anv: emit DrawID if needed
[mesa.git] / src / intel / vulkan / genX_cmd_buffer.c
index 4649dae1a11b33377f28a9cd1b9bca9c4e6f8b17..303ced982dced7704d90139c8d04bed428db8993 100644 (file)
@@ -55,8 +55,6 @@ genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
 {
    struct anv_device *device = cmd_buffer->device;
 
-/* XXX: Do we need this on more than just BDW? */
-#if (GEN_GEN >= 8)
    /* Emit a render target cache flush.
     *
     * This isn't documented anywhere in the PRM.  However, it seems to be
@@ -65,9 +63,10 @@ genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
     * clear depth, reset state base address, and then go render stuff.
     */
    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
+      pc.DCFlushEnable = true;
       pc.RenderTargetCacheFlushEnable = true;
+      pc.CommandStreamerStallEnable = true;
    }
-#endif
 
    anv_batch_emit(&cmd_buffer->batch, GENX(STATE_BASE_ADDRESS), sba) {
       sba.GeneralStateBaseAddress = (struct anv_address) { NULL, 0 };
@@ -148,6 +147,8 @@ genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
     */
    anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
       pc.TextureCacheInvalidationEnable = true;
+      pc.ConstantCacheInvalidationEnable = true;
+      pc.StateCacheInvalidationEnable = true;
    }
 }
 
@@ -242,6 +243,21 @@ color_attachment_compute_aux_usage(struct anv_device *device,
           render_area.extent.height != iview->extent.height)
          att_state->fast_clear = false;
 
+      if (GEN_GEN <= 7) {
+         /* On gen7, we can't do multi-LOD or multi-layer fast-clears.  We
+          * technically can, but it comes with crazy restrictions that we
+          * don't want to deal with now.
+          */
+         if (iview->isl.base_level > 0 ||
+             iview->isl.base_array_layer > 0 ||
+             iview->isl.array_len > 1)
+            att_state->fast_clear = false;
+      }
+
+      /* On Broadwell and earlier, we can only handle 0/1 clear colors */
+      if (GEN_GEN <= 8 && !att_state->clear_color_is_zero_one)
+         att_state->fast_clear = false;
+
       if (att_state->fast_clear) {
          memcpy(fast_clear_color->u32, att_state->clear_value.color.uint32,
                 sizeof(fast_clear_color->u32));
@@ -256,18 +272,26 @@ color_attachment_compute_aux_usage(struct anv_device *device,
       att_state->input_aux_usage = ISL_AUX_USAGE_CCS_E;
    } else if (att_state->fast_clear) {
       att_state->aux_usage = ISL_AUX_USAGE_CCS_D;
-      /* From the Sky Lake PRM, RENDER_SURFACE_STATE::AuxiliarySurfaceMode:
-       *
-       *    "If Number of Multisamples is MULTISAMPLECOUNT_1, AUX_CCS_D
-       *    setting is only allowed if Surface Format supported for Fast
-       *    Clear. In addition, if the surface is bound to the sampling
-       *    engine, Surface Format must be supported for Render Target
-       *    Compression for surfaces bound to the sampling engine."
-       *
-       * In other words, we can't sample from a fast-cleared image if it
-       * doesn't also support color compression.
-       */
-      att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
+      if (GEN_GEN >= 9) {
+         /* From the Sky Lake PRM, RENDER_SURFACE_STATE::AuxiliarySurfaceMode:
+          *
+          *    "If Number of Multisamples is MULTISAMPLECOUNT_1, AUX_CCS_D
+          *    setting is only allowed if Surface Format supported for Fast
+          *    Clear. In addition, if the surface is bound to the sampling
+          *    engine, Surface Format must be supported for Render Target
+          *    Compression for surfaces bound to the sampling engine."
+          *
+          * In other words, we can't sample from a fast-cleared image if it
+          * doesn't also support color compression.
+          */
+         att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
+      } else if (GEN_GEN == 8) {
+         /* Broadwell can sample from fast-cleared images */
+         att_state->input_aux_usage = ISL_AUX_USAGE_CCS_D;
+      } else {
+         /* Ivy Bridge and Haswell cannot */
+         att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
+      }
    } else {
       att_state->aux_usage = ISL_AUX_USAGE_NONE;
       att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
@@ -280,13 +304,93 @@ need_input_attachment_state(const struct anv_render_pass_attachment *att)
    if (!(att->usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT))
       return false;
 
-   /* We only allocate input attachment states for color and depth surfaces.
-    * Stencil doesn't allow compression so we can just use the texture surface
-    * state from the view
+   /* We only allocate input attachment states for color surfaces. Compression
+    * is not yet enabled for depth textures and stencil doesn't allow
+    * compression so we can just use the texture surface state from the view.
+    */
+   return vk_format_is_color(att->format);
+}
+
+static enum isl_aux_usage
+layout_to_hiz_usage(VkImageLayout layout, uint8_t samples)
+{
+   switch (layout) {
+   case VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL:
+      return ISL_AUX_USAGE_HIZ;
+   case VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL:
+   case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
+      if (anv_can_sample_with_hiz(GEN_GEN, samples))
+         return ISL_AUX_USAGE_HIZ;
+      /* Fall-through */
+   case VK_IMAGE_LAYOUT_GENERAL:
+      /* This buffer could be used as a source or destination in a transfer
+       * operation. Transfer operations current don't perform HiZ-enabled reads
+       * and writes.
+       */
+   default:
+      return ISL_AUX_USAGE_NONE;
+   }
+}
+
+/* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
+ * the initial layout is undefined, the HiZ buffer and depth buffer will
+ * represent the same data at the end of this operation.
+ */
+static void
+transition_depth_buffer(struct anv_cmd_buffer *cmd_buffer,
+                        const struct anv_image *image,
+                        VkImageLayout initial_layout,
+                        VkImageLayout final_layout)
+{
+   assert(image);
+
+   if (image->aux_usage != ISL_AUX_USAGE_HIZ || final_layout == initial_layout)
+      return;
+
+   const bool hiz_enabled = layout_to_hiz_usage(initial_layout, image->samples) ==
+                            ISL_AUX_USAGE_HIZ;
+   const bool enable_hiz = layout_to_hiz_usage(final_layout, image->samples) ==
+                           ISL_AUX_USAGE_HIZ;
+
+   enum blorp_hiz_op hiz_op;
+   if (initial_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
+      /* We've already initialized the aux HiZ buffer at BindImageMemory time,
+       * so there's no need to perform a HIZ resolve or clear to avoid GPU hangs.
+       * This initial layout indicates that the user doesn't care about the data
+       * that's currently in the buffer, so resolves are not necessary except
+       * for the special case noted below.
+       */
+      hiz_op = BLORP_HIZ_OP_NONE;
+   } else if (hiz_enabled && !enable_hiz) {
+      hiz_op = BLORP_HIZ_OP_DEPTH_RESOLVE;
+   } else if (!hiz_enabled && enable_hiz) {
+      hiz_op = BLORP_HIZ_OP_HIZ_RESOLVE;
+   } else {
+      assert(hiz_enabled == enable_hiz);
+      /* If the same buffer will be used, no resolves are necessary except for
+       * the special case noted below.
+       */
+      hiz_op = BLORP_HIZ_OP_NONE;
+   }
+
+   if (hiz_op != BLORP_HIZ_OP_NONE)
+      anv_gen8_hiz_op_resolve(cmd_buffer, image, hiz_op);
+
+   /* Images that have sampling with HiZ enabled cause all shader sampling to
+    * load data with the HiZ buffer. Therefore, in the case of transitioning to
+    * the general layout - which currently routes all writes to the depth
+    * buffer - we must ensure that the HiZ buffer remains consistent with the
+    * depth buffer by performing an additional HIZ resolve if the operation
+    * required by this transition was not already a HiZ resolve.
     */
-   return vk_format_is_color(att->format) || vk_format_has_depth(att->format);
+   if (final_layout == VK_IMAGE_LAYOUT_GENERAL &&
+       anv_can_sample_with_hiz(GEN_GEN, image->samples) &&
+       hiz_op != BLORP_HIZ_OP_HIZ_RESOLVE) {
+      anv_gen8_hiz_op_resolve(cmd_buffer, image, BLORP_HIZ_OP_HIZ_RESOLVE);
+   }
 }
 
+
 /**
  * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
  */
@@ -407,6 +511,7 @@ genX(cmd_buffer_setup_attachments)(struct anv_cmd_buffer *cmd_buffer,
             }
          }
 
+         state->attachments[i].current_layout = att->initial_layout;
          state->attachments[i].pending_clear_aspects = clear_aspects;
          if (clear_aspects)
             state->attachments[i].clear_value = begin->pClearValues[i];
@@ -423,6 +528,7 @@ genX(cmd_buffer_setup_attachments)(struct anv_cmd_buffer *cmd_buffer,
 
             struct isl_view view = iview->isl;
             view.usage |= ISL_SURF_USAGE_RENDER_TARGET_BIT;
+            view.swizzle = anv_swizzle_for_render(view.swizzle);
             isl_surf_fill_state(isl_dev,
                                 state->attachments[i].color_rt_state.map,
                                 .surf = &iview->image->color_surface.isl,
@@ -436,23 +542,21 @@ genX(cmd_buffer_setup_attachments)(struct anv_cmd_buffer *cmd_buffer,
                                   state->attachments[i].aux_usage,
                                   state->attachments[i].color_rt_state);
          } else {
-            state->attachments[i].aux_usage = ISL_AUX_USAGE_NONE;
+            if (iview->image->aux_usage == ISL_AUX_USAGE_HIZ) {
+               state->attachments[i].aux_usage =
+                  layout_to_hiz_usage(att->initial_layout, iview->image->samples);
+            } else {
+               state->attachments[i].aux_usage = ISL_AUX_USAGE_NONE;
+            }
             state->attachments[i].input_aux_usage = ISL_AUX_USAGE_NONE;
          }
 
          if (need_input_attachment_state(&pass->attachments[i])) {
-            const struct isl_surf *surf;
-            if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
-               surf = &iview->image->color_surface.isl;
-            } else {
-               surf = &iview->image->depth_surface.isl;
-            }
-
             struct isl_view view = iview->isl;
             view.usage |= ISL_SURF_USAGE_TEXTURE_BIT;
             isl_surf_fill_state(isl_dev,
                                 state->attachments[i].input_att_state.map,
-                                .surf = surf,
+                                .surf = &iview->image->color_surface.isl,
                                 .view = &view,
                                 .aux_surf = &iview->image->aux_surface.isl,
                                 .aux_usage = state->attachments[i].input_aux_usage,
@@ -674,7 +778,7 @@ genX(cmd_buffer_config_l3)(struct anv_cmd_buffer *cmd_buffer,
    assert(!urb_low_bw || cfg->n[GEN_L3P_URB] == cfg->n[GEN_L3P_SLM]);
 
    /* Minimum number of ways that can be allocated to the URB. */
-   const unsigned n0_urb = (devinfo->is_baytrail ? 32 : 0);
+   MAYBE_UNUSED const unsigned n0_urb = devinfo->is_baytrail ? 32 : 0;
    assert(cfg->n[GEN_L3P_URB] >= n0_urb);
 
    uint32_t l3sqcr1, l3cr2, l3cr3;
@@ -843,6 +947,13 @@ void genX(CmdPipelineBarrier)(
    for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
       src_flags |= pImageMemoryBarriers[i].srcAccessMask;
       dst_flags |= pImageMemoryBarriers[i].dstAccessMask;
+      ANV_FROM_HANDLE(anv_image, image, pImageMemoryBarriers[i].image);
+      if (pImageMemoryBarriers[i].subresourceRange.aspectMask &
+          VK_IMAGE_ASPECT_DEPTH_BIT) {
+         transition_depth_buffer(cmd_buffer, image,
+                                 pImageMemoryBarriers[i].oldLayout,
+                                 pImageMemoryBarriers[i].newLayout);
+      }
    }
 
    enum anv_pipe_bits pipe_bits = 0;
@@ -1068,9 +1179,9 @@ emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
 
       case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
          assert(stage == MESA_SHADER_FRAGMENT);
-         if (desc->image_view->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) {
-            /* For stencil input attachments, we treat it like any old texture
-             * that a user may have bound.
+         if (desc->image_view->aspect_mask != VK_IMAGE_ASPECT_COLOR_BIT) {
+            /* For depth and stencil input attachments, we treat it like any
+             * old texture that a user may have bound.
              */
             surface_state = desc->image_view->sampler_surface_state;
             assert(surface_state.alloc_size);
@@ -1078,9 +1189,9 @@ emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
                                   desc->image_view->image->aux_usage,
                                   surface_state);
          } else {
-            /* For depth and color input attachments, we create the surface
-             * state at vkBeginRenderPass time so that we can include aux
-             * and clear color information.
+            /* For color input attachments, we create the surface state at
+             * vkBeginRenderPass time so that we can include aux and clear
+             * color information.
              */
             assert(binding->input_attachment_index < subpass->input_count);
             const unsigned subpass_att = binding->input_attachment_index;
@@ -1476,29 +1587,37 @@ genX(cmd_buffer_flush_state)(struct anv_cmd_buffer *cmd_buffer)
 }
 
 static void
-emit_base_vertex_instance_bo(struct anv_cmd_buffer *cmd_buffer,
-                             struct anv_bo *bo, uint32_t offset)
+emit_vertex_bo(struct anv_cmd_buffer *cmd_buffer,
+               struct anv_bo *bo, uint32_t offset,
+               uint32_t size, uint32_t index)
 {
    uint32_t *p = anv_batch_emitn(&cmd_buffer->batch, 5,
                                  GENX(3DSTATE_VERTEX_BUFFERS));
 
    GENX(VERTEX_BUFFER_STATE_pack)(&cmd_buffer->batch, p + 1,
       &(struct GENX(VERTEX_BUFFER_STATE)) {
-         .VertexBufferIndex = 32, /* Reserved for this */
+         .VertexBufferIndex = index,
          .AddressModifyEnable = true,
          .BufferPitch = 0,
 #if (GEN_GEN >= 8)
          .MemoryObjectControlState = GENX(MOCS),
          .BufferStartingAddress = { bo, offset },
-         .BufferSize = 8
+         .BufferSize = size
 #else
          .VertexBufferMemoryObjectControlState = GENX(MOCS),
          .BufferStartingAddress = { bo, offset },
-         .EndAddress = { bo, offset + 8 },
+         .EndAddress = { bo, offset + size },
 #endif
       });
 }
 
+static void
+emit_base_vertex_instance_bo(struct anv_cmd_buffer *cmd_buffer,
+                             struct anv_bo *bo, uint32_t offset)
+{
+   emit_vertex_bo(cmd_buffer, bo, offset, 8, ANV_SVGS_VB_INDEX);
+}
+
 static void
 emit_base_vertex_instance(struct anv_cmd_buffer *cmd_buffer,
                           uint32_t base_vertex, uint32_t base_instance)
@@ -1516,6 +1635,22 @@ emit_base_vertex_instance(struct anv_cmd_buffer *cmd_buffer,
       &cmd_buffer->device->dynamic_state_block_pool.bo, id_state.offset);
 }
 
+static void
+emit_draw_index(struct anv_cmd_buffer *cmd_buffer, uint32_t draw_index)
+{
+   struct anv_state state =
+      anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, 4, 4);
+
+   ((uint32_t *)state.map)[0] = draw_index;
+
+   if (!cmd_buffer->device->info.has_llc)
+      anv_state_clflush(state);
+
+   emit_vertex_bo(cmd_buffer,
+                  &cmd_buffer->device->dynamic_state_block_pool.bo,
+                  state.offset, 4, ANV_DRAWID_VB_INDEX);
+}
+
 void genX(CmdDraw)(
     VkCommandBuffer                             commandBuffer,
     uint32_t                                    vertexCount,
@@ -1531,6 +1666,8 @@ void genX(CmdDraw)(
 
    if (vs_prog_data->uses_basevertex || vs_prog_data->uses_baseinstance)
       emit_base_vertex_instance(cmd_buffer, firstVertex, firstInstance);
+   if (vs_prog_data->uses_drawid)
+      emit_draw_index(cmd_buffer, 0);
 
    anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
       prim.VertexAccessType         = SEQUENTIAL;
@@ -1559,6 +1696,8 @@ void genX(CmdDrawIndexed)(
 
    if (vs_prog_data->uses_basevertex || vs_prog_data->uses_baseinstance)
       emit_base_vertex_instance(cmd_buffer, vertexOffset, firstInstance);
+   if (vs_prog_data->uses_drawid)
+      emit_draw_index(cmd_buffer, 0);
 
    anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
       prim.VertexAccessType         = RANDOM;
@@ -1597,6 +1736,8 @@ void genX(CmdDrawIndirect)(
 
    if (vs_prog_data->uses_basevertex || vs_prog_data->uses_baseinstance)
       emit_base_vertex_instance_bo(cmd_buffer, bo, bo_offset + 8);
+   if (vs_prog_data->uses_drawid)
+      emit_draw_index(cmd_buffer, 0);
 
    emit_lrm(&cmd_buffer->batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
    emit_lrm(&cmd_buffer->batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
@@ -1630,6 +1771,8 @@ void genX(CmdDrawIndexedIndirect)(
    /* TODO: We need to stomp base vertex to 0 somehow */
    if (vs_prog_data->uses_basevertex || vs_prog_data->uses_baseinstance)
       emit_base_vertex_instance_bo(cmd_buffer, bo, bo_offset + 12);
+   if (vs_prog_data->uses_drawid)
+      emit_draw_index(cmd_buffer, 0);
 
    emit_lrm(&cmd_buffer->batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
    emit_lrm(&cmd_buffer->batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
@@ -2010,6 +2153,51 @@ genX(cmd_buffer_emit_gen7_depth_flush)(struct anv_cmd_buffer *cmd_buffer)
    }
 }
 
+static uint32_t
+depth_stencil_surface_type(enum isl_surf_dim dim)
+{
+   switch (dim) {
+   case ISL_SURF_DIM_1D:
+      if (GEN_GEN >= 9) {
+         /* From the Sky Lake PRM, 3DSTATAE_DEPTH_BUFFER::SurfaceType
+          *
+          *    Programming Notes:
+          *    The Surface Type of the depth buffer must be the same as the
+          *    Surface Type of the render target(s) (defined in
+          *    SURFACE_STATE), unless either the depth buffer or render
+          *    targets are SURFTYPE_NULL (see exception below for SKL).  1D
+          *    surface type not allowed for depth surface and stencil surface.
+          *
+          *    Workaround:
+          *    If depth/stencil is enabled with 1D render target,
+          *    depth/stencil surface type needs to be set to 2D surface type
+          *    and height set to 1. Depth will use (legacy) TileY and stencil
+          *    will use TileW. For this case only, the Surface Type of the
+          *    depth buffer can be 2D while the Surface Type of the render
+          *    target(s) are 1D, representing an exception to a programming
+          *    note above.
+          */
+         return SURFTYPE_2D;
+      } else {
+         return SURFTYPE_1D;
+      }
+   case ISL_SURF_DIM_2D:
+      return SURFTYPE_2D;
+   case ISL_SURF_DIM_3D:
+      if (GEN_GEN >= 9) {
+         /* The Sky Lake docs list the value for 3D as "Reserved".  However,
+          * they have the exact same layout as 2D arrays on gen9+, so we can
+          * just use 2D here.
+          */
+         return SURFTYPE_2D;
+      } else {
+         return SURFTYPE_3D;
+      }
+   default:
+      unreachable("Invalid surface dimension");
+   }
+}
+
 static void
 cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
 {
@@ -2019,7 +2207,9 @@ cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
       anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
    const struct anv_image *image = iview ? iview->image : NULL;
    const bool has_depth = image && (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT);
-   const bool has_hiz = image != NULL && anv_image_has_hiz(image);
+   const uint32_t ds = cmd_buffer->state.subpass->depth_stencil_attachment;
+   const bool has_hiz = image != NULL &&
+      cmd_buffer->state.attachments[ds].aux_usage == ISL_AUX_USAGE_HIZ;
    const bool has_stencil =
       image && (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT);
 
@@ -2031,15 +2221,11 @@ cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
    /* Emit 3DSTATE_DEPTH_BUFFER */
    if (has_depth) {
       anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_DEPTH_BUFFER), db) {
-         db.SurfaceType                   = SURFTYPE_2D;
+         db.SurfaceType                   =
+            depth_stencil_surface_type(image->depth_surface.isl.dim);
          db.DepthWriteEnable              = true;
          db.StencilWriteEnable            = has_stencil;
-
-         if (cmd_buffer->state.pass->subpass_count == 1) {
-            db.HierarchicalDepthBufferEnable = has_hiz;
-         } else {
-            anv_finishme("Multiple-subpass HiZ not implemented");
-         }
+         db.HierarchicalDepthBufferEnable = has_hiz;
 
          db.SurfaceFormat = isl_surf_get_depth_format(&device->isl_dev,
                                                       &image->depth_surface.isl);
@@ -2054,14 +2240,17 @@ cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
          db.Height               = image->extent.height - 1;
          db.Width                = image->extent.width - 1;
          db.LOD                  = iview->isl.base_level;
-         db.Depth                = image->array_size - 1; /* FIXME: 3-D */
          db.MinimumArrayElement  = iview->isl.base_array_layer;
 
+         assert(image->depth_surface.isl.dim != ISL_SURF_DIM_3D);
+         db.Depth =
+         db.RenderTargetViewExtent =
+            iview->isl.array_len - iview->isl.base_array_layer - 1;
+
 #if GEN_GEN >= 8
          db.SurfaceQPitch =
             isl_surf_get_array_pitch_el_rows(&image->depth_surface.isl) >> 2;
 #endif
-         db.RenderTargetViewExtent = 1 - 1;
       }
    } else {
       /* Even when no depth buffer is present, the hardware requires that
@@ -2083,10 +2272,15 @@ cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
        * be combined with a stencil buffer so we use D32_FLOAT instead.
        */
       anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_DEPTH_BUFFER), db) {
-         db.SurfaceType          = SURFTYPE_2D;
+         if (has_stencil) {
+            db.SurfaceType       =
+               depth_stencil_surface_type(image->stencil_surface.isl.dim);
+         } else {
+            db.SurfaceType       = SURFTYPE_2D;
+         }
          db.SurfaceFormat        = D32_FLOAT;
-         db.Width                = fb->width - 1;
-         db.Height               = fb->height - 1;
+         db.Width                = MAX2(fb->width, 1) - 1;
+         db.Height               = MAX2(fb->height, 1) - 1;
          db.StencilWriteEnable   = has_stencil;
       }
    }
@@ -2107,11 +2301,14 @@ cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
           *    - SURFTYPE_1D: distance in pixels between array slices
           *    - SURFTYPE_2D/CUBE: distance in rows between array slices
           *    - SURFTYPE_3D: distance in rows between R - slices
+          *
+          * Unfortunately, the docs aren't 100% accurate here.  They fail to
+          * mention that the 1-D rule only applies to linear 1-D images.
+          * Since depth and HiZ buffers are always tiled, they are treated as
+          * 2-D images.  Prior to Sky Lake, this field is always in rows.
           */
          hdb.SurfaceQPitch =
-            image->aux_surface.isl.dim == ISL_SURF_DIM_1D ?
-               isl_surf_get_array_pitch_el(&image->aux_surface.isl) >> 2 :
-               isl_surf_get_array_pitch_el_rows(&image->aux_surface.isl) >> 2;
+            isl_surf_get_array_pitch_sa_rows(&image->aux_surface.isl) >> 2;
 #endif
       }
    } else {
@@ -2155,10 +2352,7 @@ cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
    anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CLEAR_PARAMS), cp) {
       if (has_hiz) {
          cp.DepthClearValueValid = true;
-         const uint32_t ds =
-            cmd_buffer->state.subpass->depth_stencil_attachment;
-         cp.DepthClearValue =
-            cmd_buffer->state.attachments[ds].clear_value.depthStencil.depth;
+         cp.DepthClearValue = ANV_HZ_FC_VAL;
       }
    }
 }
@@ -2171,9 +2365,22 @@ genX(cmd_buffer_set_subpass)(struct anv_cmd_buffer *cmd_buffer,
 
    cmd_buffer->state.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
 
+   const struct anv_image_view *iview =
+      anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
+
+   if (iview && iview->image->aux_usage == ISL_AUX_USAGE_HIZ) {
+      const uint32_t ds = subpass->depth_stencil_attachment;
+      transition_depth_buffer(cmd_buffer, iview->image,
+                              cmd_buffer->state.attachments[ds].current_layout,
+                              cmd_buffer->state.subpass->depth_stencil_layout);
+      cmd_buffer->state.attachments[ds].current_layout =
+         cmd_buffer->state.subpass->depth_stencil_layout;
+      cmd_buffer->state.attachments[ds].aux_usage =
+         layout_to_hiz_usage(cmd_buffer->state.subpass->depth_stencil_layout,
+                             iview->image->samples);
+   }
+
    cmd_buffer_emit_depth_stencil(cmd_buffer);
-   genX(cmd_buffer_emit_hz_op)(cmd_buffer, BLORP_HIZ_OP_HIZ_RESOLVE);
-   genX(cmd_buffer_emit_hz_op)(cmd_buffer, BLORP_HIZ_OP_DEPTH_CLEAR);
 
    anv_cmd_buffer_clear_subpass(cmd_buffer);
 }
@@ -2205,6 +2412,20 @@ void genX(CmdNextSubpass)(
 
    assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);
 
+   const struct anv_image_view *iview =
+      anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
+
+   if (iview && iview->image->aux_usage == ISL_AUX_USAGE_HIZ) {
+      const uint32_t ds = cmd_buffer->state.subpass->depth_stencil_attachment;
+
+      if (cmd_buffer->state.subpass - cmd_buffer->state.pass->subpasses ==
+          cmd_buffer->state.pass->attachments[ds].last_subpass_idx) {
+         transition_depth_buffer(cmd_buffer, iview->image,
+                                 cmd_buffer->state.attachments[ds].current_layout,
+                                 cmd_buffer->state.pass->attachments[ds].final_layout);
+      }
+   }
+
    anv_cmd_buffer_resolve_subpass(cmd_buffer);
    genX(cmd_buffer_set_subpass)(cmd_buffer, cmd_buffer->state.subpass + 1);
 }
@@ -2214,7 +2435,20 @@ void genX(CmdEndRenderPass)(
 {
    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 
-   genX(cmd_buffer_emit_hz_op)(cmd_buffer, BLORP_HIZ_OP_DEPTH_RESOLVE);
+   const struct anv_image_view *iview =
+      anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
+
+   if (iview && iview->image->aux_usage == ISL_AUX_USAGE_HIZ) {
+      const uint32_t ds = cmd_buffer->state.subpass->depth_stencil_attachment;
+
+      if (cmd_buffer->state.subpass - cmd_buffer->state.pass->subpasses ==
+          cmd_buffer->state.pass->attachments[ds].last_subpass_idx) {
+         transition_depth_buffer(cmd_buffer, iview->image,
+                                 cmd_buffer->state.attachments[ds].current_layout,
+                                 cmd_buffer->state.pass->attachments[ds].final_layout);
+      }
+   }
+
    anv_cmd_buffer_resolve_subpass(cmd_buffer);
 
 #ifndef NDEBUG