tx258/lammps-gran-kokkos
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Override any OpenCL fast math JIT settings for born/coul/wolf{/cs}/gpu to resolve numerical deviations seen with some OpenCL implementations.

Browse Source
This commit is contained in:
Mike Brown
2021-08-28 17:01:58 -07:00
parent 4d1a6cbd73
commit 89556f0bcb
6 changed files with 29 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
lib/gpu/lal_device.cpp
View File

@ -420,6 +420,16 @@ int DeviceT::set_ocl_params(std::string s_config, std::string extra_args) {
return 0;
}
template <class numtyp, class acctyp>
std::string DeviceT::compile_string_nofast() {
std::string no_fast = _ocl_compile_string;
size_t p = no_fast.find("-cl-fast-relaxed-math ");
if (p != std::string::npos) no_fast.erase(p,22);
p = no_fast.find("-DFAST_MATH=");
if (p != std::string::npos) no_fast[p + 12]='0';
return no_fast;
}
template <class numtyp, class acctyp>
int DeviceT::init(Answer<numtyp,acctyp> &ans, const bool charge,
const bool rot, const int nlocal,