portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-13 01:05:36 +00:00
Code Releases Activity

SYCL updates -- operator = giving trouble on Aurora.

Browse Source 
SYCL reduction is failing intermittently with SVM interface - returns
zero, expect non-zero.
Think I need to remove ALL dependence on SVM.
This commit is contained in:
Peter Boyle 2024-09-04 13:53:48 +00:00
parent aa67a5b095
commit 622f78ebea
2 changed files with 47 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
Grid/lattice/Lattice_base.h
View File

@ -236,10 +236,13 @@ public:
template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){ template<class sobj> inline Lattice<vobj> & operator = (const sobj & r){
vobj vtmp; vobj vtmp;
vtmp = r; vtmp = r;
#if defined(GRID_HIP) || defined(GRID_CUDA) || defined (GRID_SYCL) #if 0
deviceVector<vobj> vvtmp(1);
acceleratorPut(vvtmp[0],vtmp);
vobj *vvtmp_p = & vvtmp[0];
auto me = View(AcceleratorWrite); auto me = View(AcceleratorWrite);
accelerator_for(ss,me.size(),vobj::Nsimd(),{ accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto stmp=coalescedRead(vtmp); auto stmp=coalescedRead(*vvtmp_p);
coalescedWrite(me[ss],stmp); coalescedWrite(me[ss],stmp);
}); });
#else #else

46
Grid/lattice/Lattice_reduction_sycl.h
View File

@ -4,16 +4,36 @@ NAMESPACE_BEGIN(Grid);
// Possibly promote to double and sum // Possibly promote to double and sum
///////////////////////////////////////////////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////////////////////////////////////////////////
template <class vobj> template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites) inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites)
{ {
typedef typename vobj::scalar_object sobj; typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_objectD sobjD; typedef typename vobj::scalar_objectD sobjD;
#if 1
sobj identity; zeroit(identity);
sobj ret; zeroit(ret);
Integer nsimd= vobj::Nsimd();
{
sycl::buffer<sobj, 1> abuff(&ret, {1});
theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
auto Reduction = cl::sycl::reduction(abuff,cgh,identity,std::plus<>());
cgh.parallel_for(cl::sycl::range<1>{osites},
Reduction,
[=] (cl::sycl::id<1> item, auto &sum) {
auto osite = item[0];
sum +=Reduce(lat[osite]);
});
});
}
sobjD dret; convertType(dret,ret);
return dret;
#else
static Vector<sobj> mysum; static Vector<sobj> mysum;
mysum.resize(1); mysum.resize(1);
sobj *mysum_p = & mysum[0]; sobj *mysum_p = & mysum[0];
sobj identity; zeroit(identity); sobj identity; zeroit(identity);
mysum[0] = identity; acceleratorPut(mysum[0],identity);
sobj ret ; sobj ret ;
Integer nsimd= vobj::Nsimd(); Integer nsimd= vobj::Nsimd();
@ -33,6 +53,7 @@ inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer os
// free(mysum,*theGridAccelerator); // free(mysum,*theGridAccelerator);
sobjD dret; convertType(dret,ret); sobjD dret; convertType(dret,ret);
return dret; return dret;
#endif
} }
template <class vobj> template <class vobj>
@ -76,12 +97,28 @@ inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osite
template<class Word> Word svm_xor(Word *vec,uint64_t L) template<class Word> Word svm_xor(Word *vec,uint64_t L)
{ {
Word xorResult; xorResult = 0; #if 1
Word identity; identity=0;
Word ret = 0;
{
sycl::buffer<Word, 1> abuff(&ret, {1});
theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
auto Reduction = cl::sycl::reduction(abuff,cgh,identity,std::bit_xor<>());
cgh.parallel_for(cl::sycl::range<1>{L},
Reduction,
[=] (cl::sycl::id<1> index, auto &sum) {
sum ^=vec[index];
});
});
}
theGridAccelerator->wait();
return ret;
#else
static Vector<Word> d_sum; static Vector<Word> d_sum;
d_sum.resize(1); d_sum.resize(1);
Word *d_sum_p=&d_sum[0]; Word *d_sum_p=&d_sum[0];
Word identity; identity=0; Word identity; identity=0;
d_sum[0] = identity; acceleratorPut(d_sum[0],identity);
const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() }); const cl::sycl::property_list PropList ({ cl::sycl::property::reduction::initialize_to_identity() });
theGridAccelerator->submit([&](cl::sycl::handler &cgh) { theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
auto Reduction = cl::sycl::reduction(d_sum_p,identity,std::bit_xor<>(),PropList); auto Reduction = cl::sycl::reduction(d_sum_p,identity,std::bit_xor<>(),PropList);
@ -92,9 +129,10 @@ template<class Word> Word svm_xor(Word *vec,uint64_t L)
}); });
}); });
theGridAccelerator->wait(); theGridAccelerator->wait();
Word ret = d_sum[0]; Word ret = acceleratorGet(d_sum[0]);
// free(d_sum,*theGridAccelerator); // free(d_sum,*theGridAccelerator);
return ret; return ret;
#endif
} }
NAMESPACE_END(Grid); NAMESPACE_END(Grid);