIdrawable *dPriv)
 419 {
 420    struct gl_framebuffer *fb = dPriv->driverPrivate;
 421    struct brw_context *brw = pDRICtx->driverPrivate;
 422    struct gl_context *ctx = &brw->ctx;
 423    struct intel_renderbuffer *rb;
 424    struct gl_texture_object *texObj;
 425    struct gl_texture_image *texImage;
 426    mesa_format texFormat = MESA_FORMAT_NONE;
 427    struct intel_mipmap_tree *mt;
 428    GLenum internal_format = 0;
 429
 430    texObj = _mesa_get_current_tex_object(ctx, target);
 431
 432    if (!texObj)
 433       return;
 434
 435    if (dPriv->lastStamp != dPriv->dri2.stamp ||
 436        !pDRICtx->driScreenPriv->dri2.useInvalidate)
 437       intel_update_renderbuffers(pDRICtx, dPriv);
 438
 439    rb = intel_get_renderbuffer(fb, BUFFER_FRONT_LEFT);
 440    /* If the miptree isn't set, then intel_update_renderbuffers was unable
 441     * to get the BO for the drawable from the window system.
 442     */
 443    if (!rb || !rb->mt)
 444       return;
 445
 446    if (rb->mt->cpp == 4) {
 447       if (texture_format == __DRI_TEXTURE_FORMAT_RGB) {
 448          internal_format = GL_RGB;
 449          texFormat = MESA_FORMAT_B8G8R8X8_UNORM;
 450       }
 451       else {
 452          internal_format = GL_RGBA;
 453          texFormat = MESA_FORMAT_B8G8R8A8_UNORM;
 454       }
 455    } else if (rb->mt->cpp == 2) {
 456       internal_format = GL_RGB;
 457       texFormat = MESA_FORMAT_B5G6R5_UNORM;
 458    }
 459
 460    intel_miptree_make_shareable(brw, rb->mt);
 461    mt = intel_miptree_create_for_bo(brw, rb->mt->bo, texFormat, 0,
 462                                     rb->Base.Base.Width,
 463                                     rb->Base.Base.Height,
 464                                     1, rb->mt->surf.row_pitch,
 465                                     MIPTREE_CREATE_DEFAULT);
 466    if (mt == NULL)
 467        return;
 468    mt->target = target;
 469
 470    _mesa_lock_texture(&brw->ctx, texObj);
 471    texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
 472    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 473    intel_miptree_release(&mt);
 474    _mesa_unlock_texture(&brw->ctx, texObj);
 475 }
 476
 477 static GLboolean
 478 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
 479                                   struct gl_renderbuffer *rb,
 480                                   struct gl_texture_image *image)
 481 {
 482    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 483    struct intel_texture_image *intel_image = intel_texture_image(image);
 484    struct gl_texture_object *texobj = image->TexObject;
 485    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 486
 487    /* We can only handle RB allocated with AllocRenderbufferStorage, or
 488     * window-system renderbuffers.
 489     */
 490    assert(!rb->TexImage);
 491
 492    if (!irb->mt)
 493       return false;
 494
 495    _mesa_lock_texture(ctx, texobj);
 496    _mesa_init_teximage_fields(ctx, image,
 497                               rb->Width, rb->Height, 1,
 498                               0, rb->InternalFormat, rb->Format);
 499    image->NumSamples = rb->NumSamples;
 500
 501    intel_miptree_reference(&intel_image->mt, irb->mt);
 502
 503    /* Immediately validate the image to the object. */
 504    intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
 505
 506    intel_texobj->needs_validate = true;
 507    _mesa_unlock_texture(ctx, texobj);
 508
 509    return true;
 510 }
 511
 512 void
 513 intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 514 {
 515    /* The old interface didn't have the format argument, so copy our
 516     * implementation's behavior at the time.
 517     */
 518    intelSetTexBuffer2(pDRICtx, target, __DRI_TEXTURE_FORMAT_RGBA, dPriv);
 519 }
 520
 521 static void
 522 intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
 523                               struct gl_texture_object *texObj,
 524                               struct gl_texture_image *texImage,
 525                               GLeglImageOES image_handle)
 526 {
 527    struct brw_context *brw = brw_context(ctx);
 528    struct intel_mipmap_tree *mt;
 529    __DRIscreen *dri_screen = brw->screen->driScrnPriv;
 530    __DRIimage *image;
 531
 532    image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
 533                                                   dri_screen->loaderPrivate);
 534    if (image == NULL)
 535       return;
 536
 537    /* We support external textures only for EGLImages created with
 538     * EGL_EXT_image_dma_buf_import. We may lift that restriction in the future.
 539     */
 540    if (target == GL_TEXTURE_EXTERNAL_OES && !image->dma_buf_imported) {
 541       _mesa_error(ctx, GL_INVALID_OPERATION,
 542             "glEGLImageTargetTexture2DOES(external target is enabled only "
 543                "for images created with EGL_EXT_image_dma_buf_import");
 544       return;
 545    }
 546
 547    /* Disallow depth/stencil textures: we don't have a way to pass the
 548     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
 549     */
 550    if (image->has_depthstencil) {
 551       _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
 552       return;
 553    }
 554
 555    mt = intel_miptree_create_for_dri_image(brw, image, target,
 556                                            ISL_COLORSPACE_NONE, false);
 557    if (mt == NULL)
 558       return;
 559
 560    struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
 561    intel_texobj->planar_format = image->planar_format;
 562
 563    const GLenum internal_format =
 564       image->internal_format != 0 ?
 565       image->internal_format : _mesa_get_format_base_format(mt->format);
 566    intel_set_texture_image_mt(brw, texImage, internal_format, mt);
 567    intel_miptree_release(&mt);
 568 }
 569
 570 /**
 571  * \brief A fast path for glGetTexImage.
 572  *
 573  * \see intel_readpixels_tiled_memcpy()
 574  */
 575 static bool
 576 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
 577                                   struct gl_texture_image *texImage,
 578                                   GLint xoffset, GLint yoffset,
 579                                   GLsizei width, GLsizei height,
 580                                   GLenum format, GLenum type,
 581                                   GLvoid *pixels,
 582                                   const struct gl_pixelstore_attrib *packing)
 583 {
 584    struct brw_context *brw = brw_context(ctx);
 585    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 586    struct intel_texture_image *image = intel_texture_image(texImage);
 587    int dst_pitch;
 588
 589    /* The miptree's buffer. */
 590    struct brw_bo *bo;
 591
 592    uint32_t cpp;
 593    mem_copy_fn mem_copy = NULL;
 594
 595    /* This fastpath is restricted to specific texture types:
 596     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 597     * more types.
 598     *
 599     * FINISHME: The restrictions below on packing alignment and packing row
 600     * length are likely unneeded now because we calculate the destination stride
 601     * with _mesa_image_row_stride. However, before removing the restrictions
 602     * we need tests.
 603     */
 604    if (!devinfo->has_llc ||
 605        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 606        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 607          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 608        pixels == NULL ||
 609        _mesa_is_bufferobj(packing->BufferObj) ||
 610        packing->Alignment > 4 ||
 611        packing->SkipPixels > 0 ||
 612        packing->SkipRows > 0 ||
 613        (packing->RowLength != 0 && packing->RowLength != width) ||
 614        packing->SwapBytes ||
 615        packing->LsbFirst ||
 616        packing->Invert)
 617       return false;
 618
 619    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
 620     * function doesn't set the last channel to 1. Note this checks BaseFormat
 621     * rather than TexFormat in case the RGBX format is being simulated with an
 622     * RGBA format.
 623     */
 624    if (texImage->_BaseFormat == GL_RGB)
 625       return false;
 626
 627    if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp))
 628       return false;
 629
 630    /* If this is a nontrivial texture view, let another path handle it instead. */
 631    if (texImage->TexObject->MinLayer)
 632       return false;
 633
 634    if (!image->mt ||
 635        (image->mt->surf.tiling != ISL_TILING_X &&
 636         image->mt->surf.tiling != ISL_TILING_Y0)) {
 637       /* The algorithm is written only for X- or Y-tiled memory. */
 638       return false;
 639    }
 640
 641    /* tiled_to_linear() assumes that if the object is swizzled, it is using
 642     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 643     * true on gen5 and above.
 644     *
 645     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 646     * parts of the memory aren't swizzled at all. Userspace just can't handle
 647     * that.
 648     */
 649    if (devinfo->gen < 5 && brw->has_swizzling)
 650       return false;
 651
 652    int level = texImage->Level + texImage->TexObject->MinLevel;
 653
 654    /* Since we are going to write raw data to the miptree, we need to resolve
 655     * any pending fast color clears before we start.
 656     */
 657    assert(image->mt->surf.logical_level0_px.depth == 1);
 658    assert(image->mt->surf.logical_level0_px.array_len == 1);
 659
 660    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 661
 662    bo = image->mt->bo;
 663
 664    if (brw_batch_references(&brw->batch, bo)) {
 665       perf_debug("Flushing before mapping a referenced bo.\n");
 666       intel_batchbuffer_flush(brw);
 667    }
 668
 669    void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
 670    if (map == NULL) {
 671       DBG("%s: failed to map bo\n", __func__);
 672       return false;
 673    }
 674
 675    dst_pitch = _mesa_image_row_stride(packing, width, format, type);
 676
 677    DBG("%s: level=%d x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 678        "mesa_format=0x%x tiling=%d "
 679        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
 680        __func__, texImage->Level, xoffset, yoffset, width, height,
 681        format, type, texImage->TexFormat, image->mt->surf.tiling,
 682        packing->Alignment, packing->RowLength, packing->SkipPixels,
 683        packing->SkipRows);
 684
 685    /* Adjust x and y offset based on miplevel */
 686    unsigned level_x, level_y;
 687    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 688    xoffset += level_x;
 689    yoffset += level_y;
 690
 691    tiled_to_linear(
 692       xoffset * cpp, (xoffset + width) * cpp,
 693       yoffset, yoffset + height,
 694       pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
 695       map,
 696       dst_pitch, image->mt->surf.row_pitch,
 697       brw->has_swizzling,
 698       image->mt->surf.tiling,
 699       mem_copy
 700    );
 701
 702    brw_bo_unmap(bo);
 703    return true;
 704 }
 705
 706 static void
 707 intel_get_tex_sub_image(struct gl_context *ctx,
 708                         GLint xoffset, GLint yoffset, GLint zoffset,
 709                         GLsizei width, GLsizei height, GLint depth,
 710                         GLenum format, GLenum type, GLvoid *pixels,
 711                         struct gl_texture_image *texImage)
 712 {
 713    struct brw_context *brw = brw_context(ctx);
 714    bool ok;
 715
 716    DBG("%s\n", __func__);
 717
 718    if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
 719       if (_mesa_meta_pbo_GetTexSubImage(ctx, 3, texImage,
 720                                         xoffset, yoffset, zoffset,
 721                                         width, height, depth, format, type,
 722                                         pixels, &ctx->Pack)) {
 723          /* Flush to guarantee coherency between the render cache and other
 724           * caches the PBO could potentially be bound to after this point.
 725           * See the related comment in intelReadPixels() for a more detailed
 726           * explanation.
 727           */
 728          brw_emit_mi_flush(brw);
 729          return;
 730       }
 731
 732       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
 733    }
 734
 735    ok = intel_gettexsubimage_tiled_memcpy(ctx, texImage, xoffset, yoffset,
 736                                           width, height,
 737                                           format, type, pixels, &ctx->Pack);
 738
 739    if(ok)
 740       return;
 741
 742    _mesa_meta_GetTexSubImage(ctx, xoffset, yoffset, zoffset,
 743                              width, height, depth,
 744                              format, type, pixels, texImage);
 745
 746    DBG("%s - DONE\n", __func__);
 747 }
 748
 749 static void
 750 flush_astc_denorms(struct gl_context *ctx, GLuint dims,
 751                    struct gl_texture_image *texImage,
 752                    GLint xoffset, GLint yoffset, GLint zoffset,
 753                    GLsizei width, GLsizei height, GLsizei depth)
 754 {
 755    struct compressed_pixelstore store;
 756    _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
 757                                        width, height, depth,
 758                                        &ctx->Unpack, &store);
 759
 760    for (int slice = 0; slice < store.CopySlices; slice++) {
 761
 762       /* Map dest texture buffer */
 763       GLubyte *dstMap;
 764       GLint dstRowStride;
 765       ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
 766                                   xoffset, yoffset, width, height,
 767                                   GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
 768                                   &dstMap, &dstRowStride);
 769       if (!dstMap)
 770          continue;
 771
 772       for (int i = 0; i < store.CopyRowsPerSlice; i++) {
 773
 774          /* An ASTC block is stored in little endian mode. The byte that
 775           * contains bits 0..7 is stored at the lower address in memory.
 776           */
 777          struct astc_void_extent {
 778             uint16_t header : 12;
 779             uint16_t dontcare[3];
 780             uint16_t R;
 781             uint16_t G;
 782             uint16_t B;
 783             uint16_t A;
 784          } *blocks = (struct astc_void_extent*) dstMap;
 785
 786          /* Iterate over every copied block in the row */
 787          for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
 788
 789             /* Check if the header matches that of an LDR void-extent block */
 790             if (blocks[j].header == 0xDFC) {
 791
 792                /* Flush UNORM16 values that would be denormalized */
 793                if (blocks[j].A < 4) blocks[j].A = 0;
 794                if (blocks[j].B < 4) blocks[j].B = 0;
 795                if (blocks[j].G < 4) blocks[j].G = 0;
 796                if (blocks[j].R < 4) blocks[j].R = 0;
 797             }
 798          }
 799
 800          dstMap += dstRowStride;
 801       }
 802
 803       ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
 804    }
 805 }
 806
 807
 808 static void
 809 intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
 810                         struct gl_texture_image *texImage,
 811                         GLint xoffset, GLint yoffset, GLint zoffset,
 812                         GLsizei width, GLsizei height, GLsizei depth,
 813                         GLenum format,
 814                         GLsizei imageSize, const GLvoid *data)
 815 {
 816    /* Upload the compressed data blocks */
 817    _mesa_store_compressed_texsubimage(ctx, dims, texImage,
 818                                       xoffset, yoffset, zoffset,
 819                                       width, height, depth,
 820                                       format, imageSize, data);
 821
 822    /* Fix up copied ASTC blocks if necessary */
 823    GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
 824                                                         texImage->TexFormat);
 825    bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
 826                         !_mesa_is_srgb_format(gl_format);
 827    struct brw_context *brw = (struct brw_context*) ctx;
 828    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 829    if (devinfo->gen == 9 && is_linear_astc)
 830       flush_astc_denorms(ctx, dims, texImage,
 831                          xoffset, yoffset, zoffset,
 832                          width, height, depth);
 833 }
 834
 835 void
 836 intelInitTextureImageFuncs(struct dd_function_table *functions)
 837 {
 838    functions->TexImage = intelTexImage;
 839    functions->TexSubImage = intelTexSubImage;
 840    functions->CompressedTexSubImage = intelCompressedTexSubImage;
 841    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
 842    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
 843    functions->GetTexSubImage = intel_get_tex_sub_image;
 844 }