[Bf-committers] How about adding boost uBLAS library toblender/extern?

[Mathias Panzenböck]

Timothy Baldridge schrieb:
>> According to a Prof. of mine[1] loading/storing from/to SIMD registers is in
>> fact slow and should be avoided.
> 
> That would be a assumption that I would consider do be dangerous to assume.
> 
> I once benchmarked the difference in using standard 32-bit registers
> for copying data from one portion of memory (64MB chunk). The standard
> registers clocked in at about 2GB/sec while I hit almost 12GB/sec
> using 4x SIMD registers as I could load 128 bits at once.
> 
> Now what your prof probably said was that it was slower to load a
> 128bit register than a 32bit register so you shouldn't use SSE unless
> you are doing SIMD. So in this case, we want to use SIMD as we are
> working with 3-4 32bit floats that will just "happen" to fit perfectly
> in a 128bit SIMD register.
> 
> Timothy
> 

Maybe he meant it that way.