[Linuxsampler-devel] optimizations

[Vladimir S. <ha...@gm...>]

Hi Everybody,

It may seem like LS has been quiet for a little while . . .
But a lot of stuff has been happening behind the scenes.
Christian and I have been working on optimizations with the goal of
making LS nothing short of state of the art when it comes to
efficiency.

We focused on improving efficiency of real time calculations.
The project consisted of two major developments:
1) Reducing the amount of work and conditionals per sample point
depending on the state of the voice.
2) Developing MMX/SSE(1) implementation for most calculation intensive
routines (as well as providing a framework for further implementations
such as SSE2/3, altivec, etc.

To acomplish task number 1 without introducing branches into the "fast
path" we've introduced a concept which we called "synthesis mode". An
entire segment is processed in the same mode.
Function to process the entire mode is generated using highly
templated CPP code.
Ideally (depending on your compiler and options) an entire segment
will be processed without a single call with only one branch (loop).
So there is a function for each "synthesis mode" and there is a
function pointer that is set for a voice before the segment is
processed.

To acomplish task number 2 we had to write a few X86 assembly routines
and do a lot of benchmarking, testing, etc. For some modes we
acomplished speeds several times faster than CPP implementation.
Further optimization will be made in some areas.
We hit a number of issues related to FPU/MMX state problems of Pentium
and also lots of funny GCC related things. We really wanted to support
pre-SSE2 processors though.

Christian is working hard right now on checking everything into the CVS.
I'd like to point out that there is still some work to be done (as usual) like:
1) Some cleanup, putting #if ARCH_X86 in place, etc
2) Integrating benchmarking into LS. For now we have separate
benchmarking suite that we've used but it's not very user friendly and
up to date, so we are not releasing it at this time. Benchmarking will
allow users to do wonderful things such as comparing hardware to pick
best fit for LS, etc.
3) cubic interpolation for MMXSSE needs to be implemented at some point
4) lots of other things

There will be subsequent submits to the CVS in the near future so
please stay tuned!

-- 
Regards,
Vladimir