What steps will reproduce the problem?
1.Simply run the attached source code. Input data are attached as well.
2.
3.
What is the expected output? What do you see instead?
>> We expect the A*x=b problem can be solved by CG method successfully and very
quickly.
What version of the product are you using? On what operating system?
>> The testing environment is as follows:
CUDA: NVIDIA CUDA Toolkit 4.0(32 bit)
SDK: NVIDIA GPU Computing SDK 4.0
CUSP: cusp v0.2.0
GPU: NVIDIA GeForce GTX465(1G)
CPU: Intel(R) Core(TM) i5-2500 CPU @3.30GHz
Memory(RAM): 4G
Windows: Windows 7 Ultimate Service Pack 1-32bit
Please provide any additional information below.
>> We are using cusp in a research project to solve linear system of equations,
ie, A*x=b. A is non-symmetric sparse matrix. CUSP works fine for small size of
A (eg, 180*180). But when we test it for medium-sized A (eg, 5237*5237 with
33703 non-zero elements), the BiConjugate Gradient Stabilized method does not
converge. The GMRES method converges yet very slowly. We intend to use the
BiConjugate Gradient Stabilized method, however, we don't know how the
pre-conditioner should be set. Any advice on that will be appreciated.
Some additional information: the same problem has been tested on MATLAB, where
Conjugate Gradient method can solve but very slowly. Back slash (\) operator
however can solve the equations in no time. I believe in our case, underneath
back slash operator, MATLAB used LU decomposition as the internal solver.
Attached are the matrix A and vector b in mtx format, as well as the CUSP
source code used for testing. I appreciate if you can look into the issue and
help us solve it. Thanks!
Original issue reported on code.google.com by rbo2...@gmail.com on 22 Nov 2011 at 6:20
Original issue reported on code.google.com by
rbo2...@gmail.comon 22 Nov 2011 at 6:20Attachments: