posted 09 April 2007 11:52            

Internal processing of the DSP is 56 bits.
When data is shared between them, it is 24 bits.
The resolution is really really good!

posted 10 April 2007 13:13            

Maybe I'm wrong, but my concept about how Kyma manages the prototypes it's like "pieces of code", so when we insert a prototype in the path it's like we insert some lines of code in a script. When we do Play the script it's loaded in Capybara. So, in my opinion, the only time we have 24 bit it's at the digital output (and input). All the audio processing "between" prototypes it's 56 bit. A prototype it's a "visual metaphore" of some lines of code.

posted 10 April 2007 16:14            

your metaphor sounds right, but unfortunately it is not how DSP 56300 processors works. Their arithmetic and logical unit (ALU) have 24 bits registers and 56 bits accumulators. Roughly, it means that internal processing in done in 56 bits, and errors are "accumulated" in those 56 bits accumulators. The 56 bits processing is used internally for a single assembly call (correct me if I'm wrong), which means that a single prototype which uses many assembly calls, is always 24 bits on input and output.

posted 14 April 2007 02:51            

quote:

---

Originally posted by tuscland:
56 bits processing is used internally for a single assembly call (correct me if I'm wrong), which means that a single prototype which uses many assembly calls, is always 24 bits on input and output.

But there's no need to worry because 24 bits is really more than enough, when you consider the quality of the algorithms that Kurt and Carla have written. This is what really cares.

---

This is correct! Although depending on your algorithm it could well be you need to save a 24 bits as temperary result and continue with this. As long as the programmer is aware of this and designs his algorithm correctly there is no audible problem. In audio coding one has +/- 4 or 5 bits per subsample and one can not hear the difference...
For most algorithms 24 bits is more then enough, although I often run into the question, should this be done in 24 bit or 48 bit. For example my LowPass2nd_order uses 48 bit delays in the feedback of the filter to allow high Q values up to 128 (=+/-42dB). If a programmer would program this with 24 bits, the filter would not output anything!

As a user of kyma you should not worry about this! 24 bit => 120 dB, in practical terms this means with 20 dB headroom a signal of 100 dB has noise on level at threshold of our hearing. It means this will 100% be masked! Use your ears.

For a programmer of the algorithms it would however be handy to have 48 bits or even floating point numbers, the fixed point can be a p...in the a.. In other architectures it is just handy to work with another amount of bits. The commercial guys just promote this without knowing if it helps at all.