ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

GPU Package: Use __shfl_xor_sync starting with CUDA 9

Browse Source
This commit is contained in:
Richard Berger
2018-08-09 02:33:35 -04:00
parent 985318203d
commit 6d60075d3b
1 changed files with 21 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
lib/gpu/lal_preprocessor.h
View File

@ -119,6 +119,8 @@
#define BLOCK_ELLIPSE 128
#define MAX_SHARED_TYPES 11
#if (__CUDACC_VER_MAJOR__ < 9)
#ifdef _SINGLE_SINGLE
#define shfl_xor __shfl_xor
#else
@ -132,6 +134,25 @@ ucl_inline double shfl_xor(double var, int laneMask, int width) {
}
#endif
#else
#ifdef _SINGLE_SINGLE
ucl_inline double shfl_xor(double var, int laneMask, int width) {
return __shfl_xor_sync(0xffffffff, var, laneMask, width);
}
#else
ucl_inline double shfl_xor(double var, int laneMask, int width) {
int2 tmp;
tmp.x = __double2hiint(var);
tmp.y = __double2loint(var);
tmp.x = __shfl_xor_sync(0xffffffff,tmp.x,laneMask,width);
tmp.y = __shfl_xor_sync(0xffffffff,tmp.y,laneMask,width);
return __hiloint2double(tmp.x,tmp.y);
}
#endif
#endif
#endif
#endif