portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-05-29 13:34:17 +01:00
Code Releases Activity

Deprecate UVM

Browse Source
This commit is contained in:
Peter Boyle
2024-09-17 13:34:27 +00:00
parent 11be10d2c0
commit 066544281f
75 changed files with 668 additions and 1082 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Grid/lattice/Lattice_slicesum_core.h
+62 -19
View File
@@ -21,9 +21,18 @@ NAMESPACE_BEGIN(Grid);
#if defined(GRID_CUDA) || defined(GRID_HIP)
template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
template<class vobj>
inline void sliceSumReduction_cub_small(const vobj *Data,
std::vector<vobj> &lvSum,
const int rd,
const int e1,
const int e2,
const int stride,
const int ostride,
const int Nsimd)
{
size_t subvol_size = e1*e2;
commVector<vobj> reduction_buffer(rd*subvol_size);
deviceVector<vobj> reduction_buffer(rd*subvol_size);
auto rb_p = &reduction_buffer[0];
vobj zero_init;
zeroit(zero_init);
@@ -94,7 +103,15 @@ template<class vobj> inline void sliceSumReduction_cub_small(const vobj *Data, V
#if defined(GRID_SYCL)
template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data, Vector <vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
template<class vobj>
inline void sliceSumReduction_sycl_small(const vobj *Data,
std::vector <vobj> &lvSum,
const int &rd,
const int &e1,
const int &e2,
const int &stride,
const int &ostride,
const int &Nsimd)
{
size_t subvol_size = e1*e2;
@@ -105,7 +122,7 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
mysum[r] = vobj_zero;
}
commVector<vobj> reduction_buffer(rd*subvol_size);
deviceVector<vobj> reduction_buffer(rd*subvol_size);
auto rb_p = &reduction_buffer[0];
@@ -144,14 +161,23 @@ template<class vobj> inline void sliceSumReduction_sycl_small(const vobj *Data,
}
#endif
template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd) {
template<class vobj>
inline void sliceSumReduction_large(const vobj *Data,
std::vector<vobj> &lvSum,
const int rd,
const int e1,
const int e2,
const int stride,
const int ostride,
const int Nsimd)
{
typedef typename vobj::vector_type vector;
const int words = sizeof(vobj)/sizeof(vector);
const int osites = rd*e1*e2;
commVector<vector>buffer(osites);
deviceVector<vector>buffer(osites);
vector *dat = (vector *)Data;
vector *buf = &buffer[0];
Vector<vector> lvSum_small(rd);
std::vector<vector> lvSum_small(rd);
vector *lvSum_ptr = (vector *)&lvSum[0];
for (int w = 0; w < words; w++) {
@@ -168,13 +194,18 @@ template<class vobj> inline void sliceSumReduction_large(const vobj *Data, Vecto
for (int r = 0; r < rd; r++) {
lvSum_ptr[w+words*r]=lvSum_small[r];
}
}
}
template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
template<class vobj>
inline void sliceSumReduction_gpu(const Lattice<vobj> &Data,
std::vector<vobj> &lvSum,
const int rd,
const int e1,
const int e2,
const int stride,
const int ostride,
const int Nsimd)
{
autoView(Data_v, Data, AcceleratorRead); //reduction libraries cannot deal with large vobjs so we split into small/large case.
if constexpr (sizeof(vobj) <= 256) {
@@ -192,7 +223,15 @@ template<class vobj> inline void sliceSumReduction_gpu(const Lattice<vobj> &Data
}
template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
template<class vobj>
inline void sliceSumReduction_cpu(const Lattice<vobj> &Data,
std::vector<vobj> &lvSum,
const int &rd,
const int &e1,
const int &e2,
const int &stride,
const int &ostride,
const int &Nsimd)
{
// sum over reduced dimension planes, breaking out orthog dir
// Parallel over orthog direction
@@ -208,16 +247,20 @@ template<class vobj> inline void sliceSumReduction_cpu(const Lattice<vobj> &Data
});
}
template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int &rd, const int &e1, const int &e2, const int &stride, const int &ostride, const int &Nsimd)
template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data,
std::vector<vobj> &lvSum,
const int &rd,
const int &e1,
const int &e2,
const int &stride,
const int &ostride,
const int &Nsimd)
{
#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
#if defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)
sliceSumReduction_gpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
#else
#else
sliceSumReduction_cpu(Data, lvSum, rd, e1, e2, stride, ostride, Nsimd);
#endif
#endif
}