Scarfski/tracy
1
0
Fork 0
mirror of https://github.com/wolfpld/tracy.git synced 2026-09-15 06:44:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add counter value collection

Browse Source
This commit is contained in:
Eric Eaton
2025-07-01 15:16:03 -07:00
parent 86de39704f
commit 2e49bdf4cb
6 changed files with 105 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
profiler/src/profiler/TracyView_ZoneInfo.cpp
View File

@@ -1579,6 +1579,11 @@ void View::DrawGpuInfoWindow()
TextFocused( "Delay to execution:", TimeToString( AdjustGpuTime( ev.GpuStart(), begin, drift ) - ev.CpuStart() ) );
}
TextFocused( "Query ID:", RealToString(ev.query_id) );
for (int i = 0; i < ev.note_count; i++ ) {
TextFocused( RealToString(ev.note_ids[i]), RealToString(ev.note_vals[i]) );
}
ImGui::Separator();
std::vector<const GpuEvent*> zoneTrace;
@@ -2046,6 +2051,11 @@ void View::ZoneTooltip( const GpuEvent& ev )
TextFocused( "Delay to execution:", TimeToString( AdjustGpuTime( ev.GpuStart(), begin, drift ) - ev.CpuStart() ) );
}
TextFocused( "Query ID:", RealToString(ev.query_id) );
for (int i = 0; i < ev.note_count; i++ ) {
TextFocused( RealToString(ev.note_ids[i]), RealToString(ev.note_vals[i]) );
}
ImGui::EndTooltip();
}

58
public/client/TracyRocprof.cpp
View File

@@ -28,9 +28,11 @@
} \
}
#define USE_CALIBRATION 0
namespace
{
using kernel_symbol_data_t = rocprofiler_callback_tracing_code_object_kernel_symbol_register_data_t;
struct ToolData
@@ -47,6 +49,7 @@ struct ToolData
std::unordered_map<rocprofiler_kernel_id_t, kernel_symbol_data_t> client_kernels;
std::unordered_map<rocprofiler_dispatch_id_t, int64_t> launch_start_times;
std::unordered_map<rocprofiler_dispatch_id_t, int64_t> launch_end_times;
std::unordered_map<rocprofiler_dispatch_id_t, uint16_t> dispatch_query_id;
std::unique_ptr<tracy::Thread> cal_thread;
std::mutex mut{};
};
@@ -69,7 +72,7 @@ uint8_t gpu_context_allocate(ToolData * data) {
clock_gettime(CLOCK_BOOTTIME, &ts);
uint64_t cpu_timestamp = Profiler::GetTime();
uint64_t gpu_timestamp = ((uint64_t)ts.tv_sec * 1000000000) + ts.tv_nsec;
bool is_calibrated = true;
bool is_calibrated = USE_CALIBRATION;
float timestamp_period = 1.0f;
data->previous_cpu_time = cpu_timestamp;
@@ -140,14 +143,16 @@ uint64_t kernel_src_loc(ToolData *data, uint64_t kernel_id) {
void record_interval(ToolData *data, rocprofiler_timestamp_t start_timestamp,
rocprofiler_timestamp_t end_timestamp,
uint64_t src_loc, rocprofiler_dispatch_id_t dispatch_id) {
uint16_t query_id = 0;
uint8_t context_id = data->context_id;
{
auto _lk = std::unique_lock{data->mut};
query_id = data->query_id;
data->query_id++;
if (dispatch_id != UINT64_MAX)
data->dispatch_query_id[dispatch_id] = query_id;
}
uint64_t cpu_start_time = 0, cpu_end_time = 0;
@@ -228,6 +233,7 @@ record_callback(rocprofiler_dispatch_counting_service_data_t dispatch_data,
ToolData * data = static_cast<ToolData*>(callback_data);
if (!data->init) return;
double sum;
for(size_t i = 0; i < record_count; ++i) {
int64_t profilerTime = Profiler::GetTime();
TracyLfqPrepare( QueueType::PlotDataDouble );
@@ -235,6 +241,33 @@ record_callback(rocprofiler_dispatch_counting_service_data_t dispatch_data,
MemWrite( &item->plotDataDouble.time, profilerTime );
MemWrite( &item->plotDataDouble.val, record_data[i].counter_value );
TracyLfqCommit;
sum += record_data[i].counter_value;
}
uint16_t query_id = 0;
{
auto _lk = std::unique_lock{data->mut};
if (data->dispatch_query_id.count(dispatch_data.dispatch_info.dispatch_id)) {
query_id = data->dispatch_query_id[dispatch_data.dispatch_info.dispatch_id];
} else {
fprintf(stderr, "oops\n");
}
}
if (record_count > 0) {
//fprintf(stderr, "a %lu\n", dispatch_data.dispatch_info.dispatch_id);
auto* item = tracy::Profiler::QueueSerial();
tracy::MemWrite(&item->hdr.type, tracy::QueueType::GpuZoneAnnotation);
auto _counter_id = rocprofiler_counter_id_t{};
ROCPROFILER_CALL(rocprofiler_query_record_counter_id(record_data[0].id, &_counter_id),
"query record counter id");
tracy::MemWrite(&item->zoneAnnotation.noteId, _counter_id.handle);
fprintf(stderr, "note %lu\n", _counter_id.handle);
tracy::MemWrite(&item->zoneAnnotation.queryId, query_id);
tracy::MemWrite(&item->zoneAnnotation.value, sum);
tracy::MemWrite(&item->zoneAnnotation.context, data->context_id);
tracy::Profiler::QueueSerialFinish();
}
}
@@ -253,7 +286,7 @@ dispatch_callback(rocprofiler_dispatch_counting_service_data_t dispatch_data,
assert(callback_data != nullptr);
ToolData * data = static_cast<ToolData*>(callback_data);
if (!data->init) return;
/**
* This simple example uses the same profile counter set for all agents.
* We store this in a cache to prevent constructing many identical profile counter
@@ -416,13 +449,15 @@ void calibration_thread(void * ptr) {
data->init = true;
#if USE_CALIBRATION
while (data->init) {
timespec ts;
// HSA performs a linear interpolation of GPU time to CLOCK_BOOTTIME. However, this is subject to network time updates and can drift relative to tracy's clock.
// HSA performs a linear interpolation of GPU time to CLOCK_BOOTTIME. However, this is
// subject to network time updates and can drift relative to tracy's clock.
clock_gettime(CLOCK_BOOTTIME, &ts);
int64_t cpu_timestamp = Profiler::GetTime();
int64_t gpu_timestamp = ts.tv_nsec + ts.tv_sec * 1e9L;
if (cpu_timestamp > data->previous_cpu_time) {
auto* item = tracy::Profiler::QueueSerial();
tracy::MemWrite(&item->hdr.type, tracy::QueueType::GpuCalibration);
@@ -436,6 +471,7 @@ void calibration_thread(void * ptr) {
sleep(1);
}
#endif
}
int tool_init( rocprofiler_client_finalize_t fini_func, void* user_data )
@@ -468,7 +504,7 @@ int tool_init( rocprofiler_client_finalize_t fini_func, void* user_data )
tool_callback_tracing_callback,
user_data),
"callback tracing service failed to configure");
ROCPROFILER_CALL(
rocprofiler_configure_callback_tracing_service(get_client_ctx(),
ROCPROFILER_CALLBACK_TRACING_MEMORY_COPY,
@@ -477,7 +513,7 @@ int tool_init( rocprofiler_client_finalize_t fini_func, void* user_data )
tool_callback_tracing_callback,
user_data),
"callback tracing service failed to configure");
ROCPROFILER_CALL( rocprofiler_start_context( get_client_ctx() ), "start context" );
return 0;
@@ -485,7 +521,7 @@ int tool_init( rocprofiler_client_finalize_t fini_func, void* user_data )
void tool_fini( void* tool_data_v ) {
rocprofiler_stop_context(get_client_ctx());
ToolData * data = static_cast<ToolData*>(tool_data_v);
data->init = false;
data->cal_thread.reset();
@@ -514,4 +550,4 @@ extern "C"
return &cfg;
}
}
}

15
public/common/TracyQueue.hpp
View File

@@ -111,6 +111,7 @@ enum class QueueType : uint8_t
SecondStringData,
MemNamePayload,
ThreadGroupHint,
GpuZoneAnnotation,
StringData,
ThreadName,
PlotName,
@@ -331,7 +332,7 @@ struct QueuePlotDataInt : public QueuePlotDataBase
int64_t val;
};
struct QueuePlotDataFloat : public QueuePlotDataBase
struct QueuePlotDataFloat : public QueuePlotDataBase
{
float val;
};
@@ -446,6 +447,14 @@ struct QueueGpuZoneEnd
uint8_t context;
};
struct QueueGpuZoneAnnotation
{
int64_t noteId;
double value;
uint16_t queryId;
uint8_t context;
};
struct QueueGpuTime
{
int64_t gpuTime;
@@ -467,7 +476,7 @@ struct QueueGpuTimeSync
int64_t cpuTime;
uint8_t context;
};
struct QueueGpuContextName
{
uint8_t context;
@@ -789,6 +798,7 @@ struct QueueItem
QueueSourceCodeNotAvailable sourceCodeNotAvailable;
QueueFiberEnter fiberEnter;
QueueFiberLeave fiberLeave;
QueueGpuZoneAnnotation zoneAnnotation;
};
};
#pragma pack( pop )
@@ -900,6 +910,7 @@ static constexpr size_t QueueDataSize[] = {
sizeof( QueueHeader ), // second string data
sizeof( QueueHeader ) + sizeof( QueueMemNamePayload ),
sizeof( QueueHeader ) + sizeof( QueueThreadGroupHint ),
sizeof( QueueHeader ) + sizeof( QueueGpuZoneAnnotation ), // GPU zone annotation
// keep all QueueStringTransfer below
sizeof( QueueHeader ) + sizeof( QueueStringTransfer ), // string data
sizeof( QueueHeader ) + sizeof( QueueStringTransfer ), // thread name

4
server/TracyEvent.hpp
View File

@@ -412,6 +412,10 @@ struct GpuEvent
uint64_t _gpuStart_child1;
uint64_t _gpuEnd_child2;
Int24 callstack;
uint16_t query_id;
int64_t note_ids[10];
double note_vals[10];
uint8_t note_count;
};
enum { GpuEventSize = sizeof( GpuEvent ) };

32
server/TracyWorker.cpp
View File

@@ -741,7 +741,7 @@ Worker::Worker( FileRead& f, EventType::Type eventMask, bool bgTasks, bool allow
{
m_data.stringData.reserve_exact( sz, m_slab );
}
for( uint64_t i=0; i<sz; i++ )
{
uint64_t ptr, ssz;
@@ -4622,6 +4622,9 @@ bool Worker::Process( const QueueItem& ev )
case QueueType::GpuContextName:
ProcessGpuContextName( ev.gpuContextName );
break;
case QueueType::GpuZoneAnnotation:
ProcessGpuZoneAnnotation( ev.zoneAnnotation );
break;
case QueueType::MemAlloc:
ProcessMemAlloc( ev.memAlloc );
break;
@@ -5758,6 +5761,7 @@ void Worker::ProcessGpuZoneBeginImplCommon( GpuEvent* zone, const QueueGpuZoneBe
zone->SetGpuEnd( -1 );
zone->callstack.SetVal( 0 );
zone->SetChild( -1 );
zone->query_id = ev.queryId;
uint64_t ztid;
if( ctx->thread == 0 )
@@ -5981,7 +5985,7 @@ void Worker::ProcessGpuCalibration( const QueueGpuCalibration& ev )
ctx->calibratedGpuTime = gpuTime;
ctx->calibratedCpuTime = TscTime( ev.cpuTime );
}
void Worker::ProcessGpuTimeSync( const QueueGpuTimeSync& ev )
{
auto ctx = m_gpuCtxMap[ev.context];
@@ -6013,6 +6017,22 @@ void Worker::ProcessGpuContextName( const QueueGpuContextName& ev )
ctx->name = StringIdx( idx );
}
void Worker::ProcessGpuZoneAnnotation( const QueueGpuZoneAnnotation& ev )
{
auto ctx = m_gpuCtxMap[ev.context];
assert( ctx );
// TODO: Get thread ID properly
// TODO: Search for the query from the back
assert( ctx->threadData.size() );
assert( ctx->threadData.begin()->second.timeline.size() );
auto & zone = ctx->threadData.begin()->second.timeline.back();
assert( zone->query_id == ev.queryId );
assert( zone->note_count < 10 );
zone->note_ids[zone->note_count] = ev.noteId;
zone->note_vals[zone->note_count] = ev.value;
zone->note_count++;
}
MemEvent* Worker::ProcessMemAllocImpl( MemData& memdata, const QueueMemAlloc& ev )
{
if( memdata.active.find( ev.ptr ) != memdata.active.end() )
@@ -7788,6 +7808,10 @@ void Worker::ReadTimeline( FileRead& f, Vector<short_ptr<GpuEvent>>& _vec, uint6
refGpuTime += tgpu;
zone->SetCpuEnd( refTime );
zone->SetGpuEnd( refGpuTime );
f.Read(zone->query_id);
f.Read(zone->note_count);
f.Read(zone->note_ids);
f.Read(zone->note_vals);
}
while( ++zone != end );
}
@@ -8511,6 +8535,10 @@ void Worker::WriteTimelineImpl( FileWrite& f, const V& vec, int64_t& refTime, in
WriteTimeOffset( f, refTime, v.CpuEnd() );
WriteTimeOffset( f, refGpuTime, v.GpuEnd() );
f.Write( &v.query_id , sizeof(v.query_id) );
f.Write( &v.note_count , sizeof(v.note_count) );
f.Write( &v.note_ids , sizeof(v.note_ids) );
f.Write( &v.note_vals , sizeof(v.note_vals) );
}
}

1
server/TracyWorker.hpp
View File

@@ -740,6 +740,7 @@ private:
tracy_force_inline void ProcessGpuCalibration( const QueueGpuCalibration& ev );
tracy_force_inline void ProcessGpuTimeSync( const QueueGpuTimeSync& ev );
tracy_force_inline void ProcessGpuContextName( const QueueGpuContextName& ev );
tracy_force_inline void ProcessGpuZoneAnnotation( const QueueGpuZoneAnnotation& ev );
tracy_force_inline MemEvent* ProcessMemAlloc( const QueueMemAlloc& ev );
tracy_force_inline MemEvent* ProcessMemAllocNamed( const QueueMemAlloc& ev );
tracy_force_inline MemEvent* ProcessMemFree( const QueueMemFree& ev );