Hi David
thanks for the info.
a modern mobile phone varies about 8:1 in what they can do, and also
they are not designed for continuous all core grunt without a thermal
issue, ---they're designed for lots of grunt for very short periods to
satisfy UI requirements (that is being responsive to touch commands and
provide complex UIs).
The only thing that hard on a phone that can run continuously without
havign a thermal meltdown is the video decoder/encoder which is a
dedicated hardware subsystem on the device.
So,I will write to Jean-Marc and and ask some questions. I am looking at
some nextgen codec hardware build and might see just how much is
required. Seems like it will be the top of the available envelope.
Exciting though, a new approach to things...
cheers
On 1/04/2019 9:53 PM, David Rowe wrote:
About 3 GLOPs and 1.2E6 floating point weights, Jean-Marc has recently
summarised here:
https://people.xiph.org/~jm/demo/lpcnet_codec/
Runs in real time OK on a modern mobile phone or x86 from the last
decade (AVX or better).
- David
On 01/04/19 20:17, glen english wrote:
Hi David
Great work on LPCnet/codec2
have you got any idea yet on how many FP ADDS and FP MULS and FPMACS,
(or integer precision?) or sizes of matricies that need solving (how
sparse ????) , that it will take for a minimal and runnable LPCNET so
far? I think the last you you wrote was no good on Pi 3 (unsurprising)
but 20% of a Haswell at 3 GHz ? Also, what size memory are you
regularly accessing (to get an idea of the data-set size that influences
cache misses and cost) ?
cheers
glen
_______________________________________________
Freetel-codec2 mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/freetel-codec2
_______________________________________________
Freetel-codec2 mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/freetel-codec2
_______________________________________________
Freetel-codec2 mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/freetel-codec2
