[dxvk] Implement better priority system for background shader compiles

Reduces the number of workers that perform background optimization,
which may reduce the performance impact when encountering a large
number of new pipelines at once.
This commit is contained in:
Philip Rebohle 2023-01-13 14:20:27 +01:00
parent cc9266edaa
commit c978e62ec8
4 changed files with 103 additions and 118 deletions

View File

@ -938,7 +938,7 @@ namespace dxvk {
// If necessary, compile an optimized pipeline variant
if (!instance->fastHandle.load())
m_workers->compileGraphicsPipeline(this, state);
m_workers->compileGraphicsPipeline(this, state, DxvkPipelinePriority::Low);
// Only store pipelines in the state cache that cannot benefit
// from pipeline libraries, or if that feature is disabled.

View File

@ -21,40 +21,28 @@ namespace dxvk {
void DxvkPipelineWorkers::compilePipelineLibrary(
DxvkShaderPipelineLibrary* library,
DxvkPipelinePriority priority) {
std::unique_lock lock(m_queueLock);
std::unique_lock lock(m_lock);
this->startWorkers();
m_pendingTasks += 1;
PipelineLibraryEntry e = { };
e.pipelineLibrary = library;
if (priority == DxvkPipelinePriority::High) {
m_queuedLibrariesPrioritized.push(e);
m_queueCondPrioritized.notify_one();
} else {
m_queuedLibraries.push(e);
}
m_queueCond.notify_one();
m_buckets[uint32_t(priority)].queue.emplace(library);
notifyWorkers(priority);
}
void DxvkPipelineWorkers::compileGraphicsPipeline(
DxvkGraphicsPipeline* pipeline,
const DxvkGraphicsPipelineStateInfo& state) {
std::unique_lock lock(m_queueLock);
const DxvkGraphicsPipelineStateInfo& state,
DxvkPipelinePriority priority) {
std::unique_lock lock(m_lock);
this->startWorkers();
pipeline->acquirePipeline();
m_pendingTasks += 1;
PipelineEntry e = { };
e.graphicsPipeline = pipeline;
e.graphicsState = state;
m_queuedPipelines.push(e);
m_queueCond.notify_one();
m_buckets[uint32_t(priority)].queue.emplace(pipeline, state);
notifyWorkers(priority);
}
@ -64,14 +52,15 @@ namespace dxvk {
void DxvkPipelineWorkers::stopWorkers() {
{ std::unique_lock lock(m_queueLock);
{ std::unique_lock lock(m_lock);
if (!m_workersRunning)
return;
m_workersRunning = false;
m_queueCond.notify_all();
m_queueCondPrioritized.notify_all();
for (uint32_t i = 0; i < m_buckets.size(); i++)
m_buckets[i].cond.notify_all();
}
for (auto& worker : m_workers)
@ -81,8 +70,23 @@ namespace dxvk {
}
void DxvkPipelineWorkers::notifyWorkers(DxvkPipelinePriority priority) {
uint32_t index = uint32_t(priority);
// If any workers are idle in a suitable set, notify the corresponding
// condition variable. If all workers are busy anyway, we know that the
// job is going to be picked up at some point anyway.
for (uint32_t i = index; i < m_buckets.size(); i++) {
if (m_buckets[i].idleWorkers) {
m_buckets[i].cond.notify_one();
break;
}
}
}
void DxvkPipelineWorkers::startWorkers() {
if (!m_workersRunning) {
if (!std::exchange(m_workersRunning, true)) {
// Use all available cores by default
uint32_t workerCount = dxvk::thread::hardware_concurrency();
@ -98,102 +102,74 @@ namespace dxvk {
// Number of workers that can process pipeline pipelines with normal
// priority. Any other workers can only build high-priority pipelines.
uint32_t npWorkerCount = m_device->canUseGraphicsPipelineLibrary()
? std::max(((workerCount - 1) * 5) / 7, 1u)
: workerCount;
uint32_t hpWorkerCount = workerCount - npWorkerCount;
uint32_t npWorkerCount = std::max(((workerCount - 1) * 5) / 7, 1u);
uint32_t lpWorkerCount = std::max(((workerCount - 1) * 2) / 7, 1u);
Logger::info(str::format("DXVK: Using ", npWorkerCount, " + ", hpWorkerCount, " compiler threads"));
m_workers.resize(npWorkerCount + hpWorkerCount);
m_workers.reserve(workerCount);
// Set worker flag so that they don't exit immediately
m_workersRunning = true;
for (size_t i = 0; i < workerCount; i++) {
DxvkPipelinePriority priority = DxvkPipelinePriority::Normal;
for (size_t i = 0; i < m_workers.size(); i++) {
m_workers[i] = i >= npWorkerCount
? dxvk::thread([this] { runWorkerPrioritized(); })
: dxvk::thread([this] { runWorker(); });
m_workers[i].set_priority(ThreadPriority::Lowest);
if (m_device->canUseGraphicsPipelineLibrary()) {
if (i >= npWorkerCount)
priority = DxvkPipelinePriority::High;
else if (i < lpWorkerCount)
priority = DxvkPipelinePriority::Low;
}
m_workers.emplace_back([this, priority] {
runWorker(priority);
});
}
Logger::info(str::format("DXVK: Using ", workerCount, " compiler threads"));
}
}
void DxvkPipelineWorkers::runWorker() {
env::setThreadName("dxvk-shader");
void DxvkPipelineWorkers::runWorker(DxvkPipelinePriority maxPriority) {
static const std::array<char, 3> suffixes = { 'h', 'n', 'l' };
const uint32_t maxPriorityIndex = uint32_t(maxPriority);
env::setThreadName(str::format("dxvk-shader-", suffixes.at(maxPriorityIndex)));
while (true) {
std::optional<PipelineEntry> p;
std::optional<PipelineLibraryEntry> l;
PipelineEntry entry;
{ std::unique_lock lock(m_queueLock);
{ std::unique_lock lock(m_lock);
auto& bucket = m_buckets[maxPriorityIndex];
m_queueCond.wait(lock, [this] {
return !m_workersRunning
|| !m_queuedLibrariesPrioritized.empty()
|| !m_queuedLibraries.empty()
|| !m_queuedPipelines.empty();
bucket.idleWorkers += 1;
bucket.cond.wait(lock, [this, maxPriorityIndex, &entry] {
// Attempt to fetch a work item from the
// highest-priority queue that is not empty
for (uint32_t i = 0; i <= maxPriorityIndex; i++) {
if (!m_buckets[i].queue.empty()) {
entry = m_buckets[i].queue.front();
m_buckets[i].queue.pop();
return true;
}
}
return !m_workersRunning;
});
if (!m_workersRunning) {
// Skip pending work, exiting early is
// more important in this case.
break;
} else if (!m_queuedLibrariesPrioritized.empty()) {
l = m_queuedLibrariesPrioritized.front();
m_queuedLibrariesPrioritized.pop();
} else if (!m_queuedLibraries.empty()) {
l = m_queuedLibraries.front();
m_queuedLibraries.pop();
} else if (!m_queuedPipelines.empty()) {
p = m_queuedPipelines.front();
m_queuedPipelines.pop();
}
}
if (l) {
if (l->pipelineLibrary)
l->pipelineLibrary->compilePipeline();
m_pendingTasks -= 1;
}
if (p) {
if (p->graphicsPipeline) {
p->graphicsPipeline->compilePipeline(p->graphicsState);
p->graphicsPipeline->releasePipeline();
}
m_pendingTasks -= 1;
}
}
}
void DxvkPipelineWorkers::runWorkerPrioritized() {
env::setThreadName("dxvk-shader-p");
while (true) {
PipelineLibraryEntry l = { };
{ std::unique_lock lock(m_queueLock);
m_queueCondPrioritized.wait(lock, [this] {
return !m_workersRunning
|| !m_queuedLibrariesPrioritized.empty();
});
bucket.idleWorkers -= 1;
// Skip pending work, exiting early is
// more important in this case.
if (!m_workersRunning)
break;
l = m_queuedLibrariesPrioritized.front();
m_queuedLibrariesPrioritized.pop();
}
if (l.pipelineLibrary)
l.pipelineLibrary->compilePipeline();
m_pendingTasks -= 1;
if (entry.pipelineLibrary) {
entry.pipelineLibrary->compilePipeline();
m_pendingTasks -= 1;
} else if (entry.graphicsPipeline) {
entry.graphicsPipeline->compilePipeline(entry.graphicsState);
entry.graphicsPipeline->releasePipeline();
m_pendingTasks -= 1;
}
}
}

View File

@ -38,8 +38,9 @@ namespace dxvk {
* \brief Pipeline priority
*/
enum class DxvkPipelinePriority : uint32_t {
Normal = 0,
High = 1,
High = 0,
Normal = 1,
Low = 2,
};
/**
@ -78,7 +79,8 @@ namespace dxvk {
*/
void compileGraphicsPipeline(
DxvkGraphicsPipeline* pipeline,
const DxvkGraphicsPipelineStateInfo& state);
const DxvkGraphicsPipelineStateInfo& state,
DxvkPipelinePriority priority);
/**
* \brief Checks whether workers are busy
@ -97,34 +99,41 @@ namespace dxvk {
private:
struct PipelineEntry {
PipelineEntry()
: pipelineLibrary(nullptr), graphicsPipeline(nullptr) { }
PipelineEntry(DxvkShaderPipelineLibrary* l)
: pipelineLibrary(l), graphicsPipeline(nullptr) { }
PipelineEntry(DxvkGraphicsPipeline* p, const DxvkGraphicsPipelineStateInfo& s)
: pipelineLibrary(nullptr), graphicsPipeline(p), graphicsState(s) { }
DxvkShaderPipelineLibrary* pipelineLibrary;
DxvkGraphicsPipeline* graphicsPipeline;
DxvkGraphicsPipelineStateInfo graphicsState;
};
struct PipelineLibraryEntry {
DxvkShaderPipelineLibrary* pipelineLibrary;
struct PipelineBucket {
dxvk::condition_variable cond;
std::queue<PipelineEntry> queue;
uint32_t idleWorkers = 0;
};
DxvkDevice* m_device;
std::atomic<uint64_t> m_pendingTasks = { 0ull };
dxvk::mutex m_queueLock;
dxvk::condition_variable m_queueCond;
dxvk::condition_variable m_queueCondPrioritized;
std::queue<PipelineLibraryEntry> m_queuedLibrariesPrioritized;
std::queue<PipelineLibraryEntry> m_queuedLibraries;
std::queue<PipelineEntry> m_queuedPipelines;
dxvk::mutex m_lock;
std::array<PipelineBucket, 3> m_buckets;
bool m_workersRunning = false;
std::vector<dxvk::thread> m_workers;
void notifyWorkers(DxvkPipelinePriority priority);
void startWorkers();
void runWorker();
void runWorkerPrioritized();
void runWorker(DxvkPipelinePriority maxPriority);
};

View File

@ -453,7 +453,7 @@ namespace dxvk {
if (!pipeline)
pipeline = m_pipeManager->createGraphicsPipeline(item.gp);
m_pipeWorkers->compileGraphicsPipeline(pipeline, entry.gpState);
m_pipeWorkers->compileGraphicsPipeline(pipeline, entry.gpState, DxvkPipelinePriority::Normal);
} break;
case DxvkStateCacheEntryType::PipelineLibrary: {