-
Twisted-clover fermions do not currently support the fused exterior dslash policy (i.e., where a single kernel is used to update all boundaries instead of separate kernels for each of x, y, z and t di…
-
When I try to build with
`BUILD_FATLINK=yes`
there seem to be some missing includes / namespace issues:
`
interface_quda.cpp: In function ‘void computeKSLinkQuda(void*, void*, void*, void*, double*,…
-
CUDA 5.0 will have a separate device linker, this will allow us to put all the kernels in dslash_quda.cu (e.g., dslash, force, link fattening) into separate files since the constants and textures will…
-
I just wanted to see whether the issues in #158 can be reproduced with the quad tests and noted for a single GPU QUDA build:
[mwagner@cream tests]$ ./hisq_paths_force_test --gauge-order milc --prec …
-
QUDA is sorely lacking an eigenvector solver. We need to investigate what type of solver is most appropriate for including in QUDA. What are the relative merits of Ritz and Lanczos solvers?
Ritz …
-
The compile time of QUDA, especially in multi-GPU mode, hinders development. We should really work out how to enable parallel building of QUDA, namely, how to split parallelize dslash_quda.cu. This …
-
Currently resIncrease only counts subsequent increases of the L2 norm. That does not catch alternating increase / decrease scenarios.
-
Hi All,
Mike asked me to point out what needs to be done for the clover program to work correctly on GPUs. -- Basically, what are the pieces that our 'outside of QDP++' that QDP-JIT cannot handle.
…
-
We noticed that the number of iterations reported back to MILC is wrong:
```
MultiShift CG: Converged after 9123 iterations
...
WARNING: Check number of iterations to be reported to MILC 4682860 468…
-
It seems like the tune cache does not save the CUDA version used to compile QUDA.
As register usage of Kernels is likely to change with different CUDA versions using an tune cache for QUDA compiled wi…