jwetzl / cudalbfgs Goto Github PK

Hi there,

I happened to find that in your kernel function void strongWolfePhase1(bool second_iter) (in linesearch_gpu.h), the variable status is not initialized. So if the old value stored in device memory happened to be 2, bad things will happen.

I wrote a program with cudalbfgs and tested it with 550Ti and GTX 760. The result with former card looks like normal, but result with GTX 760 is incorrect(most of the values are zero). So I am wondering do I have to be aware of something when using different card with different compute compatibility?

so far, I had the 'cmake ..' and make done!

how to do the foillowing??

• build a reference implementetation on CPU with
  either float or double precision (requires Eigen),
• build test cases,
• enable error checking, verbose output and timing
• build example projects that demonstrate how the
  library is used (cf. /projects directory).

I try the quadratic example, using size_t n = 5000, the result shows "Line search failed"

There is a problem with cuda_compute_capability.c. It is caused by:

    if (major == 2 && minor == 1)
    {
        // There is no --arch compute_21 flag for nvcc, so force minor to 0
        minor = 0;
    }

See, the problem is that some Fermi cards do support Compute v2.1. In fact, Compute v2.1 exists (see https://en.wikipedia.org/wiki/CUDA#Supported_GPUs) but the way to activate that would be through setting the flags as -arch compute_20 -code sm_21.

The script currently assumes that whatever compute_xx is, sm should be also followed by the same number and be set as sm_xx. I've ran into problems with sm_20 on a machine that supports sm_21 before. For instance, I vaguely recall that numerical computations were more accurate with sm_21 than sm_20 (on the Caffe library if I recall). Considering the large number of CMake scripts out there that rely on this script, I hope the issue is fixed :)

Unfortunately, my knowledge about CMake is rather limited, otherwise I would've fixed it and submitted a pull request.