daschaich / susy Goto Github PK

Once the truncated links are deprecated we can implement a bit more code reuse.

A bit more will also be allowed by implementing 2d_Q04 dimensional reduction center stabilization, as well as the spectral density and eigenmode number computations, for example in params.h.

Need to figure out some way to test truncated code for correctness before merging the truncated_dev branch into master...

To help users keep track of how much memory is really being used, the usage due to the following FIELD_ALLOCS should be reported along the lines of the main fields and the pfaffian matrix at setup.c#L220:

1. The eigenvectors at setup.c#L470

2. The stochastic sources for the mode number computation at setup.c#L499

3. The SU(2) matrix for gauge-fixing hits, which was just moved to setup.c#L457 on the gather_dev branch

The current method of reading in runtime parameters from the command line seemed to lead to some trouble interfacing with performance monitoring tools. It may be more modern to read in from a file (a la QHMC), at least as an option.

This is mainly cosmetic, but becoming more worthwhile as the number of Twist_Fermion routines grows.

Currently the eigensolver works with PRIMME v1.2.2. Changes to the PRIMME interface since then are currently incompatible with eig.c/complex.h.

This should be a straightforward (if minor) performance improvement for I/O: In */generic/io_lat.c, there are currently lots of NUMLINK * sizeof(fmatrix) multiplications inside loops over the lattice volume, which can be brought out front and stored in a constant data_size variable, as in the 2d_SQCD_dev branch.

May take some work, but should give worthwhile performance boost --- good first step would be to check benefits in MILC's staggered multi-CG using QUDA...

Hi there,

When I run the program, it encounters an "Division by Zero" error. The error-inducing inputs are:

nx: 3
ny: 1
nz: 1
nt: 3
PBC: 5
iseed: 1
Nroot: 3
Norder: 1

The error locates in function setup_layout() inside susy/4d_Q16/generic/layout_hyper_prime.c. I fixes the error by checking if the divisor is 0, which is achieved via inserting the following code segment before the division operation, i.e., before line 115.

if (squaresize[XUP] == 0 ||
squaresize[XUP] == 0 ||
squaresize[ZUP] == 0 ||
squaresize[TUP] == 0 )
{
Abort();
}

Thank you!

Hi there!

In my recent use of this program, I found three segmentation fault bugs due to the inappropriate use of memory allocation routine malloc(). Below I list a verified fix for each:

1. Line 131 in susy/4d_Q16/susy/update_o.c
   Twist_Fermion **src = malloc(Nroot * sizeof(**src));
   should be changed to
   Twist_Fermion **src = malloc(Nroot * sizeof(Twist_Fermion *));

2. Line 13 in susy/4d_Q16/susy/update_o.c
   Twist_Fermion ***psim = malloc(Nroot * sizeof(***psim));
   should be changed to
   Twist_Fermion ***psim = malloc(Nroot * sizeof(Twist_Fermion **));

3. Line 44 in susy/4d_Q16/susy/congrad_multi.c
   Twist_Fermion **pm = malloc(Norder * sizeof(**pm));
   should be changed to
   Twist_Fermion **pm = malloc(Norder * sizeof(Twist_Fermion *));

Thank you!

Is there a way to rearrange the stochastic mode number computation so that it does the inversions for all Omega simultaneously? Currently we have separate calls to the multi-shift solver for each Omega, each with the same source(s).