Dear all,

Is there any kind of standard benchmark patch which I can use to index the performance of a certain Pure Data implementation? I would like to perform the benchmark on our embedded Pure Data version.

Regards,
Dinne
www.blue-sense.com

No, there is no standard patch. But to help you achieve this note there are three important branches to a Pd benchmark.

In each case create a small abstraction and replicate it as many times you can while watching the CPU usage (with both Pd internal timer and top from the shell)

1. Arithmetic SD in register benchmark

This is a signal domain patch, for synthesis etc, which can be computed only using arithmetic and trig and which is presumed to execute only using registers. Something like f(t) = a1 cos (kwt +p) + a2 cos (k2wt + p2) + a3 cos (k3wt + p3)... a simple additive synthesiser.

The next test is memory intensive.

2. DRAM IO test

You want an abstraction that creates a $0- named array writes noise to it and reads it back out (in the signal domain). The cost of [noise] and [tabread4~] are constant and negligable so this tests the memory IO bandwidth.

3. MD compuation

The last test is for message domain computation (because this is scheduled differently). Use [until] to perform an iterative procedure like adding up a bunch of numbers or finding a square root by Newtons method.

Those three should give you a good all round indicator of how the build and processor compare to some standard like a 1GHz Intel i386.

You didn't say what the board/uC is. It may be possible to calculate the expected performance using a few metrics from that data and compare that with what you actualy see.