wsi/metal: Backend addition for drivers built on top of Metal

Reviewed-By: Aleksi Sapon <aleksi.sapon@autodesk.com>
Reviewed-By: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37493>

src/vulkan/wsi/wsi_common.c

@@ -390,6 +390,21 @@ get_blit_type(const struct wsi_device *wsi,
          container_of(params, const struct wsi_dxgi_image_params, base);
       return wsi_dxgi_image_needs_blit(wsi, dxgi_params, device);
    }
+#endif
+#if defined(VK_USE_PLATFORM_METAL_EXT)
+   case WSI_IMAGE_TYPE_METAL: {
+      /* Due to mismatches between WSI and Metal, we require rendering into an
+       * intermediate texture and later blit that texture to the display
+       * texture. There is not much we can do about this since users can get
+       * the VkImages before acquiring the display image for command buffer
+       * recording as long as they acquire the image before submission. Metal
+       * on the other hand will only give us a texture handle after acquiring
+       * which leads to us having to provide an intermediate texture just for
+       * this. We could move the acquisition to the first usage of the VkImage
+       * but that's something we can contemplate if the performace gain is
+       * considerable. */
+      return WSI_SWAPCHAIN_IMAGE_BLIT;
+   }
 #endif
    default:
       UNREACHABLE("Invalid image type");
@@ -424,6 +439,13 @@ configure_image(const struct wsi_swapchain *chain,
          container_of(params, const struct wsi_dxgi_image_params, base);
       return wsi_dxgi_configure_image(chain, pCreateInfo, dxgi_params, info);
    }
+#endif
+#if defined(VK_USE_PLATFORM_METAL_EXT)
+   case WSI_IMAGE_TYPE_METAL: {
+      const struct wsi_metal_image_params *metal_params =
+         container_of(params, const struct wsi_metal_image_params, base);
+      return wsi_metal_configure_image(chain, pCreateInfo, metal_params, info);
+   }
 #endif
    default:
       UNREACHABLE("Invalid image type");

src/vulkan/wsi/wsi_common.h

@@ -137,6 +137,14 @@ struct wsi_device {
                                       VkDeviceMemory *out);
    } win32;
 
+   struct {
+      VkResult (*bind_drawable_to_vkimage)(VkImage image, void *drawable);
+      /* Does not transfer the ownership to the command buffer. The command
+       * buffer must retain the drawable to ensure lifespan. */
+      void (*encode_drawable_present)(VkCommandBuffer cmd, void *drawable);
+   } metal;
+
+
    bool sw;
 
    /* Set to true if the implementation is ok with linear WSI images. */

src/vulkan/wsi/wsi_common_metal.c

@@ -300,6 +300,149 @@ wsi_metal_swapchain_get_wsi_image(struct wsi_swapchain *wsi_chain,
    return &chain->images[image_index].base;
 }
 
+static VkResult
+wsi_cmd_blit_image_to_image(const struct wsi_swapchain *chain,
+                            const struct wsi_image_info *info,
+                            struct wsi_image *image)
+{
+   /* Should only be called from non-software backends */
+   assert(!chain->wsi->sw);
+   
+   const struct wsi_device *wsi = chain->wsi;
+   struct wsi_metal_image *metal_image = container_of(image, struct wsi_metal_image, base);
+   VkResult result;
+   int queue_count = chain->blit.queue != NULL ? 1 : wsi->queue_family_count;
+
+   for (uint32_t i = 0; i < queue_count; i++) {
+      if (!chain->cmd_pools[i])
+         continue;
+
+      /* We need to cycle command buffers since the MTLTexture backing the presentable
+       * VkImage changes every time it's acquired. We only have one command buffer
+       * per blit since we only submit to a single queue which is the blit queue. */
+      wsi->FreeCommandBuffers(chain->device, chain->cmd_pools[i], 1u,
+                              &image->blit.cmd_buffers[i + queue_count]);
+
+      /* Store the command buffer in flight */
+      image->blit.cmd_buffers[i + queue_count] = image->blit.cmd_buffers[i];
+
+      const VkCommandBufferAllocateInfo cmd_buffer_info = {
+         .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
+         .pNext = NULL,
+         .commandPool = chain->cmd_pools[0],
+         .level = VK_COMMAND_BUFFER_LEVEL_PRIMARY,
+         .commandBufferCount = 1,
+      };
+      result = wsi->AllocateCommandBuffers(chain->device, &cmd_buffer_info,
+                                          &image->blit.cmd_buffers[i]);
+      if (result != VK_SUCCESS)
+         return result;
+
+      const VkCommandBufferBeginInfo begin_info = {
+         .sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+      };
+      wsi->BeginCommandBuffer(image->blit.cmd_buffers[i], &begin_info);
+
+      VkImageMemoryBarrier img_mem_barriers[] = {
+         {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = NULL,
+            .srcAccessMask = 0,
+            .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR,
+            .newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = image->image,
+            .subresourceRange = {
+               .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+               .baseMipLevel = 0,
+               .levelCount = 1,
+               .baseArrayLayer = 0,
+               .layerCount = 1,
+            },
+         },
+         {
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = NULL,
+            .srcAccessMask = 0,
+            .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_UNDEFINED,
+            .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = image->blit.image,
+            .subresourceRange = {
+               .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+               .baseMipLevel = 0,
+               .levelCount = 1,
+               .baseArrayLayer = 0,
+               .layerCount = 1,
+            },
+         },
+      };
+      const uint32_t img_mem_barrier_count = 2;
+      wsi->CmdPipelineBarrier(image->blit.cmd_buffers[i],
+                              VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                              VK_PIPELINE_STAGE_TRANSFER_BIT,
+                              0,
+                              0, NULL,
+                              0, NULL,
+                              img_mem_barrier_count, img_mem_barriers);
+
+      struct VkImageCopy image_copy = {
+         .srcSubresource = {
+            .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            .mipLevel = 0,
+            .baseArrayLayer = 0,
+            .layerCount = 1,
+         },
+         .srcOffset = { .x = 0, .y = 0, .z = 0 },
+         .dstSubresource = {
+            .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+            .mipLevel = 0,
+            .baseArrayLayer = 0,
+            .layerCount = 1,
+         },
+         .dstOffset = { .x = 0, .y = 0, .z = 0 },
+         .extent = info->create.extent,
+      };
+
+      wsi->CmdCopyImage(image->blit.cmd_buffers[i],
+                        image->image,
+                        VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                        image->blit.image,
+                        VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                        1, &image_copy);
+
+      img_mem_barriers[0].srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
+      img_mem_barriers[0].dstAccessMask = 0;
+      img_mem_barriers[0].oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
+      img_mem_barriers[0].newLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
+      img_mem_barriers[1].srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
+      img_mem_barriers[1].dstAccessMask = 0;
+      img_mem_barriers[1].oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
+      img_mem_barriers[1].newLayout = VK_IMAGE_LAYOUT_PRESENT_SRC_KHR;
+      wsi->CmdPipelineBarrier(image->blit.cmd_buffers[i],
+                              VK_PIPELINE_STAGE_TRANSFER_BIT,
+                              VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
+                              0,
+                              0, NULL,
+                              0, NULL,
+                              img_mem_barrier_count, img_mem_barriers);
+
+      result = wsi->EndCommandBuffer(image->blit.cmd_buffers[i]);
+
+      wsi->metal.encode_drawable_present(image->blit.cmd_buffers[i], metal_image->drawable);
+   }
+
+   /* Release the drawable since command buffers should have retained the drawable. */
+   wsi_metal_release_drawable(metal_image->drawable);
+   metal_image->drawable = NULL;
+
+   return result;
+}
+
 static VkResult
 wsi_metal_swapchain_acquire_next_image(struct wsi_swapchain *wsi_chain,
                                        const VkAcquireNextImageInfoKHR *info,
@@ -320,8 +463,16 @@ wsi_metal_swapchain_acquire_next_image(struct wsi_swapchain *wsi_chain,
       CAMetalDrawable *drawable = wsi_metal_layer_acquire_drawable(chain->surface->pLayer);
       if (drawable) {
          uint32_t i = (chain->current_image_index++) % chain->base.image_count;
+         struct wsi_metal_image *image = &chain->images[i];
          *image_index = i;
-         chain->images[i].drawable = drawable;
+         image->drawable = drawable;
+         if (!wsi_chain->wsi->sw) {
+            chain->base.wsi->metal.bind_drawable_to_vkimage(image->base.blit.image,
+                                                            image->drawable);
+            /* Since present images will only be backed by MTLTextures after acquisition,
+             * we need to re-record the command buffer so it uses the new drawable. */
+            wsi_cmd_blit_image_to_image(wsi_chain, &wsi_chain->image_info, &image->base);
+         }
          return VK_SUCCESS;
       }
 
@@ -346,11 +497,13 @@ wsi_metal_swapchain_queue_present(struct wsi_swapchain *wsi_chain,
 
    struct wsi_metal_image *image = &chain->images[image_index];
 
-   wsi_metal_layer_blit_and_present(chain->blit_context,
-      &image->drawable,
-      image->base.cpu_map,
-      chain->extent.width, chain->extent.height,
-      image->base.row_pitches[0]);
+   if (wsi_chain->wsi->sw) {
+      wsi_metal_layer_blit_and_present(chain->blit_context,
+         &image->drawable,
+         image->base.cpu_map,
+         chain->extent.width, chain->extent.height,
+         image->base.row_pitches[0]);
+   }
 
    uint32_t width = 0u, height = 0u;
    wsi_metal_layer_size(chain->surface->pLayer, &width, &height);
@@ -358,6 +511,62 @@ wsi_metal_swapchain_queue_present(struct wsi_swapchain *wsi_chain,
    return is_optimal ? VK_SUCCESS : VK_SUBOPTIMAL_KHR;
 }
 
+static void
+wsi_metal_destroy_image(const struct wsi_metal_swapchain *metal_chain,
+                        struct wsi_metal_image *metal_image)
+{
+   const struct wsi_swapchain *chain = &metal_chain->base;
+   const struct wsi_device *wsi = chain->wsi;
+   struct wsi_image *image = &metal_image->base;
+
+   /* Software backends can just call common and return */
+   if (wsi->sw) {
+      wsi_destroy_image(chain, image);
+      return;
+   }
+
+   /* Required since we allocate 2 per queue */
+   if (image->blit.cmd_buffers) {
+      int cmd_buffer_count =
+         chain->blit.queue != NULL ? 2 : wsi->queue_family_count * 2;
+
+      for (uint32_t i = 0; i < cmd_buffer_count; i++) {
+         if (!chain->cmd_pools[i])
+            continue;
+         wsi->FreeCommandBuffers(chain->device, chain->cmd_pools[i],
+                                 1, &image->blit.cmd_buffers[i]);
+      }
+      vk_free(&chain->alloc, image->blit.cmd_buffers);
+      image->blit.cmd_buffers = NULL;
+   }
+
+   wsi_destroy_image(chain, image);
+}
+
+static VkResult
+wsi_metal_create_image(const struct wsi_metal_swapchain *metal_chain,
+                       const struct wsi_image_info *info,
+                       struct wsi_metal_image *metal_image)
+{
+   const struct wsi_swapchain *chain = &metal_chain->base;
+   const struct wsi_device *wsi = chain->wsi;
+   struct wsi_image *image = &metal_image->base;
+
+   VkResult result = wsi_create_image(chain, info, image);
+
+   /* Software backends can just call common and return */
+   if (wsi->sw || result != VK_SUCCESS)
+      return result;
+
+   /* Create VkImages to handle binding at acquisition. */
+   result = wsi->CreateImage(chain->device, &chain->image_info.create,
+                             &chain->alloc, &image->blit.image);
+   if (result != VK_SUCCESS)
+      wsi_metal_destroy_image(metal_chain, metal_image);
+
+   return result;
+}
+
 static VkResult
 wsi_metal_swapchain_destroy(struct wsi_swapchain *wsi_chain,
                                const VkAllocationCallbacks *pAllocator)
@@ -372,11 +581,12 @@ wsi_metal_swapchain_destroy(struct wsi_swapchain *wsi_chain,
          image->drawable = NULL;
       }
 
-      if (image->base.image != VK_NULL_HANDLE)
-         wsi_destroy_image(&chain->base, &image->base);
+      if (image != VK_NULL_HANDLE)
+         wsi_metal_destroy_image(chain, image);
    }
 
-   wsi_destroy_metal_layer_blit_context(chain->blit_context);
+   if (chain->base.wsi->sw)
+      wsi_destroy_metal_layer_blit_context(chain->blit_context);
 
    wsi_swapchain_finish(&chain->base);
 
@@ -417,12 +627,22 @@ wsi_metal_surface_create_swapchain(VkIcdSurfaceBase *icd_surface,
    if (chain == NULL)
       return VK_ERROR_OUT_OF_HOST_MEMORY;
 
+   /* Software drivers cannot render to an MTLTexture as of now. Rendering to
+    * MTLTexture could be supported, but outside of the scope of adding a
+    * Metal backend that uses MTLTexture as render target. The software path
+    * will render to a CPU texture, and blit it to the presentation MTLTexture
+    * at the last moment. */
+   const bool is_sw_driver = wsi_device->sw;
    struct wsi_cpu_image_params cpu_params = {
       .base.image_type = WSI_IMAGE_TYPE_CPU,
    };
+   struct wsi_base_image_params metal_params = {
+      .image_type = WSI_IMAGE_TYPE_METAL,
+   };
+   struct wsi_base_image_params *params = is_sw_driver ? &cpu_params.base : &metal_params;
 
    result = wsi_swapchain_init(wsi_device, &chain->base, device,
-                               pCreateInfo, &cpu_params.base, pAllocator);
+                               pCreateInfo, params, pAllocator);
    if (result != VK_SUCCESS)
       goto fail_chain_alloc;
 
@@ -439,8 +659,8 @@ wsi_metal_surface_create_swapchain(VkIcdSurfaceBase *icd_surface,
 
    uint32_t created_image_count = 0;
    for (uint32_t i = 0; i < chain->base.image_count; i++) {
-      result = wsi_create_image(&chain->base, &chain->base.image_info,
-                                &chain->images[i].base);
+      result = wsi_metal_create_image(chain, &chain->base.image_info,
+                                      &chain->images[i]);
       if (result != VK_SUCCESS)
          goto fail_init_images;
 
@@ -448,7 +668,8 @@ wsi_metal_surface_create_swapchain(VkIcdSurfaceBase *icd_surface,
       created_image_count++;
    }
 
-   chain->blit_context = wsi_create_metal_layer_blit_context();
+   if (is_sw_driver)
+      chain->blit_context = wsi_create_metal_layer_blit_context();
 
    *swapchain_out = &chain->base;
 
@@ -456,7 +677,7 @@ wsi_metal_surface_create_swapchain(VkIcdSurfaceBase *icd_surface,
 
 fail_init_images:
    for (uint32_t i = 0; i < created_image_count; i++)
-      wsi_destroy_image(&chain->base, &chain->images[i].base);
+      wsi_metal_destroy_image(chain, &chain->images[i]);
 
    wsi_swapchain_finish(&chain->base);
 
@@ -537,3 +758,79 @@ wsi_CreateMetalSurfaceEXT(
    *pSurface = VkIcdSurfaceBase_to_handle(&surface->base);
    return VK_SUCCESS;
 }
+
+static VkResult
+wsi_metal_create_mem(const struct wsi_swapchain *chain,
+                     const struct wsi_image_info *info,
+                     struct wsi_image *image)
+{
+   assert(chain->blit.type == WSI_SWAPCHAIN_IMAGE_BLIT);
+
+   const struct wsi_device *wsi = chain->wsi;
+
+   VkMemoryRequirements requirements;
+   wsi->GetImageMemoryRequirements(chain->device, image->image, &requirements);
+
+   struct wsi_memory_allocate_info memory_wsi_info = {
+      .sType = VK_STRUCTURE_TYPE_WSI_MEMORY_ALLOCATE_INFO_MESA,
+      .pNext = NULL,
+      .implicit_sync = false,
+   };
+   VkMemoryDedicatedAllocateInfo image_mem_dedicated_info = {
+      .sType = VK_STRUCTURE_TYPE_MEMORY_DEDICATED_ALLOCATE_INFO,
+      .pNext = &memory_wsi_info,
+      .image = image->image,
+      .buffer = VK_NULL_HANDLE,
+   };
+   VkMemoryAllocateInfo image_mem_info = {
+      .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
+      .pNext = &image_mem_dedicated_info,
+      .allocationSize = requirements.size,
+      .memoryTypeIndex = requirements.memoryTypeBits,
+   };
+
+   return wsi->AllocateMemory(chain->device, &image_mem_info,
+                              &chain->alloc, &image->memory);
+}
+
+static VkResult
+wsi_metal_allocate_command_buffer(const struct wsi_swapchain *chain,
+                                  const struct wsi_image_info *info,
+                                  struct wsi_image *image)
+{
+   const struct wsi_device *wsi = chain->wsi;
+   /* We need to create 2 command buffers per queue to be able to ping pong the blit.
+    * The first queue_family_count will store the next blit command,
+    * and the remaining will store the ones in flight. */
+   int cmd_buffer_count =
+      chain->blit.queue != NULL ? 2 : wsi->queue_family_count * 2;
+   image->blit.cmd_buffers =
+      vk_zalloc(&chain->alloc,
+                sizeof(VkCommandBuffer) * cmd_buffer_count, 8,
+                VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+   return image->blit.cmd_buffers ? VK_SUCCESS : VK_ERROR_OUT_OF_HOST_MEMORY;
+}
+
+/* Common utilities required by wsi_common.c */
+VkResult
+wsi_metal_configure_image(const struct wsi_swapchain *chain,
+                          const VkSwapchainCreateInfoKHR *pCreateInfo,
+                          const struct wsi_metal_image_params *params,
+                          struct wsi_image_info *info)
+{
+   VkResult result =
+      wsi_configure_image(chain, pCreateInfo, 0, info);
+   if (result != VK_SUCCESS)
+      return result;
+
+   if (chain->blit.type != WSI_SWAPCHAIN_NO_BLIT) {
+      info->create.usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+      info->wsi.blit_src = true;
+      info->finish_create = wsi_metal_allocate_command_buffer;
+      info->select_image_memory_type = wsi_select_device_memory_type;
+      info->create_mem = wsi_metal_create_mem;
+   }
+
+   return VK_SUCCESS;
+}

src/vulkan/wsi/wsi_common_private.h

@@ -48,6 +48,7 @@ enum wsi_image_type {
    WSI_IMAGE_TYPE_CPU,
    WSI_IMAGE_TYPE_DRM,
    WSI_IMAGE_TYPE_DXGI,
+   WSI_IMAGE_TYPE_METAL,
 };
 
 struct wsi_base_image_params {
@@ -496,6 +497,22 @@ void wsi_headless_finish_wsi(struct wsi_device *wsi_device,
 VK_DEFINE_NONDISP_HANDLE_CASTS(wsi_swapchain, base, VkSwapchainKHR,
                                VK_OBJECT_TYPE_SWAPCHAIN_KHR)
 
+#if defined(VK_USE_PLATFORM_METAL_EXT)
+struct wsi_metal_image_params {
+   struct wsi_base_image_params base;
+   /* Software implementations like lavapipe cannot render to an MTLTexture
+    * directly and therefore require a blit
+    */
+   bool can_render_to_texture;
+};
+
+VkResult
+wsi_metal_configure_image(const struct wsi_swapchain *chain,
+                          const VkSwapchainCreateInfoKHR *pCreateInfo,
+                          const struct wsi_metal_image_params *params,
+                          struct wsi_image_info *info);
+#endif /* defined(VK_USE_PLATFORM_METAL_EXT) */
+
 #ifdef __cplusplus
 }
 #endif