yum-archive/TaSTT-Whisper

git clone https://git.yummers.dev/yum-archive/TaSTT-Whisper

KonstantinExperimental, alternative busy wait implementationcacec672023-01-21

#include "stdafx.h"
#include "GpuProfiler.h"
#include "GpuProfilerSimple.h"
using namespace DirectCompute;

inline void GpuProfiler::sProfilerData::reset()
{
	_mm_storeu_si128( ( __m128i* ) & callsPending, _mm_setzero_si128() );
}

inline void GpuProfiler::sProfilerData::addPending( int64_t time )
{
	callsPending++;
	timePending += time;
}

inline void GpuProfiler::sProfilerData::dropPending()
{
	callsPending = 0;
	timePending = 0;
}

inline void GpuProfiler::sProfilerData::makeTime( uint64_t freq )
{
	dest->count += callsPending;
	dest->totalTicks += ::makeTime( timePending, freq );
	callsPending = 0;
	timePending = 0;
}

HRESULT GpuProfiler::Queue::create()
{
	ID3D11Device* const dev = device();

	CD3D11_QUERY_DESC desc{ D3D11_QUERY_TIMESTAMP };
	for( Entry& e : queue )
	{
		CHECK( dev->CreateQuery( &desc, &e.query ) );
		e.block = nullptr;
		e.event = eEvent::None;
		e.shader = EmptyShader;
	}
	return S_OK;
}

namespace
{
	static uint64_t getTimestamp( ID3D11Query* query, const DelayExecution& delay )
	{
		ID3D11DeviceContext* const ctx = context();

		uint64_t res = 0;
		while( true )
		{
			const HRESULT hr = ctx->GetData( query, &res, sizeof( uint64_t ), 0 );
			check( hr );
			if( S_OK == hr )
				return res;
			delay.delay();
		}
	}

	static D3D11_QUERY_DATA_TIMESTAMP_DISJOINT waitForDisjointData( ID3D11Query* query )
	{
		ID3D11DeviceContext* const ctx = context();
		ctx->End( query );

		D3D11_QUERY_DATA_TIMESTAMP_DISJOINT res;
		while( true )
		{
			const HRESULT hr = ctx->GetData( query, &res, sizeof( D3D11_QUERY_DATA_TIMESTAMP_DISJOINT ), 0 );
			check( hr );
			if( S_OK == hr )
				return res;
			Sleep( 1 );
		}
	}
}

void GpuProfiler::Queue::Entry::join( GpuProfiler& owner )
{
	assert( nullptr != block );

	uint64_t res = getTimestamp( query, owner.delay );
#if PROFILER_COLLECT_TAGS
	block->haveTimestamp( event, shader, tag, res, owner );
#else
	block->haveTimestamp( event, shader, 0, res, owner );
#endif
	block = nullptr;
	event = eEvent::None;
	shader = EmptyShader;
}

void GpuProfiler::Queue::submit( BlockState* block, eEvent evt, uint16_t shader, uint16_t tag )
{
	// if( evt == GpuProfiler::eEvent::Shader && shader == 0 ) __debugbreak();
	assert( nullptr != block );

	Entry& e = queue[ nextEntry ];
	if( nullptr != e.block )
		e.join( owner );

	e.block = block;
	e.event = evt;
	e.shader = shader;
#if PROFILER_COLLECT_TAGS
	e.tag = tag;
#endif
	context()->End( e.query );
	nextEntry = ( nextEntry + 1 ) % queueLength;
}

void GpuProfiler::Queue::join()
{
	while( true )
	{
		Entry& e = queue[ nextEntry ];
		if( nullptr == e.block )
			return;
		e.join( owner );
		nextEntry = ( nextEntry + 1 ) % queueLength;
	}
}

static inline uint32_t makeTagKey( uint16_t cs, uint16_t tag )
{
	uint32_t r = cs;
	r = r << 16;
	r |= tag;
	return r;
}

void GpuProfiler::BlockState::completePrevShader( uint64_t time, GpuProfiler& profiler )
{
	if( shaderStart == -1 )
		return;
	assert( prevShader != EmptyShader );
	const int64_t elapsed = (int64_t)time - shaderStart;

	sProfilerData* dest = nullptr;
	auto* p = profiler.results.Lookup( prevShader );
	if( nullptr != p )
		dest = &p->m_value;
	else
	{
		sProfilerData& res = profiler.results[ prevShader ];
		res.dest = &profiler.dest.measure( (eComputeShader)prevShader );
		dest = &res;
	}
	dest->addPending( elapsed );

#if PROFILER_COLLECT_TAGS
	if( 0 != prevShaderTag )
	{
		const uint32_t key = makeTagKey( prevShader, prevShaderTag );
		auto* pt = profiler.resultsTagged.Lookup( key );
		if( nullptr != pt )
			dest = &pt->m_value;
		else
		{
			sProfilerData& res = profiler.resultsTagged[ key ];
			res.dest = &profiler.dest.measure( (eComputeShader)prevShader, prevShaderTag );
			dest = &res;
		}
		dest->addPending( elapsed );
	}
#endif
	prevShader = EmptyShader;
	prevShaderTag = 0;
	shaderStart = -1;
}

void GpuProfiler::BlockState::haveTimestamp( eEvent evt, uint16_t cs, uint16_t tag, uint64_t time, GpuProfiler& profiler )
{
	switch( evt )
	{
	case eEvent::BlockStart:
		assert( -1 == timeStart );
		assert( -1 == shaderStart );
		assert( cs == EmptyShader );
		timeStart = (int64_t)time;
		if( nullptr != parentBlock )
			parentBlock->completePrevShader( time, profiler );
		return;
	case eEvent::BlockEnd:
		assert( -1 != timeStart );
		assert( cs == EmptyShader );
		completePrevShader( time, profiler );
		destBlock->addPending( (int64_t)time - timeStart );
		timeStart = -1;
		return;
	case eEvent::Shader:
		assert( cs != EmptyShader );
		// if( cs == (uint16_t)0 ) __debugbreak();
		completePrevShader( time, profiler );
		prevShader = cs;
		prevShaderTag = tag;
		shaderStart = (int64_t)time;
		return;
	}
	assert( false );
}

HRESULT GpuProfiler::create( size_t maxDepth )
{
	CD3D11_QUERY_DESC desc{ D3D11_QUERY_TIMESTAMP_DISJOINT };
	CHECK( device()->CreateQuery( &desc, &disjoint ) );
	CHECK( queries.create() );
	stack.reserve( maxDepth );
	return S_OK;
}

void GpuProfiler::blockStart( eProfilerBlock which )
{
	BlockState* parentBlock;
	if( stack.empty() )
	{
		context()->Begin( disjoint );
		parentBlock = nullptr;
	}
	else
		parentBlock = *stack.rbegin();

	BlockState* bs = nullptr;
	auto p = blockStates.Lookup( which );
	if( nullptr != p )
		bs = &p->m_value;
	else
	{
		BlockState& block = blockStates[ which ];
		block.destBlock = &results[ (uint16_t)which ];
		block.destBlock->dest = &dest.measure( which );
		bs = &block;
	}
	bs->parentBlock = parentBlock;
	queries.submit( bs, eEvent::BlockStart );
	stack.push_back( bs );
}

void GpuProfiler::blockEnd()
{
	assert( !stack.empty() );
	BlockState* const bs = *stack.rbegin();
	queries.submit( bs, eEvent::BlockEnd );
	stack.pop_back();

	if( !stack.empty() )
		return;

	const D3D11_QUERY_DATA_TIMESTAMP_DISJOINT dtsd = waitForDisjointData( disjoint );
	queries.join();

	if( !dtsd.Disjoint )
	{
		// Fortunately, these timers appear to be relatively high resolution.
		// Specifically, on the iGPU inside Ryzen 7 5700G that frequency is 1E+8 = 100 MHz
		// On nVidia 1080Ti, that frequency is 1E+9 = 1 GHz
		const uint64_t freq = dtsd.Frequency;
		resultsMakeTime( freq );
	}
	else
	{
		// Something occurred in between the query's ID3D11DeviceContext::Begin and ID3D11DeviceContext::End calls 
		// that caused the timestamp counter to become discontinuous or disjoint, such as unplugging the AC cord on a laptop, overheating, or throttling up/down due to laptop savings events.
		// The timestamp returned by ID3D11DeviceContext::GetData for a timestamp query is only reliable if Disjoint is FALSE.
		resultsDropPending();
	}
}

void GpuProfiler::computeShader( eComputeShader cs )
{
	assert( !stack.empty() );
	if( !profileShaders )
		return;

	BlockState* const bs = *stack.rbegin();
#if PROFILER_COLLECT_TAGS
	queries.submit( bs, eEvent::Shader, (uint16_t)cs, m_nextTag );
	m_nextTag = 0;
#else
	queries.submit( bs, eEvent::Shader, (uint16_t)cs );
#endif
}

void GpuProfiler::resultsDropPending()
{
	for( POSITION pos = results.GetStartPosition(); nullptr != pos; )
		results.GetNextValue( pos ).dropPending();
#if PROFILER_COLLECT_TAGS
	for( POSITION pos = resultsTagged.GetStartPosition(); nullptr != pos; )
		resultsTagged.GetNextValue( pos ).dropPending();
#endif
}

void GpuProfiler::resultsMakeTime( uint64_t freq )
{
	for( POSITION pos = results.GetStartPosition(); nullptr != pos; )
		results.GetNextValue( pos ).makeTime( freq );
#if PROFILER_COLLECT_TAGS
	for( POSITION pos = resultsTagged.GetStartPosition(); nullptr != pos; )
		resultsTagged.GetNextValue( pos ).makeTime( freq );
#endif
}

void GpuProfiler::resultsReset()
{
	for( POSITION pos = results.GetStartPosition(); nullptr != pos; )
		results.GetNextValue( pos ).reset();
#if PROFILER_COLLECT_TAGS
	for( POSITION pos = resultsTagged.GetStartPosition(); nullptr != pos; )
		resultsTagged.GetNextValue( pos ).reset();
#endif
}

#if PROFILER_COLLECT_TAGS
uint16_t __declspec( noinline ) GpuProfiler::setNextTag( const char* name )
{
	uint16_t tag = dest.makeTagId( name );
	m_nextTag = tag;
	return tag;
}
#endif

HRESULT GpuProfilerSimple::create()
{
	ID3D11Device* const dev = device();

	CD3D11_QUERY_DESC desc{ D3D11_QUERY_TIMESTAMP_DISJOINT };
	CHECK( dev->CreateQuery( &desc, &disjoint ) );

	desc.Query = D3D11_QUERY_TIMESTAMP;
	CHECK( dev->CreateQuery( &desc, &begin ) );
	CHECK( dev->CreateQuery( &desc, &end ) );

	context()->Begin( disjoint );
	context()->End( begin );
	return S_OK;
}

HRESULT GpuProfilerSimple::time( uint64_t& rdi ) const
{
	context()->End( end );

	try
	{
		const D3D11_QUERY_DATA_TIMESTAMP_DISJOINT dtsd = waitForDisjointData( disjoint );
		const uint64_t t2 = getTimestamp( end, delay );
		const uint64_t t1 = getTimestamp( begin, delay );

		if( !dtsd.Disjoint )
		{
			rdi = makeTime( t2 - t1, dtsd.Frequency );
			return S_OK;
		}
		else
		{
			// Something occurred in between the query's ID3D11DeviceContext::Begin and ID3D11DeviceContext::End calls 
			// that caused the timestamp counter to become discontinuous or disjoint, such as unplugging the AC cord on a laptop, overheating, or throttling up/down due to laptop savings events.
			// The timestamp returned by ID3D11DeviceContext::GetData for a timestamp query is only reliable if Disjoint is FALSE.
			rdi = -1;
			return S_FALSE;
		}
	}
	catch( HRESULT hr )
	{
		return hr;
	}
}