mirrors/tracy
1
0
Fork 0
mirror of https://github.com/wolfpld/tracy synced 2025-05-01 13:13:53 +00:00
Code Issues Releases Wiki Activity

reworking clock calibration

Browse Source
This commit is contained in:
Marcos Slomp 2023-09-11 12:08:15 -07:00
parent 6454b0bd65
commit f9d36060df
1 changed files with 61 additions and 50 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

111
public/tracy/TracyD3D12.hpp
View File

@ -73,8 +73,37 @@ namespace tracy
ID3D12Fence* m_payloadFence = nullptr; ID3D12Fence* m_payloadFence = nullptr;
std::queue<D3D12QueryPayload> m_payloadQueue; std::queue<D3D12QueryPayload> m_payloadQueue;
int64_t m_prevCalibration = 0; UINT64 m_prevCalibrationTicksCPU = 0;
int64_t m_qpcToNs = int64_t{ 1000000000 / GetFrequencyQpc() };
void RecalibrateClocks()
{
UINT64 cpuTimestamp;
UINT64 gpuTimestamp;
if (FAILED(m_queue->GetClockCalibration(&gpuTimestamp, &cpuTimestamp)))
{
assert(false && "failed to obtain queue clock calibration counters.");
return;
}
int64_t cpuDeltaTicks = cpuTimestamp - m_prevCalibrationTicksCPU;
if (cpuDeltaTicks > 0)
{
static const int64_t nanosecodsPerTick = int64_t(1000000000) / GetFrequencyQpc();
int64_t cpuDeltaNS = cpuDeltaTicks * nanosecodsPerTick;
// Save the device cpu timestamp, not the Tracy profiler timestamp:
m_prevCalibrationTicksCPU = cpuTimestamp;
cpuTimestamp = Profiler::GetTime();
auto* item = Profiler::QueueSerial();
MemWrite(&item->hdr.type, QueueType::GpuCalibration);
MemWrite(&item->gpuCalibration.gpuTime, gpuTimestamp);
MemWrite(&item->gpuCalibration.cpuTime, cpuTimestamp);
MemWrite(&item->gpuCalibration.cpuDelta, cpuDeltaNS);
MemWrite(&item->gpuCalibration.context, GetId());
SubmitQueueItem(item);
}
}
tracy_force_inline void SubmitQueueItem(tracy::QueueItem* item) tracy_force_inline void SubmitQueueItem(tracy::QueueItem* item)
{ {
@ -99,26 +128,6 @@ namespace tracy
assert(Success && featureData.CopyQueueTimestampQueriesSupported && "Platform does not support profiling of copy queues."); assert(Success && featureData.CopyQueueTimestampQueriesSupported && "Platform does not support profiling of copy queues.");
} }
uint64_t timestampFrequency;
if (FAILED(queue->GetTimestampFrequency(&timestampFrequency)))
{
assert(false && "Failed to get timestamp frequency.");
}
uint64_t cpuTimestamp;
uint64_t gpuTimestamp;
if (FAILED(queue->GetClockCalibration(&gpuTimestamp, &cpuTimestamp)))
{
assert(false && "Failed to get queue clock calibration.");
}
// Save the device cpu timestamp, not the profiler's timestamp.
m_prevCalibration = cpuTimestamp * m_qpcToNs;
cpuTimestamp = Profiler::GetTime();
D3D12_QUERY_HEAP_DESC heapDesc{}; D3D12_QUERY_HEAP_DESC heapDesc{};
heapDesc.Type = queue->GetDesc().Type == D3D12_COMMAND_LIST_TYPE_COPY ? D3D12_QUERY_HEAP_TYPE_COPY_QUEUE_TIMESTAMP : D3D12_QUERY_HEAP_TYPE_TIMESTAMP; heapDesc.Type = queue->GetDesc().Type == D3D12_COMMAND_LIST_TYPE_COPY ? D3D12_QUERY_HEAP_TYPE_COPY_QUEUE_TIMESTAMP : D3D12_QUERY_HEAP_TYPE_TIMESTAMP;
heapDesc.Count = m_queryLimit; heapDesc.Count = m_queryLimit;
@ -162,6 +171,33 @@ namespace tracy
assert(false && "Failed to create payload fence."); assert(false && "Failed to create payload fence.");
} }
float period = [queue]()
{
uint64_t timestampFrequency;
if (FAILED(queue->GetTimestampFrequency(&timestampFrequency)))
{
return 0.0f;
}
return static_cast<float>( 1E+09 / static_cast<double>(timestampFrequency) );
}();
if (period == 0.0f)
{
assert(false && "Failed to get timestamp frequency.");
}
uint64_t cpuTimestamp;
uint64_t gpuTimestamp;
if (FAILED(queue->GetClockCalibration(&gpuTimestamp, &cpuTimestamp)))
{
assert(false && "Failed to get queue clock calibration.");
}
// Save the device cpu timestamp, not the profiler's timestamp.
m_prevCalibrationTicksCPU = cpuTimestamp;
cpuTimestamp = Profiler::GetTime();
// all checked: ready to roll // all checked: ready to roll
m_contextId = GetGpuCtxCounter().fetch_add(1); m_contextId = GetGpuCtxCounter().fetch_add(1);
@ -169,8 +205,8 @@ namespace tracy
MemWrite(&item->hdr.type, QueueType::GpuNewContext); MemWrite(&item->hdr.type, QueueType::GpuNewContext);
MemWrite(&item->gpuNewContext.cpuTime, cpuTimestamp); MemWrite(&item->gpuNewContext.cpuTime, cpuTimestamp);
MemWrite(&item->gpuNewContext.gpuTime, gpuTimestamp); MemWrite(&item->gpuNewContext.gpuTime, gpuTimestamp);
memset(&item->gpuNewContext.thread, 0, sizeof(item->gpuNewContext.thread)); MemWrite(&item->gpuNewContext.thread, decltype(item->gpuNewContext.thread)(0)); // #TODO: why 0 instead of GetThreadHandle()?
MemWrite(&item->gpuNewContext.period, 1E+09f / static_cast<float>(timestampFrequency)); MemWrite(&item->gpuNewContext.period, period);
MemWrite(&item->gpuNewContext.context, GetId()); MemWrite(&item->gpuNewContext.context, GetId());
MemWrite(&item->gpuNewContext.flags, GpuContextCalibration); MemWrite(&item->gpuNewContext.flags, GpuContextCalibration);
MemWrite(&item->gpuNewContext.type, GpuContextType::Direct3D12); MemWrite(&item->gpuNewContext.type, GpuContextType::Direct3D12);
@ -271,32 +307,7 @@ namespace tracy
m_readbackBuffer->Unmap(0, nullptr); m_readbackBuffer->Unmap(0, nullptr);
// Recalibrate to account for drift. // Recalibrate to account for drift.
RecalibrateClocks();
uint64_t cpuTimestamp;
uint64_t gpuTimestamp;
if (FAILED(m_queue->GetClockCalibration(&gpuTimestamp, &cpuTimestamp)))
{
assert(false && "Failed to get queue clock calibration.");
}
cpuTimestamp *= m_qpcToNs;
const auto cpuDelta = cpuTimestamp - m_prevCalibration;
if (cpuDelta > 0)
{
m_prevCalibration = cpuTimestamp;
cpuTimestamp = Profiler::GetTime();
auto* item = Profiler::QueueSerial();
MemWrite(&item->hdr.type, QueueType::GpuCalibration);
MemWrite(&item->gpuCalibration.gpuTime, gpuTimestamp);
MemWrite(&item->gpuCalibration.cpuTime, cpuTimestamp);
MemWrite(&item->gpuCalibration.cpuDelta, cpuDelta);
MemWrite(&item->gpuCalibration.context, GetId());
Profiler::QueueSerialFinish();
}
} }
private: private: