i965: Workaround the gen9 hw astc5x5 sampler bug
[mesa.git] / src / mesa / drivers / dri / i965 / intel_mipmap_tree.c
index 1a797326a8233cf3f448298ebf3c1ba4ded474fe..36681352ba73c2e4c12512d529ba82bf0ed654da 100644 (file)
@@ -510,7 +510,7 @@ free_aux_state_map(enum isl_aux_state **state)
 }
 
 static bool
-need_to_retile_as_linear(struct brw_context *brw, unsigned row_pitch,
+need_to_retile_as_linear(struct brw_context *brw, unsigned blt_pitch,
                          enum isl_tiling tiling, unsigned samples)
 {
    if (samples > 1)
@@ -519,13 +519,9 @@ need_to_retile_as_linear(struct brw_context *brw, unsigned row_pitch,
    if (tiling == ISL_TILING_LINEAR)
       return false;
 
-    /* If the width is much smaller than a tile, don't bother tiling. */
-   if (row_pitch < 64)
-      return true;
-
-   if (ALIGN(row_pitch, 512) >= 32768) {
-      perf_debug("row pitch %u too large to blit, falling back to untiled",
-                 row_pitch);
+   if (blt_pitch >= 32768) {
+      perf_debug("blt pitch %u too large to blit, falling back to untiled",
+                 blt_pitch);
       return true;
    }
 
@@ -605,7 +601,7 @@ make_surface(struct brw_context *brw, GLenum target, mesa_format format,
    bool is_depth_stencil =
       mt->surf.usage & (ISL_SURF_USAGE_STENCIL_BIT | ISL_SURF_USAGE_DEPTH_BIT);
    if (!is_depth_stencil) {
-      if (need_to_retile_as_linear(brw, mt->surf.row_pitch,
+      if (need_to_retile_as_linear(brw, intel_miptree_blt_pitch(mt),
                                    mt->surf.tiling, mt->surf.samples)) {
          init_info.tiling_flags = 1u << ISL_TILING_LINEAR;
          if (!isl_surf_init_s(&brw->isl_dev, &mt->surf, &init_info))
@@ -627,6 +623,7 @@ make_surface(struct brw_context *brw, GLenum target, mesa_format format,
    if (!bo) {
       mt->bo = brw_bo_alloc_tiled(brw->bufmgr, "isl-miptree",
                                   mt->surf.size,
+                                  BRW_MEMZONE_OTHER,
                                   isl_tiling_to_i915_tiling(
                                      mt->surf.tiling),
                                   mt->surf.row_pitch, alloc_flags);
@@ -652,28 +649,21 @@ fail:
    return NULL;
 }
 
-static bool
-make_separate_stencil_surface(struct brw_context *brw,
-                              struct intel_mipmap_tree *mt)
+/* Return the usual surface usage flags for the given format. */
+static isl_surf_usage_flags_t
+mt_surf_usage(mesa_format format)
 {
-   mt->stencil_mt = make_surface(brw, mt->target, MESA_FORMAT_S_UINT8,
-                                 0, mt->surf.levels - 1,
-                                 mt->surf.logical_level0_px.width,
-                                 mt->surf.logical_level0_px.height,
-                                 mt->surf.dim == ISL_SURF_DIM_3D ?
-                                    mt->surf.logical_level0_px.depth :
-                                    mt->surf.logical_level0_px.array_len,
-                                 mt->surf.samples, ISL_TILING_W_BIT,
-                                 ISL_SURF_USAGE_STENCIL_BIT |
-                                 ISL_SURF_USAGE_TEXTURE_BIT,
-                                 BO_ALLOC_BUSY, 0, NULL);
-
-   if (!mt->stencil_mt)
-      return false;
-
-   mt->stencil_mt->r8stencil_needs_update = true;
-
-   return true;
+   switch(_mesa_get_format_base_format(format)) {
+   case GL_DEPTH_COMPONENT:
+      return ISL_SURF_USAGE_DEPTH_BIT | ISL_SURF_USAGE_TEXTURE_BIT;
+   case GL_DEPTH_STENCIL:
+      return ISL_SURF_USAGE_DEPTH_BIT | ISL_SURF_USAGE_STENCIL_BIT |
+             ISL_SURF_USAGE_TEXTURE_BIT;
+   case GL_STENCIL_INDEX:
+      return ISL_SURF_USAGE_STENCIL_BIT | ISL_SURF_USAGE_TEXTURE_BIT;
+   default:
+      return ISL_SURF_USAGE_RENDER_TARGET_BIT | ISL_SURF_USAGE_TEXTURE_BIT;
+   }
 }
 
 static struct intel_mipmap_tree *
@@ -689,75 +679,48 @@ miptree_create(struct brw_context *brw,
                enum intel_miptree_create_flags flags)
 {
    const struct gen_device_info *devinfo = &brw->screen->devinfo;
+   const uint32_t alloc_flags =
+      (flags & MIPTREE_CREATE_BUSY || num_samples > 1) ? BO_ALLOC_BUSY : 0;
+   isl_tiling_flags_t tiling_flags = ISL_TILING_ANY_MASK;
 
-   if (format == MESA_FORMAT_S_UINT8)
-      return make_surface(brw, target, format, first_level, last_level,
-                          width0, height0, depth0, num_samples,
-                          ISL_TILING_W_BIT,
-                          ISL_SURF_USAGE_STENCIL_BIT |
-                          ISL_SURF_USAGE_TEXTURE_BIT,
-                          BO_ALLOC_BUSY,
-                          0,
-                          NULL);
+   /* TODO: This used to be because there wasn't BLORP to handle Y-tiling. */
+   if (devinfo->gen < 6 && _mesa_is_format_color_format(format))
+      tiling_flags &= ~ISL_TILING_Y0_BIT;
 
-   const GLenum base_format = _mesa_get_format_base_format(format);
-   if ((base_format == GL_DEPTH_COMPONENT ||
-        base_format == GL_DEPTH_STENCIL) &&
-       !(flags & MIPTREE_CREATE_LINEAR)) {
+   mesa_format mt_fmt;
+   if (_mesa_is_format_color_format(format)) {
+      mt_fmt = intel_lower_compressed_format(brw, format);
+   } else {
       /* Fix up the Z miptree format for how we're splitting out separate
-       * stencil.  Gen7 expects there to be no stencil bits in its depth buffer.
+       * stencil. Gen7 expects there to be no stencil bits in its depth buffer.
        */
-      const mesa_format depth_only_format =
-         intel_depth_format_for_depthstencil_format(format);
-      struct intel_mipmap_tree *mt = make_surface(
-         brw, target, devinfo->gen >= 6 ? depth_only_format : format,
-         first_level, last_level,
-         width0, height0, depth0, num_samples, ISL_TILING_Y0_BIT,
-         ISL_SURF_USAGE_DEPTH_BIT | ISL_SURF_USAGE_TEXTURE_BIT,
-         BO_ALLOC_BUSY, 0, NULL);
-
-      if (needs_separate_stencil(brw, mt, format) &&
-          !make_separate_stencil_surface(brw, mt)) {
-         intel_miptree_release(&mt);
-         return NULL;
-      }
-
-      if (!(flags & MIPTREE_CREATE_NO_AUX))
-         intel_miptree_choose_aux_usage(brw, mt);
-
-      return mt;
+      mt_fmt = (devinfo->gen < 6) ? format :
+               intel_depth_format_for_depthstencil_format(format);
    }
 
-   mesa_format tex_format = format;
-   mesa_format etc_format = MESA_FORMAT_NONE;
-   uint32_t alloc_flags = 0;
-
-   format = intel_lower_compressed_format(brw, format);
-
-   etc_format = (format != tex_format) ? tex_format : MESA_FORMAT_NONE;
-
-   if (flags & MIPTREE_CREATE_BUSY)
-      alloc_flags |= BO_ALLOC_BUSY;
-
-   isl_tiling_flags_t tiling_flags = (flags & MIPTREE_CREATE_LINEAR) ?
-      ISL_TILING_LINEAR_BIT : ISL_TILING_ANY_MASK;
-
-   /* TODO: This used to be because there wasn't BLORP to handle Y-tiling. */
-   if (devinfo->gen < 6)
-      tiling_flags &= ~ISL_TILING_Y0_BIT;
+   struct intel_mipmap_tree *mt =
+      make_surface(brw, target, mt_fmt, first_level, last_level,
+                   width0, height0, depth0, num_samples,
+                   tiling_flags, mt_surf_usage(mt_fmt),
+                   alloc_flags, 0, NULL);
 
-   struct intel_mipmap_tree *mt = make_surface(
-                                     brw, target, format,
-                                     first_level, last_level,
-                                     width0, height0, depth0,
-                                     num_samples, tiling_flags,
-                                     ISL_SURF_USAGE_RENDER_TARGET_BIT |
-                                     ISL_SURF_USAGE_TEXTURE_BIT,
-                                     alloc_flags, 0, NULL);
-   if (!mt)
+   if (mt == NULL)
       return NULL;
 
-   mt->etc_format = etc_format;
+   if (needs_separate_stencil(brw, mt, format)) {
+      mt->stencil_mt =
+         make_surface(brw, target, MESA_FORMAT_S_UINT8, first_level, last_level,
+                      width0, height0, depth0, num_samples,
+                      ISL_TILING_W_BIT, mt_surf_usage(MESA_FORMAT_S_UINT8),
+                      alloc_flags, 0, NULL);
+      if (mt->stencil_mt == NULL) {
+         intel_miptree_release(&mt);
+         return NULL;
+      }
+   }
+
+   mt->etc_format = (_mesa_is_format_color_format(format) && mt_fmt != format) ?
+                    format : MESA_FORMAT_NONE;
 
    if (!(flags & MIPTREE_CREATE_NO_AUX))
       intel_miptree_choose_aux_usage(brw, mt);
@@ -821,12 +784,11 @@ intel_miptree_create_for_bo(struct brw_context *brw,
 
    if ((base_format == GL_DEPTH_COMPONENT ||
         base_format == GL_DEPTH_STENCIL)) {
-      const mesa_format depth_only_format =
+      const mesa_format mt_fmt = (devinfo->gen < 6) ? format :
          intel_depth_format_for_depthstencil_format(format);
-      mt = make_surface(brw, target,
-                        devinfo->gen >= 6 ? depth_only_format : format,
+      mt = make_surface(brw, target, mt_fmt,
                         0, 0, width, height, depth, 1, ISL_TILING_Y0_BIT,
-                        ISL_SURF_USAGE_DEPTH_BIT | ISL_SURF_USAGE_TEXTURE_BIT,
+                        mt_surf_usage(mt_fmt),
                         0, pitch, bo);
       if (!mt)
          return NULL;
@@ -841,8 +803,7 @@ intel_miptree_create_for_bo(struct brw_context *brw,
       mt = make_surface(brw, target, MESA_FORMAT_S_UINT8,
                         0, 0, width, height, depth, 1,
                         ISL_TILING_W_BIT,
-                        ISL_SURF_USAGE_STENCIL_BIT |
-                        ISL_SURF_USAGE_TEXTURE_BIT,
+                        mt_surf_usage(MESA_FORMAT_S_UINT8),
                         0, pitch, bo);
       if (!mt)
          return NULL;
@@ -864,16 +825,10 @@ intel_miptree_create_for_bo(struct brw_context *brw,
     */
    assert(pitch >= 0);
 
-   /* The BO already has a tiling format and we shouldn't confuse the lower
-    * layers by making it try to find a tiling format again.
-    */
-   assert((flags & MIPTREE_CREATE_LINEAR) == 0);
-
    mt = make_surface(brw, target, format,
                      0, 0, width, height, depth, 1,
                      1lu << tiling,
-                     ISL_SURF_USAGE_RENDER_TARGET_BIT |
-                     ISL_SURF_USAGE_TEXTURE_BIT,
+                     mt_surf_usage(format),
                      0, pitch, bo);
    if (!mt)
       return NULL;
@@ -926,8 +881,10 @@ miptree_create_for_planar_image(struct brw_context *brw,
                                      image->strides[index],
                                      tiling,
                                      MIPTREE_CREATE_NO_AUX);
-      if (mt == NULL)
+      if (mt == NULL) {
+         intel_miptree_release(&planar_mt);
          return NULL;
+      }
 
       mt->target = target;
 
@@ -990,7 +947,8 @@ create_ccs_buf_for_image(struct brw_context *brw,
       mt->aux_buf->clear_color_bo =
          brw_bo_alloc_tiled(brw->bufmgr, "clear_color_bo",
                             brw->isl_dev.ss.clear_color_state_size,
-                            I915_TILING_NONE, 0, BO_ALLOC_ZEROED);
+                            BRW_MEMZONE_OTHER, I915_TILING_NONE, 0,
+                            BO_ALLOC_ZEROED);
       if (!mt->aux_buf->clear_color_bo) {
          free(mt->aux_buf);
          mt->aux_buf = NULL;
@@ -1011,7 +969,7 @@ struct intel_mipmap_tree *
 intel_miptree_create_for_dri_image(struct brw_context *brw,
                                    __DRIimage *image, GLenum target,
                                    mesa_format format,
-                                   bool is_winsys_image)
+                                   bool allow_internal_aux)
 {
    uint32_t bo_tiling, bo_swizzle;
    brw_bo_get_tiling(image->bo, &bo_tiling, &bo_swizzle);
@@ -1054,7 +1012,7 @@ intel_miptree_create_for_dri_image(struct brw_context *brw,
     * other hand, have no resolve point so we can't have aux without a
     * modifier.
     */
-   if (!is_winsys_image)
+   if (!allow_internal_aux)
       mt_create_flags |= MIPTREE_CREATE_NO_AUX;
 
    /* If we have a modifier which specifies aux, don't create one yet */
@@ -1103,7 +1061,7 @@ intel_miptree_create_for_dri_image(struct brw_context *brw,
        * as part of the flush operation.
        */
       mt->supports_fast_clear =
-         is_winsys_image || mod_info->supports_clear_color;
+         allow_internal_aux || mod_info->supports_clear_color;
 
       /* We don't know the actual state of the surface when we get it but we
        * can make a pretty good guess based on the modifier.  What we do know
@@ -1565,6 +1523,7 @@ intel_miptree_copy_slice(struct brw_context *brw,
                          unsigned dst_level, unsigned dst_layer)
 
 {
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
    mesa_format format = src_mt->format;
    unsigned width = minify(src_mt->surf.phys_level0_sa.width,
                            src_level - src_mt->first_level);
@@ -1577,6 +1536,32 @@ intel_miptree_copy_slice(struct brw_context *brw,
    assert(_mesa_get_srgb_format_linear(src_mt->format) ==
           _mesa_get_srgb_format_linear(dst_mt->format));
 
+   DBG("validate blit mt %s %p %d,%d -> mt %s %p %d,%d (%dx%d)\n",
+       _mesa_get_format_name(src_mt->format),
+       src_mt, src_level, src_layer,
+       _mesa_get_format_name(dst_mt->format),
+       dst_mt, dst_level, dst_layer,
+       width, height);
+
+   if (devinfo->gen >= 6) {
+      /* On gen6 and above, we just use blorp.  It's faster than the blitter
+       * and can handle everything without software fallbacks.
+       */
+      brw_blorp_copy_miptrees(brw,
+                              src_mt, src_level, src_layer,
+                              dst_mt, dst_level, dst_layer,
+                              0, 0, 0, 0, width, height);
+
+      if (src_mt->stencil_mt) {
+         assert(dst_mt->stencil_mt);
+         brw_blorp_copy_miptrees(brw,
+                                 src_mt->stencil_mt, src_level, src_layer,
+                                 dst_mt->stencil_mt, dst_level, dst_layer,
+                                 0, 0, 0, 0, width, height);
+      }
+      return;
+   }
+
    if (dst_mt->compressed) {
       unsigned int i, j;
       _mesa_get_format_block_size(dst_mt->format, &i, &j);
@@ -1584,17 +1569,8 @@ intel_miptree_copy_slice(struct brw_context *brw,
       width = ALIGN_NPOT(width, i) / i;
    }
 
-   /* If it's a packed depth/stencil buffer with separate stencil, the blit
-    * below won't apply since we can't do the depth's Y tiling or the
-    * stencil's W tiling in the blitter.
-    */
-   if (src_mt->stencil_mt) {
-      intel_miptree_copy_slice_sw(brw,
-                                  src_mt, src_level, src_layer,
-                                  dst_mt, dst_level, dst_layer,
-                                  width, height);
-      return;
-   }
+   /* Gen4-5 doesn't support separate stencil */
+   assert(!src_mt->stencil_mt);
 
    uint32_t dst_x, dst_y, src_x, src_y;
    intel_miptree_get_image_offset(dst_mt, dst_level, dst_layer,
@@ -1701,8 +1677,8 @@ intel_alloc_aux_buffer(struct brw_context *brw,
     * trying to recalculate based on different format block sizes.
     */
    buf->bo = brw_bo_alloc_tiled(brw->bufmgr, "aux-miptree", size,
-                                I915_TILING_Y, aux_surf->row_pitch,
-                                alloc_flags);
+                                BRW_MEMZONE_OTHER, I915_TILING_Y,
+                                aux_surf->row_pitch, alloc_flags);
    if (!buf->bo) {
       free(buf);
       return NULL;
@@ -1796,7 +1772,7 @@ intel_miptree_alloc_aux(struct brw_context *brw,
    enum isl_aux_state initial_state;
    uint8_t memset_value;
    struct isl_surf aux_surf;
-   MAYBE_UNUSED bool aux_surf_ok;
+   MAYBE_UNUSED bool aux_surf_ok = false;
 
    switch (mt->aux_usage) {
    case ISL_AUX_USAGE_NONE:
@@ -1804,15 +1780,11 @@ intel_miptree_alloc_aux(struct brw_context *brw,
       aux_surf_ok = true;
       break;
    case ISL_AUX_USAGE_HIZ:
-      assert(!_mesa_is_format_color_format(mt->format));
-
       initial_state = ISL_AUX_STATE_AUX_INVALID;
+      memset_value = 0;
       aux_surf_ok = isl_surf_get_hiz_surf(&brw->isl_dev, &mt->surf, &aux_surf);
       break;
    case ISL_AUX_USAGE_MCS:
-      assert(_mesa_is_format_color_format(mt->format));
-      assert(brw->screen->devinfo.gen >= 7); /* MCS only used on Gen7+ */
-
       /* From the Ivy Bridge PRM, Vol 2 Part 1 p326:
        *
        *     When MCS buffer is enabled and bound to MSRT, it is required that
@@ -1831,8 +1803,6 @@ intel_miptree_alloc_aux(struct brw_context *brw,
       break;
    case ISL_AUX_USAGE_CCS_D:
    case ISL_AUX_USAGE_CCS_E:
-      assert(_mesa_is_format_color_format(mt->format));
-
       /* When CCS_E is used, we need to ensure that the CCS starts off in a
        * valid state.  From the Sky Lake PRM, "MCS Buffer for Render
        * Target(s)":
@@ -2451,11 +2421,13 @@ intel_miptree_finish_write(struct brw_context *brw,
                            uint32_t start_layer, uint32_t num_layers,
                            enum isl_aux_usage aux_usage)
 {
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
    num_layers = miptree_layer_range_length(mt, level, start_layer, num_layers);
 
    switch (mt->aux_usage) {
    case ISL_AUX_USAGE_NONE:
-      /* Nothing to do */
+      if (mt->format == MESA_FORMAT_S_UINT8 && devinfo->gen <= 7)
+         mt->r8stencil_needs_update = true;
       break;
 
    case ISL_AUX_USAGE_MCS:
@@ -2570,8 +2542,19 @@ can_texture_with_ccs(struct brw_context *brw,
 enum isl_aux_usage
 intel_miptree_texture_aux_usage(struct brw_context *brw,
                                 struct intel_mipmap_tree *mt,
-                                enum isl_format view_format)
+                                enum isl_format view_format,
+                                enum gen9_astc5x5_wa_tex_type astc5x5_wa_bits)
 {
+   assert(brw->screen->devinfo.gen == 9 || astc5x5_wa_bits == 0);
+
+   /* On gen9, ASTC 5x5 textures cannot live in the sampler cache along side
+    * CCS or HiZ compressed textures.  See gen9_apply_astc5x5_wa_flush() for
+    * details.
+    */
+   if ((astc5x5_wa_bits & GEN9_ASTC5X5_WA_TEX_TYPE_ASTC5x5) &&
+       mt->aux_usage != ISL_AUX_USAGE_MCS)
+      return ISL_AUX_USAGE_NONE;
+
    switch (mt->aux_usage) {
    case ISL_AUX_USAGE_HIZ:
       if (intel_miptree_sample_with_hiz(brw, mt))
@@ -2629,10 +2612,12 @@ intel_miptree_prepare_texture(struct brw_context *brw,
                               struct intel_mipmap_tree *mt,
                               enum isl_format view_format,
                               uint32_t start_level, uint32_t num_levels,
-                              uint32_t start_layer, uint32_t num_layers)
+                              uint32_t start_layer, uint32_t num_layers,
+                              enum gen9_astc5x5_wa_tex_type astc5x5_wa_bits)
 {
    enum isl_aux_usage aux_usage =
-      intel_miptree_texture_aux_usage(brw, mt, view_format);
+      intel_miptree_texture_aux_usage(brw, mt, view_format, astc5x5_wa_bits);
+
    bool clear_supported = aux_usage != ISL_AUX_USAGE_NONE;
 
    /* Clear color is specified as ints or floats and the conversion is done by
@@ -2943,7 +2928,7 @@ intel_update_r8stencil(struct brw_context *brw,
    assert(devinfo->gen >= 7);
    struct intel_mipmap_tree *src =
       mt->format == MESA_FORMAT_S_UINT8 ? mt : mt->stencil_mt;
-   if (!src || devinfo->gen >= 8 || !src->r8stencil_needs_update)
+   if (!src || devinfo->gen >= 8)
       return;
 
    assert(src->surf.size > 0);
@@ -2967,6 +2952,9 @@ intel_update_r8stencil(struct brw_context *brw,
       assert(mt->r8stencil_mt);
    }
 
+   if (src->r8stencil_needs_update == false)
+      return;
+
    struct intel_mipmap_tree *dst = mt->r8stencil_mt;
 
    for (int level = src->first_level; level <= src->last_level; level++) {
@@ -3040,6 +3028,9 @@ intel_miptree_map_gtt(struct brw_context *brw,
    y /= bh;
    x /= bw;
 
+   intel_miptree_access_raw(brw, mt, level, slice,
+                            map->mode & GL_MAP_WRITE_BIT);
+
    base = intel_miptree_map_raw(brw, mt, map->mode);
 
    if (base == NULL)
@@ -3074,16 +3065,23 @@ intel_miptree_unmap_blit(struct brw_context *brw,
                         unsigned int level,
                         unsigned int slice)
 {
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
    struct gl_context *ctx = &brw->ctx;
 
    intel_miptree_unmap_raw(map->linear_mt);
 
    if (map->mode & GL_MAP_WRITE_BIT) {
-      bool ok = intel_miptree_copy(brw,
-                                   map->linear_mt, 0, 0, 0, 0,
-                                   mt, level, slice, map->x, map->y,
-                                   map->w, map->h);
-      WARN_ONCE(!ok, "Failed to blit from linear temporary mapping");
+      if (devinfo->gen >= 6) {
+         brw_blorp_copy_miptrees(brw, map->linear_mt, 0, 0,
+                                 mt, level, slice,
+                                 0, 0, map->x, map->y, map->w, map->h);
+      } else {
+         bool ok = intel_miptree_copy(brw,
+                                      map->linear_mt, 0, 0, 0, 0,
+                                      mt, level, slice, map->x, map->y,
+                                      map->w, map->h);
+         WARN_ONCE(!ok, "Failed to blit from linear temporary mapping");
+      }
    }
 
    intel_miptree_release(&map->linear_mt);
@@ -3095,12 +3093,13 @@ intel_miptree_map_blit(struct brw_context *brw,
                       struct intel_miptree_map *map,
                       unsigned int level, unsigned int slice)
 {
-   map->linear_mt = intel_miptree_create(brw, GL_TEXTURE_2D, mt->format,
-                                         /* first_level */ 0,
-                                         /* last_level */ 0,
-                                         map->w, map->h, 1,
-                                         /* samples */ 1,
-                                         MIPTREE_CREATE_LINEAR);
+   const struct gen_device_info *devinfo = &brw->screen->devinfo;
+   map->linear_mt = make_surface(brw, GL_TEXTURE_2D, mt->format,
+                                 0, 0, map->w, map->h, 1, 1,
+                                 ISL_TILING_LINEAR_BIT,
+                                 ISL_SURF_USAGE_RENDER_TARGET_BIT |
+                                 ISL_SURF_USAGE_TEXTURE_BIT,
+                                 0, 0, NULL);
 
    if (!map->linear_mt) {
       fprintf(stderr, "Failed to allocate blit temporary\n");
@@ -3114,12 +3113,18 @@ intel_miptree_map_blit(struct brw_context *brw,
     * temporary buffer back out.
     */
    if (!(map->mode & GL_MAP_INVALIDATE_RANGE_BIT)) {
-      if (!intel_miptree_copy(brw,
-                              mt, level, slice, map->x, map->y,
-                              map->linear_mt, 0, 0, 0, 0,
-                              map->w, map->h)) {
-         fprintf(stderr, "Failed to blit\n");
-         goto fail;
+      if (devinfo->gen >= 6) {
+         brw_blorp_copy_miptrees(brw, mt, level, slice,
+                                 map->linear_mt, 0, 0,
+                                 map->x, map->y, 0, 0, map->w, map->h);
+      } else {
+         if (!intel_miptree_copy(brw,
+                                 mt, level, slice, map->x, map->y,
+                                 map->linear_mt, 0, 0, 0, 0,
+                                 map->w, map->h)) {
+            fprintf(stderr, "Failed to blit\n");
+            goto fail;
+         }
       }
    }
 
@@ -3164,6 +3169,8 @@ intel_miptree_map_movntdqa(struct brw_context *brw,
    assert(map->mode & GL_MAP_READ_BIT);
    assert(!(map->mode & GL_MAP_WRITE_BIT));
 
+   intel_miptree_access_raw(brw, mt, level, slice, false);
+
    DBG("%s: %d,%d %dx%d from mt %p (%s) %d,%d = %p/%d\n", __func__,
        map->x, map->y, map->w, map->h,
        mt, _mesa_get_format_name(mt->format),
@@ -3258,6 +3265,9 @@ intel_miptree_map_s8(struct brw_context *brw,
    if (!map->buffer)
       return;
 
+   intel_miptree_access_raw(brw, mt, level, slice,
+                            map->mode & GL_MAP_WRITE_BIT);
+
    /* One of either READ_BIT or WRITE_BIT or both is set.  READ_BIT implies no
     * INVALIDATE_RANGE_BIT.  WRITE_BIT needs the original values read in unless
     * invalidate is set, since we'll be writing the whole rectangle from our
@@ -3319,7 +3329,7 @@ intel_miptree_unmap_etc(struct brw_context *brw,
    else
       _mesa_unpack_etc2_format(dst, mt->surf.row_pitch,
                                map->ptr, map->stride,
-                               map->w, map->h, mt->etc_format);
+                              map->w, map->h, mt->etc_format, true);
 
    intel_miptree_unmap_raw(mt);
    free(map->buffer);
@@ -3340,6 +3350,8 @@ intel_miptree_map_etc(struct brw_context *brw,
    assert(map->mode & GL_MAP_WRITE_BIT);
    assert(map->mode & GL_MAP_INVALIDATE_RANGE_BIT);
 
+   intel_miptree_access_raw(brw, mt, level, slice, true);
+
    map->stride = _mesa_format_row_stride(mt->etc_format, map->w);
    map->buffer = malloc(_mesa_format_image_size(mt->etc_format,
                                                 map->w, map->h, 1));
@@ -3433,6 +3445,11 @@ intel_miptree_map_depthstencil(struct brw_context *brw,
    if (!map->buffer)
       return;
 
+   intel_miptree_access_raw(brw, z_mt, level, slice,
+                            map->mode & GL_MAP_WRITE_BIT);
+   intel_miptree_access_raw(brw, s_mt, level, slice,
+                            map->mode & GL_MAP_WRITE_BIT);
+
    /* One of either READ_BIT or WRITE_BIT or both is set.  READ_BIT implies no
     * INVALIDATE_RANGE_BIT.  WRITE_BIT needs the original values read in unless
     * invalidate is set, since we'll be writing the whole rectangle from our
@@ -3538,21 +3555,18 @@ intel_miptree_release_map(struct intel_mipmap_tree *mt,
 
 static bool
 can_blit_slice(struct intel_mipmap_tree *mt,
-               unsigned int level, unsigned int slice)
+               const struct intel_miptree_map *map)
 {
    /* See intel_miptree_blit() for details on the 32k pitch limit. */
-   if (mt->surf.row_pitch >= 32768)
-      return false;
-
-   return true;
+   const unsigned src_blt_pitch = intel_miptree_blt_pitch(mt);
+   const unsigned dst_blt_pitch = ALIGN(map->w * mt->cpp, 64);
+   return src_blt_pitch < 32768 && dst_blt_pitch < 32768;
 }
 
 static bool
 use_intel_mipree_map_blit(struct brw_context *brw,
                           struct intel_mipmap_tree *mt,
-                          GLbitfield mode,
-                          unsigned int level,
-                          unsigned int slice)
+                          const struct intel_miptree_map *map)
 {
    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 
@@ -3560,19 +3574,19 @@ use_intel_mipree_map_blit(struct brw_context *brw,
       /* It's probably not worth swapping to the blit ring because of
        * all the overhead involved.
        */
-       !(mode & GL_MAP_WRITE_BIT) &&
+       !(map->mode & GL_MAP_WRITE_BIT) &&
        !mt->compressed &&
        (mt->surf.tiling == ISL_TILING_X ||
         /* Prior to Sandybridge, the blitter can't handle Y tiling */
         (devinfo->gen >= 6 && mt->surf.tiling == ISL_TILING_Y0) ||
         /* Fast copy blit on skl+ supports all tiling formats. */
         devinfo->gen >= 9) &&
-       can_blit_slice(mt, level, slice))
+       can_blit_slice(mt, map))
       return true;
 
    if (mt->surf.tiling != ISL_TILING_LINEAR &&
        mt->bo->size >= brw->max_gtt_map_object_size) {
-      assert(can_blit_slice(mt, level, slice));
+      assert(can_blit_slice(mt, map));
       return true;
    }
 
@@ -3614,9 +3628,6 @@ intel_miptree_map(struct brw_context *brw,
       return;
    }
 
-   intel_miptree_access_raw(brw, mt, level, slice,
-                            map->mode & GL_MAP_WRITE_BIT);
-
    if (mt->format == MESA_FORMAT_S_UINT8) {
       intel_miptree_map_s8(brw, mt, map, level, slice);
    } else if (mt->etc_format != MESA_FORMAT_NONE &&
@@ -3624,7 +3635,7 @@ intel_miptree_map(struct brw_context *brw,
       intel_miptree_map_etc(brw, mt, map, level, slice);
    } else if (mt->stencil_mt && !(mode & BRW_MAP_DIRECT_BIT)) {
       intel_miptree_map_depthstencil(brw, mt, map, level, slice);
-   } else if (use_intel_mipree_map_blit(brw, mt, mode, level, slice)) {
+   } else if (use_intel_mipree_map_blit(brw, mt, map)) {
       intel_miptree_map_blit(brw, mt, map, level, slice);
 #if defined(USE_SSE41)
    } else if (!(mode & GL_MAP_WRITE_BIT) &&