Skip to content
Permalink
Browse files

Vulkan: Simplify command buffer fence tracking

  • Loading branch information...
stenzek committed Mar 17, 2019
1 parent f3fadd7 commit 427dd45151eddc920d681c4711a3cdbd7e363f82
@@ -54,7 +54,6 @@ bool CommandBufferManager::CreateCommandBuffers()
{
resources.init_command_buffer_used = false;
resources.semaphore_used = false;
resources.needs_fence_wait = false;

VkCommandPoolCreateInfo pool_info = {VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO, nullptr, 0,
g_vulkan_context->GetGraphicsQueueFamilyIndex()};
@@ -211,43 +210,61 @@ void CommandBufferManager::WaitForWorkerThreadIdle()
m_submit_semaphore.Post();
}

void CommandBufferManager::WaitForGPUIdle()
void CommandBufferManager::WaitForFenceCounter(u64 fence_counter)
{
WaitForWorkerThreadIdle();
vkDeviceWaitIdle(g_vulkan_context->GetDevice());
}
if (m_completed_fence_counter >= fence_counter)
return;

void CommandBufferManager::WaitForFence(VkFence fence)
{
// Find the command buffer that this fence corresponds to.
u32 command_buffer_index = 0;
for (; command_buffer_index < static_cast<u32>(m_frame_resources.size()); command_buffer_index++)
// Find the first command buffer which covers this counter value.
u32 index = (m_current_frame + 1) % NUM_COMMAND_BUFFERS;
while (index != m_current_frame)
{
if (m_frame_resources[command_buffer_index].fence == fence)
if (m_frame_resources[index].fence_counter >= fence_counter)
break;

index = (index + 1) % NUM_COMMAND_BUFFERS;
}
ASSERT(command_buffer_index < m_frame_resources.size());

// Has this command buffer already been waited for?
if (!m_frame_resources[command_buffer_index].needs_fence_wait)
return;
ASSERT(index != m_current_frame);
WaitForCommandBufferCompletion(index);
}

void CommandBufferManager::WaitForCommandBufferCompletion(u32 index)
{
// Ensure this command buffer has been submitted.
WaitForWorkerThreadIdle();

// Wait for this command buffer to be completed.
VkResult res =
vkWaitForFences(g_vulkan_context->GetDevice(), 1,
&m_frame_resources[command_buffer_index].fence, VK_TRUE, UINT64_MAX);
VkResult res = vkWaitForFences(g_vulkan_context->GetDevice(), 1, &m_frame_resources[index].fence,
VK_TRUE, UINT64_MAX);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkWaitForFences failed: ");

// Immediately fire callbacks and cleanups, since the commands has been completed.
m_frame_resources[command_buffer_index].needs_fence_wait = false;
OnCommandBufferExecuted(command_buffer_index);
// Clean up any resources for command buffers between the last known completed buffer and this
// now-completed command buffer. If we use >2 buffers, this may be more than one buffer.
const u64 now_completed_counter = m_frame_resources[index].fence_counter;
u32 cleanup_index = (m_current_frame + 1) % NUM_COMMAND_BUFFERS;
while (cleanup_index != m_current_frame)
{
FrameResources& resources = m_frame_resources[cleanup_index];
if (resources.fence_counter > now_completed_counter)
break;

if (resources.fence_counter > m_completed_fence_counter)
{
for (auto& it : resources.cleanup_resources)
it();
resources.cleanup_resources.clear();
}

cleanup_index = (cleanup_index + 1) % NUM_COMMAND_BUFFERS;
}

m_completed_fence_counter = now_completed_counter;
}

void CommandBufferManager::SubmitCommandBuffer(bool submit_on_worker_thread,
bool wait_for_completion,
VkSwapchainKHR present_swap_chain,
uint32_t present_image_index)
{
@@ -263,16 +280,13 @@ void CommandBufferManager::SubmitCommandBuffer(bool submit_on_worker_thread,
}
}

// This command buffer now has commands, so can't be re-used without waiting.
resources.needs_fence_wait = true;

// Grab the semaphore before submitting command buffer either on-thread or off-thread.
// This prevents a race from occurring where a second command buffer is executed
// before the worker thread has woken and executed the first one yet.
m_submit_semaphore.Wait();

// Submitting off-thread?
if (m_use_threaded_submission && submit_on_worker_thread)
if (m_use_threaded_submission && submit_on_worker_thread && !wait_for_completion)
{
// Push to the pending submit queue.
{
@@ -287,6 +301,8 @@ void CommandBufferManager::SubmitCommandBuffer(bool submit_on_worker_thread,
{
// Pass through to normal submission path.
SubmitCommandBuffer(m_current_frame, present_swap_chain, present_image_index);
if (wait_for_completion)
WaitForCommandBufferCompletion(m_current_frame);
}

// Switch to next cmdbuffer.
@@ -365,39 +381,15 @@ void CommandBufferManager::SubmitCommandBuffer(u32 command_buffer_index,
m_submit_semaphore.Post();
}

void CommandBufferManager::OnCommandBufferExecuted(u32 index)
{
FrameResources& resources = m_frame_resources[index];

// Fire fence tracking callbacks.
for (auto iter = m_fence_callbacks.begin(); iter != m_fence_callbacks.end();)
{
auto backup_iter = iter++;
backup_iter->second(resources.fence);
}

// Clean up all objects pending destruction on this command buffer
for (auto& it : resources.cleanup_resources)
it();
resources.cleanup_resources.clear();
}

void CommandBufferManager::BeginCommandBuffer()
{
// Move to the next command buffer.
m_current_frame = (m_current_frame + 1) % NUM_COMMAND_BUFFERS;
FrameResources& resources = m_frame_resources[m_current_frame];
const u32 next_buffer_index = (m_current_frame + 1) % NUM_COMMAND_BUFFERS;
FrameResources& resources = m_frame_resources[next_buffer_index];

// Wait for the GPU to finish with all resources for this command buffer.
if (resources.needs_fence_wait)
{
VkResult res =
vkWaitForFences(g_vulkan_context->GetDevice(), 1, &resources.fence, true, UINT64_MAX);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkWaitForFences failed: ");

OnCommandBufferExecuted(m_current_frame);
}
if (resources.fence_counter > m_completed_fence_counter)
WaitForCommandBufferCompletion(next_buffer_index);

// Reset fence to unsignaled before starting.
VkResult res = vkResetFences(g_vulkan_context->GetDevice(), 1, &resources.fence);
@@ -427,6 +419,8 @@ void CommandBufferManager::BeginCommandBuffer()
// Reset upload command buffer state
resources.init_command_buffer_used = false;
resources.semaphore_used = false;
resources.fence_counter = m_next_fence_counter++;
m_current_frame = next_buffer_index;
}

void CommandBufferManager::DeferBufferDestruction(VkBuffer object)
@@ -471,19 +465,5 @@ void CommandBufferManager::DeferImageViewDestruction(VkImageView object)
[object]() { vkDestroyImageView(g_vulkan_context->GetDevice(), object, nullptr); });
}

void CommandBufferManager::AddFenceSignaledCallback(const void* key, FenceSignaledCallback callback)
{
// Shouldn't be adding twice.
ASSERT(m_fence_callbacks.find(key) == m_fence_callbacks.end());
m_fence_callbacks.emplace(key, std::move(callback));
}

void CommandBufferManager::RemoveFenceSignaledCallback(const void* key)
{
auto iter = m_fence_callbacks.find(key);
ASSERT(iter != m_fence_callbacks.end());
m_fence_callbacks.erase(iter);
}

std::unique_ptr<CommandBufferManager> g_command_buffer_mgr;
} // namespace Vulkan
@@ -51,9 +51,15 @@ class CommandBufferManager
// Allocates a descriptors set from the pool reserved for the current frame.
VkDescriptorSet AllocateDescriptorSet(VkDescriptorSetLayout set_layout);

// Fence "counters" are used to track which commands have been completed by the GPU.
// If the last completed fence counter is greater or equal to N, it means that the work
// associated counter N has been completed by the GPU. The value of N to associate with
// commands can be retreived by calling GetCurrentFenceCounter().
u64 GetCompletedFenceCounter() const { return m_completed_fence_counter; }

// Gets the fence that will be signaled when the currently executing command buffer is
// queued and executed. Do not wait for this fence before the buffer is executed.
VkFence GetCurrentCommandBufferFence() const { return m_frame_resources[m_current_frame].fence; }
u64 GetCurrentFenceCounter() const { return m_frame_resources[m_current_frame].fence_counter; }

// Returns the semaphore for the current command buffer, which can be used to ensure the
// swap chain image is ready before the command buffer executes.
@@ -66,15 +72,11 @@ class CommandBufferManager
// Ensure that the worker thread has submitted any previous command buffers and is idle.
void WaitForWorkerThreadIdle();

// Ensure that the worker thread has both submitted all commands, and the GPU has caught up.
// Use with caution, huge performance penalty.
void WaitForGPUIdle();

// Wait for a fence to be completed.
// Also invokes callbacks for completion.
void WaitForFence(VkFence fence);
void WaitForFenceCounter(u64 fence_counter);

void SubmitCommandBuffer(bool submit_on_worker_thread,
void SubmitCommandBuffer(bool submit_on_worker_thread, bool wait_for_completion,
VkSwapchainKHR present_swap_chain = VK_NULL_HANDLE,
uint32_t present_image_index = 0xFFFFFFFF);

@@ -90,25 +92,17 @@ class CommandBufferManager
void DeferImageDestruction(VkImage object);
void DeferImageViewDestruction(VkImageView object);

// Instruct the manager to fire the specified callback when a fence is flagged to be signaled.
// This happens when command buffers are executed, and can be tested if signaled, which means
// that all commands up to the point when the callback was fired have completed.
using FenceSignaledCallback = std::function<void(VkFence)>;
void AddFenceSignaledCallback(const void* key, FenceSignaledCallback callback);
void RemoveFenceSignaledCallback(const void* key);

private:
bool CreateCommandBuffers();
void DestroyCommandBuffers();

bool CreateSubmitThread();

void WaitForCommandBufferCompletion(u32 command_buffer_index);
void SubmitCommandBuffer(u32 command_buffer_index, VkSwapchainKHR present_swap_chain,
u32 present_image_index);
void BeginCommandBuffer();

void OnCommandBufferExecuted(u32 index);

struct FrameResources
{
// [0] - Init (upload) command buffer, [1] - draw command buffer
@@ -117,19 +111,19 @@ class CommandBufferManager
VkDescriptorPool descriptor_pool = VK_NULL_HANDLE;
VkFence fence = VK_NULL_HANDLE;
VkSemaphore semaphore = VK_NULL_HANDLE;
u64 fence_counter = 0;
bool init_command_buffer_used = false;
bool semaphore_used = false;
bool needs_fence_wait = false;

std::vector<std::function<void()>> cleanup_resources;
};

u64 m_next_fence_counter = 1;
u64 m_completed_fence_counter = 0;

std::array<FrameResources, NUM_COMMAND_BUFFERS> m_frame_resources;
u32 m_current_frame;

// callbacks when a fence point is set
std::map<const void*, FenceSignaledCallback> m_fence_callbacks;

// Threaded command buffer execution
// Semaphore determines when a command buffer can be queued
Common::Semaphore m_submit_semaphore;
@@ -24,8 +24,6 @@ PerfQuery::PerfQuery() = default;

PerfQuery::~PerfQuery()
{
g_command_buffer_mgr->RemoveFenceSignaledCallback(this);

if (m_query_pool != VK_NULL_HANDLE)
vkDestroyQueryPool(g_vulkan_context->GetDevice(), m_query_pool, nullptr);
}
@@ -49,9 +47,6 @@ bool PerfQuery::Initialize()
return false;
}

g_command_buffer_mgr->AddFenceSignaledCallback(
this, std::bind(&PerfQuery::OnFenceSignaled, this, std::placeholders::_1));

return true;
}

@@ -113,7 +108,7 @@ void PerfQuery::ResetQuery()

for (auto& entry : m_query_buffer)
{
entry.pending_fence = VK_NULL_HANDLE;
entry.fence_counter = 0;
entry.available = false;
entry.active = false;
}
@@ -217,7 +212,7 @@ void PerfQuery::QueueCopyQueryResults(u32 start_index, u32 query_count)
{
u32 index = start_index + i;
ActiveQuery& entry = m_query_buffer[index];
entry.pending_fence = g_command_buffer_mgr->GetCurrentCommandBufferFence();
entry.fence_counter = g_command_buffer_mgr->GetCurrentFenceCounter();
entry.available = true;
entry.active = false;
}
@@ -261,8 +256,10 @@ void PerfQuery::FlushQueries()
QueueCopyQueryResults(copy_start_index, copy_count);
}

void PerfQuery::OnFenceSignaled(VkFence fence)
void PerfQuery::ProcessPendingResults()
{
const u64 completed_fence_counter = g_command_buffer_mgr->GetCurrentFenceCounter();

// Need to save these since ProcessResults will modify them.
u32 query_read_pos = m_query_read_pos;
u32 query_count = m_query_count;
@@ -273,7 +270,7 @@ void PerfQuery::OnFenceSignaled(VkFence fence)
for (u32 i = 0; i < query_count; i++)
{
u32 index = (query_read_pos + i) % PERF_QUERY_BUFFER_SIZE;
if (m_query_buffer[index].pending_fence != fence)
if (m_query_buffer[index].fence_counter > completed_fence_counter)
{
// These should be grouped together, at the end.
break;
@@ -314,8 +311,8 @@ void PerfQuery::ProcessResults(u32 start_index, u32 query_count)
ActiveQuery& entry = m_query_buffer[index];

// Should have a fence associated with it (waiting for a result).
ASSERT(entry.pending_fence != VK_NULL_HANDLE);
entry.pending_fence = VK_NULL_HANDLE;
ASSERT(entry.fence_counter != 0);
entry.fence_counter = 0;
entry.available = false;
entry.active = false;

@@ -340,9 +337,11 @@ void PerfQuery::NonBlockingPartialFlush()
return;

// Submit a command buffer in the background if the front query is not bound to one.
// Ideally this will complete before the buffer fills.
if (m_query_buffer[m_query_read_pos].pending_fence == VK_NULL_HANDLE)
ActiveQuery& entry = m_query_buffer[m_query_read_pos];
if (entry.fence_counter == g_command_buffer_mgr->GetCurrentFenceCounter())
Renderer::GetInstance()->ExecuteCommandBuffer(true, false);

ProcessPendingResults();
}

void PerfQuery::BlockingPartialFlush()
@@ -352,17 +351,9 @@ void PerfQuery::BlockingPartialFlush()

// If the first pending query is needing command buffer execution, do that.
ActiveQuery& entry = m_query_buffer[m_query_read_pos];
if (entry.pending_fence == VK_NULL_HANDLE)
{
// This will callback OnCommandBufferQueued which will set the fence on the entry.
// We wait for completion, which will also call OnCommandBufferExecuted, and clear the fence.
if (entry.fence_counter == g_command_buffer_mgr->GetCurrentFenceCounter())
Renderer::GetInstance()->ExecuteCommandBuffer(false, true);
}
else
{
// The command buffer has been submitted, but is awaiting completion.
// Wait for the fence to complete, which will call OnCommandBufferExecuted.
g_command_buffer_mgr->WaitForFence(entry.pending_fence);
}

ProcessPendingResults();
}
} // namespace Vulkan

0 comments on commit 427dd45

Please sign in to comment.
You can’t perform that action at this time.