portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-10-31 03:54:33 +00:00
Code Releases Activity

SYCL reduction

Browse Source
This commit is contained in:
Peter Boyle
2022-11-08 12:45:25 -08:00
parent 477ebf24f4
commit 82e959f66c
1 changed files with 12 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
Grid/lattice/Lattice_reduction.h
View File

@@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
#if defined(GRID_CUDA)||defined(GRID_HIP) #if defined(GRID_CUDA)||defined(GRID_HIP)
#include <Grid/lattice/Lattice_reduction_gpu.h> #include <Grid/lattice/Lattice_reduction_gpu.h>
#endif #endif
#if defined(GRID_SYCL)
#include <Grid/lattice/Lattice_reduction_sycl.h>
#endif
NAMESPACE_BEGIN(Grid); NAMESPACE_BEGIN(Grid);
@@ -127,7 +130,7 @@ inline Double max(const Double *arg, Integer osites)
template<class vobj> template<class vobj>
inline typename vobj::scalar_object sum(const vobj *arg, Integer osites) inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
{ {
#if defined(GRID_CUDA)||defined(GRID_HIP) #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
return sum_gpu(arg,osites); return sum_gpu(arg,osites);
#else #else
return sum_cpu(arg,osites); return sum_cpu(arg,osites);
@@ -136,7 +139,7 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
template<class vobj> template<class vobj>
inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites) inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
{ {
#if defined(GRID_CUDA)||defined(GRID_HIP) #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
return sumD_gpu(arg,osites); return sumD_gpu(arg,osites);
#else #else
return sumD_cpu(arg,osites); return sumD_cpu(arg,osites);
@@ -145,7 +148,7 @@ inline typename vobj::scalar_objectD sumD(const vobj *arg, Integer osites)
template<class vobj> template<class vobj>
inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites) inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
{ {
#if defined(GRID_CUDA)||defined(GRID_HIP) #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
return sumD_gpu_large(arg,osites); return sumD_gpu_large(arg,osites);
#else #else
return sumD_cpu(arg,osites); return sumD_cpu(arg,osites);
@@ -155,13 +158,13 @@ inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
template<class vobj> template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg) inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{ {
#if defined(GRID_CUDA)||defined(GRID_HIP)
autoView( arg_v, arg, AcceleratorRead);
Integer osites = arg.Grid()->oSites(); Integer osites = arg.Grid()->oSites();
auto ssum= sum_gpu(&arg_v[0],osites); #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
typename vobj::scalar_object ssum;
autoView( arg_v, arg, AcceleratorRead);
ssum= sum_gpu(&arg_v[0],osites);
#else #else
autoView(arg_v, arg, CpuRead); autoView(arg_v, arg, CpuRead);
Integer osites = arg.Grid()->oSites();
auto ssum= sum_cpu(&arg_v[0],osites); auto ssum= sum_cpu(&arg_v[0],osites);
#endif #endif
arg.Grid()->GlobalSum(ssum); arg.Grid()->GlobalSum(ssum);
@@ -171,7 +174,7 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
template<class vobj> template<class vobj>
inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg) inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
{ {
#if defined(GRID_CUDA)||defined(GRID_HIP) #if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
autoView( arg_v, arg, AcceleratorRead); autoView( arg_v, arg, AcceleratorRead);
Integer osites = arg.Grid()->oSites(); Integer osites = arg.Grid()->oSites();
auto ssum= sum_gpu_large(&arg_v[0],osites); auto ssum= sum_gpu_large(&arg_v[0],osites);
@@ -235,11 +238,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
typedef decltype(innerProductD(vobj(),vobj())) inner_t; typedef decltype(innerProductD(vobj(),vobj())) inner_t;
Vector<inner_t> inner_tmp(sites); Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0]; auto inner_tmp_v = &inner_tmp[0];
{ {
autoView( left_v , left, AcceleratorRead); autoView( left_v , left, AcceleratorRead);
autoView( right_v,right, AcceleratorRead); autoView( right_v,right, AcceleratorRead);
// This code could read coalesce
// GPU - SIMT lane compliance... // GPU - SIMT lane compliance...
accelerator_for( ss, sites, 1,{ accelerator_for( ss, sites, 1,{
auto x_l = left_v[ss]; auto x_l = left_v[ss];