From 154a7653f9705ebee095b8f10609fd0f3f9ca394 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Fri, 8 Jan 2021 02:37:02 -0300 Subject: vk_fence_manager: Use timeline semaphores instead of spin waits With timeline semaphores we can avoid creating objects. Instead of creating an event, grab the current tick from the scheduler and flush the current command buffer. When the fence has to be queried/waited, we can do so against the master semaphore instead of spinning on an event. If Vulkan supported NVN like events or fences, we could signal from the command buffer and wait for that without splitting things in two separate command buffers. --- .../renderer_vulkan/vk_fence_manager.cpp | 54 ++++++---------------- src/video_core/renderer_vulkan/vk_fence_manager.h | 15 ++---- src/video_core/renderer_vulkan/vk_rasterizer.cpp | 3 +- 3 files changed, 18 insertions(+), 54 deletions(-) (limited to 'src/video_core') diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.cpp b/src/video_core/renderer_vulkan/vk_fence_manager.cpp index 4c5bc0aa1..6cd00884d 100644 --- a/src/video_core/renderer_vulkan/vk_fence_manager.cpp +++ b/src/video_core/renderer_vulkan/vk_fence_manager.cpp @@ -3,7 +3,6 @@ // Refer to the license.txt file included. #include -#include #include "video_core/renderer_vulkan/vk_buffer_cache.h" #include "video_core/renderer_vulkan/vk_fence_manager.h" @@ -14,13 +13,11 @@ namespace Vulkan { -InnerFence::InnerFence(const Device& device_, VKScheduler& scheduler_, u32 payload_, - bool is_stubbed_) - : FenceBase{payload_, is_stubbed_}, device{device_}, scheduler{scheduler_} {} +InnerFence::InnerFence(VKScheduler& scheduler_, u32 payload_, bool is_stubbed_) + : FenceBase{payload_, is_stubbed_}, scheduler{scheduler_} {} -InnerFence::InnerFence(const Device& device_, VKScheduler& scheduler_, GPUVAddr address_, - u32 payload_, bool is_stubbed_) - : FenceBase{address_, payload_, is_stubbed_}, device{device_}, scheduler{scheduler_} {} +InnerFence::InnerFence(VKScheduler& scheduler_, GPUVAddr address_, u32 payload_, bool is_stubbed_) + : FenceBase{address_, payload_, is_stubbed_}, scheduler{scheduler_} {} InnerFence::~InnerFence() = default; @@ -28,63 +25,38 @@ void InnerFence::Queue() { if (is_stubbed) { return; } - ASSERT(!event); - - event = device.GetLogical().CreateEvent(); - ticks = scheduler.CurrentTick(); - - scheduler.RequestOutsideRenderPassOperationContext(); - scheduler.Record([event = *event](vk::CommandBuffer cmdbuf) { - cmdbuf.SetEvent(event, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT); - }); + // Get the current tick so we can wait for it + wait_tick = scheduler.CurrentTick(); + scheduler.Flush(); } bool InnerFence::IsSignaled() const { if (is_stubbed) { return true; } - ASSERT(event); - return IsEventSignalled(); + return scheduler.IsFree(wait_tick); } void InnerFence::Wait() { if (is_stubbed) { return; } - ASSERT(event); - - if (ticks >= scheduler.CurrentTick()) { - scheduler.Flush(); - } - while (!IsEventSignalled()) { - std::this_thread::yield(); - } -} - -bool InnerFence::IsEventSignalled() const { - switch (const VkResult result = event.GetStatus()) { - case VK_EVENT_SET: - return true; - case VK_EVENT_RESET: - return false; - default: - throw vk::Exception(result); - } + scheduler.Wait(wait_tick); } VKFenceManager::VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_, Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_, VKBufferCache& buffer_cache_, VKQueryCache& query_cache_, - const Device& device_, VKScheduler& scheduler_) + VKScheduler& scheduler_) : GenericFenceManager{rasterizer_, gpu_, texture_cache_, buffer_cache_, query_cache_}, - device{device_}, scheduler{scheduler_} {} + scheduler{scheduler_} {} Fence VKFenceManager::CreateFence(u32 value, bool is_stubbed) { - return std::make_shared(device, scheduler, value, is_stubbed); + return std::make_shared(scheduler, value, is_stubbed); } Fence VKFenceManager::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) { - return std::make_shared(device, scheduler, addr, value, is_stubbed); + return std::make_shared(scheduler, addr, value, is_stubbed); } void VKFenceManager::QueueFence(Fence& fence) { diff --git a/src/video_core/renderer_vulkan/vk_fence_manager.h b/src/video_core/renderer_vulkan/vk_fence_manager.h index 6b51e4587..9c5e5aa8f 100644 --- a/src/video_core/renderer_vulkan/vk_fence_manager.h +++ b/src/video_core/renderer_vulkan/vk_fence_manager.h @@ -28,10 +28,8 @@ class VKScheduler; class InnerFence : public VideoCommon::FenceBase { public: - explicit InnerFence(const Device& device_, VKScheduler& scheduler_, u32 payload_, - bool is_stubbed_); - explicit InnerFence(const Device& device_, VKScheduler& scheduler_, GPUVAddr address_, - u32 payload_, bool is_stubbed_); + explicit InnerFence(VKScheduler& scheduler_, u32 payload_, bool is_stubbed_); + explicit InnerFence(VKScheduler& scheduler_, GPUVAddr address_, u32 payload_, bool is_stubbed_); ~InnerFence(); void Queue(); @@ -41,12 +39,8 @@ public: void Wait(); private: - bool IsEventSignalled() const; - - const Device& device; VKScheduler& scheduler; - vk::Event event; - u64 ticks = 0; + u64 wait_tick = 0; }; using Fence = std::shared_ptr; @@ -58,7 +52,7 @@ public: explicit VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_, Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_, VKBufferCache& buffer_cache_, VKQueryCache& query_cache_, - const Device& device_, VKScheduler& scheduler_); + VKScheduler& scheduler_); protected: Fence CreateFence(u32 value, bool is_stubbed) override; @@ -68,7 +62,6 @@ protected: void WaitFence(Fence& fence) override; private: - const Device& device; VKScheduler& scheduler; }; diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 93fbea510..ce3db49bd 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -428,8 +428,7 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra buffer_cache(*this, gpu_memory, cpu_memory_, device, memory_manager, scheduler, stream_buffer, staging_pool), query_cache{*this, maxwell3d, gpu_memory, device, scheduler}, - fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, device, - scheduler), + fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, scheduler), wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) { scheduler.SetQueryCache(query_cache); if (device.UseAsynchronousShaders()) { -- cgit v1.2.3