[sldev] [VWR] Vectorization on i386?

[Dzonatas]

Callum Lerwick wrote:
> You seem confused about what I mean when I say "enabled". The whole
> point of this thread is the SSE code is *there*, its getting compiled
> into the viewer. Compiling the SSE code is *not* disabled under gcc.
> Only under MSVC. The comments in llv4math.h and
> llviewerjointmesh_sse2.cpp indicate this is the intention. SSE is not
> compiled at all on "Windows". Though this would be another case of
> saying "Windows" when you really mean MSVC...
>
> By "enabled" in this thread I mean enabling the SSE code at *runtime*.


There is a way to enable this into the runtime on MSVC. It's true that 
MSVC leaks SSE, but it is also true that it can be isolated. For some 
reason, there is a patch to isolated it, but it never has been applied.  
It has actually been tested many times, so it works -- even on the 
machines that it doesn't allow SL to run on when not isolated.

The only worry is that the implementation used STL strings to aid in 
isolation, but the STL strings need to be replaced with C strings to 
make it work on Vista. (don't ask why)

There is also another patch that speeds up the avatar skinning by up to 
an additional 10% just by untangling a few headers and a couple of 
in-line code moves. It is found with code related to the 
llpolyskeletaldistortion patch:

https://jira.secondlife.com/browse/VWR-1812


..."This is too complicated and fragile for us to take, I'm afraid."  ? 
*shrugs* ?

-- 
Power to Change the Void