anv/image: Simplify setup of HiZ sampler surface state
[mesa.git] / src / intel / vulkan / anv_image.c
index e59ef4dbb6759ebe5d7786bc7b5e282e5f5bbc31..16a60833e022cf947e70bdbdfd39a219aaffa633 100644 (file)
@@ -180,6 +180,11 @@ make_surface(const struct anv_device *dev,
    /* Add a HiZ surface to a depth buffer that will be used for rendering.
     */
    if (aspect == VK_IMAGE_ASPECT_DEPTH_BIT) {
+      /* We don't advertise that depth buffers could be used as storage
+       * images.
+       */
+       assert(!(image->usage & VK_IMAGE_USAGE_STORAGE_BIT));
+
       /* Allow the user to control HiZ enabling. Disable by default on gen7
        * because resolves are not currently implemented pre-BDW.
        */
@@ -209,8 +214,9 @@ make_surface(const struct anv_device *dev,
          anv_finishme("Test gen8 multisampled HiZ");
       } else {
          assert(image->aux_surface.isl.size == 0);
-         isl_surf_get_hiz_surf(&dev->isl_dev, &image->depth_surface.isl,
-                               &image->aux_surface.isl);
+         ok = isl_surf_get_hiz_surf(&dev->isl_dev, &image->depth_surface.isl,
+                                    &image->aux_surface.isl);
+         assert(ok);
          add_surface(image, &image->aux_surface);
          image->aux_usage = ISL_AUX_USAGE_HIZ;
       }
@@ -237,6 +243,15 @@ make_surface(const struct anv_device *dev,
             }
          }
       }
+   } else if (aspect == VK_IMAGE_ASPECT_COLOR_BIT && vk_info->samples > 1) {
+      assert(image->aux_surface.isl.size == 0);
+      assert(!(vk_info->usage & VK_IMAGE_USAGE_STORAGE_BIT));
+      ok = isl_surf_get_mcs_surf(&dev->isl_dev, &anv_surf->isl,
+                                 &image->aux_surface.isl);
+      if (ok) {
+         add_surface(image, &image->aux_surface);
+         image->aux_usage = ISL_AUX_USAGE_MCS;
+      }
    }
 
    return VK_SUCCESS;
@@ -422,6 +437,139 @@ void anv_GetImageSubresourceLayout(
    }
 }
 
+/**
+ * This function determines the optimal buffer to use for device
+ * accesses given a VkImageLayout and other pieces of information needed to
+ * make that determination. This does not determine the optimal buffer to
+ * use during a resolve operation.
+ *
+ * NOTE: Some layouts do not support device access.
+ *
+ * @param devinfo The device information of the Intel GPU.
+ * @param image The image that may contain a collection of buffers.
+ * @param aspects The aspect(s) of the image to be accessed.
+ * @param layout The current layout of the image aspect(s).
+ *
+ * @return The primary buffer that should be used for the given layout.
+ */
+enum isl_aux_usage
+anv_layout_to_aux_usage(const struct gen_device_info * const devinfo,
+                        const struct anv_image * const image,
+                        const VkImageAspectFlags aspects,
+                        const VkImageLayout layout)
+{
+   /* Validate the inputs. */
+
+   /* The devinfo is needed as the optimal buffer varies across generations. */
+   assert(devinfo != NULL);
+
+   /* The layout of a NULL image is not properly defined. */
+   assert(image != NULL);
+
+   /* The aspects must be a subset of the image aspects. */
+   assert(aspects & image->aspects && aspects <= image->aspects);
+
+   /* Determine the optimal buffer. */
+
+   /* If there is no auxiliary surface allocated, we must use the one and only
+    * main buffer.
+    */
+   if (image->aux_surface.isl.size == 0)
+      return ISL_AUX_USAGE_NONE;
+
+   /* All images that use an auxiliary surface are required to be tiled. */
+   assert(image->tiling == VK_IMAGE_TILING_OPTIMAL);
+
+   /* On BDW+, when clearing the stencil aspect of a depth stencil image,
+    * the HiZ buffer allows us to record the clear with a relatively small
+    * number of packets. Prior to BDW, the HiZ buffer provides no known benefit
+    * to the stencil aspect.
+    */
+   if (devinfo->gen < 8 && aspects == VK_IMAGE_ASPECT_STENCIL_BIT)
+      return ISL_AUX_USAGE_NONE;
+
+   const bool color_aspect = aspects == VK_IMAGE_ASPECT_COLOR_BIT;
+
+   /* The following switch currently only handles depth stencil aspects.
+    * TODO: Handle the color aspect.
+    */
+   if (color_aspect)
+      return image->aux_usage;
+
+   switch (layout) {
+
+   /* Invalid Layouts */
+
+   /* According to the Vulkan Spec, the following layouts are valid only as
+    * initial layouts in a layout transition and don't support device access.
+    */
+   case VK_IMAGE_LAYOUT_UNDEFINED:
+   case VK_IMAGE_LAYOUT_PREINITIALIZED:
+   case VK_IMAGE_LAYOUT_RANGE_SIZE:
+   case VK_IMAGE_LAYOUT_MAX_ENUM:
+      unreachable("Invalid image layout for device access.");
+
+
+   /* Transfer Layouts
+    *
+    * This buffer could be a depth buffer used in a transfer operation. BLORP
+    * currently doesn't use HiZ for transfer operations so we must use the main
+    * buffer for this layout. TODO: Enable HiZ in BLORP.
+    */
+   case VK_IMAGE_LAYOUT_GENERAL:
+   case VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL:
+   case VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL:
+      return ISL_AUX_USAGE_NONE;
+
+
+   /* Sampling Layouts */
+   case VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL:
+      assert(!color_aspect);
+      /* Fall-through */
+   case VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL:
+      if (anv_can_sample_with_hiz(devinfo, aspects, image->samples))
+         return ISL_AUX_USAGE_HIZ;
+      else
+         return ISL_AUX_USAGE_NONE;
+
+   case VK_IMAGE_LAYOUT_PRESENT_SRC_KHR:
+      assert(color_aspect);
+
+      /* On SKL+, the render buffer can be decompressed by the presentation
+       * engine. Support for this feature has not yet landed in the wider
+       * ecosystem. TODO: Update this code when support lands.
+       *
+       * From the BDW PRM, Vol 7, Render Target Resolve:
+       *
+       *    If the MCS is enabled on a non-multisampled render target, the
+       *    render target must be resolved before being used for other
+       *    purposes (display, texture, CPU lock) The clear value from
+       *    SURFACE_STATE is written into pixels in the render target
+       *    indicated as clear in the MCS.
+       *
+       * Pre-SKL, the render buffer must be resolved before being used for
+       * presentation. We can infer that the auxiliary buffer is not used.
+       */
+      return ISL_AUX_USAGE_NONE;
+
+
+   /* Rendering Layouts */
+   case VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL:
+      assert(color_aspect);
+      unreachable("Color images are not yet supported.");
+
+   case VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL:
+      assert(!color_aspect);
+      return ISL_AUX_USAGE_HIZ;
+   }
+
+   /* If the layout isn't recognized in the exhaustive switch above, the
+    * VkImageLayout value is not defined in vulkan.h.
+    */
+   unreachable("layout is not a VkImageLayout enumeration member.");
+}
+
+
 static struct anv_state
 alloc_surface_state(struct anv_device *device)
 {
@@ -534,24 +682,6 @@ anv_CreateImageView(VkDevice _device,
       iview->isl.usage = 0;
    }
 
-   /* If the HiZ buffer can be sampled from, set the constant clear color.
-    * If it cannot, disable the isl aux usage flag.
-    */
-   float red_clear_color = 0.0f;
-   enum isl_aux_usage surf_usage = image->aux_usage;
-   if (image->aux_usage == ISL_AUX_USAGE_HIZ) {
-      if (iview->aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT &&
-          anv_can_sample_with_hiz(device->info.gen, image->samples)) {
-         /* When a HiZ buffer is sampled on gen9+, ensure that
-          * the constant fast clear value is set in the surface state.
-          */
-         if (device->info.gen >= 9)
-            red_clear_color = ANV_HZ_FC_VAL;
-      } else {
-         surf_usage = ISL_AUX_USAGE_NONE;
-      }
-   }
-
    /* Input attachment surfaces for color are allocated and filled
     * out at BeginRenderPass time because they need compression information.
     * Compression is not yet enabled for depth textures and stencil doesn't
@@ -563,6 +693,18 @@ anv_CreateImageView(VkDevice _device,
         !(iview->aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT))) {
       iview->sampler_surface_state = alloc_surface_state(device);
 
+      /* Select the optimal aux_usage for sampling. */
+      const enum isl_aux_usage surf_usage =
+         anv_layout_to_aux_usage(&device->info, image, iview->aspect_mask,
+                                 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
+
+      /* If this is a HiZ buffer we can sample from with a programmable clear
+       * value (SKL+), define the clear value to the optimal constant.
+       */
+      const float red_clear_color = surf_usage == ISL_AUX_USAGE_HIZ &&
+                                    device->info.gen >= 9 ?
+                                    ANV_HZ_FC_VAL : 0.0f;
+
       struct isl_view view = iview->isl;
       view.usage |= ISL_SURF_USAGE_TEXTURE_BIT;
       isl_surf_fill_state(&device->isl_dev,
@@ -574,8 +716,7 @@ anv_CreateImageView(VkDevice _device,
                           .aux_usage = surf_usage,
                           .mocs = device->default_mocs);
 
-      if (!device->info.has_llc)
-         anv_state_clflush(iview->sampler_surface_state);
+      anv_state_flush(device, iview->sampler_surface_state);
    } else {
       iview->sampler_surface_state.alloc_size = 0;
    }
@@ -583,19 +724,37 @@ anv_CreateImageView(VkDevice _device,
    /* NOTE: This one needs to go last since it may stomp isl_view.format */
    if (image->usage & VK_IMAGE_USAGE_STORAGE_BIT) {
       iview->storage_surface_state = alloc_surface_state(device);
+      iview->writeonly_storage_surface_state = alloc_surface_state(device);
+
+      struct isl_view view = iview->isl;
+      view.usage |= ISL_SURF_USAGE_STORAGE_BIT;
+
+      /* Write-only accesses always used a typed write instruction and should
+       * therefore use the real format.
+       */
+      isl_surf_fill_state(&device->isl_dev,
+                          iview->writeonly_storage_surface_state.map,
+                          .surf = &surface->isl,
+                          .view = &view,
+                          .aux_surf = &image->aux_surface.isl,
+                          .aux_usage = image->aux_usage,
+                          .mocs = device->default_mocs);
 
       if (isl_has_matching_typed_storage_image_format(&device->info,
                                                       format.isl_format)) {
-         struct isl_view view = iview->isl;
-         view.usage |= ISL_SURF_USAGE_STORAGE_BIT;
+         /* Typed surface reads support a very limited subset of the shader
+          * image formats.  Translate it into the closest format the hardware
+          * supports.
+          */
          view.format = isl_lower_storage_image_format(&device->info,
                                                       format.isl_format);
+
          isl_surf_fill_state(&device->isl_dev,
                              iview->storage_surface_state.map,
                              .surf = &surface->isl,
                              .view = &view,
                              .aux_surf = &image->aux_surface.isl,
-                             .aux_usage = surf_usage,
+                             .aux_usage = image->aux_usage,
                              .mocs = device->default_mocs);
       } else {
          anv_fill_buffer_surface_state(device, iview->storage_surface_state,
@@ -608,10 +767,11 @@ anv_CreateImageView(VkDevice _device,
                                 &iview->storage_image_param,
                                 &surface->isl, &iview->isl);
 
-      if (!device->info.has_llc)
-         anv_state_clflush(iview->storage_surface_state);
+      anv_state_flush(device, iview->storage_surface_state);
+      anv_state_flush(device, iview->writeonly_storage_surface_state);
    } else {
       iview->storage_surface_state.alloc_size = 0;
+      iview->writeonly_storage_surface_state.alloc_size = 0;
    }
 
    *pView = anv_image_view_to_handle(iview);
@@ -639,6 +799,11 @@ anv_DestroyImageView(VkDevice _device, VkImageView _iview,
                           iview->storage_surface_state);
    }
 
+   if (iview->writeonly_storage_surface_state.alloc_size > 0) {
+      anv_state_pool_free(&device->surface_state_pool,
+                          iview->writeonly_storage_surface_state);
+   }
+
    vk_free2(&device->alloc, pAllocator, iview);
 }
 
@@ -682,6 +847,7 @@ anv_CreateBufferView(VkDevice _device,
 
    if (buffer->usage & VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT) {
       view->storage_surface_state = alloc_surface_state(device);
+      view->writeonly_storage_surface_state = alloc_surface_state(device);
 
       enum isl_format storage_format =
          isl_has_matching_typed_storage_image_format(&device->info,
@@ -695,11 +861,18 @@ anv_CreateBufferView(VkDevice _device,
                                     (storage_format == ISL_FORMAT_RAW ? 1 :
                                      isl_format_get_layout(storage_format)->bpb / 8));
 
+      /* Write-only accesses should use the original format. */
+      anv_fill_buffer_surface_state(device, view->writeonly_storage_surface_state,
+                                    view->format,
+                                    view->offset, view->range,
+                                    isl_format_get_layout(view->format)->bpb / 8);
+
       isl_buffer_fill_image_param(&device->isl_dev,
                                   &view->storage_image_param,
                                   view->format, view->range);
    } else {
       view->storage_surface_state = (struct anv_state){ 0 };
+      view->writeonly_storage_surface_state = (struct anv_state){ 0 };
    }
 
    *pView = anv_buffer_view_to_handle(view);
@@ -725,6 +898,10 @@ anv_DestroyBufferView(VkDevice _device, VkBufferView bufferView,
       anv_state_pool_free(&device->surface_state_pool,
                           view->storage_surface_state);
 
+   if (view->writeonly_storage_surface_state.alloc_size > 0)
+      anv_state_pool_free(&device->surface_state_pool,
+                          view->writeonly_storage_surface_state);
+
    vk_free2(&device->alloc, pAllocator, view);
 }