Re: [Help-glpk] GLPK re-endrant

[Francois Galea]

Ali Baharev a écrit :
> Hi,
> 
> Once upon a time, I had a discussion with one of the GPULib developers
> and he told me GPUs were not good for the simplex method, because the
> gain is not so significant for sparse matrix-vector operations. (Or at
> least it was what i understood.)

Moreover, the RAM dedicated for GPU's is different, as it is not totally 
data-safe. Some bits may change during time, potentially introducing 
errors in computations. You cannot implement an exact solver using such 
hardware.

In addition, GPUs were historically been designed for computing on 
single precision computations. Double precision has recently been 
included on most modern GPUs, though.

Finally, modern CPU's also have SIMD instructions (SSE/SSE2/... on 
Intel's, Altivec on POWER's) which can be used for data-parallel 
operations on a single core. C compilers provide the possibility to uses 
such instructions directly in C, through special "function calls" called 
'intrinsics'. Though, already-existing software such as GLPK would 
require heavy re-rengineering (most data structures and algorithms would 
require to be redesigned) to support such a programming model.


-- 
François Galea
INRIA - Saclay - Batiment N
Parc Orsay Université
4 rue J. Monod F-91893 Orsay Cedex
Tél. : +33 1 72 92 59 24