Merge pull request #6953 from ameerj/anv-semaphore

renderer_vulkan: Wait on present semaphore at queue submit

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -164,7 +164,8 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
         blit_screen.Recreate();
     }
     const VkSemaphore render_semaphore = blit_screen.DrawToSwapchain(*framebuffer, use_accelerated);
-    scheduler.Flush(render_semaphore);
+    const VkSemaphore present_semaphore = swapchain.CurrentPresentSemaphore();
+    scheduler.Flush(render_semaphore, present_semaphore);
     scheduler.WaitWorker();
     swapchain.Present(render_semaphore);
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -55,14 +55,14 @@ VKScheduler::~VKScheduler() {
     worker_thread.join();
 }
 
-void VKScheduler::Flush(VkSemaphore semaphore) {
+void VKScheduler::Flush(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
-    SubmitExecution(semaphore);
+    SubmitExecution(signal_semaphore, wait_semaphore);
     AllocateNewContext();
 }
 
-void VKScheduler::Finish(VkSemaphore semaphore) {
+void VKScheduler::Finish(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
     const u64 presubmit_tick = CurrentTick();
-    SubmitExecution(semaphore);
+    SubmitExecution(signal_semaphore, wait_semaphore);
     WaitWorker();
     Wait(presubmit_tick);
     AllocateNewContext();
@@ -171,37 +171,41 @@ void VKScheduler::AllocateWorkerCommandBuffer() {
     });
 }
 
-void VKScheduler::SubmitExecution(VkSemaphore semaphore) {
+void VKScheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
     EndPendingOperations();
     InvalidateState();
 
     const u64 signal_value = master_semaphore->NextTick();
-    Record([semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
+    Record([signal_semaphore, wait_semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
         cmdbuf.End();
-
-        const u32 num_signal_semaphores = semaphore ? 2U : 1U;
-
-        const u64 wait_value = signal_value - 1;
-        const VkPipelineStageFlags wait_stage_mask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
-
         const VkSemaphore timeline_semaphore = master_semaphore->Handle();
+
+        const u32 num_signal_semaphores = signal_semaphore ? 2U : 1U;
         const std::array signal_values{signal_value, u64(0)};
-        const std::array signal_semaphores{timeline_semaphore, semaphore};
+        const std::array signal_semaphores{timeline_semaphore, signal_semaphore};
+
+        const u32 num_wait_semaphores = wait_semaphore ? 2U : 1U;
+        const std::array wait_values{signal_value - 1, u64(1)};
+        const std::array wait_semaphores{timeline_semaphore, wait_semaphore};
+        static constexpr std::array<VkPipelineStageFlags, 2> wait_stage_masks{
+            VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+        };
 
         const VkTimelineSemaphoreSubmitInfoKHR timeline_si{
             .sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR,
             .pNext = nullptr,
-            .waitSemaphoreValueCount = 1,
+            .waitSemaphoreValueCount = num_wait_semaphores,
-            .pWaitSemaphoreValues = &wait_value,
+            .pWaitSemaphoreValues = wait_values.data(),
             .signalSemaphoreValueCount = num_signal_semaphores,
             .pSignalSemaphoreValues = signal_values.data(),
         };
         const VkSubmitInfo submit_info{
             .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
             .pNext = &timeline_si,
-            .waitSemaphoreCount = 1,
+            .waitSemaphoreCount = num_wait_semaphores,
-            .pWaitSemaphores = &timeline_semaphore,
+            .pWaitSemaphores = wait_semaphores.data(),
-            .pWaitDstStageMask = &wait_stage_mask,
+            .pWaitDstStageMask = wait_stage_masks.data(),
             .commandBufferCount = 1,
             .pCommandBuffers = cmdbuf.address(),
             .signalSemaphoreCount = num_signal_semaphores,

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -34,10 +34,10 @@ public:
     ~VKScheduler();
 
     /// Sends the current execution context to the GPU.
-    void Flush(VkSemaphore semaphore = nullptr);
+    void Flush(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
 
     /// Sends the current execution context to the GPU and waits for it to complete.
-    void Finish(VkSemaphore semaphore = nullptr);
+    void Finish(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
 
     /// Waits for the worker thread to finish executing everything. After this function returns it's
     /// safe to touch worker resources.
@@ -191,7 +191,7 @@ private:
 
     void AllocateWorkerCommandBuffer();
 
-    void SubmitExecution(VkSemaphore semaphore);
+    void SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore);
 
     void AllocateNewContext();
 

src/video_core/renderer_vulkan/vk_swapchain.cpp

@@ -107,14 +107,12 @@ void VKSwapchain::AcquireNextImage() {
 }
 
 void VKSwapchain::Present(VkSemaphore render_semaphore) {
-    const VkSemaphore present_semaphore{*present_semaphores[frame_index]};
-    const std::array<VkSemaphore, 2> semaphores{present_semaphore, render_semaphore};
     const auto present_queue{device.GetPresentQueue()};
     const VkPresentInfoKHR present_info{
         .sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
         .pNext = nullptr,
-        .waitSemaphoreCount = render_semaphore ? 2U : 1U,
+        .waitSemaphoreCount = render_semaphore ? 1U : 0U,
-        .pWaitSemaphores = semaphores.data(),
+        .pWaitSemaphores = &render_semaphore,
         .swapchainCount = 1,
         .pSwapchains = swapchain.address(),
         .pImageIndices = &image_index,

src/video_core/renderer_vulkan/vk_swapchain.h

@@ -72,6 +72,10 @@ public:
         return image_format;
     }
 
+    VkSemaphore CurrentPresentSemaphore() const {
+        return *present_semaphores[frame_index];
+    }
+
 private:
     void CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities, u32 width, u32 height,
                          bool srgb);