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 #include "main/glthread.h"
  18
  19 #include "drivers/common/meta.h"
  20
  21 #include "intel_mipmap_tree.h"
  22 #include "intel_buffer_objects.h"
  23 #include "intel_batchbuffer.h"
  24 #include "intel_tex.h"
  25 #include "intel_fbo.h"
  26 #include "intel_image.h"
  27 #include "brw_context.h"
  28 #include "brw_blorp.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 static bool
 131 intel_texsubimage_blorp(struct brw_context *brw, GLuint dims,
 132                         struct gl_texture_image *tex_image,
 133                         unsigned x, unsigned y, unsigned z,
 134                         unsigned width, unsigned height, unsigned depth,
 135                         GLenum format, GLenum type, const void *pixels,
 136                         const struct gl_pixelstore_attrib *packing)
 137 {
 138    struct intel_texture_image *intel_image = intel_texture_image(tex_image);
 139    const unsigned mt_level = tex_image->Level + tex_image->TexObject->MinLevel;
 140    const unsigned mt_z = tex_image->TexObject->MinLayer + tex_image->Face + z;
 141
 142    /* The blorp path can't understand crazy format hackery */
 143    if (_mesa_base_tex_format(&brw->ctx, tex_image->InternalFormat) !=
 144        _mesa_get_format_base_format(tex_image->TexFormat))
 145       return false;
 146
 147    return brw_blorp_upload_miptree(brw, intel_image->mt, tex_image->TexFormat,
 148                                    mt_level, x, y, mt_z, width, height, depth,
 149                                    tex_image->TexObject->Target, format, type,
 150                                    pixels, packing);
 151 }
 152
 153 /**
 154  * \brief A fast path for glTexImage and glTexSubImage.
 155  *
 156  * This fast path is taken when the texture format is BGRA, RGBA,
 157  * A or L and when the texture memory is X- or Y-tiled.  It uploads
 158  * the texture data by mapping the texture memory without a GTT fence, thus
 159  * acquiring a tiled view of the memory, and then copying sucessive
 160  * spans within each tile.
 161  *
 162  * This is a performance win over the conventional texture upload path because
 163  * it avoids the performance penalty of writing through the write-combine
 164  * buffer. In the conventional texture upload path,
 165  * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
 166  * fence, thus acquiring a linear view of the memory, then each row in the
 167  * image is memcpy'd. In this fast path, we replace each row's copy with
 168  * a sequence of copies over each linear span in tile.
 169  *
 170  * One use case is Google Chrome's paint rectangles.  Chrome (as
 171  * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
 172  * Each page's content is initially uploaded with glTexImage2D and damaged
 173  * regions are updated with glTexSubImage2D. On some workloads, the
 174  * performance gain of this fastpath on Sandybridge is over 5x.
 175  */
 176 static bool
 177 intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
 178                                GLuint dims,
 179                                struct gl_texture_image *texImage,
 180                                GLint xoffset, GLint yoffset, GLint zoffset,
 181                                GLsizei width, GLsizei height, GLsizei depth,
 182                                GLenum format, GLenum type,
 183                                const GLvoid *pixels,
 184                                const struct gl_pixelstore_attrib *packing)
 185 {
 186    struct brw_context *brw = brw_context(ctx);
 187    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 188    struct intel_texture_image *image = intel_texture_image(texImage);
 189    int src_pitch;
 190
 191    /* The miptree's buffer. */
 192    struct brw_bo *bo;
 193
 194    uint32_t cpp;
 195    isl_memcpy_type copy_type;
 196
 197    /* This fastpath is restricted to specific texture types:
 198     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 199     * more types.
 200     *
 201     * FINISHME: The restrictions below on packing alignment and packing row
 202     * length are likely unneeded now because we calculate the source stride
 203     * with _mesa_image_row_stride. However, before removing the restrictions
 204     * we need tests.
 205     */
 206    if (!devinfo->has_llc ||
 207        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 208        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 209          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 210        pixels == NULL ||
 211        _mesa_is_bufferobj(packing->BufferObj) ||
 212        packing->Alignment > 4 ||
 213        packing->SkipPixels > 0 ||
 214        packing->SkipRows > 0 ||
 215        (packing->RowLength != 0 && packing->RowLength != width) ||
 216        packing->SwapBytes ||
 217        packing->LsbFirst ||
 218        packing->Invert)
 219       return false;
 220
 221    /* Only a simple blit, no scale, bias or other mapping. */
 222    if (ctx->_ImageTransferState)
 223       return false;
 224
 225    copy_type = intel_miptree_get_memcpy_type(texImage->TexFormat, format, type,
 226                                              &cpp);
 227    if (copy_type == ISL_MEMCPY_INVALID)
 228       return false;
 229
 230    /* If this is a nontrivial texture view, let another path handle it instead. */
 231    if (texImage->TexObject->MinLayer)
 232       return false;
 233
 234    if (!image->mt ||
 235        (image->mt->surf.tiling != ISL_TILING_X &&
 236         image->mt->surf.tiling != ISL_TILING_Y0)) {
 237       /* The algorithm is written only for X- or Y-tiled memory. */
 238       return false;
 239    }
 240
 241    /* linear_to_tiled() assumes that if the object is swizzled, it is using
 242     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 243     * true on gen5 and above.
 244     *
 245     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 246     * parts of the memory aren't swizzled at all. Userspace just can't handle
 247     * that.
 248     */
 249    if (devinfo->gen < 5 && brw->has_swizzling)
 250       return false;
 251
 252    int level = texImage->Level + texImage->TexObject->MinLevel;
 253
 254    /* Since we are going to write raw data to the miptree, we need to resolve
 255     * any pending fast color clears before we start.
 256     */
 257    assert(image->mt->surf.logical_level0_px.depth == 1);
 258    assert(image->mt->surf.logical_level0_px.array_len == 1);
 259
 260    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 261
 262    bo = image->mt->bo;
 263
 264    if (brw_batch_references(&brw->batch, bo)) {
 265       perf_debug("Flushing before mapping a referenced bo.\n");
 266       intel_batchbuffer_flush(brw);
 267    }
 268
 269    void *map = brw_bo_map(brw, bo, MAP_WRITE | MAP_RAW);
 270    if (map == NULL) {
 271       DBG("%s: failed to map bo\n", __func__);
 272       return false;
 273    }
 274
 275    src_pitch = _mesa_image_row_stride(packing, width, format, type);
 276
 277    /* We postponed printing this message until having committed to executing
 278     * the function.
 279     */
 280    DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 281        "mesa_format=0x%x tiling=%d "
 282        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d) ",
 283        __func__, texImage->Level, xoffset, yoffset, width, height,
 284        format, type, texImage->TexFormat, image->mt->surf.tiling,
 285        packing->Alignment, packing->RowLength, packing->SkipPixels,
 286        packing->SkipRows);
 287
 288    /* Adjust x and y offset based on miplevel */
 289    unsigned level_x, level_y;
 290    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 291    xoffset += level_x;
 292    yoffset += level_y;
 293
 294    isl_memcpy_linear_to_tiled(
 295       xoffset * cpp, (xoffset + width) * cpp,
 296       yoffset, yoffset + height,
 297       map,
 298       pixels,
 299       image->mt->surf.row_pitch_B, src_pitch,
 300       brw->has_swizzling,
 301       image->mt->surf.tiling,
 302       copy_type
 303    );
 304
 305    brw_bo_unmap(bo);
 306    return true;
 307 }
 308
 309
 310 static void
 311 intel_upload_tex(struct gl_context * ctx,
 312                  GLuint dims,
 313                  struct gl_texture_image *texImage,
 314                  GLint xoffset, GLint yoffset, GLint zoffset,
 315                  GLsizei width, GLsizei height, GLsizei depth,
 316                  GLenum format, GLenum type,
 317                  const GLvoid * pixels,
 318                  const struct gl_pixelstore_attrib *packing)
 319 {
 320    struct brw_context *brw = brw_context(ctx);
 321    struct intel_mipmap_tree *mt = intel_texture_image(texImage)->mt;
 322    bool ok;
 323
 324    /* Check that there is actually data to store. */
 325    if (pixels == NULL && !_mesa_is_bufferobj(packing->BufferObj))
 326       return;
 327
 328    bool tex_busy = mt && brw_bo_busy(mt->bo);
 329
 330    if (_mesa_is_bufferobj(packing->BufferObj) || tex_busy ||
 331        mt->aux_usage == ISL_AUX_USAGE_CCS_E) {
 332       ok = intel_texsubimage_blorp(brw, dims, texImage,
 333                                    xoffset, yoffset, zoffset,
 334                                    width, height, depth, format, type,
 335                                    pixels, packing);
 336       if (ok)
 337          return;
 338    }
 339
 340    ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
 341                                        xoffset, yoffset, zoffset,
 342                                        width, height, depth,
 343                                        format, type, pixels, packing);
 344    if (ok)
 345      return;
 346
 347    _mesa_store_texsubimage(ctx, dims, texImage,
 348                            xoffset, yoffset, zoffset,
 349                            width, height, depth,
 350                            format, type, pixels, packing);
 351 }
 352
 353
 354 static void
 355 intelTexImage(struct gl_context * ctx,
 356               GLuint dims,
 357               struct gl_texture_image *texImage,
 358               GLenum format, GLenum type, const void *pixels,
 359               const struct gl_pixelstore_attrib *unpack)
 360 {
 361    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 362        __func__, _mesa_get_format_name(texImage->TexFormat),
 363        _mesa_enum_to_string(texImage->TexObject->Target),
 364        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 365        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 366
 367    /* Allocate storage for texture data. */
 368    if (!ctx->Driver.AllocTextureImageBuffer(ctx, texImage)) {
 369       _mesa_error(ctx, GL_OUT_OF_MEMORY, "glTexImage%uD", dims);
 370       return;
 371    }
 372
 373    assert(intel_texture_image(texImage)->mt);
 374
 375    intel_upload_tex(ctx, dims, texImage, 0, 0, 0,
 376                     texImage->Width, texImage->Height, texImage->Depth,
 377                     format, type, pixels, unpack);
 378 }
 379
 380
 381 static void
 382 intelTexSubImage(struct gl_context * ctx,
 383                  GLuint dims,
 384                  struct gl_texture_image *texImage,
 385                  GLint xoffset, GLint yoffset, GLint zoffset,
 386                  GLsizei width, GLsizei height, GLsizei depth,
 387                  GLenum format, GLenum type,
 388                  const GLvoid * pixels,
 389                  const struct gl_pixelstore_attrib *packing)
 390 {
 391    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
 392        __func__, _mesa_get_format_name(texImage->TexFormat),
 393        _mesa_enum_to_string(texImage->TexObject->Target),
 394        _mesa_enum_to_string(format), _mesa_enum_to_string(type),
 395        texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 396
 397    intel_upload_tex(ctx, dims, texImage, xoffset, yoffset, zoffset,
 398                     width, height, depth, format, type, pixels, packing);
 399 }
 400
 401
 402 static void
 403 intel_set_texture_image_mt(struct brw_context *brw,
 404                            struct gl_texture_image *image,
 405                            GLenum internal_format,
 406                            mesa_format 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, 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_B / mt->cpp;
 423    assert(mt->surf.row_pitch_B % 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    mesa_format texFormat = MESA_FORMAT_NONE;
 444    GLenum internal_format = 0;
 445
 446    _mesa_glthread_finish(ctx);
 447
 448    texObj = _mesa_get_current_tex_object(ctx, target);
 449
 450    if (!texObj)
 451       return;
 452
 453    if (dPriv->lastStamp != dPriv->dri2.stamp ||
 454        !pDRICtx->driScreenPriv->dri2.useInvalidate)
 455       intel_update_renderbuffers(pDRICtx, dPriv);
 456
 457    rb = intel_get_renderbuffer(fb, BUFFER_FRONT_LEFT);
 458    /* If the miptree isn't set, then intel_update_renderbuffers was unable
 459     * to get the BO for the drawable from the window system.
 460     */
 461    if (!rb || !rb->mt)
 462       return;
 463
 464    /* Neither the EGL and GLX texture_from_pixmap specs say anything about
 465     * sRGB.  They are both from a time where sRGB was considered an extra
 466     * encoding step you did as part of rendering/blending and not a format.
 467     * Even though we have concept of sRGB visuals, X has classically assumed
 468     * that your data is just bits and sRGB rendering is entirely a client-side
 469     * rendering construct.  The assumption is that the result of BindTexImage
 470     * is a texture with a linear format even if it was rendered with sRGB
 471     * encoding enabled.
 472     */
 473    texFormat = _mesa_get_srgb_format_linear(intel_rb_format(rb));
 474
 475    if (rb->mt->cpp == 4) {
 476       /* The extra texture_format parameter indicates whether the alpha
 477        * channel should be respected or ignored.  If we set internal_format to
 478        * GL_RGB, the texture handling code is smart enough to swap the format
 479        * or apply a swizzle if the underlying format is RGBA so we don't need
 480        * to stomp it to RGBX or anything like that.
 481        */
 482       if (texture_format == __DRI_TEXTURE_FORMAT_RGB)
 483          internal_format = GL_RGB;
 484       else
 485          internal_format = GL_RGBA;
 486    } else if (rb->mt->cpp == 2) {
 487       internal_format = GL_RGB;
 488    }
 489
 490    intel_miptree_finish_external(brw, rb->mt);
 491
 492    _mesa_lock_texture(&brw->ctx, texObj);
 493    texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
 494    intel_set_texture_image_mt(brw, texImage, internal_format,
 495                               texFormat, rb->mt);
 496    _mesa_unlock_texture(&brw->ctx, texObj);
 497 }
 498
 499 void
 500 intelReleaseTexBuffer(__DRIcontext *pDRICtx, GLint target,
 501                       __DRIdrawable *dPriv)
 502 {
 503    struct brw_context *brw = pDRICtx->driverPrivate;
 504    struct gl_context *ctx = &brw->ctx;
 505    struct gl_texture_object *tex_obj;
 506    struct intel_texture_object *intel_tex;
 507
 508    tex_obj = _mesa_get_current_tex_object(ctx, target);
 509    if (!tex_obj)
 510       return;
 511
 512    _mesa_lock_texture(&brw->ctx, tex_obj);
 513
 514    intel_tex = intel_texture_object(tex_obj);
 515    if (!intel_tex->mt) {
 516       _mesa_unlock_texture(&brw->ctx, tex_obj);
 517       return;
 518    }
 519
 520    /* The intel_miptree_prepare_external below as well as the finish_external
 521     * above in intelSetTexBuffer2 *should* do nothing.  The BindTexImage call
 522     * from both GLX and EGL has TexImage2D and not TexSubImage2D semantics so
 523     * the texture is not immutable.  This means that the user cannot create a
 524     * texture view of the image with a different format.  Since the only three
 525     * formats available when using BindTexImage are all UNORM, we can never
 526     * end up with an sRGB format being used for texturing and so we shouldn't
 527     * get any format-related resolves when texturing from it.
 528     *
 529     * While very unlikely, it is possible that the client could use the bound
 530     * texture with GL_ARB_image_load_store.  In that case, we'll do a resolve
 531     * but that's not actually a problem as it just means that we lose
 532     * compression on this texture until the next time it's used as a render
 533     * target.
 534     *
 535     * The only other way we could end up with an unexpected aux usage would be
 536     * if we rendered to the image from the same context as we have it bound as
 537     * a texture between BindTexImage and ReleaseTexImage.  However, the spec
 538     * clearly calls this case out and says you shouldn't do that.  It doesn't
 539     * explicitly prevent binding the texture to a framebuffer but it says the
 540     * results of trying to render to it while bound are undefined.
 541     *
 542     * Just to keep everything safe and sane, we do a prepare_external but it
 543     * should be a no-op in almost all cases.  On the off chance that someone
 544     * ever triggers this, we should at least warn them.
 545     */
 546    if (intel_tex->mt->aux_buf &&
 547        intel_miptree_get_aux_state(intel_tex->mt, 0, 0) !=
 548        isl_drm_modifier_get_default_aux_state(intel_tex->mt->drm_modifier)) {
 549       _mesa_warning(ctx, "Aux state changed between BindTexImage and "
 550                          "ReleaseTexImage.  Most likely someone tried to draw "
 551                          "to the pixmap bound in BindTexImage or used it with "
 552                          "image_load_store.");
 553    }
 554
 555    intel_miptree_prepare_external(brw, intel_tex->mt);
 556
 557    _mesa_unlock_texture(&brw->ctx, tex_obj);
 558 }
 559
 560 static GLboolean
 561 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
 562                                   struct gl_renderbuffer *rb,
 563                                   struct gl_texture_image *image)
 564 {
 565    struct intel_renderbuffer *irb = intel_renderbuffer(rb);
 566    struct intel_texture_image *intel_image = intel_texture_image(image);
 567    struct gl_texture_object *texobj = image->TexObject;
 568    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
 569
 570    /* We can only handle RB allocated with AllocRenderbufferStorage, or
 571     * window-system renderbuffers.
 572     */
 573    assert(!rb->TexImage);
 574
 575    if (!irb->mt)
 576       return false;
 577
 578    _mesa_lock_texture(ctx, texobj);
 579    _mesa_init_teximage_fields(ctx, image,
 580                               rb->Width, rb->Height, 1,
 581                               0, rb->InternalFormat, rb->Format);
 582    image->NumSamples = rb->NumSamples;
 583
 584    intel_miptree_reference(&intel_image->mt, irb->mt);
 585
 586    /* Immediately validate the image to the object. */
 587    intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
 588
 589    intel_texobj->needs_validate = true;
 590    _mesa_unlock_texture(ctx, texobj);
 591
 592    return true;
 593 }
 594
 595 void
 596 intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 597 {
 598    /* The old interface didn't have the format argument, so copy our
 599     * implementation's behavior at the time.
 600     */
 601    intelSetTexBuffer2(pDRICtx, target, __DRI_TEXTURE_FORMAT_RGBA, dPriv);
 602 }
 603
 604 static void
 605 intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
 606                               struct gl_texture_object *texObj,
 607                               struct gl_texture_image *texImage,
 608                               GLeglImageOES image_handle)
 609 {
 610    struct brw_context *brw = brw_context(ctx);
 611    struct intel_mipmap_tree *mt;
 612    __DRIscreen *dri_screen = brw->screen->driScrnPriv;
 613    __DRIimage *image;
 614
 615    image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
 616                                                   dri_screen->loaderPrivate);
 617    if (image == NULL)
 618       return;
 619
 620    /* Disallow depth/stencil textures: we don't have a way to pass the
 621     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
 622     */
 623    if (image->has_depthstencil) {
 624       _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
 625       return;
 626    }
 627
 628    mt = intel_miptree_create_for_dri_image(brw, image, target, image->format,
 629                                            false);
 630    if (mt == NULL)
 631       return;
 632
 633    struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
 634    intel_texobj->planar_format = image->planar_format;
 635
 636    const GLenum internal_format =
 637       image->internal_format != 0 ?
 638       image->internal_format : _mesa_get_format_base_format(mt->format);
 639    intel_set_texture_image_mt(brw, texImage, internal_format, mt->format, mt);
 640    intel_miptree_release(&mt);
 641 }
 642
 643 static bool
 644 intel_gettexsubimage_blorp(struct brw_context *brw,
 645                            struct gl_texture_image *tex_image,
 646                            unsigned x, unsigned y, unsigned z,
 647                            unsigned width, unsigned height, unsigned depth,
 648                            GLenum format, GLenum type, const void *pixels,
 649                            const struct gl_pixelstore_attrib *packing)
 650 {
 651    struct intel_texture_image *intel_image = intel_texture_image(tex_image);
 652    const unsigned mt_level = tex_image->Level + tex_image->TexObject->MinLevel;
 653    const unsigned mt_z = tex_image->TexObject->MinLayer + tex_image->Face + z;
 654
 655    /* The blorp path can't understand crazy format hackery */
 656    if (_mesa_base_tex_format(&brw->ctx, tex_image->InternalFormat) !=
 657        _mesa_get_format_base_format(tex_image->TexFormat))
 658       return false;
 659
 660    return brw_blorp_download_miptree(brw, intel_image->mt,
 661                                      tex_image->TexFormat, SWIZZLE_XYZW,
 662                                      mt_level, x, y, mt_z,
 663                                      width, height, depth,
 664                                      tex_image->TexObject->Target,
 665                                      format, type, false, pixels, packing);
 666 }
 667
 668 /**
 669  * \brief A fast path for glGetTexImage.
 670  *
 671  * \see intel_readpixels_tiled_memcpy()
 672  */
 673 static bool
 674 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
 675                                   struct gl_texture_image *texImage,
 676                                   GLint xoffset, GLint yoffset,
 677                                   GLsizei width, GLsizei height,
 678                                   GLenum format, GLenum type,
 679                                   GLvoid *pixels,
 680                                   const struct gl_pixelstore_attrib *packing)
 681 {
 682    struct brw_context *brw = brw_context(ctx);
 683    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 684    struct intel_texture_image *image = intel_texture_image(texImage);
 685    int dst_pitch;
 686
 687    /* The miptree's buffer. */
 688    struct brw_bo *bo;
 689
 690    uint32_t cpp;
 691    isl_memcpy_type copy_type;
 692
 693    /* This fastpath is restricted to specific texture types:
 694     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
 695     * more types.
 696     *
 697     * FINISHME: The restrictions below on packing alignment and packing row
 698     * length are likely unneeded now because we calculate the destination stride
 699     * with _mesa_image_row_stride. However, before removing the restrictions
 700     * we need tests.
 701     */
 702    if (!devinfo->has_llc ||
 703        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
 704        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
 705          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
 706        pixels == NULL ||
 707        _mesa_is_bufferobj(packing->BufferObj) ||
 708        packing->Alignment > 4 ||
 709        packing->SkipPixels > 0 ||
 710        packing->SkipRows > 0 ||
 711        (packing->RowLength != 0 && packing->RowLength != width) ||
 712        packing->SwapBytes ||
 713        packing->LsbFirst ||
 714        packing->Invert)
 715       return false;
 716
 717    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
 718     * function doesn't set the last channel to 1. Note this checks BaseFormat
 719     * rather than TexFormat in case the RGBX format is being simulated with an
 720     * RGBA format.
 721     */
 722    if (texImage->_BaseFormat == GL_RGB)
 723       return false;
 724
 725    copy_type = intel_miptree_get_memcpy_type(texImage->TexFormat, format, type,
 726                                              &cpp);
 727    if (copy_type == ISL_MEMCPY_INVALID)
 728       return false;
 729
 730    /* If this is a nontrivial texture view, let another path handle it instead. */
 731    if (texImage->TexObject->MinLayer)
 732       return false;
 733
 734    if (!image->mt ||
 735        (image->mt->surf.tiling != ISL_TILING_X &&
 736         image->mt->surf.tiling != ISL_TILING_Y0)) {
 737       /* The algorithm is written only for X- or Y-tiled memory. */
 738       return false;
 739    }
 740
 741    /* tiled_to_linear() assumes that if the object is swizzled, it is using
 742     * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
 743     * true on gen5 and above.
 744     *
 745     * The killer on top is that some gen4 have an L-shaped swizzle mode, where
 746     * parts of the memory aren't swizzled at all. Userspace just can't handle
 747     * that.
 748     */
 749    if (devinfo->gen < 5 && brw->has_swizzling)
 750       return false;
 751
 752    int level = texImage->Level + texImage->TexObject->MinLevel;
 753
 754    /* Since we are going to write raw data to the miptree, we need to resolve
 755     * any pending fast color clears before we start.
 756     */
 757    assert(image->mt->surf.logical_level0_px.depth == 1);
 758    assert(image->mt->surf.logical_level0_px.array_len == 1);
 759
 760    intel_miptree_access_raw(brw, image->mt, level, 0, true);
 761
 762    bo = image->mt->bo;
 763
 764    if (brw_batch_references(&brw->batch, bo)) {
 765       perf_debug("Flushing before mapping a referenced bo.\n");
 766       intel_batchbuffer_flush(brw);
 767    }
 768
 769    void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
 770    if (map == NULL) {
 771       DBG("%s: failed to map bo\n", __func__);
 772       return false;
 773    }
 774
 775    dst_pitch = _mesa_image_row_stride(packing, width, format, type);
 776
 777    DBG("%s: level=%d x,y=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
 778        "mesa_format=0x%x tiling=%d "
 779        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
 780        __func__, texImage->Level, xoffset, yoffset, width, height,
 781        format, type, texImage->TexFormat, image->mt->surf.tiling,
 782        packing->Alignment, packing->RowLength, packing->SkipPixels,
 783        packing->SkipRows);
 784
 785    /* Adjust x and y offset based on miplevel */
 786    unsigned level_x, level_y;
 787    intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
 788    xoffset += level_x;
 789    yoffset += level_y;
 790
 791    isl_memcpy_tiled_to_linear(
 792       xoffset * cpp, (xoffset + width) * cpp,
 793       yoffset, yoffset + height,
 794       pixels,
 795       map,
 796       dst_pitch, image->mt->surf.row_pitch_B,
 797       brw->has_swizzling,
 798       image->mt->surf.tiling,
 799       copy_type
 800    );
 801
 802    brw_bo_unmap(bo);
 803    return true;
 804 }
 805
 806 static void
 807 intel_get_tex_sub_image(struct gl_context *ctx,
 808                         GLint xoffset, GLint yoffset, GLint zoffset,
 809                         GLsizei width, GLsizei height, GLint depth,
 810                         GLenum format, GLenum type, GLvoid *pixels,
 811                         struct gl_texture_image *texImage)
 812 {
 813    struct brw_context *brw = brw_context(ctx);
 814    bool ok;
 815
 816    DBG("%s\n", __func__);
 817
 818    if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
 819       if (intel_gettexsubimage_blorp(brw, texImage,
 820                                      xoffset, yoffset, zoffset,
 821                                      width, height, depth, format, type,
 822                                      pixels, &ctx->Pack))
 823          return;
 824
 825       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
 826    }
 827
 828    ok = intel_gettexsubimage_tiled_memcpy(ctx, texImage, xoffset, yoffset,
 829                                           width, height,
 830                                           format, type, pixels, &ctx->Pack);
 831
 832    if(ok)
 833       return;
 834
 835    _mesa_meta_GetTexSubImage(ctx, xoffset, yoffset, zoffset,
 836                              width, height, depth,
 837                              format, type, pixels, texImage);
 838
 839    DBG("%s - DONE\n", __func__);
 840 }
 841
 842 static void
 843 flush_astc_denorms(struct gl_context *ctx, GLuint dims,
 844                    struct gl_texture_image *texImage,
 845                    GLint xoffset, GLint yoffset, GLint zoffset,
 846                    GLsizei width, GLsizei height, GLsizei depth)
 847 {
 848    struct compressed_pixelstore store;
 849    _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
 850                                        width, height, depth,
 851                                        &ctx->Unpack, &store);
 852
 853    for (int slice = 0; slice < store.CopySlices; slice++) {
 854
 855       /* Map dest texture buffer */
 856       GLubyte *dstMap;
 857       GLint dstRowStride;
 858       ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
 859                                   xoffset, yoffset, width, height,
 860                                   GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
 861                                   &dstMap, &dstRowStride);
 862       if (!dstMap)
 863          continue;
 864
 865       for (int i = 0; i < store.CopyRowsPerSlice; i++) {
 866
 867          /* An ASTC block is stored in little endian mode. The byte that
 868           * contains bits 0..7 is stored at the lower address in memory.
 869           */
 870          struct astc_void_extent {
 871             uint16_t header : 12;
 872             uint16_t dontcare[3];
 873             uint16_t R;
 874             uint16_t G;
 875             uint16_t B;
 876             uint16_t A;
 877          } *blocks = (struct astc_void_extent*) dstMap;
 878
 879          /* Iterate over every copied block in the row */
 880          for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
 881
 882             /* Check if the header matches that of an LDR void-extent block */
 883             if (blocks[j].header == 0xDFC) {
 884
 885                /* Flush UNORM16 values that would be denormalized */
 886                if (blocks[j].A < 4) blocks[j].A = 0;
 887                if (blocks[j].B < 4) blocks[j].B = 0;
 888                if (blocks[j].G < 4) blocks[j].G = 0;
 889                if (blocks[j].R < 4) blocks[j].R = 0;
 890             }
 891          }
 892
 893          dstMap += dstRowStride;
 894       }
 895
 896       ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
 897    }
 898 }
 899
 900
 901 static void
 902 intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
 903                         struct gl_texture_image *texImage,
 904                         GLint xoffset, GLint yoffset, GLint zoffset,
 905                         GLsizei width, GLsizei height, GLsizei depth,
 906                         GLenum format,
 907                         GLsizei imageSize, const GLvoid *data)
 908 {
 909    /* Upload the compressed data blocks */
 910    _mesa_store_compressed_texsubimage(ctx, dims, texImage,
 911                                       xoffset, yoffset, zoffset,
 912                                       width, height, depth,
 913                                       format, imageSize, data);
 914
 915    /* Fix up copied ASTC blocks if necessary */
 916    GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
 917                                                         texImage->TexFormat);
 918    bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
 919                         !_mesa_is_srgb_format(gl_format);
 920    struct brw_context *brw = (struct brw_context*) ctx;
 921    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 922    if (devinfo->gen == 9 && !gen_device_info_is_9lp(devinfo) && is_linear_astc)
 923       flush_astc_denorms(ctx, dims, texImage,
 924                          xoffset, yoffset, zoffset,
 925                          width, height, depth);
 926 }
 927
 928 void
 929 intelInitTextureImageFuncs(struct dd_function_table *functions)
 930 {
 931    functions->TexImage = intelTexImage;
 932    functions->TexSubImage = intelTexSubImage;
 933    functions->CompressedTexSubImage = intelCompressedTexSubImage;
 934    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
 935    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
 936    functions->GetTexSubImage = intel_get_tex_sub_image;
 937 }