i965/i915: Add colorspace support to YUV sampling
[mesa.git] / src / mesa / drivers / dri / i965 / intel_tex_image.c
index 3611280012ef31132f07dc26db3d7c4e9e35357f..c0a4a9c2ca9163364830fe1201266f37c393a35c 100644 (file)
@@ -1,11 +1,11 @@
 
-#include "main/glheader.h"
 #include "main/macros.h"
 #include "main/mtypes.h"
 #include "main/enums.h"
 #include "main/bufferobj.h"
 #include "main/context.h"
 #include "main/formats.h"
+#include "main/glformats.h"
 #include "main/image.h"
 #include "main/pbo.h"
 #include "main/renderbuffer.h"
@@ -14,6 +14,7 @@
 #include "main/texobj.h"
 #include "main/teximage.h"
 #include "main/texstore.h"
+#include "main/glthread.h"
 
 #include "drivers/common/meta.h"
 
 #include "intel_buffer_objects.h"
 #include "intel_batchbuffer.h"
 #include "intel_tex.h"
-#include "intel_blit.h"
 #include "intel_fbo.h"
 #include "intel_image.h"
-#include "intel_tiled_memcpy.h"
 #include "brw_context.h"
+#include "brw_blorp.h"
 
 #define FILE_DEBUG_FLAG DEBUG_TEXTURE
 
+/* Make sure one doesn't end up shrinking base level zero unnecessarily.
+ * Determining the base level dimension by shifting higher level dimension
+ * ends up in off-by-one value in case base level has NPOT size (for example,
+ * 293 != 146 << 1).
+ * Choose the original base level dimension when shifted dimensions agree.
+ * Otherwise assume real resize is intended and use the new shifted value.
+ */
+static unsigned
+get_base_dim(unsigned old_base_dim, unsigned new_level_dim, unsigned level)
+{
+   const unsigned old_level_dim = old_base_dim >> level;
+   const unsigned new_base_dim = new_level_dim << level;
+
+   return old_level_dim == new_level_dim ? old_base_dim : new_base_dim;
+}
+
 /* Work back from the specified level of the image to the baselevel and create a
  * miptree of that size.
  */
@@ -36,24 +52,52 @@ struct intel_mipmap_tree *
 intel_miptree_create_for_teximage(struct brw_context *brw,
                                  struct intel_texture_object *intelObj,
                                  struct intel_texture_image *intelImage,
-                                  uint32_t layout_flags)
+                                  enum intel_miptree_create_flags flags)
 {
    GLuint lastLevel;
    int width, height, depth;
-   GLuint i;
-
-   intel_miptree_get_dimensions_for_image(&intelImage->base.Base,
-                                          &width, &height, &depth);
+   unsigned old_width = 0, old_height = 0, old_depth = 0;
+   const struct intel_mipmap_tree *old_mt = intelObj->mt;
+   const unsigned level = intelImage->base.Base.Level;
+
+   intel_get_image_dims(&intelImage->base.Base, &width, &height, &depth);
+
+   if (old_mt) {
+      old_width = old_mt->surf.logical_level0_px.width;
+      old_height = old_mt->surf.logical_level0_px.height;
+      old_depth = old_mt->surf.dim == ISL_SURF_DIM_3D ?
+                     old_mt->surf.logical_level0_px.depth :
+                     old_mt->surf.logical_level0_px.array_len;
+   }
 
    DBG("%s\n", __func__);
 
    /* Figure out image dimensions at start level. */
-   for (i = intelImage->base.Base.Level; i > 0; i--) {
-      width <<= 1;
-      if (height != 1)
-         height <<= 1;
-      if (depth != 1)
-         depth <<= 1;
+   switch(intelObj->base.Target) {
+   case GL_TEXTURE_2D_MULTISAMPLE:
+   case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+   case GL_TEXTURE_RECTANGLE:
+   case GL_TEXTURE_EXTERNAL_OES:
+      assert(level == 0);
+      break;
+   case GL_TEXTURE_3D:
+      depth = old_mt ? get_base_dim(old_depth, depth, level) :
+                       depth << level;
+      /* Fall through */
+   case GL_TEXTURE_2D:
+   case GL_TEXTURE_2D_ARRAY:
+   case GL_TEXTURE_CUBE_MAP:
+   case GL_TEXTURE_CUBE_MAP_ARRAY:
+      height = old_mt ? get_base_dim(old_height, height, level) :
+                        height << level;
+      /* Fall through */
+   case GL_TEXTURE_1D:
+   case GL_TEXTURE_1D_ARRAY:
+      width = old_mt ? get_base_dim(old_width, width, level) :
+                       width << level;
+      break;
+   default:
+      unreachable("Unexpected target");
    }
 
    /* Guess a reasonable value for lastLevel.  This is probably going
@@ -79,10 +123,235 @@ intel_miptree_create_for_teximage(struct brw_context *brw,
                               width,
                               height,
                               depth,
-                               intelImage->base.Base.NumSamples,
-                               layout_flags | MIPTREE_LAYOUT_ALLOC_ANY_TILED);
+                               MAX2(intelImage->base.Base.NumSamples, 1),
+                               flags);
 }
 
+static bool
+intel_texsubimage_blorp(struct brw_context *brw, GLuint dims,
+                        struct gl_texture_image *tex_image,
+                        unsigned x, unsigned y, unsigned z,
+                        unsigned width, unsigned height, unsigned depth,
+                        GLenum format, GLenum type, const void *pixels,
+                        const struct gl_pixelstore_attrib *packing)
+{
+   struct intel_texture_image *intel_image = intel_texture_image(tex_image);
+   const unsigned mt_level = tex_image->Level + tex_image->TexObject->MinLevel;
+   const unsigned mt_z = tex_image->TexObject->MinLayer + tex_image->Face + z;
+
+   /* The blorp path can't understand crazy format hackery */
+   if (_mesa_base_tex_format(&brw->ctx, tex_image->InternalFormat) !=
+       _mesa_get_format_base_format(tex_image->TexFormat))
+      return false;
+
+   return brw_blorp_upload_miptree(brw, intel_image->mt, tex_image->TexFormat,
+                                   mt_level, x, y, mt_z, width, height, depth,
+                                   tex_image->TexObject->Target, format, type,
+                                   pixels, packing);
+}
+
+/**
+ * \brief A fast path for glTexImage and glTexSubImage.
+ *
+ * This fast path is taken when the texture format is BGRA, RGBA,
+ * A or L and when the texture memory is X- or Y-tiled.  It uploads
+ * the texture data by mapping the texture memory without a GTT fence, thus
+ * acquiring a tiled view of the memory, and then copying sucessive
+ * spans within each tile.
+ *
+ * This is a performance win over the conventional texture upload path because
+ * it avoids the performance penalty of writing through the write-combine
+ * buffer. In the conventional texture upload path,
+ * texstore.c:store_texsubimage(), the texture memory is mapped through a GTT
+ * fence, thus acquiring a linear view of the memory, then each row in the
+ * image is memcpy'd. In this fast path, we replace each row's copy with
+ * a sequence of copies over each linear span in tile.
+ *
+ * One use case is Google Chrome's paint rectangles.  Chrome (as
+ * of version 21) renders each page as a tiling of 256x256 GL_BGRA textures.
+ * Each page's content is initially uploaded with glTexImage2D and damaged
+ * regions are updated with glTexSubImage2D. On some workloads, the
+ * performance gain of this fastpath on Sandybridge is over 5x.
+ */
+static bool
+intel_texsubimage_tiled_memcpy(struct gl_context * ctx,
+                               GLuint dims,
+                               struct gl_texture_image *texImage,
+                               GLint xoffset, GLint yoffset, GLint zoffset,
+                               GLsizei width, GLsizei height, GLsizei depth,
+                               GLenum format, GLenum type,
+                               const GLvoid *pixels,
+                               const struct gl_pixelstore_attrib *packing)
+{
+   struct brw_context *brw = brw_context(ctx);
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
+   struct intel_texture_image *image = intel_texture_image(texImage);
+   int src_pitch;
+
+   /* The miptree's buffer. */
+   struct brw_bo *bo;
+
+   uint32_t cpp;
+   isl_memcpy_type copy_type;
+
+   /* This fastpath is restricted to specific texture types:
+    * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
+    * more types.
+    *
+    * FINISHME: The restrictions below on packing alignment and packing row
+    * length are likely unneeded now because we calculate the source stride
+    * with _mesa_image_row_stride. However, before removing the restrictions
+    * we need tests.
+    */
+   if (!devinfo->has_llc ||
+       !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
+       !(texImage->TexObject->Target == GL_TEXTURE_2D ||
+         texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
+       pixels == NULL ||
+       packing->BufferObj ||
+       packing->Alignment > 4 ||
+       packing->SkipPixels > 0 ||
+       packing->SkipRows > 0 ||
+       (packing->RowLength != 0 && packing->RowLength != width) ||
+       packing->SwapBytes ||
+       packing->LsbFirst ||
+       packing->Invert)
+      return false;
+
+   /* Only a simple blit, no scale, bias or other mapping. */
+   if (ctx->_ImageTransferState)
+      return false;
+
+   copy_type = intel_miptree_get_memcpy_type(texImage->TexFormat, format, type,
+                                             &cpp);
+   if (copy_type == ISL_MEMCPY_INVALID)
+      return false;
+
+   /* If this is a nontrivial texture view, let another path handle it instead. */
+   if (texImage->TexObject->MinLayer)
+      return false;
+
+   if (!image->mt ||
+       (image->mt->surf.tiling != ISL_TILING_X &&
+        image->mt->surf.tiling != ISL_TILING_Y0)) {
+      /* The algorithm is written only for X- or Y-tiled memory. */
+      return false;
+   }
+
+   /* linear_to_tiled() assumes that if the object is swizzled, it is using
+    * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
+    * true on gen5 and above.
+    *
+    * The killer on top is that some gen4 have an L-shaped swizzle mode, where
+    * parts of the memory aren't swizzled at all. Userspace just can't handle
+    * that.
+    */
+   if (devinfo->gen < 5 && brw->has_swizzling)
+      return false;
+
+   int level = texImage->Level + texImage->TexObject->MinLevel;
+
+   /* Since we are going to write raw data to the miptree, we need to resolve
+    * any pending fast color clears before we start.
+    */
+   assert(image->mt->surf.logical_level0_px.depth == 1);
+   assert(image->mt->surf.logical_level0_px.array_len == 1);
+
+   intel_miptree_access_raw(brw, image->mt, level, 0, true);
+
+   bo = image->mt->bo;
+
+   if (brw_batch_references(&brw->batch, bo)) {
+      perf_debug("Flushing before mapping a referenced bo.\n");
+      intel_batchbuffer_flush(brw);
+   }
+
+   void *map = brw_bo_map(brw, bo, MAP_WRITE | MAP_RAW);
+   if (map == NULL) {
+      DBG("%s: failed to map bo\n", __func__);
+      return false;
+   }
+
+   src_pitch = _mesa_image_row_stride(packing, width, format, type);
+
+   /* We postponed printing this message until having committed to executing
+    * the function.
+    */
+   DBG("%s: level=%d offset=(%d,%d) (w,h)=(%d,%d) format=0x%x type=0x%x "
+       "mesa_format=0x%x tiling=%d "
+       "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d) ",
+       __func__, texImage->Level, xoffset, yoffset, width, height,
+       format, type, texImage->TexFormat, image->mt->surf.tiling,
+       packing->Alignment, packing->RowLength, packing->SkipPixels,
+       packing->SkipRows);
+
+   /* Adjust x and y offset based on miplevel */
+   unsigned level_x, level_y;
+   intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
+   xoffset += level_x;
+   yoffset += level_y;
+
+   isl_memcpy_linear_to_tiled(
+      xoffset * cpp, (xoffset + width) * cpp,
+      yoffset, yoffset + height,
+      map,
+      pixels,
+      image->mt->surf.row_pitch_B, src_pitch,
+      brw->has_swizzling,
+      image->mt->surf.tiling,
+      copy_type
+   );
+
+   brw_bo_unmap(bo);
+   return true;
+}
+
+
+static void
+intel_upload_tex(struct gl_context * ctx,
+                 GLuint dims,
+                 struct gl_texture_image *texImage,
+                 GLint xoffset, GLint yoffset, GLint zoffset,
+                 GLsizei width, GLsizei height, GLsizei depth,
+                 GLenum format, GLenum type,
+                 const GLvoid * pixels,
+                 const struct gl_pixelstore_attrib *packing)
+{
+   struct brw_context *brw = brw_context(ctx);
+   struct intel_mipmap_tree *mt = intel_texture_image(texImage)->mt;
+   bool ok;
+
+   /* Check that there is actually data to store. */
+   if (pixels == NULL && !packing->BufferObj)
+      return;
+
+   bool tex_busy = mt &&
+      (brw_batch_references(&brw->batch, mt->bo) || brw_bo_busy(mt->bo));
+
+   if (packing->BufferObj || tex_busy ||
+       mt->aux_usage == ISL_AUX_USAGE_CCS_E) {
+      ok = intel_texsubimage_blorp(brw, dims, texImage,
+                                   xoffset, yoffset, zoffset,
+                                   width, height, depth, format, type,
+                                   pixels, packing);
+      if (ok)
+         return;
+   }
+
+   ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
+                                       xoffset, yoffset, zoffset,
+                                       width, height, depth,
+                                       format, type, pixels, packing);
+   if (ok)
+     return;
+
+   _mesa_store_texsubimage(ctx, dims, texImage,
+                           xoffset, yoffset, zoffset,
+                           width, height, depth,
+                           format, type, pixels, packing);
+}
+
+
 static void
 intelTexImage(struct gl_context * ctx,
               GLuint dims,
@@ -90,11 +359,6 @@ intelTexImage(struct gl_context * ctx,
               GLenum format, GLenum type, const void *pixels,
               const struct gl_pixelstore_attrib *unpack)
 {
-   struct intel_texture_image *intelImage = intel_texture_image(texImage);
-   bool ok;
-
-   bool tex_busy = intelImage->mt && drm_intel_bo_busy(intelImage->mt->bo);
-
    DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
        __func__, _mesa_get_format_name(texImage->TexFormat),
        _mesa_enum_to_string(texImage->TexObject->Target),
@@ -107,100 +371,65 @@ intelTexImage(struct gl_context * ctx,
       return;
    }
 
-   assert(intelImage->mt);
+   assert(intel_texture_image(texImage)->mt);
 
-   ok = _mesa_meta_pbo_TexSubImage(ctx, dims, texImage, 0, 0, 0,
-                                   texImage->Width, texImage->Height,
-                                   texImage->Depth,
-                                   format, type, pixels,
-                                   false /*allocate_storage*/,
-                                   tex_busy, unpack);
-   if (ok)
-      return;
+   intel_upload_tex(ctx, dims, texImage, 0, 0, 0,
+                    texImage->Width, texImage->Height, texImage->Depth,
+                    format, type, pixels, unpack);
+}
 
-   ok = intel_texsubimage_tiled_memcpy(ctx, dims, texImage,
-                                       0, 0, 0, /*x,y,z offsets*/
-                                       texImage->Width,
-                                       texImage->Height,
-                                       texImage->Depth,
-                                       format, type, pixels, unpack,
-                                       false /*allocate_storage*/);
-   if (ok)
-      return;
 
-   DBG("%s: upload image %dx%dx%d pixels %p\n",
-       __func__, texImage->Width, texImage->Height, texImage->Depth,
-       pixels);
+static void
+intelTexSubImage(struct gl_context * ctx,
+                 GLuint dims,
+                 struct gl_texture_image *texImage,
+                 GLint xoffset, GLint yoffset, GLint zoffset,
+                 GLsizei width, GLsizei height, GLsizei depth,
+                 GLenum format, GLenum type,
+                 const GLvoid * pixels,
+                 const struct gl_pixelstore_attrib *packing)
+{
+   DBG("%s mesa_format %s target %s format %s type %s level %d %dx%dx%d\n",
+       __func__, _mesa_get_format_name(texImage->TexFormat),
+       _mesa_enum_to_string(texImage->TexObject->Target),
+       _mesa_enum_to_string(format), _mesa_enum_to_string(type),
+       texImage->Level, texImage->Width, texImage->Height, texImage->Depth);
 
-   _mesa_store_teximage(ctx, dims, texImage,
-                        format, type, pixels, unpack);
+   intel_upload_tex(ctx, dims, texImage, xoffset, yoffset, zoffset,
+                    width, height, depth, format, type, pixels, packing);
 }
 
 
-/**
- * Binds a BO to a texture image, as if it was uploaded by glTexImage2D().
- *
- * Used for GLX_EXT_texture_from_pixmap and EGL image extensions,
- */
 static void
-intel_set_texture_image_bo(struct gl_context *ctx,
+intel_set_texture_image_mt(struct brw_context *brw,
                            struct gl_texture_image *image,
-                           drm_intel_bo *bo,
-                           GLenum target,
-                           GLenum internalFormat,
+                           GLenum internal_format,
                            mesa_format format,
-                           uint32_t offset,
-                           GLuint width, GLuint height,
-                           GLuint pitch,
-                           GLuint tile_x, GLuint tile_y,
-                           uint32_t layout_flags)
+                           struct intel_mipmap_tree *mt)
+
 {
-   struct brw_context *brw = brw_context(ctx);
-   struct intel_texture_image *intel_image = intel_texture_image(image);
    struct gl_texture_object *texobj = image->TexObject;
    struct intel_texture_object *intel_texobj = intel_texture_object(texobj);
-   uint32_t draw_x, draw_y;
+   struct intel_texture_image *intel_image = intel_texture_image(image);
 
    _mesa_init_teximage_fields(&brw->ctx, image,
-                             width, height, 1,
-                             0, internalFormat, format);
-
-   ctx->Driver.FreeTextureImageBuffer(ctx, image);
-
-   intel_image->mt = intel_miptree_create_for_bo(brw, bo, image->TexFormat,
-                                                 0, width, height, 1, pitch,
-                                                 layout_flags);
-   if (intel_image->mt == NULL)
-       return;
-   intel_image->mt->target = target;
-   intel_image->mt->total_width = width;
-   intel_image->mt->total_height = height;
-   intel_image->mt->level[0].slice[0].x_offset = tile_x;
-   intel_image->mt->level[0].slice[0].y_offset = tile_y;
-
-   intel_miptree_get_tile_offsets(intel_image->mt, 0, 0, &draw_x, &draw_y);
-
-   /* From "OES_EGL_image" error reporting. We report GL_INVALID_OPERATION
-    * for EGL images from non-tile aligned sufaces in gen4 hw and earlier which has
-    * trouble resolving back to destination image due to alignment issues.
-    */
-   if (!brw->has_surface_tile_offset &&
-       (draw_x != 0 || draw_y != 0)) {
-      _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
-      intel_miptree_release(&intel_image->mt);
-      return;
-   }
+                              mt->surf.logical_level0_px.width,
+                              mt->surf.logical_level0_px.height, 1,
+                              0, internal_format, format);
+
+   brw->ctx.Driver.FreeTextureImageBuffer(&brw->ctx, image);
 
    intel_texobj->needs_validate = true;
+   intel_image->base.RowStride = mt->surf.row_pitch_B / mt->cpp;
+   assert(mt->surf.row_pitch_B % mt->cpp == 0);
 
-   intel_image->mt->offset = offset;
-   assert(pitch % intel_image->mt->cpp == 0);
-   intel_image->base.RowStride = pitch / intel_image->mt->cpp;
+   intel_miptree_reference(&intel_image->mt, mt);
 
    /* Immediately validate the image to the object. */
-   intel_miptree_reference(&intel_texobj->mt, intel_image->mt);
+   intel_miptree_reference(&intel_texobj->mt, mt);
 }
 
+
 void
 intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
                   GLint texture_format,
@@ -212,8 +441,10 @@ intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
    struct intel_renderbuffer *rb;
    struct gl_texture_object *texObj;
    struct gl_texture_image *texImage;
-   int level = 0, internalFormat = 0;
    mesa_format texFormat = MESA_FORMAT_NONE;
+   GLenum internal_format = 0;
+
+   _mesa_glthread_finish(ctx);
 
    texObj = _mesa_get_current_tex_object(ctx, target);
 
@@ -231,32 +462,102 @@ intelSetTexBuffer2(__DRIcontext *pDRICtx, GLint target,
    if (!rb || !rb->mt)
       return;
 
+   /* Neither the EGL and GLX texture_from_pixmap specs say anything about
+    * sRGB.  They are both from a time where sRGB was considered an extra
+    * encoding step you did as part of rendering/blending and not a format.
+    * Even though we have concept of sRGB visuals, X has classically assumed
+    * that your data is just bits and sRGB rendering is entirely a client-side
+    * rendering construct.  The assumption is that the result of BindTexImage
+    * is a texture with a linear format even if it was rendered with sRGB
+    * encoding enabled.
+    */
+   texFormat = _mesa_get_srgb_format_linear(intel_rb_format(rb));
+
    if (rb->mt->cpp == 4) {
-      if (texture_format == __DRI_TEXTURE_FORMAT_RGB) {
-         internalFormat = GL_RGB;
-         texFormat = MESA_FORMAT_B8G8R8X8_UNORM;
-      }
-      else {
-         internalFormat = GL_RGBA;
-         texFormat = MESA_FORMAT_B8G8R8A8_UNORM;
-      }
+      /* The extra texture_format parameter indicates whether the alpha
+       * channel should be respected or ignored.  If we set internal_format to
+       * GL_RGB, the texture handling code is smart enough to swap the format
+       * or apply a swizzle if the underlying format is RGBA so we don't need
+       * to stomp it to RGBX or anything like that.
+       */
+      if (texture_format == __DRI_TEXTURE_FORMAT_RGB)
+         internal_format = GL_RGB;
+      else
+         internal_format = GL_RGBA;
    } else if (rb->mt->cpp == 2) {
-      internalFormat = GL_RGB;
-      texFormat = MESA_FORMAT_B5G6R5_UNORM;
+      internal_format = GL_RGB;
    }
 
+   intel_miptree_finish_external(brw, rb->mt);
+
    _mesa_lock_texture(&brw->ctx, texObj);
-   texImage = _mesa_get_tex_image(ctx, texObj, target, level);
-   intel_miptree_make_shareable(brw, rb->mt);
-   intel_set_texture_image_bo(ctx, texImage, rb->mt->bo, target,
-                              internalFormat, texFormat, 0,
-                              rb->Base.Base.Width,
-                              rb->Base.Base.Height,
-                              rb->mt->pitch,
-                              0, 0, 0);
+   texImage = _mesa_get_tex_image(ctx, texObj, target, 0);
+   intel_set_texture_image_mt(brw, texImage, internal_format,
+                              texFormat, rb->mt);
    _mesa_unlock_texture(&brw->ctx, texObj);
 }
 
+void
+intelReleaseTexBuffer(__DRIcontext *pDRICtx, GLint target,
+                      __DRIdrawable *dPriv)
+{
+   struct brw_context *brw = pDRICtx->driverPrivate;
+   struct gl_context *ctx = &brw->ctx;
+   struct gl_texture_object *tex_obj;
+   struct intel_texture_object *intel_tex;
+
+   tex_obj = _mesa_get_current_tex_object(ctx, target);
+   if (!tex_obj)
+      return;
+
+   _mesa_lock_texture(&brw->ctx, tex_obj);
+
+   intel_tex = intel_texture_object(tex_obj);
+   if (!intel_tex->mt) {
+      _mesa_unlock_texture(&brw->ctx, tex_obj);
+      return;
+   }
+
+   /* The intel_miptree_prepare_external below as well as the finish_external
+    * above in intelSetTexBuffer2 *should* do nothing.  The BindTexImage call
+    * from both GLX and EGL has TexImage2D and not TexSubImage2D semantics so
+    * the texture is not immutable.  This means that the user cannot create a
+    * texture view of the image with a different format.  Since the only three
+    * formats available when using BindTexImage are all UNORM, we can never
+    * end up with an sRGB format being used for texturing and so we shouldn't
+    * get any format-related resolves when texturing from it.
+    *
+    * While very unlikely, it is possible that the client could use the bound
+    * texture with GL_ARB_image_load_store.  In that case, we'll do a resolve
+    * but that's not actually a problem as it just means that we lose
+    * compression on this texture until the next time it's used as a render
+    * target.
+    *
+    * The only other way we could end up with an unexpected aux usage would be
+    * if we rendered to the image from the same context as we have it bound as
+    * a texture between BindTexImage and ReleaseTexImage.  However, the spec
+    * clearly calls this case out and says you shouldn't do that.  It doesn't
+    * explicitly prevent binding the texture to a framebuffer but it says the
+    * results of trying to render to it while bound are undefined.
+    *
+    * Just to keep everything safe and sane, we do a prepare_external but it
+    * should be a no-op in almost all cases.  On the off chance that someone
+    * ever triggers this, we should at least warn them.
+    */
+   if (intel_tex->mt->aux_buf &&
+       intel_miptree_get_aux_state(intel_tex->mt, 0, 0) !=
+       isl_drm_modifier_get_default_aux_state(intel_tex->mt->drm_modifier)) {
+      _mesa_warning(ctx, "Aux state changed between BindTexImage and "
+                         "ReleaseTexImage.  Most likely someone tried to draw "
+                         "to the pixmap bound in BindTexImage or used it with "
+                         "image_load_store.");
+   }
+
+   intel_miptree_prepare_external(brw, intel_tex->mt);
+
+   _mesa_unlock_texture(&brw->ctx, tex_obj);
+}
+
 static GLboolean
 intel_bind_renderbuffer_tex_image(struct gl_context *ctx,
                                   struct gl_renderbuffer *rb,
@@ -302,31 +603,22 @@ intelSetTexBuffer(__DRIcontext *pDRICtx, GLint target, __DRIdrawable *dPriv)
 }
 
 static void
-intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
-                             struct gl_texture_object *texObj,
-                             struct gl_texture_image *texImage,
-                             GLeglImageOES image_handle)
+intel_image_target_texture(struct gl_context *ctx, GLenum target,
+                           struct gl_texture_object *texObj,
+                           struct gl_texture_image *texImage,
+                           GLeglImageOES image_handle,
+                           bool storage)
 {
    struct brw_context *brw = brw_context(ctx);
-   __DRIscreen *screen;
+   struct intel_mipmap_tree *mt;
+   __DRIscreen *dri_screen = brw->screen->driScrnPriv;
    __DRIimage *image;
 
-   screen = brw->intelScreen->driScrnPriv;
-   image = screen->dri2.image->lookupEGLImage(screen, image_handle,
-                                             screen->loaderPrivate);
+   image = dri_screen->dri2.image->lookupEGLImage(dri_screen, image_handle,
+                                                  dri_screen->loaderPrivate);
    if (image == NULL)
       return;
 
-   /* We support external textures only for EGLImages created with
-    * EGL_EXT_image_dma_buf_import. We may lift that restriction in the future.
-    */
-   if (target == GL_TEXTURE_EXTERNAL_OES && !image->dma_buf_imported) {
-      _mesa_error(ctx, GL_INVALID_OPERATION,
-            "glEGLImageTargetTexture2DOES(external target is enabled only "
-               "for images created with EGL_EXT_image_dma_buf_import");
-      return;
-   }
-
    /* Disallow depth/stencil textures: we don't have a way to pass the
     * separate stencil miptree of a GL_DEPTH_STENCIL texture through.
     */
@@ -335,18 +627,95 @@ intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
       return;
    }
 
-   /* Disable creation of the texture's aux buffers because the driver exposes
-    * no EGL API to manage them. That is, there is no API for resolving the aux
-    * buffer's content to the main buffer nor for invalidating the aux buffer's
-    * content.
+   mt = intel_miptree_create_for_dri_image(brw, image, target, image->format,
+                                           false);
+   if (mt == NULL)
+      return;
+
+   struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
+   intel_texobj->planar_format = image->planar_format;
+   intel_texobj->yuv_color_space = image->yuv_color_space;
+
+   GLenum internal_format =
+      image->internal_format != 0 ?
+      image->internal_format : _mesa_get_format_base_format(mt->format);
+
+   /* Fix the internal format when _mesa_get_format_base_format(mt->format)
+    * isn't a valid one for that particular format.
+    */
+   if (brw->mesa_format_supports_render[image->format]) {
+      if (image->format == MESA_FORMAT_R10G10B10A2_UNORM ||
+          image->format == MESA_FORMAT_R10G10B10X2_UNORM ||
+          image->format == MESA_FORMAT_B10G10R10A2_UNORM ||
+          image->format == MESA_FORMAT_B10G10R10X2_UNORM)
+         internal_format = GL_RGB10_A2;
+   }
+
+   /* Guess sized internal format for dma-bufs, as specified by
+    * EXT_EGL_image_storage.
     */
-   intel_set_texture_image_bo(ctx, texImage, image->bo,
-                              target, image->internal_format,
-                              image->format, image->offset,
-                              image->width,  image->height,
-                              image->pitch,
-                              image->tile_x, image->tile_y,
-                              MIPTREE_LAYOUT_DISABLE_AUX);
+   if (storage && target == GL_TEXTURE_2D && image->imported_dmabuf) {
+      internal_format = driGLFormatToSizedInternalGLFormat(image->format);
+      if (internal_format == GL_NONE) {
+         _mesa_error(ctx, GL_INVALID_OPERATION, __func__);
+         return;
+      }
+   }
+
+   intel_set_texture_image_mt(brw, texImage, internal_format, mt->format, mt);
+   intel_miptree_release(&mt);
+}
+
+static void
+intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
+                              struct gl_texture_object *texObj,
+                              struct gl_texture_image *texImage,
+                              GLeglImageOES image_handle)
+{
+   intel_image_target_texture(ctx, target, texObj, texImage, image_handle,
+                              false);
+}
+
+static void
+intel_image_target_tex_storage(struct gl_context *ctx, GLenum target,
+                              struct gl_texture_object *texObj,
+                              struct gl_texture_image *texImage,
+                              GLeglImageOES image_handle)
+{
+   struct intel_texture_object *intel_texobj = intel_texture_object(texObj);
+   intel_image_target_texture(ctx, target, texObj, texImage, image_handle,
+                              true);
+
+   /* The miptree is in a validated state, so no need to check later. */
+   intel_texobj->needs_validate = false;
+   intel_texobj->validated_first_level = 0;
+   intel_texobj->validated_last_level = 0;
+   intel_texobj->_Format = texImage->TexFormat;
+}
+
+static bool
+intel_gettexsubimage_blorp(struct brw_context *brw,
+                           struct gl_texture_image *tex_image,
+                           unsigned x, unsigned y, unsigned z,
+                           unsigned width, unsigned height, unsigned depth,
+                           GLenum format, GLenum type, const void *pixels,
+                           const struct gl_pixelstore_attrib *packing)
+{
+   struct intel_texture_image *intel_image = intel_texture_image(tex_image);
+   const unsigned mt_level = tex_image->Level + tex_image->TexObject->MinLevel;
+   const unsigned mt_z = tex_image->TexObject->MinLayer + tex_image->Face + z;
+
+   /* The blorp path can't understand crazy format hackery */
+   if (_mesa_base_tex_format(&brw->ctx, tex_image->InternalFormat) !=
+       _mesa_get_format_base_format(tex_image->TexFormat))
+      return false;
+
+   return brw_blorp_download_miptree(brw, intel_image->mt,
+                                     tex_image->TexFormat, SWIZZLE_XYZW,
+                                     mt_level, x, y, mt_z,
+                                     width, height, depth,
+                                     tex_image->TexObject->Target,
+                                     format, type, false, pixels, packing);
 }
 
 /**
@@ -354,7 +723,7 @@ intel_image_target_texture_2d(struct gl_context *ctx, GLenum target,
  *
  * \see intel_readpixels_tiled_memcpy()
  */
-bool
+static bool
 intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
                                   struct gl_texture_image *texImage,
                                   GLint xoffset, GLint yoffset,
@@ -364,16 +733,15 @@ intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
                                   const struct gl_pixelstore_attrib *packing)
 {
    struct brw_context *brw = brw_context(ctx);
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
    struct intel_texture_image *image = intel_texture_image(texImage);
    int dst_pitch;
 
    /* The miptree's buffer. */
-   drm_intel_bo *bo;
-
-   int error = 0;
+   struct brw_bo *bo;
 
    uint32_t cpp;
-   mem_copy_fn mem_copy = NULL;
+   isl_memcpy_type copy_type;
 
    /* This fastpath is restricted to specific texture types:
     * a 2D BGRA, RGBA, L8 or A8 texture. It could be generalized to support
@@ -384,12 +752,12 @@ intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
     * with _mesa_image_row_stride. However, before removing the restrictions
     * we need tests.
     */
-   if (!brw->has_llc ||
+   if (!devinfo->has_llc ||
        !(type == GL_UNSIGNED_BYTE || type == GL_UNSIGNED_INT_8_8_8_8_REV) ||
        !(texImage->TexObject->Target == GL_TEXTURE_2D ||
          texImage->TexObject->Target == GL_TEXTURE_RECTANGLE) ||
        pixels == NULL ||
-       _mesa_is_bufferobj(packing->BufferObj) ||
+       packing->BufferObj ||
        packing->Alignment > 4 ||
        packing->SkipPixels > 0 ||
        packing->SkipRows > 0 ||
@@ -400,14 +768,16 @@ intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
       return false;
 
    /* We can't handle copying from RGBX or BGRX because the tiled_memcpy
-    * function doesn't set the last channel to 1.
+    * function doesn't set the last channel to 1. Note this checks BaseFormat
+    * rather than TexFormat in case the RGBX format is being simulated with an
+    * RGBA format.
     */
-   if (texImage->TexFormat == MESA_FORMAT_B8G8R8X8_UNORM ||
-       texImage->TexFormat == MESA_FORMAT_R8G8B8X8_UNORM)
+   if (texImage->_BaseFormat == GL_RGB)
       return false;
 
-   if (!intel_get_memcpy(texImage->TexFormat, format, type, &mem_copy, &cpp,
-                         INTEL_DOWNLOAD))
+   copy_type = intel_miptree_get_memcpy_type(texImage->TexFormat, format, type,
+                                             &cpp);
+   if (copy_type == ISL_MEMCPY_INVALID)
       return false;
 
    /* If this is a nontrivial texture view, let another path handle it instead. */
@@ -415,26 +785,42 @@ intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
       return false;
 
    if (!image->mt ||
-       (image->mt->tiling != I915_TILING_X &&
-       image->mt->tiling != I915_TILING_Y)) {
+       (image->mt->surf.tiling != ISL_TILING_X &&
+        image->mt->surf.tiling != ISL_TILING_Y0)) {
       /* The algorithm is written only for X- or Y-tiled memory. */
       return false;
    }
 
+   /* tiled_to_linear() assumes that if the object is swizzled, it is using
+    * I915_BIT6_SWIZZLE_9_10 for X and I915_BIT6_SWIZZLE_9 for Y.  This is only
+    * true on gen5 and above.
+    *
+    * The killer on top is that some gen4 have an L-shaped swizzle mode, where
+    * parts of the memory aren't swizzled at all. Userspace just can't handle
+    * that.
+    */
+   if (devinfo->gen < 5 && brw->has_swizzling)
+      return false;
+
+   int level = texImage->Level + texImage->TexObject->MinLevel;
+
    /* Since we are going to write raw data to the miptree, we need to resolve
     * any pending fast color clears before we start.
     */
-   intel_miptree_resolve_color(brw, image->mt);
+   assert(image->mt->surf.logical_level0_px.depth == 1);
+   assert(image->mt->surf.logical_level0_px.array_len == 1);
+
+   intel_miptree_access_raw(brw, image->mt, level, 0, true);
 
    bo = image->mt->bo;
 
-   if (drm_intel_bo_references(brw->batch.bo, bo)) {
+   if (brw_batch_references(&brw->batch, bo)) {
       perf_debug("Flushing before mapping a referenced bo.\n");
       intel_batchbuffer_flush(brw);
    }
 
-   error = brw_bo_map(brw, bo, false /* write enable */, "miptree");
-   if (error) {
+   void *map = brw_bo_map(brw, bo, MAP_READ | MAP_RAW);
+   if (map == NULL) {
       DBG("%s: failed to map bo\n", __func__);
       return false;
    }
@@ -445,28 +831,28 @@ intel_gettexsubimage_tiled_memcpy(struct gl_context *ctx,
        "mesa_format=0x%x tiling=%d "
        "packing=(alignment=%d row_length=%d skip_pixels=%d skip_rows=%d)\n",
        __func__, texImage->Level, xoffset, yoffset, width, height,
-       format, type, texImage->TexFormat, image->mt->tiling,
+       format, type, texImage->TexFormat, image->mt->surf.tiling,
        packing->Alignment, packing->RowLength, packing->SkipPixels,
        packing->SkipRows);
 
-   int level = texImage->Level + texImage->TexObject->MinLevel;
-
    /* Adjust x and y offset based on miplevel */
-   xoffset += image->mt->level[level].level_x;
-   yoffset += image->mt->level[level].level_y;
+   unsigned level_x, level_y;
+   intel_miptree_get_image_offset(image->mt, level, 0, &level_x, &level_y);
+   xoffset += level_x;
+   yoffset += level_y;
 
-   tiled_to_linear(
+   isl_memcpy_tiled_to_linear(
       xoffset * cpp, (xoffset + width) * cpp,
       yoffset, yoffset + height,
-      pixels - (ptrdiff_t) yoffset * dst_pitch - (ptrdiff_t) xoffset * cpp,
-      bo->virtual,
-      dst_pitch, image->mt->pitch,
+      pixels,
+      map,
+      dst_pitch, image->mt->surf.row_pitch_B,
       brw->has_swizzling,
-      image->mt->tiling,
-      mem_copy
+      image->mt->surf.tiling,
+      copy_type
    );
 
-   drm_intel_bo_unmap(bo);
+   brw_bo_unmap(bo);
    return true;
 }
 
@@ -482,19 +868,12 @@ intel_get_tex_sub_image(struct gl_context *ctx,
 
    DBG("%s\n", __func__);
 
-   if (_mesa_is_bufferobj(ctx->Pack.BufferObj)) {
-      if (_mesa_meta_pbo_GetTexSubImage(ctx, 3, texImage,
-                                        xoffset, yoffset, zoffset,
-                                        width, height, depth, format, type,
-                                        pixels, &ctx->Pack)) {
-         /* Flush to guarantee coherency between the render cache and other
-          * caches the PBO could potentially be bound to after this point.
-          * See the related comment in intelReadPixels() for a more detailed
-          * explanation.
-          */
-         brw_emit_mi_flush(brw);
+   if (ctx->Pack.BufferObj) {
+      if (intel_gettexsubimage_blorp(brw, texImage,
+                                     xoffset, yoffset, zoffset,
+                                     width, height, depth, format, type,
+                                     pixels, &ctx->Pack))
          return;
-      }
 
       perf_debug("%s: fallback to CPU mapping in PBO case\n", __func__);
    }
@@ -513,11 +892,100 @@ intel_get_tex_sub_image(struct gl_context *ctx,
    DBG("%s - DONE\n", __func__);
 }
 
+static void
+flush_astc_denorms(struct gl_context *ctx, GLuint dims,
+                   struct gl_texture_image *texImage,
+                   GLint xoffset, GLint yoffset, GLint zoffset,
+                   GLsizei width, GLsizei height, GLsizei depth)
+{
+   struct compressed_pixelstore store;
+   _mesa_compute_compressed_pixelstore(dims, texImage->TexFormat,
+                                       width, height, depth,
+                                       &ctx->Unpack, &store);
+
+   for (int slice = 0; slice < store.CopySlices; slice++) {
+
+      /* Map dest texture buffer */
+      GLubyte *dstMap;
+      GLint dstRowStride;
+      ctx->Driver.MapTextureImage(ctx, texImage, slice + zoffset,
+                                  xoffset, yoffset, width, height,
+                                  GL_MAP_READ_BIT | GL_MAP_WRITE_BIT,
+                                  &dstMap, &dstRowStride);
+      if (!dstMap)
+         continue;
+
+      for (int i = 0; i < store.CopyRowsPerSlice; i++) {
+
+         /* An ASTC block is stored in little endian mode. The byte that
+          * contains bits 0..7 is stored at the lower address in memory.
+          */
+         struct astc_void_extent {
+            uint16_t header : 12;
+            uint16_t dontcare[3];
+            uint16_t R;
+            uint16_t G;
+            uint16_t B;
+            uint16_t A;
+         } *blocks = (struct astc_void_extent*) dstMap;
+
+         /* Iterate over every copied block in the row */
+         for (int j = 0; j < store.CopyBytesPerRow / 16; j++) {
+
+            /* Check if the header matches that of an LDR void-extent block */
+            if (blocks[j].header == 0xDFC) {
+
+               /* Flush UNORM16 values that would be denormalized */
+               if (blocks[j].A < 4) blocks[j].A = 0;
+               if (blocks[j].B < 4) blocks[j].B = 0;
+               if (blocks[j].G < 4) blocks[j].G = 0;
+               if (blocks[j].R < 4) blocks[j].R = 0;
+            }
+         }
+
+         dstMap += dstRowStride;
+      }
+
+      ctx->Driver.UnmapTextureImage(ctx, texImage, slice + zoffset);
+   }
+}
+
+
+static void
+intelCompressedTexSubImage(struct gl_context *ctx, GLuint dims,
+                        struct gl_texture_image *texImage,
+                        GLint xoffset, GLint yoffset, GLint zoffset,
+                        GLsizei width, GLsizei height, GLsizei depth,
+                        GLenum format,
+                        GLsizei imageSize, const GLvoid *data)
+{
+   /* Upload the compressed data blocks */
+   _mesa_store_compressed_texsubimage(ctx, dims, texImage,
+                                      xoffset, yoffset, zoffset,
+                                      width, height, depth,
+                                      format, imageSize, data);
+
+   /* Fix up copied ASTC blocks if necessary */
+   GLenum gl_format = _mesa_compressed_format_to_glenum(ctx,
+                                                        texImage->TexFormat);
+   bool is_linear_astc = _mesa_is_astc_format(gl_format) &&
+                        !_mesa_is_srgb_format(gl_format);
+   struct brw_context *brw = (struct brw_context*) ctx;
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
+   if (devinfo->gen == 9 && !gen_device_info_is_9lp(devinfo) && is_linear_astc)
+      flush_astc_denorms(ctx, dims, texImage,
+                         xoffset, yoffset, zoffset,
+                         width, height, depth);
+}
+
 void
 intelInitTextureImageFuncs(struct dd_function_table *functions)
 {
    functions->TexImage = intelTexImage;
+   functions->TexSubImage = intelTexSubImage;
+   functions->CompressedTexSubImage = intelCompressedTexSubImage;
    functions->EGLImageTargetTexture2D = intel_image_target_texture_2d;
+   functions->EGLImageTargetTexStorage = intel_image_target_tex_storage;
    functions->BindRenderbufferTexImage = intel_bind_renderbuffer_tex_image;
    functions->GetTexSubImage = intel_get_tex_sub_image;
 }