mirror/tracy
1
0
Fork 0
mirror of https://github.com/wolfpld/tracy.git synced 2024-11-10 02:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #603 from slomp/slomp/d3d11-refactor

Browse Source 
refactoring of the D3D11 back-end
This commit is contained in:
Bartosz Taudul 2023-10-09 23:56:20 +02:00 committed by GitHub
commit 757362f782
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 207 additions and 205 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

412
public/tracy/TracyD3D11.hpp
View File

@ -13,13 +13,13 @@
#define TracyD3D11ZoneC(ctx, name, color) #define TracyD3D11ZoneC(ctx, name, color)
#define TracyD3D11NamedZone(ctx, varname, name, active) #define TracyD3D11NamedZone(ctx, varname, name, active)
#define TracyD3D11NamedZoneC(ctx, varname, name, color, active) #define TracyD3D11NamedZoneC(ctx, varname, name, color, active)
#define TracyD3D12ZoneTransient(ctx, varname, name, active) #define TracyD3D11ZoneTransient(ctx, varname, name, active)
#define TracyD3D11ZoneS(ctx, name, depth) #define TracyD3D11ZoneS(ctx, name, depth)
#define TracyD3D11ZoneCS(ctx, name, color, depth) #define TracyD3D11ZoneCS(ctx, name, color, depth)
#define TracyD3D11NamedZoneS(ctx, varname, name, depth, active) #define TracyD3D11NamedZoneS(ctx, varname, name, depth, active)
#define TracyD3D11NamedZoneCS(ctx, varname, name, color, depth, active) #define TracyD3D11NamedZoneCS(ctx, varname, name, color, depth, active)
#define TracyD3D12ZoneTransientS(ctx, varname, name, depth, active) #define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active)
#define TracyD3D11Collect(ctx) #define TracyD3D11Collect(ctx)
@ -39,11 +39,12 @@ using TracyD3D11Ctx = void*;
#include "Tracy.hpp" #include "Tracy.hpp"
#include "../client/TracyProfiler.hpp" #include "../client/TracyProfiler.hpp"
#include "../client/TracyCallstack.hpp" #include "../client/TracyCallstack.hpp"
#include "../common/TracyAlign.hpp" #include "../common/TracyYield.hpp"
#include "../common/TracyAlloc.hpp"
#include <d3d11.h> #include <d3d11.h>
#define TracyD3D11Panic(msg, ...) do { assert(false && "TracyD3D11: " msg); TracyMessageLC("TracyD3D11: " msg, tracy::Color::Red4); __VA_ARGS__; } while(false);
namespace tracy namespace tracy
{ {
@ -51,71 +52,83 @@ class D3D11Ctx
{ {
friend class D3D11ZoneScope; friend class D3D11ZoneScope;
enum { QueryCount = 64 * 1024 }; static constexpr uint32_t MaxQueries = 64 * 1024;
enum CollectMode { POLL, BLOCK };
public: public:
D3D11Ctx( ID3D11Device* device, ID3D11DeviceContext* devicectx ) D3D11Ctx( ID3D11Device* device, ID3D11DeviceContext* devicectx )
: m_device( device )
, m_devicectx( devicectx )
, m_context( GetGpuCtxCounter().fetch_add( 1, std::memory_order_relaxed ) )
, m_head( 0 )
, m_tail( 0 )
{ {
assert( m_context != 255 ); // TODO: consider calling ID3D11Device::GetImmediateContext() instead of passing it as an argument
m_device = device;
device->AddRef();
m_immediateDevCtx = devicectx;
devicectx->AddRef();
for (int i = 0; i < QueryCount; i++)
{ {
HRESULT hr = S_OK; D3D11_QUERY_DESC desc = { };
D3D11_QUERY_DESC desc;
desc.MiscFlags = 0;
desc.Query = D3D11_QUERY_TIMESTAMP;
hr |= device->CreateQuery(&desc, &m_queries[i]);
desc.Query = D3D11_QUERY_TIMESTAMP_DISJOINT; desc.Query = D3D11_QUERY_TIMESTAMP_DISJOINT;
hr |= device->CreateQuery(&desc, &m_disjoints[i]); if (FAILED(m_device->CreateQuery(&desc, &m_disjointQuery)))
{
m_disjointMap[i] = nullptr; TracyD3D11Panic("unable to create disjoint timestamp query.", return);
}
assert(SUCCEEDED(hr));
} }
// Force query the initial GPU timestamp (pipeline stall) for (ID3D11Query*& query : m_queries)
D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint; {
UINT64 timestamp; D3D11_QUERY_DESC desc = { };
desc.Query = D3D11_QUERY_TIMESTAMP;
if (FAILED(m_device->CreateQuery(&desc, &query)))
{
TracyD3D11Panic("unable to create timestamp query.", return);
}
}
// Calibrate CPU and GPU timestamps
int64_t tcpu = 0;
int64_t tgpu = 0;
for (int attempts = 0; attempts < 50; attempts++) for (int attempts = 0; attempts < 50; attempts++)
{ {
devicectx->Begin(m_disjoints[0]); m_immediateDevCtx->Begin(m_disjointQuery);
devicectx->End(m_queries[0]); m_immediateDevCtx->End(m_queries[0]);
devicectx->End(m_disjoints[0]); m_immediateDevCtx->End(m_disjointQuery);
devicectx->Flush();
while (devicectx->GetData(m_disjoints[0], &disjoint, sizeof(disjoint), 0) == S_FALSE) int64_t tcpu0 = Profiler::GetTime();
/* Nothing */; WaitForQuery(m_disjointQuery);
int64_t tcpu1 = Profiler::GetTime();
D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint = { };
if (m_immediateDevCtx->GetData(m_disjointQuery, &disjoint, sizeof(disjoint), 0) != S_OK)
{
TracyMessageLC("TracyD3D11: unable to query GPU timestamp; retrying...", tracy::Color::Tomato);
continue;
}
if (disjoint.Disjoint) if (disjoint.Disjoint)
continue; continue;
while (devicectx->GetData(m_queries[0], &timestamp, sizeof(timestamp), 0) == S_FALSE) UINT64 timestamp = 0;
/* Nothing */; if (m_immediateDevCtx->GetData(m_queries[0], &timestamp, sizeof(timestamp), 0) != S_OK)
continue; // this should never happen, since the enclosing disjoint query succeeded
tcpu = tcpu0 + (tcpu1 - tcpu0) * 1 / 2;
tgpu = timestamp * (1000000000 / disjoint.Frequency);
break; break;
} }
int64_t tgpu = timestamp * (1000000000ull / disjoint.Frequency); // ready to roll
int64_t tcpu = Profiler::GetTime(); m_contextId = GetGpuCtxCounter().fetch_add(1);
m_immediateDevCtx->Begin(m_disjointQuery);
m_previousCheckpoint = m_nextCheckpoint = 0;
uint8_t flags = 0;
const float period = 1.f;
auto* item = Profiler::QueueSerial(); auto* item = Profiler::QueueSerial();
MemWrite( &item->hdr.type, QueueType::GpuNewContext ); MemWrite( &item->hdr.type, QueueType::GpuNewContext );
MemWrite( &item->gpuNewContext.cpuTime, tcpu ); MemWrite( &item->gpuNewContext.cpuTime, tcpu );
MemWrite( &item->gpuNewContext.gpuTime, tgpu ); MemWrite( &item->gpuNewContext.gpuTime, tgpu );
memset(&item->gpuNewContext.thread, 0, sizeof(item->gpuNewContext.thread)); MemWrite( &item->gpuNewContext.thread, uint32_t(0) ); // #TODO: why not GetThreadHandle()?
MemWrite( &item->gpuNewContext.period, period ); MemWrite( &item->gpuNewContext.period, 1.0f );
MemWrite( &item->gpuNewContext.context, m_context ); MemWrite( &item->gpuNewContext.context, m_contextId);
MemWrite( &item->gpuNewContext.flags, flags ); MemWrite( &item->gpuNewContext.flags, uint8_t(0) );
MemWrite( &item->gpuNewContext.type, GpuContextType::Direct3D11 ); MemWrite( &item->gpuNewContext.type, GpuContextType::Direct3D11 );
#ifdef TRACY_ON_DEMAND #ifdef TRACY_ON_DEMAND
@ -127,12 +140,20 @@ public:
~D3D11Ctx() ~D3D11Ctx()
{ {
for (int i = 0; i < QueryCount; i++) // collect all pending timestamps before destroying everything
do
{ {
m_queries[i]->Release(); Collect(BLOCK);
m_disjoints[i]->Release(); } while (m_previousCheckpoint != m_queryCounter);
m_disjointMap[i] = nullptr;
for (ID3D11Query* query : m_queries)
{
query->Release();
} }
m_immediateDevCtx->End(m_disjointQuery);
m_disjointQuery->Release();
m_immediateDevCtx->Release();
m_device->Release();
} }
void Name( const char* name, uint16_t len ) void Name( const char* name, uint16_t len )
@ -142,7 +163,7 @@ public:
auto item = Profiler::QueueSerial(); auto item = Profiler::QueueSerial();
MemWrite( &item->hdr.type, QueueType::GpuContextName ); MemWrite( &item->hdr.type, QueueType::GpuContextName );
MemWrite( &item->gpuContextNameFat.context, m_context ); MemWrite( &item->gpuContextNameFat.context, m_contextId );
MemWrite( &item->gpuContextNameFat.ptr, (uint64_t)ptr ); MemWrite( &item->gpuContextNameFat.ptr, (uint64_t)ptr );
MemWrite( &item->gpuContextNameFat.size, len ); MemWrite( &item->gpuContextNameFat.size, len );
#ifdef TRACY_ON_DEMAND #ifdef TRACY_ON_DEMAND
@ -151,217 +172,170 @@ public:
Profiler::QueueSerialFinish(); Profiler::QueueSerialFinish();
} }
void Collect() void Collect(CollectMode mode = POLL)
{ {
ZoneScopedC( Color::Red4 ); ZoneScopedC( Color::Red4 );
if( m_tail == m_head ) return;
#ifdef TRACY_ON_DEMAND #ifdef TRACY_ON_DEMAND
if( !GetProfiler().IsConnected() ) if( !GetProfiler().IsConnected() )
{ {
m_head = m_tail = 0; m_previousCheckpoint = m_nextCheckpoint = m_queryCounter;
return; return;
} }
#endif #endif
auto start = m_tail; if (m_previousCheckpoint == m_nextCheckpoint)
auto end = m_head + QueryCount;
auto cnt = (end - start) % QueryCount;
while (cnt > 1)
{ {
auto mid = start + cnt / 2; uintptr_t nextCheckpoint = m_queryCounter;
if (nextCheckpoint == m_nextCheckpoint)
bool available =
m_devicectx->GetData(m_disjointMap[mid % QueryCount], nullptr, 0, D3D11_ASYNC_GETDATA_DONOTFLUSH) == S_OK &&
m_devicectx->GetData(m_queries[mid % QueryCount], nullptr, 0, D3D11_ASYNC_GETDATA_DONOTFLUSH) == S_OK;
if (available)
{ {
start = mid; return;
} }
else m_nextCheckpoint = nextCheckpoint;
{ m_immediateDevCtx->End(m_disjointQuery);
end = mid;
}
cnt = (end - start) % QueryCount;
} }
start %= QueryCount; if (mode == CollectMode::BLOCK)
while (m_tail != start)
{ {
D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint; WaitForQuery(m_disjointQuery);
UINT64 time; }
m_devicectx->GetData(m_disjointMap[m_tail], &disjoint, sizeof(disjoint), 0); D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint = { };
m_devicectx->GetData(m_queries[m_tail], &time, sizeof(time), 0); if (m_immediateDevCtx->GetData(m_disjointQuery, &disjoint, sizeof(disjoint), D3D11_ASYNC_GETDATA_DONOTFLUSH) != S_OK)
{
return;
}
time *= (1000000000ull / disjoint.Frequency); if (disjoint.Disjoint == TRUE)
{
m_previousCheckpoint = m_nextCheckpoint;
TracyD3D11Panic("disjoint timestamps detected; dropping.");
return;
}
auto begin = m_previousCheckpoint;
auto end = m_nextCheckpoint;
for (auto i = begin; i != end; ++i)
{
uint32_t k = RingIndex(i);
UINT64 timestamp = 0;
if (m_immediateDevCtx->GetData(m_queries[k], &timestamp, sizeof(timestamp), 0) != S_OK)
{
TracyD3D11Panic("timestamp expected to be ready, but it was not!");
break;
}
timestamp *= (1000000000ull / disjoint.Frequency);
auto* item = Profiler::QueueSerial(); auto* item = Profiler::QueueSerial();
MemWrite(&item->hdr.type, QueueType::GpuTime); MemWrite(&item->hdr.type, QueueType::GpuTime);
MemWrite(&item->gpuTime.gpuTime, (int64_t)time); MemWrite(&item->gpuTime.gpuTime, static_cast<int64_t>(timestamp));
MemWrite(&item->gpuTime.queryId, (uint16_t)m_tail); MemWrite(&item->gpuTime.queryId, static_cast<uint16_t>(k));
MemWrite(&item->gpuTime.context, m_context); MemWrite(&item->gpuTime.context, m_contextId);
Profiler::QueueSerialFinish(); Profiler::QueueSerialFinish();
m_tail = (m_tail + 1) % QueryCount;
} }
// disjoint timestamp queries should only be invoked once per frame or less
// https://learn.microsoft.com/en-us/windows/win32/api/d3d11/ne-d3d11-d3d11_query
m_immediateDevCtx->Begin(m_disjointQuery);
m_previousCheckpoint = m_nextCheckpoint;
} }
private: private:
tracy_force_inline unsigned int NextQueryId() tracy_force_inline uint32_t RingIndex(uintptr_t index)
{ {
const auto id = m_head; index %= MaxQueries;
m_head = ( m_head + 1 ) % QueryCount; return static_cast<uint32_t>(index);
assert( m_head != m_tail );
return id;
} }
tracy_force_inline ID3D11Query* TranslateQueryId( unsigned int id ) tracy_force_inline uint32_t RingCount(uintptr_t begin, uintptr_t end)
{
// wrap-around safe: all unsigned
uintptr_t count = end - begin;
return static_cast<uint32_t>(count);
}
tracy_force_inline uint32_t NextQueryId()
{
auto id = m_queryCounter++;
if (RingCount(m_previousCheckpoint, id) >= MaxQueries)
{
TracyD3D11Panic("too many pending timestamp queries.");
// #TODO: return some sentinel value; ideally a "hidden" query index
}
return RingIndex(id);
}
tracy_force_inline ID3D11Query* GetQueryObjectFromId(uint32_t id)
{ {
return m_queries[id]; return m_queries[id];
} }
tracy_force_inline ID3D11Query* MapDisjointQueryId( unsigned int id, unsigned int disjointId ) tracy_force_inline void WaitForQuery(ID3D11Query* query)
{ {
m_disjointMap[id] = m_disjoints[disjointId]; m_immediateDevCtx->Flush();
return m_disjoints[disjointId]; while (m_immediateDevCtx->GetData(query, nullptr, 0, 0) != S_OK)
YieldThread(); // busy-wait :-( attempt to reduce power usage with _mm_pause() & friends...
} }
tracy_force_inline uint8_t GetId() const tracy_force_inline uint8_t GetContextId() const
{ {
return m_context; return m_contextId;
} }
ID3D11Device* m_device; ID3D11Device* m_device = nullptr;
ID3D11DeviceContext* m_devicectx; ID3D11DeviceContext* m_immediateDevCtx = nullptr;
ID3D11Query* m_queries[QueryCount]; ID3D11Query* m_queries[MaxQueries];
ID3D11Query* m_disjoints[QueryCount]; ID3D11Query* m_disjointQuery = nullptr;
ID3D11Query* m_disjointMap[QueryCount]; // Multiple time queries can have one disjoint query
uint8_t m_context;
unsigned int m_head; uint8_t m_contextId = 255; // NOTE: apparently, 255 means invalid id; is this documented anywhere?
unsigned int m_tail;
uintptr_t m_queryCounter = 0;
uintptr_t m_previousCheckpoint = 0;
uintptr_t m_nextCheckpoint = 0;
}; };
class D3D11ZoneScope class D3D11ZoneScope
{ {
public: public:
tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, const SourceLocationData* srcloc, bool is_active ) tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, const SourceLocationData* srcloc, bool active )
#ifdef TRACY_ON_DEMAND : D3D11ZoneScope(ctx, active)
: m_active( is_active && GetProfiler().IsConnected() )
#else
: m_active( is_active )
#endif
{ {
if( !m_active ) return; if( !m_active ) return;
m_ctx = ctx;
const auto queryId = ctx->NextQueryId();
ctx->m_devicectx->Begin(ctx->MapDisjointQueryId(queryId, queryId));
ctx->m_devicectx->End(ctx->TranslateQueryId(queryId));
m_disjointId = queryId;
auto* item = Profiler::QueueSerial(); auto* item = Profiler::QueueSerial();
MemWrite( &item->hdr.type, QueueType::GpuZoneBeginSerial ); WriteQueueItem(item, QueueType::GpuZoneBeginSerial, reinterpret_cast<uint64_t>(srcloc));
MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
MemWrite( &item->gpuZoneBegin.srcloc, (uint64_t)srcloc );
MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
Profiler::QueueSerialFinish();
} }
tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, const SourceLocationData* srcloc, int depth, bool is_active ) tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, const SourceLocationData* srcloc, int depth, bool active )
#ifdef TRACY_ON_DEMAND : D3D11ZoneScope(ctx, active)
: m_active( is_active && GetProfiler().IsConnected() )
#else
: m_active( is_active )
#endif
{ {
if( !m_active ) return; if( !m_active ) return;
m_ctx = ctx;
const auto queryId = ctx->NextQueryId(); auto* item = Profiler::QueueSerialCallstack(Callstack(depth));
ctx->m_devicectx->Begin(ctx->MapDisjointQueryId(queryId, queryId)); WriteQueueItem(item, QueueType::GpuZoneBeginCallstackSerial, reinterpret_cast<uint64_t>(srcloc));
ctx->m_devicectx->End(ctx->TranslateQueryId(queryId));
m_disjointId = queryId;
auto* item = Profiler::QueueSerial();
MemWrite( &item->hdr.type, QueueType::GpuZoneBeginCallstackSerial );
MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
MemWrite( &item->gpuZoneBegin.srcloc, (uint64_t)srcloc );
MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
MemWrite( &item->gpuZoneBegin.context, ctx->GetId() );
Profiler::QueueSerialFinish();
GetProfiler().SendCallstack( depth );
} }
tracy_force_inline D3D11ZoneScope(D3D11Ctx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, bool active) tracy_force_inline D3D11ZoneScope(D3D11Ctx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, bool active)
#ifdef TRACY_ON_DEMAND : D3D11ZoneScope(ctx, active)
: m_active(active&& GetProfiler().IsConnected())
#else
: m_active(active)
#endif
{ {
if( !m_active ) return; if( !m_active ) return;
m_ctx = ctx;
const auto queryId = ctx->NextQueryId();
ctx->m_devicectx->Begin(ctx->MapDisjointQueryId(queryId, queryId));
ctx->m_devicectx->End(ctx->TranslateQueryId(queryId));
m_disjointId = queryId;
const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz); const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz);
auto* item = Profiler::QueueSerial(); auto* item = Profiler::QueueSerial();
MemWrite(&item->hdr.type, QueueType::GpuZoneBeginAllocSrcLocSerial); WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocSerial, sourceLocation);
MemWrite(&item->gpuZoneBegin.cpuTime, Profiler::GetTime());
MemWrite(&item->gpuZoneBegin.srcloc, sourceLocation);
MemWrite(&item->gpuZoneBegin.thread, GetThreadHandle());
MemWrite(&item->gpuZoneBegin.queryId, static_cast<uint16_t>(queryId));
MemWrite(&item->gpuZoneBegin.context, ctx->GetId());
Profiler::QueueSerialFinish();
} }
tracy_force_inline D3D11ZoneScope(D3D11Ctx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, int depth, bool active) tracy_force_inline D3D11ZoneScope(D3D11Ctx* ctx, uint32_t line, const char* source, size_t sourceSz, const char* function, size_t functionSz, const char* name, size_t nameSz, int depth, bool active)
#ifdef TRACY_ON_DEMAND : D3D11ZoneScope(ctx, active)
: m_active(active&& GetProfiler().IsConnected())
#else
: m_active(active)
#endif
{ {
if( !m_active ) return; if( !m_active ) return;
m_ctx = ctx;
const auto queryId = ctx->NextQueryId();
ctx->m_devicectx->Begin(ctx->MapDisjointQueryId(queryId, queryId));
ctx->m_devicectx->End(ctx->TranslateQueryId(queryId));
m_disjointId = queryId;
const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz); const auto sourceLocation = Profiler::AllocSourceLocation(line, source, sourceSz, function, functionSz, name, nameSz);
auto* item = Profiler::QueueSerialCallstack(Callstack(depth)); auto* item = Profiler::QueueSerialCallstack(Callstack(depth));
MemWrite(&item->hdr.type, QueueType::GpuZoneBeginAllocSrcLocCallstackSerial); WriteQueueItem(item, QueueType::GpuZoneBeginAllocSrcLocCallstackSerial, sourceLocation);
MemWrite(&item->gpuZoneBegin.cpuTime, Profiler::GetTime());
MemWrite(&item->gpuZoneBegin.srcloc, sourceLocation);
MemWrite(&item->gpuZoneBegin.thread, GetThreadHandle());
MemWrite(&item->gpuZoneBegin.queryId, static_cast<uint16_t>(queryId));
MemWrite(&item->gpuZoneBegin.context, ctx->GetId());
Profiler::QueueSerialFinish();
} }
tracy_force_inline ~D3D11ZoneScope() tracy_force_inline ~D3D11ZoneScope()
@ -369,24 +343,46 @@ public:
if( !m_active ) return; if( !m_active ) return;
const auto queryId = m_ctx->NextQueryId(); const auto queryId = m_ctx->NextQueryId();
m_ctx->m_devicectx->End(m_ctx->TranslateQueryId(queryId)); m_ctx->m_immediateDevCtx->End(m_ctx->GetQueryObjectFromId(queryId));
m_ctx->m_devicectx->End(m_ctx->MapDisjointQueryId(queryId, m_disjointId));
auto* item = Profiler::QueueSerial(); auto* item = Profiler::QueueSerial();
MemWrite( &item->hdr.type, QueueType::GpuZoneEndSerial ); MemWrite( &item->hdr.type, QueueType::GpuZoneEndSerial );
MemWrite( &item->gpuZoneEnd.cpuTime, Profiler::GetTime() ); MemWrite( &item->gpuZoneEnd.cpuTime, Profiler::GetTime() );
MemWrite( &item->gpuZoneEnd.thread, GetThreadHandle() ); MemWrite( &item->gpuZoneEnd.thread, GetThreadHandle() );
MemWrite( &item->gpuZoneEnd.queryId, uint16_t( queryId ) ); MemWrite( &item->gpuZoneEnd.queryId, uint16_t( queryId ) );
MemWrite( &item->gpuZoneEnd.context, m_ctx->GetId() ); MemWrite( &item->gpuZoneEnd.context, m_ctx->GetContextId() );
Profiler::QueueSerialFinish(); Profiler::QueueSerialFinish();
} }
private: private:
tracy_force_inline D3D11ZoneScope( D3D11Ctx* ctx, bool active )
#ifdef TRACY_ON_DEMAND
: m_active( is_active && GetProfiler().IsConnected() )
#else
: m_active( active )
#endif
{
if( !m_active ) return;
m_ctx = ctx;
}
void WriteQueueItem(tracy::QueueItem* item, tracy::QueueType queueItemType, uint64_t sourceLocation)
{
const auto queryId = m_ctx->NextQueryId();
m_ctx->m_immediateDevCtx->End(m_ctx->GetQueryObjectFromId(queryId));
MemWrite( &item->hdr.type, queueItemType);
MemWrite( &item->gpuZoneBegin.cpuTime, Profiler::GetTime() );
MemWrite( &item->gpuZoneBegin.srcloc, sourceLocation );
MemWrite( &item->gpuZoneBegin.thread, GetThreadHandle() );
MemWrite( &item->gpuZoneBegin.queryId, uint16_t( queryId ) );
MemWrite( &item->gpuZoneBegin.context, m_ctx->GetContextId() );
Profiler::QueueSerialFinish();
}
const bool m_active; const bool m_active;
D3D11Ctx* m_ctx; D3D11Ctx* m_ctx;
unsigned int m_disjointId;
}; };
static inline D3D11Ctx* CreateD3D11Context( ID3D11Device* device, ID3D11DeviceContext* devicectx ) static inline D3D11Ctx* CreateD3D11Context( ID3D11Device* device, ID3D11DeviceContext* devicectx )
@ -403,38 +399,44 @@ static inline void DestroyD3D11Context( D3D11Ctx* ctx )
} }
} }
#undef TracyD3D11Panic
using TracyD3D11Ctx = tracy::D3D11Ctx*; using TracyD3D11Ctx = tracy::D3D11Ctx*;
#define TracyD3D11Context( device, devicectx ) tracy::CreateD3D11Context( device, devicectx ); #define TracyD3D11Context( device, devicectx ) tracy::CreateD3D11Context( device, devicectx );
#define TracyD3D11Destroy(ctx) tracy::DestroyD3D11Context(ctx); #define TracyD3D11Destroy(ctx) tracy::DestroyD3D11Context(ctx);
#define TracyD3D11ContextName(ctx, name, size) ctx->Name(name, size); #define TracyD3D11ContextName(ctx, name, size) ctx->Name(name, size);
#define TracyD3D11UnnamedZone ___tracy_gpu_d3d11_zone
#define TracyD3D11SrcLocSymbol TracyConcat(__tracy_gpu_d3d11_source_location,TracyLine)
#define TracyD3D11SrcLocObject(name, color) static constexpr tracy::SourceLocationData TracyD3D11SrcLocSymbol { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color };
#if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK #if defined TRACY_HAS_CALLSTACK && defined TRACY_CALLSTACK
# define TracyD3D11Zone( ctx, name ) TracyD3D11NamedZoneS( ctx, ___tracy_gpu_zone, name, TRACY_CALLSTACK, true ) # define TracyD3D11Zone( ctx, name ) TracyD3D11NamedZoneS( ctx, TracyD3D11UnnamedZone, name, TRACY_CALLSTACK, true )
# define TracyD3D11ZoneC( ctx, name, color ) TracyD3D11NamedZoneCS( ctx, ___tracy_gpu_zone, name, color, TRACY_CALLSTACK, true ) # define TracyD3D11ZoneC( ctx, name, color ) TracyD3D11NamedZoneCS( ctx, TracyD3D11UnnamedZone, name, color, TRACY_CALLSTACK, true )
# define TracyD3D11NamedZone( ctx, varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); # define TracyD3D11NamedZone( ctx, varname, name, active ) TracyD3D11SrcLocObject(name, 0); tracy::D3D11ZoneScope varname( ctx, &TracyD3D11SrcLocSymbol, TRACY_CALLSTACK, active );
# define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), TRACY_CALLSTACK, active ); # define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) TracyD3D11SrcLocObject(name, color); tracy::D3D11ZoneScope varname( ctx, &TracyD3D11SrcLocSymbol, TRACY_CALLSTACK, active );
# define TracyD3D11ZoneTransient(ctx, varname, name, active) TracyD3D11ZoneTransientS(ctx, varname, cmdList, name, TRACY_CALLSTACK, active) # define TracyD3D11ZoneTransient(ctx, varname, name, active) TracyD3D11ZoneTransientS(ctx, varname, cmdList, name, TRACY_CALLSTACK, active)
#else #else
# define TracyD3D11Zone( ctx, name ) TracyD3D11NamedZone( ctx, ___tracy_gpu_zone, name, true ) # define TracyD3D11Zone( ctx, name ) TracyD3D11NamedZone( ctx, TracyD3D11UnnamedZone, name, true )
# define TracyD3D11ZoneC( ctx, name, color ) TracyD3D11NamedZoneC( ctx, ___tracy_gpu_zone, name, color, true ) # define TracyD3D11ZoneC( ctx, name, color ) TracyD3D11NamedZoneC( ctx, TracyD3D11UnnamedZone, name, color, true )
# define TracyD3D11NamedZone( ctx, varname, name, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), active ); # define TracyD3D11NamedZone( ctx, varname, name, active ) TracyD3D11SrcLocObject(name, 0); tracy::D3D11ZoneScope varname( ctx, &TracyD3D11SrcLocSymbol, active );
# define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), active ); # define TracyD3D11NamedZoneC( ctx, varname, name, color, active ) TracyD3D11SrcLocObject(name, color); tracy::D3D11ZoneScope varname( ctx, &TracyD3D11SrcLocSymbol, active );
# define TracyD3D11ZoneTransient(ctx, varname, name, active) tracy::D3D11ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), active }; # define TracyD3D11ZoneTransient(ctx, varname, name, active) tracy::D3D11ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), active };
#endif #endif
#ifdef TRACY_HAS_CALLSTACK #ifdef TRACY_HAS_CALLSTACK
# define TracyD3D11ZoneS( ctx, name, depth ) TracyD3D11NamedZoneS( ctx, ___tracy_gpu_zone, name, depth, true ) # define TracyD3D11ZoneS( ctx, name, depth ) TracyD3D11NamedZoneS( ctx, TracyD3D11UnnamedZone, name, depth, true )
# define TracyD3D11ZoneCS( ctx, name, color, depth ) TracyD3D11NamedZoneCS( ctx, ___tracy_gpu_zone, name, color, depth, true ) # define TracyD3D11ZoneCS( ctx, name, color, depth ) TracyD3D11NamedZoneCS( ctx, TracyD3D11UnnamedZone, name, color, depth, true )
# define TracyD3D11NamedZoneS( ctx, varname, name, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, 0 }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active ); # define TracyD3D11NamedZoneS( ctx, varname, name, depth, active ) TracyD3D11SrcLocObject(name, 0); tracy::D3D11ZoneScope varname( ctx, &TracyD3D11SrcLocSymbol, depth, active );
# define TracyD3D11NamedZoneCS( ctx, varname, name, color, depth, active ) static constexpr tracy::SourceLocationData TracyConcat(__tracy_gpu_source_location,TracyLine) { name, TracyFunction, TracyFile, (uint32_t)TracyLine, color }; tracy::D3D11ZoneScope varname( ctx, &TracyConcat(__tracy_gpu_source_location,TracyLine), depth, active ); # define TracyD3D11NamedZoneCS( ctx, varname, name, color, depth, active ) TracyD3D11SrcLocObject(name, color); tracy::D3D11ZoneScope varname( ctx, &TracyD3D11SrcLocSymbol, depth, active );
# define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active) tracy::D3D11ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), depth, active }; # define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active) tracy::D3D11ZoneScope varname{ ctx, TracyLine, TracyFile, strlen(TracyFile), TracyFunction, strlen(TracyFunction), name, strlen(name), depth, active };
#else #else
# define TracyD3D11ZoneS( ctx, name, depth, active ) TracyD3D11Zone( ctx, name ) # define TracyD3D11ZoneS( ctx, name, depth, active ) TracyD3D11Zone( ctx, name )
# define TracyD3D11ZoneCS( ctx, name, color, depth, active ) TracyD3D11ZoneC( name, color ) # define TracyD3D11ZoneCS( ctx, name, color, depth, active ) TracyD3D11ZoneC( name, color )
# define TracyD3D11NamedZoneS( ctx, varname, name, depth, active ) TracyD3D11NamedZone( ctx, varname, name, active ) # define TracyD3D11NamedZoneS( ctx, varname, name, depth, active ) TracyD3D11NamedZone( ctx, varname, name, active )
# define TracyD3D11NamedZoneCS( ctx, varname, name, color, depth, active ) TracyD3D11NamedZoneC( ctx, varname, name, color, active ) # define TracyD3D11NamedZoneCS( ctx, varname, name, color, depth, active ) TracyD3D11NamedZoneC( ctx, varname, name, color, active )
# define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active) TracyD3D12ZoneTransient(ctx, varname, name, active) # define TracyD3D11ZoneTransientS(ctx, varname, name, depth, active) TracyD3D11ZoneTransient(ctx, varname, name, active)
#endif #endif
#define TracyD3D11Collect( ctx ) ctx->Collect(); #define TracyD3D11Collect( ctx ) ctx->Collect();