vk_master_semaphore: Remove waitable atomic

* These are buggy on some platforms and regular condition_variables are faster most of the time
This commit is contained in:
GPUCode 2024-01-28 23:25:56 +02:00
parent b9c9beeee5
commit 2bcbfeb861
2 changed files with 27 additions and 36 deletions

View File

@ -5,7 +5,6 @@
#include <mutex> #include <mutex>
#include "video_core/renderer_vulkan/vk_instance.h" #include "video_core/renderer_vulkan/vk_instance.h"
#include "video_core/renderer_vulkan/vk_master_semaphore.h" #include "video_core/renderer_vulkan/vk_master_semaphore.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"
namespace Vulkan { namespace Vulkan {
@ -109,23 +108,20 @@ constexpr u64 FENCE_RESERVE = 8;
MasterSemaphoreFence::MasterSemaphoreFence(const Instance& instance_) : instance{instance_} { MasterSemaphoreFence::MasterSemaphoreFence(const Instance& instance_) : instance{instance_} {
const vk::Device device{instance.GetDevice()}; const vk::Device device{instance.GetDevice()};
for (u64 i = 0; i < FENCE_RESERVE; i++) { for (u64 i = 0; i < FENCE_RESERVE; i++) {
free_queue.push(device.createFenceUnique({})); free_queue.push_back(device.createFence({}));
} }
wait_thread = std::jthread([this](std::stop_token token) { WaitThread(token); }); wait_thread = std::jthread([this](std::stop_token token) { WaitThread(token); });
} }
MasterSemaphoreFence::~MasterSemaphoreFence() = default; MasterSemaphoreFence::~MasterSemaphoreFence() {
std::ranges::for_each(free_queue, [this](auto fence) { instance.GetDevice().destroyFence(fence); });
}
void MasterSemaphoreFence::Refresh() {} void MasterSemaphoreFence::Refresh() {}
void MasterSemaphoreFence::Wait(u64 tick) { void MasterSemaphoreFence::Wait(u64 tick) {
while (true) { std::unique_lock lk{free_mutex};
u64 current_value = gpu_tick.load(std::memory_order_relaxed); free_cv.wait(lk, [&] { return gpu_tick.load(std::memory_order_relaxed) >= tick; });
if (current_value >= tick) {
return;
}
gpu_tick.wait(current_value);
}
} }
void MasterSemaphoreFence::SubmitWork(vk::CommandBuffer cmdbuf, vk::Semaphore wait, void MasterSemaphoreFence::SubmitWork(vk::CommandBuffer cmdbuf, vk::Semaphore wait,
@ -149,59 +145,58 @@ void MasterSemaphoreFence::SubmitWork(vk::CommandBuffer cmdbuf, vk::Semaphore wa
.pSignalSemaphores = &signal, .pSignalSemaphores = &signal,
}; };
vk::UniqueFence fence{GetFreeFence()}; const vk::Fence fence = GetFreeFence();
try { try {
instance.GetGraphicsQueue().submit(submit_info, *fence); instance.GetGraphicsQueue().submit(submit_info, fence);
} catch (vk::DeviceLostError& err) { } catch (vk::DeviceLostError& err) {
LOG_CRITICAL(Render_Vulkan, "Device lost during submit: {}", err.what()); LOG_CRITICAL(Render_Vulkan, "Device lost during submit: {}", err.what());
UNREACHABLE(); UNREACHABLE();
} }
std::scoped_lock lock{wait_mutex}; std::scoped_lock lock{wait_mutex};
wait_queue.push({ wait_queue.emplace(fence, signal_value);
.handle = std::move(fence),
.signal_value = signal_value,
});
wait_cv.notify_one(); wait_cv.notify_one();
} }
void MasterSemaphoreFence::WaitThread(std::stop_token token) { void MasterSemaphoreFence::WaitThread(std::stop_token token) {
const vk::Device device{instance.GetDevice()}; const vk::Device device{instance.GetDevice()};
while (!token.stop_requested()) { while (!token.stop_requested()) {
Fence fence; vk::Fence fence;
u64 signal_value;
{ {
std::unique_lock lock{wait_mutex}; std::unique_lock lock{wait_mutex};
Common::CondvarWait(wait_cv, lock, token, [this] { return !wait_queue.empty(); }); Common::CondvarWait(wait_cv, lock, token, [this] { return !wait_queue.empty(); });
if (token.stop_requested()) { if (token.stop_requested()) {
return; return;
} }
fence = std::move(wait_queue.front()); std::tie(fence, signal_value) = wait_queue.front();
wait_queue.pop(); wait_queue.pop();
} }
const vk::Result result = device.waitForFences(*fence.handle, true, WAIT_TIMEOUT); const vk::Result result = device.waitForFences(fence, true, WAIT_TIMEOUT);
if (result != vk::Result::eSuccess) { if (result != vk::Result::eSuccess) {
LOG_CRITICAL(Render_Vulkan, "Fence wait failed with error {}", vk::to_string(result)); LOG_CRITICAL(Render_Vulkan, "Fence wait failed with error {}", vk::to_string(result));
UNREACHABLE(); UNREACHABLE();
} }
device.resetFences(*fence.handle);
gpu_tick.store(fence.signal_value); device.resetFences(fence);
gpu_tick.notify_all(); gpu_tick.store(signal_value);
std::scoped_lock lock{free_mutex}; std::scoped_lock lock{free_mutex};
free_queue.push(std::move(fence.handle)); free_queue.push_back(fence);
free_cv.notify_all();
} }
} }
vk::UniqueFence MasterSemaphoreFence::GetFreeFence() { vk::Fence MasterSemaphoreFence::GetFreeFence() {
std::scoped_lock lock{free_mutex}; std::scoped_lock lock{free_mutex};
if (free_queue.empty()) { if (free_queue.empty()) {
return instance.GetDevice().createFenceUnique({}); return instance.GetDevice().createFence({});
} }
vk::UniqueFence fence{std::move(free_queue.front())}; const vk::Fence fence = free_queue.front();
free_queue.pop(); free_queue.pop_front();
return fence; return fence;
} }

View File

@ -72,6 +72,7 @@ private:
}; };
class MasterSemaphoreFence : public MasterSemaphore { class MasterSemaphoreFence : public MasterSemaphore {
using Waitable = std::pair<vk::Fence, u64>;
public: public:
explicit MasterSemaphoreFence(const Instance& instance); explicit MasterSemaphoreFence(const Instance& instance);
~MasterSemaphoreFence() override; ~MasterSemaphoreFence() override;
@ -86,20 +87,15 @@ public:
private: private:
void WaitThread(std::stop_token token); void WaitThread(std::stop_token token);
vk::UniqueFence GetFreeFence(); vk::Fence GetFreeFence();
private: private:
const Instance& instance; const Instance& instance;
std::deque<vk::Fence> free_queue;
struct Fence { std::queue<Waitable> wait_queue;
vk::UniqueFence handle;
u64 signal_value;
};
std::queue<vk::UniqueFence> free_queue;
std::queue<Fence> wait_queue;
std::mutex free_mutex; std::mutex free_mutex;
std::mutex wait_mutex; std::mutex wait_mutex;
std::condition_variable free_cv;
std::condition_variable_any wait_cv; std::condition_variable_any wait_cv;
std::jthread wait_thread; std::jthread wait_thread;
}; };