[sc-users] multiple Convolution2 UGens without CPU overload

[Sciss <contact@xxxxxxxx>]

hello,

i'm trying to understand how to have as many Convolution2 synths  
running as possible. i found out, i get much more synths or much  
bigger fftSize, when the synths carrying Convolution2 are created not  
at once but in a delayed manner. What's the formula to determine the  
best delay between creating them? in the example below i use  
(fftSize / numSynth) / s.sampleRate) seconds between two creations ...

// note: CPU load was measured on macbook pro v1 with 2 ghz
~numChannels = 8;
~fftSize = 16384;
~fftBufs = Buffer.allocConsecutive( ~numChannels * 2, s,  
~fftSize ).reshape( ~numChannels, 2 );
~path = "~/Desktop/Kalligraphie/IRs".absolutePath ++ "/";	// folder  
of impulse responses, the files should be "[ 0 ... 7 ] 
GainHPFCutMix.aif" (binaural stereo!)
~fftBufs.do({ arg chanBufs, speakCh; chanBufs.do({ arg buf, irChan;  
buf.readChannel( ~path ++ speakCh.asString ++ "GainHPFCutMix.aif", 0,  
~fftSize, channels: [ irChan ])})});

(
SynthDef( \convMix, { arg in, out, bufL, bufR, amp = 1.0;
	var loadbang;
	loadbang = Impulse.ar( 0 );
	OffsetOut.ar( out, Convolution2.ar( In.ar( in, 2 ) * amp, [ bufL,  
bufR ], loadbang ));
}).send( s );
)

~mixBusses = Array.fill( ~numChannels, { Bus.audio( s, 2 )});

// this pushes peak CPU to some 400 % (would have to use ~fftSize ==  
2048 for no audio dropouts!)
~convSynths = Array.fill( ~numChannels, { arg ch; Synth( \convMix,  
[ \in, ~mixBusses[ ch ].index, \out, 0, \bufL, ~fftBufs[ch] 
[0].bufnum, \bufR, ~fftBufs[ch][1].bufnum, \amp, 0.5 ], s,  
\addToTail )});


// ... second approach:

~convSynths.do({ arg synth; synth.free; });

// delaying the synth creations results in peak CPU at <80%  
(therefore no audio drop outs)!
(
~convSynths = Array.newClear( ~numChannels );
fork {
	~numChannels.do({ arg ch;
		~convSynths[ ch ] = Synth( \convMix, [ \in, ~mixBusses[ ch ].index,  
\out, 0, \bufL, ~fftBufs[ch][0].bufnum, \bufR, ~fftBufs[ch] 
[1].bufnum, \amp, 0.5 ], s, \addToTail );
		((~fftSize / ~numChannels) / s.sampleRate).wait;
	});
}
)


would it be possible to have a multichannel Convolution2 version,  
like i read in an 8 channel impulse response into one single buffer?  
i know there exists an optimized two-channel FFT version, so i guess  
this could be expanded to numChannels := 1 << n;  i guess that would  
also result in a great(?) CPU load improvement?

thx, -sciss-


p.s.: if i split up the stereo multi channel expansion onto two  
different delayed synths, i get peak CPU even down to 40%