cvxgrp / scs

Splitting Conic Solver
MIT License
547 stars 134 forks source link

GPU algorithm stability? #77

Open tachim opened 7 years ago

tachim commented 7 years ago

Is there any known solver instability when solving with GPU? E.g. I get the following behavior

----------------------------------------------------------------------------
    SCS v1.2.6 - Splitting Conic Solver
    (c) Brendan O'Donoghue, Stanford University, 2012-2016
----------------------------------------------------------------------------
Lin-sys: sparse-indirect GPU, nnz in A = 17734846, CG tol ~ 1/iter^(2.00)
eps = 1.00e-03, alpha = 1.50, max_iters = 2500, normalize = 1, scale = 1.00
Variables n = 9651600, constraints m = 10832400
Cones:  primal zero / dual free vars: 4234050
    linear vars: 2952000
    sd vars: 3646350, sd blks: 1
Setup time: 1.01e+00s
----------------------------------------------------------------------------
 Iter | pri res | dua res | rel gap | pri obj | dua obj | kap/tau | time (s)
----------------------------------------------------------------------------
     0| 2.07e+00  9.99e-01  1.07e-01 -2.64e+08 -2.13e+08  1.16e-08  1.40e+01 
   100| 2.59e+00  9.99e-01  2.85e-01 -1.68e+08 -9.36e+07  7.26e-09  1.38e+02 
   200| 2.59e+00  9.99e-01  2.85e-01 -1.68e+08 -9.36e+07  2.48e-09  2.61e+02 
   300|      inf       inf  2.85e-01 -1.68e+08 -9.36e+07  6.02e-09  3.83e+02 
   400|      inf       inf  2.85e-01 -1.68e+08 -9.36e+07  1.41e-09  5.05e+02 
bodono commented 7 years ago

Does the same thing happen for you when not running on GPU?