[tz] [PROPOSED] Do not assume bytes have 8 bits

Brian Inglis Brian.Inglis at SystematicSw.ab.ca
Fri Apr 23 22:40:24 UTC 2021


On 2021-04-23 13:27, Paul Eggert via tz wrote:
> * zic.c (convert, convert64): Mask bytes with 0xff before storing
> them, for portability to machines where bytes have more than 8 bits.
> Although this is surely only of theoretical interest, we might as
> well be portable.

Not uncommon on DSPs which are now being used not only for graphics but also for 
machine learning and neural network processing for audio speech recognition, 
video object recognition and classification, and natural language recognition e.g.

https://www.embecosm.com/2017/04/18/non-8-bit-char-support-in-clang-and-llvm/

The models are so large with billions or trillions of parameters that they hit 
limits on systems with 4 GPUs each having 12GB local memory, 64-128GB system 
memory, even with usual optimizations to use 16 bit floats.

-- 
Take care. Thanks, Brian Inglis, Calgary, Alberta, Canada

This email may be disturbing to some readers as it contains
too much technical detail. Reader discretion is advised.
[Data in binary units and prefixes, physical quantities in SI.]


More information about the tz mailing list