cwpearson/astaroth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Added sum reduction. NOTE: Scalar sum does not pass the automated test but vector sum does. I couldn't see anything wrong with the code itself and I strongly suspect that the failures are caused by loss of precision due to summing a huge amount of numbers of different magnitudes. However I'm not yet completely sure. Something like the Kahan summation algorithm might be useful if the errors are really caused by fp arithmetic.

Browse Source
This commit is contained in:
jpekkila
2019-07-30 14:28:18 +03:00
parent 9796d5e981
commit 69deef66fe
4 changed files with 50 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
src/core/astaroth.cu
View File

@@ -474,7 +474,7 @@ simple_final_reduce_scal(const ReductionType& rtype, const AcReal* results, cons
else if (rtype == RTYPE_MIN) {
res = min(res, results[i]);
}
else if (rtype == RTYPE_RMS || rtype == RTYPE_RMS_EXP) {
else if (rtype == RTYPE_RMS || rtype == RTYPE_RMS_EXP || rtype == RTYPE_SUM) {
res = sum(res, results[i]);
}
else {
@@ -486,7 +486,6 @@ simple_final_reduce_scal(const ReductionType& rtype, const AcReal* results, cons
const AcReal inv_n = AcReal(1.) / (grid.n.x * grid.n.y * grid.n.z);
res = sqrt(inv_n * res);
}
return res;
}

8
src/core/kernels/kernels.cuh
View File

@@ -893,6 +893,10 @@ reduce_scal(const cudaStream_t stream, const ReductionType rtype, const int3& st
kernel_filter<dexp_squared><<<bpg_filter, tpb_filter, 0, stream>>>(vtxbuf, start, end, scratchpad);
kernel_reduce<dsum><<<bpg_reduce, tpb_reduce, sizeof(AcReal) * tpb_reduce, stream>>>(scratchpad, num_elems);
kernel_reduce_block<dsum><<<1, 1, 0, stream>>>(scratchpad, bpg_reduce, tpb_reduce, reduce_result);
} else if (rtype == RTYPE_SUM) {
kernel_filter<dvalue><<<bpg_filter, tpb_filter, 0, stream>>>(vtxbuf, start, end, scratchpad);
kernel_reduce<dsum><<<bpg_reduce, tpb_reduce, sizeof(AcReal) * tpb_reduce, stream>>>(scratchpad, num_elems);
kernel_reduce_block<dsum><<<1, 1, 0, stream>>>(scratchpad, bpg_reduce, tpb_reduce, reduce_result);
} else {
ERROR("Unrecognized rtype");
}
@@ -944,6 +948,10 @@ reduce_vec(const cudaStream_t stream, const ReductionType rtype, const int3& sta
kernel_filter_vec<dexp_squared_vec><<<bpg_filter, tpb_filter, 0, stream>>>(vtxbuf0, vtxbuf1, vtxbuf2, start, end, scratchpad);
kernel_reduce<dsum><<<bpg_reduce, tpb_reduce, sizeof(AcReal) * tpb_reduce, stream>>>(scratchpad, num_elems);
kernel_reduce_block<dsum><<<1, 1, 0, stream>>>(scratchpad, bpg_reduce, tpb_reduce, reduce_result);
} else if (rtype == RTYPE_SUM) {
kernel_filter_vec<dlength_vec><<<bpg_filter, tpb_filter, 0, stream>>>(vtxbuf0, vtxbuf1, vtxbuf2, start, end, scratchpad);
kernel_reduce<dsum><<<bpg_reduce, tpb_reduce, sizeof(AcReal) * tpb_reduce, stream>>>(scratchpad, num_elems);
kernel_reduce_block<dsum><<<1, 1, 0, stream>>>(scratchpad, bpg_reduce, tpb_reduce, reduce_result);
} else {
ERROR("Unrecognized rtype");
}