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.
This commit is contained in:
ReinUsesLisp 2021-01-08 02:37:02 -03:00
parent aaf9e39f56
commit 154a7653f9
3 changed files with 18 additions and 54 deletions

View file

@ -3,7 +3,6 @@
// Refer to the license.txt file included. // Refer to the license.txt file included.
#include <memory> #include <memory>
#include <thread>
#include "video_core/renderer_vulkan/vk_buffer_cache.h" #include "video_core/renderer_vulkan/vk_buffer_cache.h"
#include "video_core/renderer_vulkan/vk_fence_manager.h" #include "video_core/renderer_vulkan/vk_fence_manager.h"
@ -14,13 +13,11 @@
namespace Vulkan { namespace Vulkan {
InnerFence::InnerFence(const Device& device_, VKScheduler& scheduler_, u32 payload_, InnerFence::InnerFence(VKScheduler& scheduler_, u32 payload_, bool is_stubbed_)
bool is_stubbed_) : FenceBase{payload_, is_stubbed_}, scheduler{scheduler_} {}
: FenceBase{payload_, is_stubbed_}, device{device_}, scheduler{scheduler_} {}
InnerFence::InnerFence(const Device& device_, VKScheduler& scheduler_, GPUVAddr address_, InnerFence::InnerFence(VKScheduler& scheduler_, GPUVAddr address_, u32 payload_, bool is_stubbed_)
u32 payload_, bool is_stubbed_) : FenceBase{address_, payload_, is_stubbed_}, scheduler{scheduler_} {}
: FenceBase{address_, payload_, is_stubbed_}, device{device_}, scheduler{scheduler_} {}
InnerFence::~InnerFence() = default; InnerFence::~InnerFence() = default;
@ -28,63 +25,38 @@ void InnerFence::Queue() {
if (is_stubbed) { if (is_stubbed) {
return; return;
} }
ASSERT(!event); // Get the current tick so we can wait for it
wait_tick = scheduler.CurrentTick();
event = device.GetLogical().CreateEvent(); scheduler.Flush();
ticks = scheduler.CurrentTick();
scheduler.RequestOutsideRenderPassOperationContext();
scheduler.Record([event = *event](vk::CommandBuffer cmdbuf) {
cmdbuf.SetEvent(event, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT);
});
} }
bool InnerFence::IsSignaled() const { bool InnerFence::IsSignaled() const {
if (is_stubbed) { if (is_stubbed) {
return true; return true;
} }
ASSERT(event); return scheduler.IsFree(wait_tick);
return IsEventSignalled();
} }
void InnerFence::Wait() { void InnerFence::Wait() {
if (is_stubbed) { if (is_stubbed) {
return; return;
} }
ASSERT(event); scheduler.Wait(wait_tick);
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);
}
} }
VKFenceManager::VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_, VKFenceManager::VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_, Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_,
VKBufferCache& buffer_cache_, VKQueryCache& query_cache_, VKBufferCache& buffer_cache_, VKQueryCache& query_cache_,
const Device& device_, VKScheduler& scheduler_) VKScheduler& scheduler_)
: GenericFenceManager{rasterizer_, gpu_, texture_cache_, buffer_cache_, query_cache_}, : GenericFenceManager{rasterizer_, gpu_, texture_cache_, buffer_cache_, query_cache_},
device{device_}, scheduler{scheduler_} {} scheduler{scheduler_} {}
Fence VKFenceManager::CreateFence(u32 value, bool is_stubbed) { Fence VKFenceManager::CreateFence(u32 value, bool is_stubbed) {
return std::make_shared<InnerFence>(device, scheduler, value, is_stubbed); return std::make_shared<InnerFence>(scheduler, value, is_stubbed);
} }
Fence VKFenceManager::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) { Fence VKFenceManager::CreateFence(GPUVAddr addr, u32 value, bool is_stubbed) {
return std::make_shared<InnerFence>(device, scheduler, addr, value, is_stubbed); return std::make_shared<InnerFence>(scheduler, addr, value, is_stubbed);
} }
void VKFenceManager::QueueFence(Fence& fence) { void VKFenceManager::QueueFence(Fence& fence) {

View file

@ -28,10 +28,8 @@ class VKScheduler;
class InnerFence : public VideoCommon::FenceBase { class InnerFence : public VideoCommon::FenceBase {
public: public:
explicit InnerFence(const Device& device_, VKScheduler& scheduler_, u32 payload_, explicit InnerFence(VKScheduler& scheduler_, u32 payload_, bool is_stubbed_);
bool is_stubbed_); explicit InnerFence(VKScheduler& scheduler_, GPUVAddr address_, u32 payload_, bool is_stubbed_);
explicit InnerFence(const Device& device_, VKScheduler& scheduler_, GPUVAddr address_,
u32 payload_, bool is_stubbed_);
~InnerFence(); ~InnerFence();
void Queue(); void Queue();
@ -41,12 +39,8 @@ public:
void Wait(); void Wait();
private: private:
bool IsEventSignalled() const;
const Device& device;
VKScheduler& scheduler; VKScheduler& scheduler;
vk::Event event; u64 wait_tick = 0;
u64 ticks = 0;
}; };
using Fence = std::shared_ptr<InnerFence>; using Fence = std::shared_ptr<InnerFence>;
@ -58,7 +52,7 @@ public:
explicit VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_, explicit VKFenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_, Tegra::MemoryManager& memory_manager_, TextureCache& texture_cache_,
VKBufferCache& buffer_cache_, VKQueryCache& query_cache_, VKBufferCache& buffer_cache_, VKQueryCache& query_cache_,
const Device& device_, VKScheduler& scheduler_); VKScheduler& scheduler_);
protected: protected:
Fence CreateFence(u32 value, bool is_stubbed) override; Fence CreateFence(u32 value, bool is_stubbed) override;
@ -68,7 +62,6 @@ protected:
void WaitFence(Fence& fence) override; void WaitFence(Fence& fence) override;
private: private:
const Device& device;
VKScheduler& scheduler; VKScheduler& scheduler;
}; };

View file

@ -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, buffer_cache(*this, gpu_memory, cpu_memory_, device, memory_manager, scheduler, stream_buffer,
staging_pool), staging_pool),
query_cache{*this, maxwell3d, gpu_memory, device, scheduler}, query_cache{*this, maxwell3d, gpu_memory, device, scheduler},
fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, device, fence_manager(*this, gpu, gpu_memory, texture_cache, buffer_cache, query_cache, scheduler),
scheduler),
wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) { wfi_event(device.GetLogical().CreateEvent()), async_shaders(emu_window_) {
scheduler.SetQueryCache(query_cache); scheduler.SetQueryCache(query_cache);
if (device.UseAsynchronousShaders()) { if (device.UseAsynchronousShaders()) {