[<<][rai][>>][..]
Mon Sep 22 15:48:13 CEST 2014

Fixed point

So, go over the M4F fixed point instructions and see how it best maps
to existing rai functions.  Then extend the primitives to a sensible
goldielocks subset.

- saturation
- count leading zeros
- mac
- mul w x w-> w (same-wordsize multiplication)
- pack / unpack
- extend
- bitfield


Notes:

- selecting multiplications: sign, top/bottom, accumulate,
- additions: 
  - Q / GE setting?
  - halved results (is this for FFT?)
  - exchange
- dual


So apart from saturation these are all easily expressed as
combinations of primitives.

Most of the effort in the DSP instructions is in routing operands and
results.  Moving stuff around...

So the question:

- how much of this can practically be done by the compiler?  are there
  any compilers accessible to me at this point that actually do a good
  job at this?

- if the "routing optimization" problem is in need for human
  assistance, maybe best to start from that point: make things into a
  "puzzle" that automates most of the red tape, but leaves the
  cleverness to the human operator.


Practically, for getting fixed point things to run on the axo, it's
probably best to work with what's there and add saturation, bit count
(rounded log 2).  then see if a postprocessing step can insert
intrinsics or leave it entirely up to the compiler.

i'm interested in the idea of "augmented assembler" but not sure if i
want to get into general compiler things like register allocation and
instruction selection.




[Reply][About]
[<<][rai][>>][..]