ICODE-ADC/lammps
4
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

7
lib/gpu/lal_base_charge.h
View File

@ -44,6 +44,7 @@ class BaseCharge {
* \param cell_size cutoff + skin
* \param gpu_split fraction of particles handled by device
* \param k_name name for the kernel for force calculation
* \param disable_fast_math override any fast math opts for kernel JIT
*
* Returns:
* - 0 if successful
@ -54,7 +55,8 @@ class BaseCharge {
int init_atomic(const int nlocal, const int nall, const int max_nbors,
const int maxspecial, const double cell_size,
const double gpu_split, FILE *screen,
const void *pair_program, const char *k_name);
const void *pair_program, const char *k_name,
const int disable_fast_math = 0);
/// Estimate the overhead for GPU context changes and CPU driver
void estimate_gpu_overhead(const int add_kernels=0);
@ -198,7 +200,8 @@ class BaseCharge {
double _gpu_overhead, _driver_overhead;
UCL_D_Vec<int> *_nbor_data;
void compile_kernels(UCL_Device &dev, const void *pair_string, const char *k);
void compile_kernels(UCL_Device &dev, const void *pair_string,
const char *k, const int disable_fast_math);
virtual int loop(const int eflag, const int vflag) = 0;
};