Skip to content

Commit

Permalink
GS/Vulkan: Purge threaded presentation
Browse files Browse the repository at this point in the history
  • Loading branch information
@stenzek
stenzek committed May 27, 2024
1 parent d94f1dd commit c94282c
Show file tree
Hide file tree
Showing 8 changed files with 37 additions and 203 deletions.
7 changes: 0 additions & 7 deletions pcsx2-qt/Settings/GraphicsSettingsWidget.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -227,7 +227,6 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.useDebugDevice, "EmuCore/GS", "UseDebugDevice", false);
SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.skipPresentingDuplicateFrames, "EmuCore/GS", "SkipDuplicateFrames", false);
SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.disableMailboxPresentation, "EmuCore/GS", "DisableMailboxPresentation", false);
SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.threadedPresentation, "EmuCore/GS", "DisableThreadedPresentation", false);
SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.exclusiveFullscreenControl, "EmuCore/GS", "ExclusiveFullscreenControl", -1, -1);
SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.overrideTextureBarriers, "EmuCore/GS", "OverrideTextureBarriers", -1, -1);
SettingWidgetBinder::BindWidgetToIntSetting(
Expand Down Expand Up @@ -335,7 +334,6 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
m_ui.useBlitSwapChain = nullptr;
m_ui.disableMailboxPresentation = nullptr;
m_ui.skipPresentingDuplicateFrames = nullptr;
m_ui.threadedPresentation = nullptr;
m_ui.overrideTextureBarriers = nullptr;
m_ui.disableFramebufferFetch = nullptr;
m_ui.disableShaderCache = nullptr;
Expand Down Expand Up @@ -767,11 +765,6 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
tr("Forces the use of FIFO over Mailbox presentation, i.e. double buffering instead of triple buffering. "
"Usually results in worse frame pacing."));

dialog->registerWidgetHelp(m_ui.threadedPresentation, tr("Disable Threaded Presentation"), tr("Unchecked"),
tr("Presents frames on the main GS thread instead of a worker thread. Used for debugging frametime issues. "
"Could reduce chance of missing a frame or reduce tearing at the expense of more erratic frame times. "
"Only applies to the Vulkan renderer."));

dialog->registerWidgetHelp(m_ui.useDebugDevice, tr("Enable Debug Device"), tr("Unchecked"),
tr("Enables API-level validation of graphics commands."));

Expand Down
7 changes: 0 additions & 7 deletions pcsx2-qt/Settings/GraphicsSettingsWidget.ui
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2052,13 +2052,6 @@
</widget>
</item>
<item row="1" column="0">
<widget class="QCheckBox" name="threadedPresentation">
<property name="text">
<string>Disable Threaded Presentation</string>
</property>
</widget>
</item>
<item row="1" column="1">
<widget class="QCheckBox" name="disableMailboxPresentation">
<property name="text">
<string>Disable Mailbox Presentation</string>
Expand Down
1 change: 0 additions & 1 deletion pcsx2/Config.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -602,7 +602,6 @@ struct Pcsx2Config
DisableShaderCache : 1,
DisableFramebufferFetch : 1,
DisableVertexShaderExpand : 1,
DisableThreadedPresentation : 1,
SkipDuplicateFrames : 1,
OsdShowMessages : 1,
OsdShowSpeed : 1,
Expand Down
180 changes: 34 additions & 146 deletions pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1008,7 +1008,6 @@ void GSDeviceVK::WaitForFenceCounter(u64 fence_counter)

void GSDeviceVK::WaitForGPUIdle()
{
WaitForPresentComplete();
vkDeviceWaitIdle(m_device);
}

Expand Down Expand Up @@ -1050,19 +1049,12 @@ void GSDeviceVK::ScanForCommandBufferCompletion()

void GSDeviceVK::WaitForCommandBufferCompletion(u32 index)
{
// We might be waiting for the buffer we just submitted to the worker thread.
if (m_queued_present.command_buffer_index == index && !m_present_done.load(std::memory_order_acquire))
{
Console.WarningFmt("Waiting for threaded submission of cmdbuffer {}", index);
WaitForPresentComplete();
}

// Wait for this command buffer to be completed.
const VkResult res = vkWaitForFences(m_device, 1, &m_frame_resources[index].fence, VK_TRUE, UINT64_MAX);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkWaitForFences failed: ");
m_last_submit_failed.store(true, std::memory_order_release);
m_last_submit_failed = true;
return;
}

Expand All @@ -1085,8 +1077,7 @@ void GSDeviceVK::WaitForCommandBufferCompletion(u32 index)
m_completed_fence_counter = now_completed_counter;
}

void GSDeviceVK::SubmitCommandBuffer(
VKSwapChain* present_swap_chain /* = nullptr */, bool submit_on_thread /* = false */)
void GSDeviceVK::SubmitCommandBuffer(VKSwapChain* present_swap_chain)
{
FrameResources& resources = m_frame_resources[m_current_frame];

Expand Down Expand Up @@ -1154,33 +1145,9 @@ void GSDeviceVK::SubmitCommandBuffer(
if (spin_cycles != 0)
WaitForSpinCompletion(m_current_frame);

std::unique_lock<std::mutex> lock(m_present_mutex);
WaitForPresentComplete(lock);

if (spin_enabled && m_optional_extensions.vk_ext_calibrated_timestamps)
resources.submit_timestamp = GetCPUTimestamp();

// Don't use threaded presentation when spinning is enabled. ScanForCommandBufferCompletion()
// calls vkGetFenceStatus(), which reads a fence that has been passed off to the thread.
if (!submit_on_thread || GSConfig.HWSpinGPUForReadbacks || !m_present_thread.joinable())
{
DoSubmitCommandBuffer(m_current_frame, present_swap_chain, spin_cycles);
if (present_swap_chain)
DoPresent(present_swap_chain);
return;
}

m_queued_present.command_buffer_index = m_current_frame;
m_queued_present.swap_chain = present_swap_chain;
m_queued_present.spin_cycles = spin_cycles;
m_present_done.store(false, std::memory_order_release);
m_present_queued_cv.notify_one();
}

void GSDeviceVK::DoSubmitCommandBuffer(u32 index, VKSwapChain* present_swap_chain, u32 spin_cycles)
{
FrameResources& resources = m_frame_resources[index];

uint32_t wait_bits = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
VkSemaphore semas[2];
VkSubmitInfo submit_info = {VK_STRUCTURE_TYPE_SUBMIT_INFO};
Expand All @@ -1197,7 +1164,7 @@ void GSDeviceVK::DoSubmitCommandBuffer(u32 index, VKSwapChain* present_swap_chai
if (spin_cycles != 0)
{
semas[0] = present_swap_chain->GetRenderingFinishedSemaphore();
semas[1] = m_spin_resources[index].semaphore;
semas[1] = m_spin_resources[m_current_frame].semaphore;
submit_info.signalSemaphoreCount = 2;
submit_info.pSignalSemaphores = semas;
}
Expand All @@ -1210,105 +1177,44 @@ void GSDeviceVK::DoSubmitCommandBuffer(u32 index, VKSwapChain* present_swap_chai
else if (spin_cycles != 0)
{
submit_info.signalSemaphoreCount = 1;
submit_info.pSignalSemaphores = &m_spin_resources[index].semaphore;
submit_info.pSignalSemaphores = &m_spin_resources[m_current_frame].semaphore;
}

const VkResult res = vkQueueSubmit(m_graphics_queue, 1, &submit_info, resources.fence);
res = vkQueueSubmit(m_graphics_queue, 1, &submit_info, resources.fence);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkQueueSubmit failed: ");
m_last_submit_failed.store(true, std::memory_order_release);
m_last_submit_failed = true;
return;
}

if (spin_cycles != 0)
SubmitSpinCommand(index, spin_cycles);
}

void GSDeviceVK::DoPresent(VKSwapChain* present_swap_chain)
{
const VkPresentInfoKHR present_info = {VK_STRUCTURE_TYPE_PRESENT_INFO_KHR, nullptr, 1,
present_swap_chain->GetRenderingFinishedSemaphorePtr(), 1, present_swap_chain->GetSwapChainPtr(),
present_swap_chain->GetCurrentImageIndexPtr(), nullptr};

present_swap_chain->ReleaseCurrentImage();

const VkResult res = vkQueuePresentKHR(m_present_queue, &present_info);
if (res != VK_SUCCESS)
{
// VK_ERROR_OUT_OF_DATE_KHR is not fatal, just means we need to recreate our swap chain.
if (res != VK_ERROR_OUT_OF_DATE_KHR && res != VK_SUBOPTIMAL_KHR)
LOG_VULKAN_ERROR(res, "vkQueuePresentKHR failed: ");

m_last_present_failed.store(true, std::memory_order_release);
return;
}

// Grab the next image as soon as possible, that way we spend less time blocked on the next
// submission. Don't care if it fails, we'll deal with that at the presentation call site.
// Credit to dxvk for the idea.
present_swap_chain->AcquireNextImage();
}

void GSDeviceVK::WaitForPresentComplete()
{
if (m_present_done.load(std::memory_order_acquire))
return;

std::unique_lock<std::mutex> lock(m_present_mutex);
WaitForPresentComplete(lock);
}

void GSDeviceVK::WaitForPresentComplete(std::unique_lock<std::mutex>& lock)
{
if (m_present_done.load(std::memory_order_acquire))
return;
SubmitSpinCommand(m_current_frame, spin_cycles);

m_present_done_cv.wait(lock, [this]() { return m_present_done.load(std::memory_order_acquire); });
}

void GSDeviceVK::PresentThread()
{
std::unique_lock<std::mutex> lock(m_present_mutex);
while (!m_present_thread_done.load(std::memory_order_acquire))
if (present_swap_chain)
{
m_present_queued_cv.wait(lock, [this]() {
return !m_present_done.load(std::memory_order_acquire) ||
m_present_thread_done.load(std::memory_order_acquire);
});
const VkPresentInfoKHR present_info = {VK_STRUCTURE_TYPE_PRESENT_INFO_KHR, nullptr, 1,
present_swap_chain->GetRenderingFinishedSemaphorePtr(), 1, present_swap_chain->GetSwapChainPtr(),
present_swap_chain->GetCurrentImageIndexPtr(), nullptr};

if (m_present_done.load(std::memory_order_acquire))
continue;

DoSubmitCommandBuffer(
m_queued_present.command_buffer_index, m_queued_present.swap_chain, m_queued_present.spin_cycles);
if (m_queued_present.swap_chain)
DoPresent(m_queued_present.swap_chain);
m_present_done.store(true, std::memory_order_release);
m_present_done_cv.notify_one();
}
}
present_swap_chain->ReleaseCurrentImage();

void GSDeviceVK::StartPresentThread()
{
pxAssert(!m_present_thread.joinable());
m_present_thread_done.store(false, std::memory_order_release);
m_present_thread = std::thread(&GSDeviceVK::PresentThread, this);
}
const VkResult res = vkQueuePresentKHR(m_present_queue, &present_info);
if (res != VK_SUCCESS)
{
// VK_ERROR_OUT_OF_DATE_KHR is not fatal, just means we need to recreate our swap chain.
if (res != VK_ERROR_OUT_OF_DATE_KHR && res != VK_SUBOPTIMAL_KHR)
LOG_VULKAN_ERROR(res, "vkQueuePresentKHR failed: ");

void GSDeviceVK::StopPresentThread()
{
if (!m_present_thread.joinable())
return;
m_last_present_failed = true;
return;
}

{
std::unique_lock<std::mutex> lock(m_present_mutex);
WaitForPresentComplete(lock);
m_present_thread_done.store(true, std::memory_order_release);
m_present_queued_cv.notify_one();
// Grab the next image as soon as possible, that way we spend less time blocked on the next
// submission. Don't care if it fails, we'll deal with that at the presentation call site.
// Credit to dxvk for the idea.
present_swap_chain->AcquireNextImage();
}

m_present_thread.join();
}

void GSDeviceVK::CommandBufferCompleted(u32 index)
Expand Down Expand Up @@ -1411,12 +1317,11 @@ void GSDeviceVK::ActivateCommandBuffer(u32 index)

void GSDeviceVK::ExecuteCommandBuffer(WaitType wait_for_completion)
{
if (m_last_submit_failed.load(std::memory_order_acquire))
if (m_last_submit_failed)
return;

// If we're waiting for completion, don't bother waking the worker thread.
const u32 current_frame = m_current_frame;
SubmitCommandBuffer();
SubmitCommandBuffer(nullptr);
MoveToNextCommandBuffer();

if (wait_for_completion != WaitType::None)
Expand All @@ -1433,16 +1338,6 @@ void GSDeviceVK::ExecuteCommandBuffer(WaitType wait_for_completion)
}
}

bool GSDeviceVK::CheckLastPresentFail()
{
return m_last_present_failed.exchange(false, std::memory_order_acq_rel);
}

bool GSDeviceVK::CheckLastSubmitFail()
{
return m_last_submit_failed.load(std::memory_order_acquire);
}

void GSDeviceVK::DeferBufferDestruction(VkBuffer object, VmaAllocation allocation)
{
FrameResources& resources = m_frame_resources[m_current_frame];
Expand Down Expand Up @@ -1809,7 +1704,7 @@ void GSDeviceVK::WaitForSpinCompletion(u32 index)
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkWaitForFences failed: ");
m_last_submit_failed.store(true, std::memory_order_release);
m_last_submit_failed = true;
return;
}
SpinCommandCompleted(index);
Expand Down Expand Up @@ -2169,7 +2064,6 @@ void GSDeviceVK::Destroy()
WaitForGPUIdle();
}

StopPresentThread();
m_swap_chain.reset();

DestroySpinResources();
Expand Down Expand Up @@ -2334,6 +2228,10 @@ GSDevice::PresentResult GSDeviceVK::BeginPresent(bool frame_skip)
{
EndRenderPass();

// Check if the device was lost.
if (m_last_submit_failed)
return PresentResult::DeviceLost;

if (frame_skip)
return PresentResult::FrameSkipped;

Expand All @@ -2344,13 +2242,6 @@ GSDevice::PresentResult GSDeviceVK::BeginPresent(bool frame_skip)
return PresentResult::FrameSkipped;
}

// Previous frame needs to be presented before we can acquire the swap chain.
WaitForPresentComplete();

// Check if the device was lost.
if (CheckLastSubmitFail())
return PresentResult::DeviceLost;

VkResult res = m_swap_chain->AcquireNextImage();
if (res != VK_SUCCESS)
{
Expand Down Expand Up @@ -2422,7 +2313,7 @@ void GSDeviceVK::EndPresent()
m_swap_chain->GetCurrentTexture()->TransitionToLayout(cmdbuffer, GSTextureVK::Layout::PresentSrc);
g_perfmon.Put(GSPerfMon::RenderPasses, 1);

SubmitCommandBuffer(m_swap_chain.get(), !m_swap_chain->IsPresentModeSynchronizing());
SubmitCommandBuffer(m_swap_chain.get());
MoveToNextCommandBuffer();

InvalidateCachedState();
Expand Down Expand Up @@ -2621,9 +2512,6 @@ bool GSDeviceVK::CreateDeviceAndSwapChain()

VKShaderCache::Create();

if (!GSConfig.DisableThreadedPresentation)
StartPresentThread();

if (surface != VK_NULL_HANDLE)
{
VkPresentModeKHR present_mode;
Expand Down Expand Up @@ -4554,7 +4442,7 @@ void GSDeviceVK::RenderBlankFrame()
cmdbuffer, sctex->GetImage(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, &s_present_clear_color.color, 1, &srr);

m_swap_chain->GetCurrentTexture()->TransitionToLayout(cmdbuffer, GSTextureVK::Layout::PresentSrc);
SubmitCommandBuffer(m_swap_chain.get(), !m_swap_chain->IsPresentModeSynchronizing());
SubmitCommandBuffer(m_swap_chain.get());
ActivateCommandBuffer((m_current_frame + 1) % NUM_COMMAND_BUFFERS);
}

Expand Down
Loading

0 comments on commit c94282c

Please sign in to comment.