[Bf-committers] How about adding boost uBLAS library toblender/extern?
markos at codex.gr
Mon Jan 26 15:37:11 CET 2009
I'll only comment on the points I disagree.
Στις Monday 26 January 2009 16:08:19 ο/η Shaul Kedem έγραψε:
> - C++ is a superset of C, which means C++ can never be faster than C..
Valid for most cases but not always, see below.
> - Template based coding is hard to debug, sometimes impossible
> (believe me, when u see gdb crash it's not a pretty sight)
> - C is simpler and compilers are better in optimizing C code
not always true. Unless you hand tune C code, a C++ compiler has better chance
of rescheduling the instructions for a given algorithm.
> It's not a fluke that linus didn't switch to C++ for linux - he has
> good reasons for it. UI is nice in C++, even some other things more
> related to data manipulation, but kernel work should remain in C, and
> if I'll take Yves' argument and reverse it - there is no reason why we
> can't learn from eigen and code the same optimizations in C,
At the start of 2008 I was determined to change the whole math stuff in
Blender - and I actually have a local tree (based on 2.45)- with every math
routine replaced with an optimised version that is SIMD optimised. In theory
it would offer a big speedup (my plan was to do SIMD optimisations first and
then port it to Cell), but that didn't work for the following reasons, this
was the reason that I gave up -or rather put this work on hold:
1. replacing all math functions and passing normal float structs in every
function, means that everytime you have to do load/save from the vector
registers to memory, which degrades performance hugely. Ideally, it would mean
that eg. to do a complex matrix manipulation, which would be really sped up by
SIMD math, unless one would write custom SIMD code in the function (and thus
ignoring the math functions completely) the speed gain would be minimal,
because the overhead of loading the matrix into registers, doing the
calculation, then saving again, then loading again (in another function),
performing another calculation, saving, etc. would counter any gain from SIMD
optimisations. I know, I did it already. The same is true for SSE, AltiVec and
Cell optimisations (my work is exactly related to SIMD architectures from SSE
and Cell to the ARM Neon).
2. Eigen uses templated matrix structures. For complex operation, it uses
internally the vector form and only loads/saves at the end of the operation.
While in practice it's still slower than hand-tuned optimised SIMD code, in
most cases where change to the code should remain minimal and as readable as
possible -you don't want loads of #ifdef SSE inside the renderer code as it
would hinder readability. In my experience, using some library like Eigen2
would increase readability, remove the necessity for separate math libraries
and also provide extra speed over the current scheme or even anything that
would not involve hand-tuned optimised SIMD code inside each routine.
Of course it's your call, just wanted to let you know that in some cases C++
*can* actually be faster than C, especially if you don't really want to go
More information about the Bf-committers