made rendering happen with one global working-buffer per thread - hopefully improves L1/L2-cache-efficiency

git-svn-id: https://lmms.svn.sf.net/svnroot/lmms/trunk/lmms@890 0778d3d1-df1d-0410-868b-ea421aaaa00d

src/core/instrument.cpp

@@ -49,14 +49,7 @@ instrument::~instrument()
 
 
 
-void instrument::play( bool )
-{
-}
-
-
-
-
-void instrument::playNote( notePlayHandle *, bool )
+void instrument::play( bool, sampleFrame * )
 {
 }
 

src/core/mixer.cpp

@@ -68,6 +68,39 @@
 sample_rate_t SAMPLE_RATES[QUALITY_LEVELS] = { 44100, 88200 } ;
 
 
+
+#define ALIGN_SIZE 64
+
+void aligned_free( void * _buf )
+{
+	if( _buf != NULL )
+	{
+		int *ptr2=(int *)_buf - 1;
+		_buf = (char *)_buf- *ptr2;
+		free(_buf);
+	}
+}
+
+void * aligned_malloc( int _bytes )
+{
+	char *ptr,*ptr2,*aligned_ptr;
+	int align_mask = ALIGN_SIZE- 1;
+	ptr=(char *)malloc(_bytes +ALIGN_SIZE+ sizeof(int));
+	if(ptr==NULL) return(NULL);
+
+	ptr2 = ptr + sizeof(int);
+	aligned_ptr = ptr2 + (ALIGN_SIZE- ((size_t)ptr2 & align_mask));
+
+
+	ptr2 = aligned_ptr - sizeof(int);
+	*((int *)ptr2)=(int)(aligned_ptr - ptr);
+
+	return(aligned_ptr);
+}
+
+
+
+
 class mixerWorkerThread : public QThread
 {
 public:
@@ -146,6 +179,9 @@ public:
 private:
 	virtual void run( void )
 	{
+		sampleFrame * working_buf = (sampleFrame *) aligned_malloc(
+						m_mixer->framesPerPeriod() *
+							sizeof( sampleFrame ) );
 		m_jobWait.acquire();
 		m_jobAccepted.acquire();
 		m_idle = TRUE;
@@ -167,7 +203,7 @@ private:
 					switch( it->type )
 					{
 						case PlayHandle:
-		it->playHandleJob->play();
+		it->playHandleJob->play( FALSE, working_buf );
 							break;
 						case AudioPortEffects:
 							{
@@ -196,6 +232,7 @@ private:
 			m_idle = TRUE;
 			m_sem->release();
 		}
+		aligned_free( working_buf );
 	}
 
 	mixer * m_mixer;
@@ -211,6 +248,7 @@ private:
 
 mixer::mixer( void ) :
 	m_framesPerPeriod( DEFAULT_BUFFER_SIZE ),
+	m_workingBuf( NULL ),
 	m_readBuf( NULL ),
 	m_writeBuf( NULL ),
 	m_cpuLoad( 0 ),
@@ -249,10 +287,14 @@ mixer::mixer( void ) :
 		m_fifo = new fifo( 1 );
 	}
 
+	m_workingBuf = (sampleFrame*) aligned_malloc( m_framesPerPeriod *
+							sizeof( sampleFrame ) );
 	for( Uint8 i = 0; i < 3; i++ )
 	{
-		m_readBuf = new surroundSampleFrame[m_framesPerPeriod];
-		
+		m_readBuf = (surroundSampleFrame*)
+			aligned_malloc( m_framesPerPeriod *
+						sizeof( surroundSampleFrame ) );
+
 		clearAudioBuffer( m_readBuf, m_framesPerPeriod );
 		m_bufferPool.push_back( m_readBuf );
 	}
@@ -287,8 +329,10 @@ mixer::~mixer()
 
 	for( Uint8 i = 0; i < 3; i++ )
 	{
-		delete[] m_bufferPool[i];
+		aligned_free( m_bufferPool[i] );
 	}
+
+	aligned_free( m_workingBuf );
 }
 
 
@@ -435,7 +479,7 @@ const surroundSampleFrame * mixer::renderNextBuffer( void )
 				playHandle * n = m_playHandles[idx];
 				if( !n->done() && n->supportsParallelizing() )
 				{
-					n->play( TRUE );
+					n->play( TRUE, m_workingBuf );
 					par_hndls.push_back( n );
 				}
 				++idx;
@@ -489,7 +533,7 @@ if( COND_NPH )
 				if( !( *it )->done() )
 				{
 
-					( *it )->play();
+					( *it )->play( FALSE, m_workingBuf );
 				}
 			}
 		}

src/core/note_play_handle.cpp

@@ -157,7 +157,8 @@ notePlayHandle::~notePlayHandle()
 
 
 
-void notePlayHandle::play( bool _try_parallelizing )
+void notePlayHandle::play( bool _try_parallelizing,
+						sampleFrame * _working_buffer )
 {
 	if( m_muted == TRUE )
 	{
@@ -177,7 +178,8 @@ void notePlayHandle::play( bool _try_parallelizing )
 	if( framesLeft() > 0 )
 	{
 		// play note!
-		m_instrumentTrack->playNote( this, _try_parallelizing );
+		m_instrumentTrack->playNote( this, _try_parallelizing,
+							_working_buffer );
 	}
 
 	if( m_released == TRUE )
@@ -238,7 +240,7 @@ void notePlayHandle::play( bool _try_parallelizing )
 	for( notePlayHandleVector::iterator it = m_subNotes.begin();
 						it != m_subNotes.end(); )
 	{
-		( *it )->play( _try_parallelizing );
+		( *it )->play( _try_parallelizing, _working_buffer );
 		if( ( *it )->done() )
 		{
 			delete *it;

src/core/preset_preview_play_handle.cpp

@@ -161,9 +161,10 @@ presetPreviewPlayHandle::~presetPreviewPlayHandle()
 
 
 
-void presetPreviewPlayHandle::play( bool _try_parallelizing )
+void presetPreviewPlayHandle::play( bool _try_parallelizing,
+						sampleFrame * _working_buffer )
 {
-	m_previewNote->play( _try_parallelizing );
+	m_previewNote->play( _try_parallelizing, _working_buffer );
 }
 
 

src/core/sample_play_handle.cpp

@@ -116,7 +116,8 @@ samplePlayHandle::~samplePlayHandle()
 
 
 
-void samplePlayHandle::play( bool /* _try_parallelizing */ )
+void samplePlayHandle::play( bool /* _try_parallelizing */,
+						sampleFrame * _working_buffer )
 {
 	//play( 0, _try_parallelizing );
 	if( framesDone() >= totalFrames() )
@@ -128,14 +129,11 @@ void samplePlayHandle::play( bool /* _try_parallelizing */ )
 	if( !( m_track && m_track->muted() )
 				&& !( m_bbTrack && m_bbTrack->muted() ) )
 	{
-		sampleFrame * buf = new sampleFrame[frames];
 		stereoVolumeVector v = { { m_volumeModel->value(),
 						m_volumeModel->value() } };
-		m_sampleBuffer->play( buf, &m_state, frames );
-		engine::getMixer()->bufferToPort( buf, frames, offset(), v,
-								m_audioPort );
-
-		delete[] buf;
+		m_sampleBuffer->play( _working_buffer, &m_state, frames );
+		engine::getMixer()->bufferToPort( _working_buffer, frames,
+						offset(), v, m_audioPort );
 	}
 
 	m_frame += frames;

src/tracks/instrument_track.cpp

@@ -338,7 +338,8 @@ void instrumentTrack::processOutEvent( const midiEvent & _me,
 
 
 
-void instrumentTrack::playNote( notePlayHandle * _n, bool _try_parallelizing )
+void instrumentTrack::playNote( notePlayHandle * _n, bool _try_parallelizing,
+						sampleFrame * _working_buffer )
 {
 	// arpeggio- and chord-widget has to do its work -> adding sub-notes
 	// for chords/arpeggios
@@ -400,7 +401,8 @@ void instrumentTrack::playNote( notePlayHandle * _n, bool _try_parallelizing )
 			}
 		}
 		// all is done, so now lets play the note!
-		m_instrument->playNote( _n, _try_parallelizing );
+		m_instrument->playNote( _n, _try_parallelizing,
+							_working_buffer );
 	}
 }