cpviolator
commented
2 years ago To give you some assurance that QUDA's treatment of the QDPJIT ordered clover term is correct, see for example this line https://github.com/lattice/quda/blob/feature/Nc_agnosticism/include/clover_field_order.h#L789
I attach a simple C++ function to demonstrate that the ordering is valid for all Nc
#include <stdio.h>
#include <iostream>
using namespace std;
int main(int argc, char **argv) {
int Nc = atoi(argv[1]);
int elems = 2*Nc*Nc-Nc;
int idtab[elems];
int idx = 0;
const int L = 2*Nc-1;
for(int col = 0; col<L; col++) {
int id = col*(col+3)/2;
idtab[idx] = id;
std::cout << idtab[idx] << ", ";
idx++;
for(int i=col+1; i<L; i++) {
id += i;
idtab[idx] = id;
std::cout << idtab[idx] << ", ";
idx++;
}
}
std::cout << endl;
bool idtab_check[elems];
for(int i=0; i<elems; i++) idtab_check[i] = false;
for(int i=0; i<elems; i++)
for(int k=0; k<elems; k++)
if(idtab[k] == i) idtab_check[k] = true;
for(int i=0; i<elems; i++)
if(idtab_check[i]) std::cout << "true at " << i << endl;
else {
std::cout << "false at " << i << endl;
exit(0);
}
}
fwinter
I've built some Nc=3,4,5 stacks using a variety of combinations of QDPXX, QDP-JIT, and a color agnostic branch of QUDA:
where in the
+QUDAcases QUDA handles the inversions only as normal.For Nc=3, all 4 stacks produce 2 Flavour clover HMC configurations (no link smearing) with excellent agreement in
Delta Hwhen using the same seed for all 4 stacks. When I move to Nc>3, only theCHROMA + QDPXXandCHROMA + QDPXX + QUDAare stable, in that theirDelta Hvalues are in agreement, and the solver converges. For bothCHROMA + QDPJITandCHROMA + QDPJIT + QUDAstacks, the solver diverges quite quickly. This isolates the problem to Chroma's QDPJIT clover term code. If QUDA's treatment of the QDPJIT ordered clover term were bad, only theCHROMA + QDPJIT + QUDAbuilds would produce poor data. Furthermore, the fact that both theCHROMA + QDPJIT + QUDAandCHROMA + QDPJITstacks produce good data for Nc=3 demonstrates that the issue is in the Nc != 3 QDPJIT code only.I'm using CHROMA devel with master merged in and some minor modifications to place guards around baryon code. QDPXX and QDPJIT are also the devel branches with similar harmless guards around baryon code. The QUDA branch is
feature/Nc_agnostiscismwhich is up-to-date with QUDA's develop branch.If someone could cast a more experienced eye than mine over the QDPJIT clover code in CHROMA to look for an error I cannot see, I would be very appreciative. I can also provide a stack reproducer, but I do not have push rights to QDPXX or QDPJIT, so I will have to fork them for downloading.