portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-18 15:57:05 +01:00
Code Releases Activity

merged upstream

Browse Source
This commit is contained in:
Christoph Lehner
2022-12-31 11:16:21 +02:00
parent 88015b0858 4ca1bf7cca
commit 458c943987
93 changed files with 17025 additions and 11646 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
Grid/DisableWarnings.h
View File

@ -44,14 +44,22 @@ directory
#ifdef __NVCC__
//disables nvcc specific warning in json.hpp
#pragma clang diagnostic ignored "-Wdeprecated-register"
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
//disables nvcc specific warning in json.hpp
#pragma nv_diag_suppress unsigned_compare_with_zero
#pragma nv_diag_suppress cast_to_qualified_type
//disables nvcc specific warning in many files
#pragma nv_diag_suppress esa_on_defaulted_function_ignored
#pragma nv_diag_suppress extra_semicolon
#else
//disables nvcc specific warning in json.hpp
#pragma diag_suppress unsigned_compare_with_zero
#pragma diag_suppress cast_to_qualified_type
//disables nvcc specific warning in many files
#pragma diag_suppress esa_on_defaulted_function_ignored
#pragma diag_suppress extra_semicolon
//Eigen only
#endif
#endif
// Disable vectorisation in Eigen on the Power8/9 and PowerPC

1
Grid/GridStd.h
View File

@ -16,6 +16,7 @@
#include <functional>
#include <stdio.h>
#include <stdlib.h>
#include <strings.h>
#include <stdio.h>
#include <signal.h>
#include <ctime>

4
Grid/Grid_Eigen_Dense.h
View File

@ -14,7 +14,11 @@
/* NVCC save and restore compile environment*/
#ifdef __NVCC__
#pragma push
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
#pragma nv_diag_suppress code_is_unreachable
#else
#pragma diag_suppress code_is_unreachable
#endif
#pragma push_macro("__CUDA_ARCH__")
#pragma push_macro("__NVCC__")
#pragma push_macro("__CUDACC__")

2
Grid/algorithms/CoarsenedMatrix.h
View File

@ -262,7 +262,7 @@ public:
autoView( Tnp_v , (*Tnp), AcceleratorWrite);
autoView( Tnm_v , (*Tnm), AcceleratorWrite);
const int Nsimd = CComplex::Nsimd();
accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
accelerator_for(ss, FineGrid->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});

2
Grid/algorithms/approx/Chebyshev.h
View File

@ -264,7 +264,7 @@ public:
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
constexpr int Nsimd = vector_type::Nsimd();
accelerator_forNB(ss, in.Grid()->oSites(), Nsimd, {
accelerator_for(ss, in.Grid()->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
coalescedWrite(Tnp_v[ss],2.0*y_v(ss)-Tnm_v(ss));
});

38
Grid/algorithms/iterative/Deflation.h
View File

@ -113,7 +113,43 @@ public:
blockPromote(guess_coarse,guess,subspace);
guess.Checkerboard() = src.Checkerboard();
};
};
void operator()(const std::vector<FineField> &src,std::vector<FineField> &guess) {
int Nevec = (int)evec_coarse.size();
int Nsrc = (int)src.size();
// make temp variables
std::vector<CoarseField> src_coarse(Nsrc,evec_coarse[0].Grid());
std::vector<CoarseField> guess_coarse(Nsrc,evec_coarse[0].Grid());
//Preporcessing
std::cout << GridLogMessage << "Start BlockProject for loop" << std::endl;
for (int j=0;j<Nsrc;j++)
{
guess_coarse[j] = Zero();
std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
blockProject(src_coarse[j],src[j],subspace);
}
//deflation set up for eigen vector batchsize 1 and source batch size equal number of sources
std::cout << GridLogMessage << "Start ProjectAccum for loop" << std::endl;
for (int i=0;i<Nevec;i++)
{
std::cout << GridLogMessage << "ProjectAccum Nvec: " << i << std::endl;
const CoarseField & tmp = evec_coarse[i];
for (int j=0;j<Nsrc;j++)
{
axpy(guess_coarse[j],TensorRemove(innerProduct(tmp,src_coarse[j])) / eval_coarse[i],tmp,guess_coarse[j]);
}
}
//postprocessing
std::cout << GridLogMessage << "Start BlockPromote for loop" << std::endl;
for (int j=0;j<Nsrc;j++)
{
std::cout << GridLogMessage << "BlockProject iter: " << j << std::endl;
blockPromote(guess_coarse[j],guess[j],subspace);
guess[j].Checkerboard() = src[j].Checkerboard();
}
};
};

2
Grid/allocator/MemoryManager.cc
View File

@ -40,7 +40,7 @@ void MemoryManager::PrintBytes(void)
//////////////////////////////////////////////////////////////////////
MemoryManager::AllocationCacheEntry MemoryManager::Entries[MemoryManager::NallocType][MemoryManager::NallocCacheMax];
int MemoryManager::Victim[MemoryManager::NallocType];
int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 2, 8, 2, 8 };
int MemoryManager::Ncache[MemoryManager::NallocType] = { 2, 8, 8, 16, 8, 16 };
uint64_t MemoryManager::CacheBytes[MemoryManager::NallocType];
//////////////////////////////////////////////////////////////////////
// Actual allocation and deallocation utils

11
Grid/allocator/MemoryManager.h
View File

@ -36,6 +36,11 @@ NAMESPACE_BEGIN(Grid);
#define GRID_ALLOC_SMALL_LIMIT (4096)
#define STRINGIFY(x) #x
#define TOSTRING(x) STRINGIFY(x)
#define FILE_LINE __FILE__ ":" TOSTRING(__LINE__)
#define AUDIT(a) MemoryManager::Audit(FILE_LINE)
/*Pinning pages is costly*/
////////////////////////////////////////////////////////////////////////////
// Advise the LatticeAccelerator class
@ -92,8 +97,9 @@ private:
static void *Insert(void *ptr,size_t bytes,AllocationCacheEntry *entries,int ncache,int &victim,uint64_t &cbytes) ;
static void *Lookup(size_t bytes,AllocationCacheEntry *entries,int ncache,uint64_t &cbytes) ;
static void PrintBytes(void);
public:
static void PrintBytes(void);
static void Audit(std::string s);
static void Init(void);
static void InitMessage(void);
static void *AcceleratorAllocate(size_t bytes);
@ -113,6 +119,8 @@ private:
static uint64_t DeviceToHostBytes;
static uint64_t HostToDeviceXfer;
static uint64_t DeviceToHostXfer;
static uint64_t DeviceEvictions;
static uint64_t DeviceDestroy;
private:
#ifndef GRID_UVM
@ -170,6 +178,7 @@ private:
public:
static void Print(void);
static void PrintAll(void);
static void PrintState( void* CpuPtr);
static int isOpen (void* CpuPtr);
static void ViewClose(void* CpuPtr,ViewMode mode);

184
Grid/allocator/MemoryManagerCache.cc
View File

@ -3,8 +3,13 @@
#warning "Using explicit device memory copies"
NAMESPACE_BEGIN(Grid);
//#define dprintf(...) printf ( __VA_ARGS__ ); fflush(stdout);
#define dprintf(...)
#define MAXLINE 512
static char print_buffer [ MAXLINE ];
#define mprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
#define dprintf(...) snprintf (print_buffer,MAXLINE, __VA_ARGS__ ); std::cout << GridLogMemory << print_buffer;
//#define dprintf(...)
////////////////////////////////////////////////////////////
@ -23,6 +28,8 @@ uint64_t MemoryManager::HostToDeviceBytes;
uint64_t MemoryManager::DeviceToHostBytes;
uint64_t MemoryManager::HostToDeviceXfer;
uint64_t MemoryManager::DeviceToHostXfer;
uint64_t MemoryManager::DeviceEvictions;
uint64_t MemoryManager::DeviceDestroy;
////////////////////////////////////
// Priority ordering for unlocked entries
@ -104,15 +111,17 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
///////////////////////////////////////////////////////////
assert(AccCache.state!=Empty);
dprintf("MemoryManager: Discard(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);
mprintf("MemoryManager: Discard(%lx) %lx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);
assert(AccCache.accLock==0);
assert(AccCache.cpuLock==0);
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.AccPtr) {
AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
DeviceDestroy++;
DeviceBytes -=AccCache.bytes;
LRUremove(AccCache);
dprintf("MemoryManager: Free(%llx) LRU %lld Total %lld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);
AccCache.AccPtr=(uint64_t) NULL;
dprintf("MemoryManager: Free(%lx) LRU %ld Total %ld\n",(uint64_t)AccCache.AccPtr,DeviceLRUBytes,DeviceBytes);
}
uint64_t CpuPtr = AccCache.CpuPtr;
EntryErase(CpuPtr);
@ -121,26 +130,36 @@ void MemoryManager::AccDiscard(AcceleratorViewEntry &AccCache)
void MemoryManager::Evict(AcceleratorViewEntry &AccCache)
{
///////////////////////////////////////////////////////////////////////////
// Make CPU consistent, remove from Accelerator, remove entry
// Cannot be locked. If allocated must be in LRU pool.
// Make CPU consistent, remove from Accelerator, remove from LRU, LEAVE CPU only entry
// Cannot be acclocked. If allocated must be in LRU pool.
//
// Nov 2022... Felix issue: Allocating two CpuPtrs, can have an entry in LRU-q with CPUlock.
// and require to evict the AccPtr copy. Eviction was a mistake in CpuViewOpen
// but there is a weakness where CpuLock entries are attempted for erase
// Take these OUT LRU queue when CPU locked?
// Cannot take out the table as cpuLock data is important.
///////////////////////////////////////////////////////////////////////////
assert(AccCache.state!=Empty);
dprintf("MemoryManager: Evict(%llx) %llx\n",(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr);
assert(AccCache.accLock==0);
assert(AccCache.cpuLock==0);
mprintf("MemoryManager: Evict cpu %lx acc %lx cpuLock %ld accLock %ld\n",
(uint64_t)AccCache.CpuPtr,(uint64_t)AccCache.AccPtr,
(uint64_t)AccCache.cpuLock,(uint64_t)AccCache.accLock);
assert(AccCache.accLock==0); // Cannot evict so logic bomb
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.state==AccDirty) {
Flush(AccCache);
}
assert(AccCache.CpuPtr!=(uint64_t)NULL);
if(AccCache.AccPtr) {
AcceleratorFree((void *)AccCache.AccPtr,AccCache.bytes);
DeviceBytes -=AccCache.bytes;
LRUremove(AccCache);
dprintf("MemoryManager: Free(%llx) footprint now %lld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);
AccCache.AccPtr=(uint64_t)NULL;
AccCache.state=CpuDirty; // CPU primary now
DeviceBytes -=AccCache.bytes;
dprintf("MemoryManager: Free(%lx) footprint now %ld \n",(uint64_t)AccCache.AccPtr,DeviceBytes);
}
uint64_t CpuPtr = AccCache.CpuPtr;
EntryErase(CpuPtr);
// uint64_t CpuPtr = AccCache.CpuPtr;
DeviceEvictions++;
// EntryErase(CpuPtr);
}
void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
{
@ -150,7 +169,7 @@ void MemoryManager::Flush(AcceleratorViewEntry &AccCache)
assert(AccCache.AccPtr!=(uint64_t)NULL);
assert(AccCache.CpuPtr!=(uint64_t)NULL);
acceleratorCopyFromDevice((void *)AccCache.AccPtr,(void *)AccCache.CpuPtr,AccCache.bytes);
dprintf("MemoryManager: Flush %llx -> %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
mprintf("MemoryManager: Flush %lx -> %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
DeviceToHostBytes+=AccCache.bytes;
DeviceToHostXfer++;
AccCache.state=Consistent;
@ -165,7 +184,7 @@ void MemoryManager::Clone(AcceleratorViewEntry &AccCache)
AccCache.AccPtr=(uint64_t)AcceleratorAllocate(AccCache.bytes);
DeviceBytes+=AccCache.bytes;
}
dprintf("MemoryManager: Clone %llx <- %llx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
mprintf("MemoryManager: Clone %lx <- %lx\n",(uint64_t)AccCache.AccPtr,(uint64_t)AccCache.CpuPtr); fflush(stdout);
acceleratorCopyToDevice((void *)AccCache.CpuPtr,(void *)AccCache.AccPtr,AccCache.bytes);
HostToDeviceBytes+=AccCache.bytes;
HostToDeviceXfer++;
@ -191,6 +210,7 @@ void MemoryManager::CpuDiscard(AcceleratorViewEntry &AccCache)
void MemoryManager::ViewClose(void* Ptr,ViewMode mode)
{
if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
dprintf("AcceleratorViewClose %lx\n",(uint64_t)Ptr);
AcceleratorViewClose((uint64_t)Ptr);
} else if( (mode==CpuRead)||(mode==CpuWrite)){
CpuViewClose((uint64_t)Ptr);
@ -202,6 +222,7 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
{
uint64_t CpuPtr = (uint64_t)_CpuPtr;
if( (mode==AcceleratorRead)||(mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard) ){
dprintf("AcceleratorViewOpen %lx\n",(uint64_t)CpuPtr);
return (void *) AcceleratorViewOpen(CpuPtr,bytes,mode,hint);
} else if( (mode==CpuRead)||(mode==CpuWrite)){
return (void *)CpuViewOpen(CpuPtr,bytes,mode,hint);
@ -212,13 +233,16 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
}
void MemoryManager::EvictVictims(uint64_t bytes)
{
assert(bytes<DeviceMaxBytes);
while(bytes+DeviceLRUBytes > DeviceMaxBytes){
if ( DeviceLRUBytes > 0){
assert(LRU.size()>0);
uint64_t victim = LRU.back();
uint64_t victim = LRU.back(); // From the LRU
auto AccCacheIterator = EntryLookup(victim);
auto & AccCache = AccCacheIterator->second;
Evict(AccCache);
} else {
return;
}
}
}
@ -241,11 +265,12 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
assert(AccCache.cpuLock==0); // Programming error
if(AccCache.state!=Empty) {
dprintf("ViewOpen found entry %llx %llx : %lld %lld\n",
dprintf("ViewOpen found entry %lx %lx : %ld %ld accLock %ld\n",
(uint64_t)AccCache.CpuPtr,
(uint64_t)CpuPtr,
(uint64_t)AccCache.bytes,
(uint64_t)bytes);
(uint64_t)bytes,
(uint64_t)AccCache.accLock);
assert(AccCache.CpuPtr == CpuPtr);
assert(AccCache.bytes ==bytes);
}
@ -280,6 +305,7 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
AccCache.state = Consistent; // Empty + AccRead => Consistent
}
AccCache.accLock= 1;
dprintf("Copied Empty entry into device accLock= %d\n",AccCache.accLock);
} else if(AccCache.state==CpuDirty ){
if(mode==AcceleratorWriteDiscard) {
CpuDiscard(AccCache);
@ -292,28 +318,30 @@ uint64_t MemoryManager::AcceleratorViewOpen(uint64_t CpuPtr,size_t bytes,ViewMod
AccCache.state = Consistent; // CpuDirty + AccRead => Consistent
}
AccCache.accLock++;
dprintf("Copied CpuDirty entry into device accLock %d\n",AccCache.accLock);
dprintf("CpuDirty entry into device ++accLock= %d\n",AccCache.accLock);
} else if(AccCache.state==Consistent) {
if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
AccCache.state = AccDirty; // Consistent + AcceleratorWrite=> AccDirty
else
AccCache.state = Consistent; // Consistent + AccRead => Consistent
AccCache.accLock++;
dprintf("Consistent entry into device accLock %d\n",AccCache.accLock);
dprintf("Consistent entry into device ++accLock= %d\n",AccCache.accLock);
} else if(AccCache.state==AccDirty) {
if((mode==AcceleratorWrite)||(mode==AcceleratorWriteDiscard))
AccCache.state = AccDirty; // AccDirty + AcceleratorWrite=> AccDirty
else
AccCache.state = AccDirty; // AccDirty + AccRead => AccDirty
AccCache.accLock++;
dprintf("AccDirty entry into device accLock %d\n",AccCache.accLock);
dprintf("AccDirty entry ++accLock= %d\n",AccCache.accLock);
} else {
assert(0);
}
// If view is opened on device remove from LRU
assert(AccCache.accLock>0);
// If view is opened on device must remove from LRU
if(AccCache.LRU_valid==1){
// must possibly remove from LRU as now locked on GPU
dprintf("AccCache entry removed from LRU \n");
LRUremove(AccCache);
}
@ -334,10 +362,12 @@ void MemoryManager::AcceleratorViewClose(uint64_t CpuPtr)
assert(AccCache.accLock>0);
AccCache.accLock--;
// Move to LRU queue if not locked and close on device
if(AccCache.accLock==0) {
dprintf("AccleratorViewClose %lx AccLock decremented to %ld move to LRU queue\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
LRUinsert(AccCache);
} else {
dprintf("AccleratorViewClose %lx AccLock decremented to %ld\n",(uint64_t)CpuPtr,(uint64_t)AccCache.accLock);
}
}
void MemoryManager::CpuViewClose(uint64_t CpuPtr)
@ -374,9 +404,10 @@ uint64_t MemoryManager::CpuViewOpen(uint64_t CpuPtr,size_t bytes,ViewMode mode,V
auto AccCacheIterator = EntryLookup(CpuPtr);
auto & AccCache = AccCacheIterator->second;
if (!AccCache.AccPtr) {
EvictVictims(bytes);
}
// CPU doesn't need to free space
// if (!AccCache.AccPtr) {
// EvictVictims(bytes);
// }
assert((mode==CpuRead)||(mode==CpuWrite));
assert(AccCache.accLock==0); // Programming error
@ -430,20 +461,28 @@ void MemoryManager::NotifyDeletion(void *_ptr)
void MemoryManager::Print(void)
{
PrintBytes();
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << "Memory Manager " << std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << DeviceBytes << " bytes allocated on device " << std::endl;
std::cout << GridLogDebug << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
std::cout << GridLogDebug << DeviceMaxBytes<< " bytes max on device " << std::endl;
std::cout << GridLogDebug << HostToDeviceXfer << " transfers to device " << std::endl;
std::cout << GridLogDebug << DeviceToHostXfer << " transfers from device " << std::endl;
std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to device " << std::endl;
std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
std::cout << GridLogMessage << "Memory Manager " << std::endl;
std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
std::cout << GridLogMessage << DeviceBytes << " bytes allocated on device " << std::endl;
std::cout << GridLogMessage << DeviceLRUBytes<< " bytes evictable on device " << std::endl;
std::cout << GridLogMessage << DeviceMaxBytes<< " bytes max on device " << std::endl;
std::cout << GridLogMessage << HostToDeviceXfer << " transfers to device " << std::endl;
std::cout << GridLogMessage << DeviceToHostXfer << " transfers from device " << std::endl;
std::cout << GridLogMessage << HostToDeviceBytes<< " bytes transfered to device " << std::endl;
std::cout << GridLogMessage << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
std::cout << GridLogMessage << DeviceEvictions << " Evictions from device " << std::endl;
std::cout << GridLogMessage << DeviceDestroy << " Destroyed vectors on device " << std::endl;
std::cout << GridLogMessage << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
}
void MemoryManager::PrintAll(void)
{
Print();
std::cout << GridLogMessage << std::endl;
std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
auto &AccCache = it->second;
@ -453,13 +492,13 @@ void MemoryManager::Print(void)
if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
if ( AccCache.state==Consistent)str = std::string("Consistent");
std::cout << GridLogDebug << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
<< "\t" << AccCache.cpuLock
<< "\t" << AccCache.accLock
<< "\t" << AccCache.LRU_valid<<std::endl;
}
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogMessage << "--------------------------------------------" << std::endl;
};
int MemoryManager::isOpen (void* _CpuPtr)
@ -473,6 +512,61 @@ int MemoryManager::isOpen (void* _CpuPtr)
return 0;
}
}
void MemoryManager::Audit(std::string s)
{
uint64_t CpuBytes=0;
uint64_t AccBytes=0;
uint64_t LruBytes1=0;
uint64_t LruBytes2=0;
uint64_t LruCnt=0;
uint64_t LockedBytes=0;
std::cout << " Memory Manager::Audit() from "<<s<<std::endl;
for(auto it=LRU.begin();it!=LRU.end();it++){
uint64_t cpuPtr = *it;
assert(EntryPresent(cpuPtr));
auto AccCacheIterator = EntryLookup(cpuPtr);
auto & AccCache = AccCacheIterator->second;
LruBytes2+=AccCache.bytes;
assert(AccCache.LRU_valid==1);
assert(AccCache.LRU_entry==it);
}
std::cout << " Memory Manager::Audit() LRU queue matches table entries "<<std::endl;
for(auto it=AccViewTable.begin();it!=AccViewTable.end();it++){
auto &AccCache = it->second;
std::string str;
if ( AccCache.state==Empty ) str = std::string("Empty");
if ( AccCache.state==CpuDirty ) str = std::string("CpuDirty");
if ( AccCache.state==AccDirty ) str = std::string("AccDirty");
if ( AccCache.state==Consistent)str = std::string("Consistent");
CpuBytes+=AccCache.bytes;
if( AccCache.AccPtr ) AccBytes+=AccCache.bytes;
if( AccCache.LRU_valid ) LruBytes1+=AccCache.bytes;
if( AccCache.LRU_valid ) LruCnt++;
if ( AccCache.cpuLock || AccCache.accLock ) {
assert(AccCache.LRU_valid==0);
std::cout << GridLogError << s<< "\n\t 0x"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
<< "\t cpuLock " << AccCache.cpuLock
<< "\t accLock " << AccCache.accLock
<< "\t LRUvalid " << AccCache.LRU_valid<<std::endl;
}
assert( AccCache.cpuLock== 0 ) ;
assert( AccCache.accLock== 0 ) ;
}
std::cout << " Memory Manager::Audit() no locked table entries "<<std::endl;
assert(LruBytes1==LruBytes2);
assert(LruBytes1==DeviceLRUBytes);
std::cout << " Memory Manager::Audit() evictable bytes matches sum over table "<<std::endl;
assert(AccBytes==DeviceBytes);
std::cout << " Memory Manager::Audit() device bytes matches sum over table "<<std::endl;
assert(LruCnt == LRU.size());
std::cout << " Memory Manager::Audit() LRU entry count matches "<<std::endl;
}
void MemoryManager::PrintState(void* _CpuPtr)
{
@ -489,8 +583,8 @@ void MemoryManager::PrintState(void* _CpuPtr)
if ( AccCache.state==EvictNext) str = std::string("EvictNext");
std::cout << GridLogMessage << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
std::cout << GridLogMessage << "0x"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\t0x"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
std::cout << GridLogMessage << "\tx"<<std::hex<<AccCache.CpuPtr<<std::dec
<< "\tx"<<std::hex<<AccCache.AccPtr<<std::dec<<"\t" <<str
<< "\t" << AccCache.cpuLock
<< "\t" << AccCache.accLock
<< "\t" << AccCache.LRU_valid<<std::endl;

4
Grid/allocator/MemoryManagerShared.cc
View File

@ -12,7 +12,10 @@ uint64_t MemoryManager::HostToDeviceBytes;
uint64_t MemoryManager::DeviceToHostBytes;
uint64_t MemoryManager::HostToDeviceXfer;
uint64_t MemoryManager::DeviceToHostXfer;
uint64_t MemoryManager::DeviceEvictions;
uint64_t MemoryManager::DeviceDestroy;
void MemoryManager::Audit(std::string s){};
void MemoryManager::ViewClose(void* AccPtr,ViewMode mode){};
void *MemoryManager::ViewOpen(void* CpuPtr,size_t bytes,ViewMode mode,ViewAdvise hint){ return CpuPtr; };
int MemoryManager::isOpen (void* CpuPtr) { return 0;}
@ -21,6 +24,7 @@ void MemoryManager::PrintState(void* CpuPtr)
std::cout << GridLogMessage << "Host<->Device memory movement not currently managed by Grid." << std::endl;
};
void MemoryManager::Print(void){};
void MemoryManager::PrintAll(void){};
void MemoryManager::NotifyDeletion(void *ptr){};
NAMESPACE_END(Grid);

6
Grid/communicator/Communicator_mpi3.cc
View File

@ -392,9 +392,9 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,bytes);
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
this->StencilSendToRecvFromComplete(list,dir);
}
// if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
// this->StencilSendToRecvFromComplete(list,dir);
// }
return off_node_bytes;
}

24165
Grid/json/json.hpp
View File

File diff suppressed because it is too large Load Diff

1
Grid/lattice/Lattice.h
View File

@ -46,3 +46,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/lattice/Lattice_unary.h>
#include <Grid/lattice/Lattice_transfer.h>
#include <Grid/lattice/Lattice_basis.h>
#include <Grid/lattice/Lattice_crc.h>

12
Grid/lattice/Lattice_base.h
View File

@ -129,7 +129,7 @@ public:
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWriteDiscard);
auto me = View(AcceleratorWrite);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto tmp = eval(ss,exprCopy);
coalescedWrite(me[ss],tmp);
@ -152,7 +152,7 @@ public:
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWriteDiscard);
auto me = View(AcceleratorWrite);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto tmp = eval(ss,exprCopy);
coalescedWrite(me[ss],tmp);
@ -174,7 +174,7 @@ public:
this->checkerboard=cb;
auto exprCopy = expr;
ExpressionViewOpen(exprCopy);
auto me = View(AcceleratorWriteDiscard);
auto me = View(AcceleratorWrite);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
auto tmp = eval(ss,exprCopy);
coalescedWrite(me[ss],tmp);
@ -245,7 +245,7 @@ public:
///////////////////////////////////////////
// user defined constructor
///////////////////////////////////////////
Lattice(GridBase *grid,ViewMode mode=AcceleratorWriteDiscard) {
Lattice(GridBase *grid,ViewMode mode=AcceleratorWrite) {
this->_grid = grid;
resize(this->_grid->oSites());
assert((((uint64_t)&this->_odata[0])&0xF) ==0);
@ -288,7 +288,7 @@ public:
typename std::enable_if<!std::is_same<robj,vobj>::value,int>::type i=0;
conformable(*this,r);
this->checkerboard = r.Checkerboard();
auto me = View(AcceleratorWriteDiscard);
auto me = View(AcceleratorWrite);
auto him= r.View(AcceleratorRead);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));
@ -303,7 +303,7 @@ public:
inline Lattice<vobj> & operator = (const Lattice<vobj> & r){
this->checkerboard = r.Checkerboard();
conformable(*this,r);
auto me = View(AcceleratorWriteDiscard);
auto me = View(AcceleratorWrite);
auto him= r.View(AcceleratorRead);
accelerator_for(ss,me.size(),vobj::Nsimd(),{
coalescedWrite(me[ss],him(ss));

55
Grid/lattice/Lattice_crc.h Normal file
View File

@ -0,0 +1,55 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/lattice/Lattice_crc.h
Copyright (C) 2021
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#pragma once
NAMESPACE_BEGIN(Grid);
template<class vobj> void DumpSliceNorm(std::string s,Lattice<vobj> &f,int mu=-1)
{
auto ff = localNorm2(f);
if ( mu==-1 ) mu = f.Grid()->Nd()-1;
typedef typename vobj::tensor_reduced normtype;
typedef typename normtype::scalar_object scalar;
std::vector<scalar> sff;
sliceSum(ff,sff,mu);
for(int t=0;t<sff.size();t++){
std::cout << s<<" "<<t<<" "<<sff[t]<<std::endl;
}
}
template<class vobj> uint32_t crc(Lattice<vobj> & buf)
{
autoView( buf_v , buf, CpuRead);
return ::crc32(0L,(unsigned char *)&buf_v[0],(size_t)sizeof(vobj)*buf.oSites());
}
#define CRC(U) std::cout << "FingerPrint "<<__FILE__ <<" "<< __LINE__ <<" "<< #U <<" "<<crc(U)<<std::endl;
NAMESPACE_END(Grid);

46
Grid/lattice/Lattice_reduction.h
View File

@ -28,6 +28,9 @@ Author: Christoph Lehner <christoph@lhnr.de>
#if defined(GRID_CUDA)||defined(GRID_HIP)
#include <Grid/lattice/Lattice_reduction_gpu.h>
#endif
#if defined(GRID_SYCL)
#include <Grid/lattice/Lattice_reduction_sycl.h>
#endif
NAMESPACE_BEGIN(Grid);
@ -127,7 +130,7 @@ inline Double max(const Double *arg, Integer osites)
template<class vobj>
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);
#else
return sum_cpu(arg,osites);
@ -136,23 +139,31 @@ inline typename vobj::scalar_object sum(const vobj *arg, Integer osites)
template<class vobj>
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);
#else
return sumD_cpu(arg,osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_objectD sumD_large(const vobj *arg, Integer osites)
{
#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
return sumD_gpu_large(arg,osites);
#else
return sumD_cpu(arg,osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_object rankSum(const Lattice<vobj> &arg)
{
#if defined(GRID_CUDA)||defined(GRID_HIP)
autoView( arg_v, arg, AcceleratorRead);
Integer osites = arg.Grid()->oSites();
#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
autoView( arg_v, arg, AcceleratorRead);
return sum_gpu(&arg_v[0],osites);
#else
autoView(arg_v, arg, CpuRead);
Integer osites = arg.Grid()->oSites();
return sum_cpu(&arg_v[0],osites);
#endif
}
@ -165,6 +176,28 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
return ssum;
}
template<class vobj>
inline typename vobj::scalar_object rankSumLarge(const Lattice<vobj> &arg)
{
#if defined(GRID_CUDA)||defined(GRID_HIP)||defined(GRID_SYCL)
autoView( arg_v, arg, AcceleratorRead);
Integer osites = arg.Grid()->oSites();
return sum_gpu_large(&arg_v[0],osites);
#else
autoView(arg_v, arg, CpuRead);
Integer osites = arg.Grid()->oSites();
return sum_cpu(&arg_v[0],osites);
#endif
}
template<class vobj>
inline typename vobj::scalar_object sum_large(const Lattice<vobj> &arg)
{
auto ssum = rankSumLarge(arg);
arg.Grid()->GlobalSum(ssum);
return ssum;
}
////////////////////////////////////////////////////////////////////////////////////////////////////
// Deterministic Reduction operations
////////////////////////////////////////////////////////////////////////////////////////////////////
@ -216,11 +249,10 @@ inline ComplexD rankInnerProduct(const Lattice<vobj> &left,const Lattice<vobj> &
typedef decltype(innerProductD(vobj(),vobj())) inner_t;
Vector<inner_t> inner_tmp(sites);
auto inner_tmp_v = &inner_tmp[0];
{
autoView( left_v , left, AcceleratorRead);
autoView( right_v,right, AcceleratorRead);
// This code could read coalesce
// GPU - SIMT lane compliance...
accelerator_for( ss, sites, 1,{
auto x_l = left_v[ss];

73
Grid/lattice/Lattice_reduction_gpu.h
View File

@ -23,7 +23,7 @@ unsigned int nextPow2(Iterator x) {
}
template <class Iterator>
void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
int getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator &threads, Iterator &blocks) {
int device;
#ifdef GRID_CUDA
@ -37,13 +37,13 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
Iterator sharedMemPerBlock = gpu_props[device].sharedMemPerBlock;
Iterator maxThreadsPerBlock = gpu_props[device].maxThreadsPerBlock;
Iterator multiProcessorCount = gpu_props[device].multiProcessorCount;
/*
std::cout << GridLogDebug << "GPU has:" << std::endl;
std::cout << GridLogDebug << "\twarpSize = " << warpSize << std::endl;
std::cout << GridLogDebug << "\tsharedMemPerBlock = " << sharedMemPerBlock << std::endl;
std::cout << GridLogDebug << "\tmaxThreadsPerBlock = " << maxThreadsPerBlock << std::endl;
std::cout << GridLogDebug << "\tmultiProcessorCount = " << multiProcessorCount << std::endl;
*/
if (warpSize != WARP_SIZE) {
std::cout << GridLogError << "The warp size of the GPU in use does not match the warp size set when compiling Grid." << std::endl;
exit(EXIT_FAILURE);
@ -53,12 +53,12 @@ void getNumBlocksAndThreads(const Iterator n, const size_t sizeofsobj, Iterator
threads = warpSize;
if ( threads*sizeofsobj > sharedMemPerBlock ) {
std::cout << GridLogError << "The object is too large for the shared memory." << std::endl;
exit(EXIT_FAILURE);
return 0;
}
while( 2*threads*sizeofsobj < sharedMemPerBlock && 2*threads <= maxThreadsPerBlock ) threads *= 2;
// keep all the streaming multiprocessors busy
blocks = nextPow2(multiProcessorCount);
return 1;
}
template <class sobj, class Iterator>
@ -198,7 +198,7 @@ __global__ void reduceKernel(const vobj *lat, sobj *buffer, Iterator n) {
// Possibly promote to double and sum
/////////////////////////////////////////////////////////////////////////////////////////////////////////
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_objectD sobj;
typedef decltype(lat) Iterator;
@ -207,7 +207,9 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
Integer size = osites*nsimd;
Integer numThreads, numBlocks;
getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
assert(ok);
Integer smemSize = numThreads * sizeof(sobj);
Vector<sobj> buffer(numBlocks);
@ -218,6 +220,54 @@ inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
auto result = buffer_v[0];
return result;
}
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
{
typedef typename vobj::vector_type vector;
typedef typename vobj::scalar_typeD scalarD;
typedef typename vobj::scalar_objectD sobj;
sobj ret;
scalarD *ret_p = (scalarD *)&ret;
const int words = sizeof(vobj)/sizeof(vector);
Vector<vector> buffer(osites);
vector *dat = (vector *)lat;
vector *buf = &buffer[0];
iScalar<vector> *tbuf =(iScalar<vector> *) &buffer[0];
for(int w=0;w<words;w++) {
accelerator_for(ss,osites,1,{
buf[ss] = dat[ss*words+w];
});
ret_p[w] = sumD_gpu_small(tbuf,osites);
}
return ret;
}
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
{
typedef typename vobj::vector_type vector;
typedef typename vobj::scalar_typeD scalarD;
typedef typename vobj::scalar_objectD sobj;
sobj ret;
Integer nsimd= vobj::Nsimd();
Integer size = osites*nsimd;
Integer numThreads, numBlocks;
int ok = getNumBlocksAndThreads(size, sizeof(sobj), numThreads, numBlocks);
if ( ok ) {
ret = sumD_gpu_small(lat,osites);
} else {
ret = sumD_gpu_large(lat,osites);
}
return ret;
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////
// Return as same precision as input performing reduction in double precision though
/////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -230,6 +280,13 @@ inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
return result;
}
template <class vobj>
inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_object sobj;
sobj result;
result = sumD_gpu_large(lat,osites);
return result;
}
NAMESPACE_END(Grid);

125
Grid/lattice/Lattice_reduction_sycl.h Normal file
View File

@ -0,0 +1,125 @@
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////////////////////////////////////////////
// Possibly promote to double and sum
/////////////////////////////////////////////////////////////////////////////////////////////////////////
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu_tensor(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_objectD sobjD;
sobj *mysum =(sobj *) malloc_shared(sizeof(sobj),*theGridAccelerator);
sobj identity; zeroit(identity);
sobj ret ;
Integer nsimd= vobj::Nsimd();
theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
auto Reduction = cl::sycl::reduction(mysum,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]);
});
});
theGridAccelerator->wait();
ret = mysum[0];
free(mysum,*theGridAccelerator);
sobjD dret; convertType(dret,ret);
return dret;
}
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu_large(const vobj *lat, Integer osites)
{
return sumD_gpu_tensor(lat,osites);
}
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu_small(const vobj *lat, Integer osites)
{
return sumD_gpu_large(lat,osites);
}
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu(const vobj *lat, Integer osites)
{
return sumD_gpu_large(lat,osites);
}
/////////////////////////////////////////////////////////////////////////////////////////////////////////
// Return as same precision as input performing reduction in double precision though
/////////////////////////////////////////////////////////////////////////////////////////////////////////
template <class vobj>
inline typename vobj::scalar_object sum_gpu(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_object sobj;
sobj result;
result = sumD_gpu(lat,osites);
return result;
}
template <class vobj>
inline typename vobj::scalar_object sum_gpu_large(const vobj *lat, Integer osites)
{
typedef typename vobj::scalar_object sobj;
sobj result;
result = sumD_gpu_large(lat,osites);
return result;
}
NAMESPACE_END(Grid);
/*
template<class Double> Double svm_reduce(Double *vec,uint64_t L)
{
Double sumResult; zeroit(sumResult);
Double *d_sum =(Double *)cl::sycl::malloc_shared(sizeof(Double),*theGridAccelerator);
Double identity; zeroit(identity);
theGridAccelerator->submit([&](cl::sycl::handler &cgh) {
auto Reduction = cl::sycl::reduction(d_sum,identity,std::plus<>());
cgh.parallel_for(cl::sycl::range<1>{L},
Reduction,
[=] (cl::sycl::id<1> index, auto &sum) {
sum +=vec[index];
});
});
theGridAccelerator->wait();
Double ret = d_sum[0];
free(d_sum,*theGridAccelerator);
std::cout << " svm_reduce finished "<<L<<" sites sum = " << ret <<std::endl;
return ret;
}
template <class vobj>
inline typename vobj::scalar_objectD sumD_gpu_repack(const vobj *lat, Integer osites)
{
typedef typename vobj::vector_type vector;
typedef typename vobj::scalar_type scalar;
typedef typename vobj::scalar_typeD scalarD;
typedef typename vobj::scalar_objectD sobjD;
sobjD ret;
scalarD *ret_p = (scalarD *)&ret;
const int nsimd = vobj::Nsimd();
const int words = sizeof(vobj)/sizeof(vector);
Vector<scalar> buffer(osites*nsimd);
scalar *buf = &buffer[0];
vector *dat = (vector *)lat;
for(int w=0;w<words;w++) {
accelerator_for(ss,osites,nsimd,{
int lane = acceleratorSIMTlane(nsimd);
buf[ss*nsimd+lane] = dat[ss*words+w].getlane(lane);
});
//Precision change at this point is to late to gain precision
ret_p[w] = svm_reduce(buf,nsimd*osites);
}
return ret;
}
*/

17
Grid/log/Log.cc
View File

@ -65,29 +65,40 @@ GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
GridLogger GridLogError (1, "Error" , GridLogColours, "RED");
GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
GridLogger GridLogMemory (1, "Memory", GridLogColours, "NORMAL");
GridLogger GridLogTracing(1, "Tracing", GridLogColours, "NORMAL");
GridLogger GridLogDebug (1, "Debug", GridLogColours, "PURPLE");
GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
GridLogger GridLogDslash (1, "Dslash", GridLogColours, "BLUE");
GridLogger GridLogIterative (1, "Iterative", GridLogColours, "BLUE");
GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
GridLogger GridLogHMC (1, "HMC", GridLogColours, "BLUE");
void GridLogConfigure(std::vector<std::string> &logstreams) {
GridLogError.Active(0);
GridLogError.Active(1);
GridLogWarning.Active(0);
GridLogMessage.Active(1); // at least the messages should be always on
GridLogMemory.Active(0);
GridLogTracing.Active(0);
GridLogIterative.Active(0);
GridLogDebug.Active(0);
GridLogPerformance.Active(0);
GridLogDslash.Active(0);
GridLogIntegrator.Active(1);
GridLogColours.Active(0);
GridLogHMC.Active(1);
for (int i = 0; i < logstreams.size(); i++) {
if (logstreams[i] == std::string("Error")) GridLogError.Active(1);
if (logstreams[i] == std::string("Tracing")) GridLogTracing.Active(1);
if (logstreams[i] == std::string("Memory")) GridLogMemory.Active(1);
if (logstreams[i] == std::string("Warning")) GridLogWarning.Active(1);
if (logstreams[i] == std::string("NoMessage")) GridLogMessage.Active(0);
if (logstreams[i] == std::string("Iterative")) GridLogIterative.Active(1);
if (logstreams[i] == std::string("Debug")) GridLogDebug.Active(1);
if (logstreams[i] == std::string("Performance")) GridLogPerformance.Active(1);
if (logstreams[i] == std::string("Integrator")) GridLogIntegrator.Active(1);
if (logstreams[i] == std::string("Dslash")) GridLogDslash.Active(1);
if (logstreams[i] == std::string("NoIntegrator"))GridLogIntegrator.Active(0);
if (logstreams[i] == std::string("NoHMC")) GridLogHMC.Active(0);
if (logstreams[i] == std::string("Colours")) GridLogColours.Active(1);
}
}

11
Grid/log/Log.h
View File

@ -138,7 +138,8 @@ public:
stream << std::setw(log.topWidth);
}
stream << log.topName << log.background()<< " : ";
stream << log.colour() << std::left;
// stream << log.colour() << std::left;
stream << std::left;
if (log.chanWidth > 0)
{
stream << std::setw(log.chanWidth);
@ -153,9 +154,9 @@ public:
stream << log.evidence()
<< now << log.background() << " : " ;
}
stream << log.colour();
// stream << log.colour();
stream << std::right;
stream.flags(f);
return stream;
} else {
return devnull;
@ -180,8 +181,12 @@ extern GridLogger GridLogWarning;
extern GridLogger GridLogMessage;
extern GridLogger GridLogDebug ;
extern GridLogger GridLogPerformance;
extern GridLogger GridLogDslash;
extern GridLogger GridLogIterative ;
extern GridLogger GridLogIntegrator ;
extern GridLogger GridLogHMC;
extern GridLogger GridLogMemory;
extern GridLogger GridLogTracing;
extern Colours GridLogColours;
std::string demangle(const char* name) ;

9
Grid/parallelIO/IldgIO.h
View File

@ -31,6 +31,7 @@ directory
#include <fstream>
#include <iomanip>
#include <iostream>
#include <string>
#include <map>
#include <pwd.h>
@ -654,7 +655,8 @@ class IldgWriter : public ScidacWriter {
// Fill ILDG header data struct
//////////////////////////////////////////////////////
ildgFormat ildgfmt ;
ildgfmt.field = std::string("su3gauge");
const std::string stNC = std::to_string( Nc ) ;
ildgfmt.field = std::string("su"+stNC+"gauge");
if ( format == std::string("IEEE32BIG") ) {
ildgfmt.precision = 32;
@ -871,7 +873,8 @@ class IldgReader : public GridLimeReader {
} else {
assert(found_ildgFormat);
assert ( ildgFormat_.field == std::string("su3gauge") );
const std::string stNC = std::to_string( Nc ) ;
assert ( ildgFormat_.field == std::string("su"+stNC+"gauge") );
///////////////////////////////////////////////////////////////////////////////////////
// Populate our Grid metadata as best we can
@ -879,7 +882,7 @@ class IldgReader : public GridLimeReader {
std::ostringstream vers; vers << ildgFormat_.version;
FieldMetaData_.hdr_version = vers.str();
FieldMetaData_.data_type = std::string("4D_SU3_GAUGE_3X3");
FieldMetaData_.data_type = std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC);
FieldMetaData_.nd=4;
FieldMetaData_.dimension.resize(4);

33
Grid/parallelIO/MetaData.h
View File

@ -6,8 +6,8 @@
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -182,8 +182,8 @@ class GaugeStatistics
public:
void operator()(Lattice<vLorentzColourMatrixD> & data,FieldMetaData &header)
{
header.link_trace=WilsonLoops<Impl>::linkTrace(data);
header.plaquette =WilsonLoops<Impl>::avgPlaquette(data);
header.link_trace = WilsonLoops<Impl>::linkTrace(data);
header.plaquette = WilsonLoops<Impl>::avgPlaquette(data);
}
};
typedef GaugeStatistics<PeriodicGimplD> PeriodicGaugeStatistics;
@ -203,20 +203,24 @@ template<> inline void PrepareMetaData<vLorentzColourMatrixD>(Lattice<vLorentzCo
//////////////////////////////////////////////////////////////////////
inline void reconstruct3(LorentzColourMatrix & cm)
{
const int x=0;
const int y=1;
const int z=2;
assert( Nc < 4 && Nc > 1 ) ;
for(int mu=0;mu<Nd;mu++){
cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
#if Nc == 2
cm(mu)()(1,0) = -adj(cm(mu)()(0,y)) ;
cm(mu)()(1,1) = adj(cm(mu)()(0,x)) ;
#else
const int x=0 , y=1 , z=2 ; // a little disinenuous labelling
cm(mu)()(2,x) = adj(cm(mu)()(0,y)*cm(mu)()(1,z)-cm(mu)()(0,z)*cm(mu)()(1,y)); //x= yz-zy
cm(mu)()(2,y) = adj(cm(mu)()(0,z)*cm(mu)()(1,x)-cm(mu)()(0,x)*cm(mu)()(1,z)); //y= zx-xz
cm(mu)()(2,z) = adj(cm(mu)()(0,x)*cm(mu)()(1,y)-cm(mu)()(0,y)*cm(mu)()(1,x)); //z= xy-yx
#endif
}
}
////////////////////////////////////////////////////////////////////////////////
// Some data types for intermediate storage
////////////////////////////////////////////////////////////////////////////////
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, 2>, Nd >;
template<typename vtype> using iLorentzColour2x3 = iVector<iVector<iVector<vtype, Nc>, Nc-1>, Nd >;
typedef iLorentzColour2x3<Complex> LorentzColour2x3;
typedef iLorentzColour2x3<ComplexF> LorentzColour2x3F;
@ -278,7 +282,6 @@ struct GaugeSimpleMunger{
template <class fobj, class sobj>
struct GaugeSimpleUnmunger {
void operator()(sobj &in, fobj &out) {
for (int mu = 0; mu < Nd; mu++) {
for (int i = 0; i < Nc; i++) {
@ -317,8 +320,8 @@ template<class fobj,class sobj>
struct Gauge3x2munger{
void operator() (fobj &in,sobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
for(int i=0;i<Nc-1;i++){
for(int j=0;j<Nc;j++){
out(mu)()(i,j) = in(mu)(i)(j);
}}
}
@ -330,8 +333,8 @@ template<class fobj,class sobj>
struct Gauge3x2unmunger{
void operator() (sobj &in,fobj &out){
for(int mu=0;mu<Nd;mu++){
for(int i=0;i<2;i++){
for(int j=0;j<3;j++){
for(int i=0;i<Nc-1;i++){
for(int j=0;j<Nc;j++){
out(mu)(i)(j) = in(mu)()(i,j);
}}
}

46
Grid/parallelIO/NerscIO.h
View File

@ -9,6 +9,7 @@
Author: Matt Spraggs <matthew.spraggs@gmail.com>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Jamie Hudspith <renwick.james.hudspth@gmail.com>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -30,6 +31,8 @@
#ifndef GRID_NERSC_IO_H
#define GRID_NERSC_IO_H
#include <string>
NAMESPACE_BEGIN(Grid);
using namespace Grid;
@ -145,15 +148,17 @@ public:
std::string format(header.floating_point);
int ieee32big = (format == std::string("IEEE32BIG"));
int ieee32 = (format == std::string("IEEE32"));
int ieee64big = (format == std::string("IEEE64BIG"));
int ieee64 = (format == std::string("IEEE64") || format == std::string("IEEE64LITTLE"));
const int ieee32big = (format == std::string("IEEE32BIG"));
const int ieee32 = (format == std::string("IEEE32"));
const int ieee64big = (format == std::string("IEEE64BIG"));
const int ieee64 = (format == std::string("IEEE64") || \
format == std::string("IEEE64LITTLE"));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
// depending on datatype, set up munger;
// munger is a function of <floating point, Real, data_type>
if ( header.data_type == std::string("4D_SU3_GAUGE") ) {
const std::string stNC = std::to_string( Nc ) ;
if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE") ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<vLorentzColourMatrixD, LorentzColour2x3F>
(Umu,file,Gauge3x2munger<LorentzColour2x3F,LorentzColourMatrix>(), offset,format,
@ -164,7 +169,7 @@ public:
(Umu,file,Gauge3x2munger<LorentzColour2x3D,LorentzColourMatrix>(),offset,format,
nersc_csum,scidac_csuma,scidac_csumb);
}
} else if ( header.data_type == std::string("4D_SU3_GAUGE_3x3") ) {
} else if ( header.data_type == std::string("4D_SU"+stNC+"_GAUGE_"+stNC+"x"+stNC) ) {
if ( ieee32 || ieee32big ) {
BinaryIO::readLatticeObject<vLorentzColourMatrixD,LorentzColourMatrixF>
(Umu,file,GaugeSimpleMunger<LorentzColourMatrixF,LorentzColourMatrix>(),offset,format,
@ -231,6 +236,7 @@ public:
header.sequence_number = sequence_number;
header.ensemble_id = ens_id;
header.ensemble_label = ens_label;
header.hdr_version = "1.0" ;
typedef LorentzColourMatrixD fobj3D;
typedef LorentzColour2x3D fobj2D;
@ -244,10 +250,14 @@ public:
uint64_t offset;
// Sod it -- always write 3x3 double
header.floating_point = std::string("IEEE64BIG");
header.data_type = std::string("4D_SU3_GAUGE_3x3");
GaugeSimpleUnmunger<fobj3D,sobj> munge;
// Sod it -- always write NcxNc double
header.floating_point = std::string("IEEE64BIG");
const std::string stNC = std::to_string( Nc ) ;
if( two_row ) {
header.data_type = std::string("4D_SU" + stNC + "_GAUGE" );
} else {
header.data_type = std::string("4D_SU" + stNC + "_GAUGE_" + stNC + "x" + stNC );
}
if ( grid->IsBoss() ) {
truncate(file);
offset = writeHeader(header,file);
@ -255,8 +265,15 @@ public:
grid->Broadcast(0,(void *)&offset,sizeof(offset));
uint32_t nersc_csum,scidac_csuma,scidac_csumb;
BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
nersc_csum,scidac_csuma,scidac_csumb);
if( two_row ) {
Gauge3x2unmunger<fobj2D,sobj> munge;
BinaryIO::writeLatticeObject<vobj,fobj2D>(Umu,file,munge,offset,header.floating_point,
nersc_csum,scidac_csuma,scidac_csumb);
} else {
GaugeSimpleUnmunger<fobj3D,sobj> munge;
BinaryIO::writeLatticeObject<vobj,fobj3D>(Umu,file,munge,offset,header.floating_point,
nersc_csum,scidac_csuma,scidac_csumb);
}
header.checksum = nersc_csum;
if ( grid->IsBoss() ) {
writeHeader(header,file);
@ -288,8 +305,7 @@ public:
header.plaquette=0.0;
MachineCharacteristics(header);
uint64_t offset;
uint64_t offset;
#ifdef RNG_RANLUX
header.floating_point = std::string("UINT64");
header.data_type = std::string("RANLUX48");
@ -329,7 +345,7 @@ public:
GridBase *grid = parallel.Grid();
uint64_t offset = readHeader(file,grid,header);
uint64_t offset = readHeader(file,grid,header);
FieldMetaData clone(header);

5
Grid/perfmon/PerfCount.cc
View File

@ -27,10 +27,13 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
#include <Grid/GridCore.h>
#include <Grid/perfmon/PerfCount.h>
#include <Grid/perfmon/Timer.h>
#include <Grid/perfmon/PerfCount.h>
NAMESPACE_BEGIN(Grid);
GridTimePoint theProgramStart = GridClock::now();
#define CacheControl(L,O,R) ((PERF_COUNT_HW_CACHE_##L)|(PERF_COUNT_HW_CACHE_OP_##O<<8)| (PERF_COUNT_HW_CACHE_RESULT_##R<<16))
#define RawConfig(A,B) (A<<8|B)
const PerformanceCounter::PerformanceCounterConfig PerformanceCounter::PerformanceCounterConfigs [] = {

14
Grid/perfmon/PerfCount.h
View File

@ -30,6 +30,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_PERFCOUNT_H
#define GRID_PERFCOUNT_H
#ifndef __SSC_START
#define __SSC_START
#define __SSC_STOP
#endif
#include <sys/time.h>
#include <ctime>
#include <chrono>
@ -72,17 +78,9 @@ static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
inline uint64_t cyclecount(void){
return 0;
}
#define __SSC_MARK(mark) __asm__ __volatile__ ("movl %0, %%ebx; .byte 0x64, 0x67, 0x90 " ::"i"(mark):"%ebx")
#define __SSC_STOP __SSC_MARK(0x110)
#define __SSC_START __SSC_MARK(0x111)
#else
#define __SSC_MARK(mark)
#define __SSC_STOP
#define __SSC_START
/*
* cycle counters arch dependent
*/

22
Grid/perfmon/Timer.h
View File

@ -35,17 +35,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
NAMESPACE_BEGIN(Grid)
// Dress the output; use std::chrono
// C++11 time facilities better?
inline double usecond(void) {
struct timeval tv;
#ifdef TIMERS_ON
gettimeofday(&tv,NULL);
#endif
return 1.0*tv.tv_usec + 1.0e6*tv.tv_sec;
}
typedef std::chrono::system_clock GridClock;
//typedef std::chrono::system_clock GridClock;
typedef std::chrono::high_resolution_clock GridClock;
typedef std::chrono::time_point<GridClock> GridTimePoint;
typedef std::chrono::seconds GridSecs;
@ -53,6 +44,15 @@ typedef std::chrono::milliseconds GridMillisecs;
typedef std::chrono::microseconds GridUsecs;
typedef std::chrono::microseconds GridTime;
extern GridTimePoint theProgramStart;
// Dress the output; use std::chrono
// C++11 time facilities better?
inline double usecond(void) {
auto usecs = std::chrono::duration_cast<GridUsecs>(GridClock::now()-theProgramStart);
return 1.0*usecs.count();
}
inline std::ostream& operator<< (std::ostream & stream, const GridSecs & time)
{
stream << time.count()<<" s";

70
Grid/perfmon/Tracing.h Normal file
View File

@ -0,0 +1,70 @@
#pragma once
NAMESPACE_BEGIN(Grid);
#ifdef GRID_TRACING_NVTX
#include <nvToolsExt.h>
class GridTracer {
public:
GridTracer(const char* name) {
nvtxRangePushA(name);
}
~GridTracer() {
nvtxRangePop();
}
};
inline void tracePush(const char *name) { nvtxRangePushA(name); }
inline void tracePop(const char *name) { nvtxRangePop(); }
inline int traceStart(const char *name) { }
inline void traceStop(int ID) { }
#endif
#ifdef GRID_TRACING_ROCTX
#include <roctracer/roctx.h>
class GridTracer {
public:
GridTracer(const char* name) {
roctxRangePushA(name);
std::cout << "roctxRangePush "<<name<<std::endl;
}
~GridTracer() {
roctxRangePop();
std::cout << "roctxRangePop "<<std::endl;
}
};
inline void tracePush(const char *name) { roctxRangePushA(name); }
inline void tracePop(const char *name) { roctxRangePop(); }
inline int traceStart(const char *name) { roctxRangeStart(name); }
inline void traceStop(int ID) { roctxRangeStop(ID); }
#endif
#ifdef GRID_TRACING_TIMER
class GridTracer {
public:
const char *name;
double elapsed;
GridTracer(const char* _name) {
name = _name;
elapsed=-usecond();
}
~GridTracer() {
elapsed+=usecond();
std::cout << GridLogTracing << name << " took " <<elapsed<< " us" <<std::endl;
}
};
inline void tracePush(const char *name) { }
inline void tracePop(const char *name) { }
inline int traceStart(const char *name) { return 0; }
inline void traceStop(int ID) { }
#endif
#ifdef GRID_TRACING_NONE
#define GRID_TRACE(name)
inline void tracePush(const char *name) { }
inline void tracePop(const char *name) { }
inline int traceStart(const char *name) { return 0; }
inline void traceStop(int ID) { }
#else
#define GRID_TRACE(name) GridTracer uniq_name_using_macros##__COUNTER__(name);
#endif
NAMESPACE_END(Grid);

4
Grid/pugixml/pugixml.cc
View File

@ -16,8 +16,12 @@
#ifdef __NVCC__
#pragma push
#if (__CUDACC_VER_MAJOR__ >= 11) && (__CUDACC_VER_MINOR__ >= 5)
#pragma nv_diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
#else
#pragma diag_suppress declared_but_not_referenced // suppress "function was declared but never referenced warning"
#endif
#endif
#include "pugixml.h"

23
Grid/qcd/QCD.h
View File

@ -451,9 +451,20 @@ template<class vobj> void pokeLorentz(vobj &lhs,const decltype(peekIndex<Lorentz
// Fermion <-> propagator assignements
//////////////////////////////////////////////
//template <class Prop, class Ferm>
#define FAST_FERM_TO_PROP
template <class Fimpl>
void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::FermionField &f, const int s, const int c)
{
#ifdef FAST_FERM_TO_PROP
autoView(p_v,p,CpuWrite);
autoView(f_v,f,CpuRead);
thread_for(idx,p_v.oSites(),{
for(int ss = 0; ss < Ns; ++ss) {
for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
p_v[idx]()(ss,s)(cc,c) = f_v[idx]()(ss)(cc); // Propagator sink index is LEFT, suitable for left mult by gauge link (e.g.)
}}
});
#else
for(int j = 0; j < Ns; ++j)
{
auto pjs = peekSpin(p, j, s);
@ -465,12 +476,23 @@ void FermToProp(typename Fimpl::PropagatorField &p, const typename Fimpl::Fermio
}
pokeSpin(p, pjs, j, s);
}
#endif
}
//template <class Prop, class Ferm>
template <class Fimpl>
void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::PropagatorField &p, const int s, const int c)
{
#ifdef FAST_FERM_TO_PROP
autoView(p_v,p,CpuRead);
autoView(f_v,f,CpuWrite);
thread_for(idx,p_v.oSites(),{
for(int ss = 0; ss < Ns; ++ss) {
for(int cc = 0; cc < Fimpl::Dimension; ++cc) {
f_v[idx]()(ss)(cc) = p_v[idx]()(ss,s)(cc,c); // LEFT index is copied across for s,c right index
}}
});
#else
for(int j = 0; j < Ns; ++j)
{
auto pjs = peekSpin(p, j, s);
@ -482,6 +504,7 @@ void PropToFerm(typename Fimpl::FermionField &f, const typename Fimpl::Propagato
}
pokeSpin(f, fj, j);
}
#endif
}
//////////////////////////////////////////////

333
Grid/qcd/action/fermion/CloverHelpers.h Normal file
View File

@ -0,0 +1,333 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverFermionImplementation.h
Copyright (C) 2017 - 2022
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
Author: Mattia Bruno <mattia.bruno@cern.ch>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#pragma once
#include <Grid/Grid.h>
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
////////////////////////////////////////////
// Standard Clover
// (4+m0) + csw * clover_term
// Exp Clover
// (4+m0) * exp(csw/(4+m0) clover_term)
// = (4+m0) + csw * clover_term + ...
////////////////////////////////////////////
NAMESPACE_BEGIN(Grid);
//////////////////////////////////
// Generic Standard Clover
//////////////////////////////////
template<class Impl>
class CloverHelpers: public WilsonCloverHelpers<Impl> {
public:
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
typedef WilsonCloverHelpers<Impl> Helpers;
static void Instantiate(CloverField& CloverTerm, CloverField& CloverTermInv, RealD csw_t, RealD diag_mass) {
GridBase *grid = CloverTerm.Grid();
CloverTerm += diag_mass;
int lvol = grid->lSites();
int DimRep = Impl::Dimension;
{
autoView(CTv,CloverTerm,CpuRead);
autoView(CTIv,CloverTermInv,CpuWrite);
thread_for(site, lvol, {
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
peekLocalSite(Qx, CTv, lcoor);
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++){
auto zz = Qx()(j, k)(a, b);
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
}
EigenInvCloverOp = EigenCloverOp.inverse();
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
pokeLocalSite(Qxinv, CTIv, lcoor);
});
}
}
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
return Helpers::Cmunu(U, lambda, mu, nu);
}
};
//////////////////////////////////
// Generic Exp Clover
//////////////////////////////////
template<class Impl>
class ExpCloverHelpers: public WilsonCloverHelpers<Impl> {
public:
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
typedef WilsonCloverHelpers<Impl> Helpers;
// Can this be avoided?
static void IdentityTimesC(const CloverField& in, RealD c) {
int DimRep = Impl::Dimension;
autoView(in_v, in, AcceleratorWrite);
accelerator_for(ss, in.Grid()->oSites(), 1, {
for (int sa=0; sa<Ns; sa++)
for (int ca=0; ca<DimRep; ca++)
in_v[ss]()(sa,sa)(ca,ca) = c;
});
}
static int getNMAX(RealD prec, RealD R) {
/* compute stop condition for exponential */
int NMAX=1;
RealD cond=R*R/2.;
while (cond*std::exp(R)>prec) {
NMAX++;
cond*=R/(double)(NMAX+1);
}
return NMAX;
}
static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
GridBase* grid = Clover.Grid();
CloverField ExpClover(grid);
int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
Clover *= (1.0/diag_mass);
// Taylor expansion, slow but generic
// Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
// qN = cN
// qn = cn + qn+1 X
std::vector<RealD> cn(NMAX+1);
cn[0] = 1.0;
for (int i=1; i<=NMAX; i++)
cn[i] = cn[i-1] / RealD(i);
ExpClover = Zero();
IdentityTimesC(ExpClover, cn[NMAX]);
for (int i=NMAX-1; i>=0; i--)
ExpClover = ExpClover * Clover + cn[i];
// prepare inverse
CloverInv = (-1.0)*Clover;
Clover = ExpClover * diag_mass;
ExpClover = Zero();
IdentityTimesC(ExpClover, cn[NMAX]);
for (int i=NMAX-1; i>=0; i--)
ExpClover = ExpClover * CloverInv + cn[i];
CloverInv = ExpClover * (1.0/diag_mass);
}
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
assert(0);
return lambda;
}
};
//////////////////////////////////
// Compact Standard Clover
//////////////////////////////////
template<class Impl>
class CompactCloverHelpers: public CompactWilsonCloverHelpers<Impl>,
public WilsonCloverHelpers<Impl> {
public:
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
INHERIT_COMPACT_CLOVER_TYPES(Impl);
typedef WilsonCloverHelpers<Impl> Helpers;
typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
Clover += diag_mass;
}
static void InvertClover(CloverField& InvClover,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle,
CloverDiagonalField& diagonalInv,
CloverTriangleField& triangleInv,
bool fixedBoundaries) {
CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
}
// TODO: implement Cmunu for better performances with compact layout, but don't do it
// here, but rather in WilsonCloverHelpers.h -> CompactWilsonCloverHelpers
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
return Helpers::Cmunu(U, lambda, mu, nu);
}
};
//////////////////////////////////
// Compact Exp Clover
//////////////////////////////////
template<class Impl>
class CompactExpCloverHelpers: public CompactWilsonCloverHelpers<Impl> {
public:
INHERIT_IMPL_TYPES(Impl);
INHERIT_CLOVER_TYPES(Impl);
INHERIT_COMPACT_CLOVER_TYPES(Impl);
template <typename vtype> using iImplClover = iScalar<iMatrix<iMatrix<vtype, Impl::Dimension>, Ns>>;
typedef CompactWilsonCloverHelpers<Impl> CompactHelpers;
// Can this be avoided?
static void IdentityTimesC(const CloverField& in, RealD c) {
int DimRep = Impl::Dimension;
autoView(in_v, in, AcceleratorWrite);
accelerator_for(ss, in.Grid()->oSites(), 1, {
for (int sa=0; sa<Ns; sa++)
for (int ca=0; ca<DimRep; ca++)
in_v[ss]()(sa,sa)(ca,ca) = c;
});
}
static int getNMAX(RealD prec, RealD R) {
/* compute stop condition for exponential */
int NMAX=1;
RealD cond=R*R/2.;
while (cond*std::exp(R)>prec) {
NMAX++;
cond*=R/(double)(NMAX+1);
}
return NMAX;
}
static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
GridBase* grid = Clover.Grid();
CloverField ExpClover(grid);
int NMAX = getNMAX(Clover, 3.*csw_t/diag_mass);
Clover *= (1.0/diag_mass);
// Taylor expansion, slow but generic
// Horner scheme: a0 + a1 x + a2 x^2 + .. = a0 + x (a1 + x(...))
// qN = cN
// qn = cn + qn+1 X
std::vector<RealD> cn(NMAX+1);
cn[0] = 1.0;
for (int i=1; i<=NMAX; i++)
cn[i] = cn[i-1] / RealD(i);
ExpClover = Zero();
IdentityTimesC(ExpClover, cn[NMAX]);
for (int i=NMAX-1; i>=0; i--)
ExpClover = ExpClover * Clover + cn[i];
// prepare inverse
CloverInv = (-1.0)*Clover;
Clover = ExpClover * diag_mass;
ExpClover = Zero();
IdentityTimesC(ExpClover, cn[NMAX]);
for (int i=NMAX-1; i>=0; i--)
ExpClover = ExpClover * CloverInv + cn[i];
CloverInv = ExpClover * (1.0/diag_mass);
}
static void InvertClover(CloverField& InvClover,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle,
CloverDiagonalField& diagonalInv,
CloverTriangleField& triangleInv,
bool fixedBoundaries) {
if (fixedBoundaries)
{
CompactHelpers::Invert(diagonal, triangle, diagonalInv, triangleInv);
}
else
{
CompactHelpers::ConvertLayout(InvClover, diagonalInv, triangleInv);
}
}
static GaugeLinkField Cmunu(std::vector<GaugeLinkField> &U, GaugeLinkField &lambda, int mu, int nu) {
assert(0);
return lambda;
}
};
NAMESPACE_END(Grid);

5
Grid/qcd/action/fermion/CompactWilsonCloverFermion.h
View File

@ -31,6 +31,7 @@
#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
#include <Grid/qcd/action/fermion/CloverHelpers.h>
NAMESPACE_BEGIN(Grid);
@ -85,7 +86,7 @@ NAMESPACE_BEGIN(Grid);
// + (2 * 1 + 4 * 1/2) triangle parts = 4 triangle parts = 60 complex words per site
// = 84 complex words per site
template<class Impl>
template<class Impl, class CloverHelpers>
class CompactWilsonCloverFermion : public WilsonFermion<Impl>,
public WilsonCloverHelpers<Impl>,
public CompactWilsonCloverHelpers<Impl> {
@ -224,7 +225,7 @@ public:
RealD csw_t;
RealD cF;
bool open_boundaries;
bool fixedBoundaries;
CloverDiagonalField Diagonal, DiagonalEven, DiagonalOdd;
CloverDiagonalField DiagonalInv, DiagonalInvEven, DiagonalInvOdd;

62
Grid/qcd/action/fermion/Fermion.h
View File

@ -138,38 +138,52 @@ typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
// Clover fermions
typedef WilsonCloverFermion<WilsonImplR> WilsonCloverFermionR;
typedef WilsonCloverFermion<WilsonImplF> WilsonCloverFermionF;
typedef WilsonCloverFermion<WilsonImplD> WilsonCloverFermionD;
template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
typedef WilsonCloverFermion<WilsonAdjImplR> WilsonCloverAdjFermionR;
typedef WilsonCloverFermion<WilsonAdjImplF> WilsonCloverAdjFermionF;
typedef WilsonCloverFermion<WilsonAdjImplD> WilsonCloverAdjFermionD;
typedef WilsonClover<WilsonImplR> WilsonCloverFermionR;
typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
typedef WilsonCloverFermion<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
typedef WilsonExpClover<WilsonImplR> WilsonExpCloverFermionR;
typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
typedef WilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
typedef WilsonClover<WilsonAdjImplR> WilsonCloverAdjFermionR;
typedef WilsonClover<WilsonAdjImplF> WilsonCloverAdjFermionF;
typedef WilsonClover<WilsonAdjImplD> WilsonCloverAdjFermionD;
typedef WilsonClover<WilsonTwoIndexSymmetricImplR> WilsonCloverTwoIndexSymmetricFermionR;
typedef WilsonClover<WilsonTwoIndexSymmetricImplF> WilsonCloverTwoIndexSymmetricFermionF;
typedef WilsonClover<WilsonTwoIndexSymmetricImplD> WilsonCloverTwoIndexSymmetricFermionD;
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplR> WilsonCloverTwoIndexAntiSymmetricFermionR;
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplF> WilsonCloverTwoIndexAntiSymmetricFermionF;
typedef WilsonClover<WilsonTwoIndexAntiSymmetricImplD> WilsonCloverTwoIndexAntiSymmetricFermionD;
// Compact Clover fermions
typedef CompactWilsonCloverFermion<WilsonImplR> CompactWilsonCloverFermionR;
typedef CompactWilsonCloverFermion<WilsonImplF> CompactWilsonCloverFermionF;
typedef CompactWilsonCloverFermion<WilsonImplD> CompactWilsonCloverFermionD;
template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion<WImpl, CompactCloverHelpers<WImpl>>;
template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
typedef CompactWilsonCloverFermion<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
typedef CompactWilsonCloverFermion<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
typedef CompactWilsonCloverFermion<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
typedef CompactWilsonClover<WilsonImplR> CompactWilsonCloverFermionR;
typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
typedef CompactWilsonCloverFermion<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
typedef CompactWilsonExpClover<WilsonImplR> CompactWilsonExpCloverFermionR;
typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
typedef CompactWilsonCloverFermion<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
typedef CompactWilsonClover<WilsonAdjImplR> CompactWilsonCloverAdjFermionR;
typedef CompactWilsonClover<WilsonAdjImplF> CompactWilsonCloverAdjFermionF;
typedef CompactWilsonClover<WilsonAdjImplD> CompactWilsonCloverAdjFermionD;
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplR> CompactWilsonCloverTwoIndexSymmetricFermionR;
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplF> CompactWilsonCloverTwoIndexSymmetricFermionF;
typedef CompactWilsonClover<WilsonTwoIndexSymmetricImplD> CompactWilsonCloverTwoIndexSymmetricFermionD;
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplR> CompactWilsonCloverTwoIndexAntiSymmetricFermionR;
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplF> CompactWilsonCloverTwoIndexAntiSymmetricFermionF;
typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonCloverTwoIndexAntiSymmetricFermionD;
// Domain Wall fermions
typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;

3
Grid/qcd/action/fermion/WilsonCloverFermion.h
View File

@ -32,6 +32,7 @@
#include <Grid/qcd/action/fermion/WilsonCloverTypes.h>
#include <Grid/qcd/action/fermion/WilsonCloverHelpers.h>
#include <Grid/qcd/action/fermion/CloverHelpers.h>
NAMESPACE_BEGIN(Grid);
@ -51,7 +52,7 @@ NAMESPACE_BEGIN(Grid);
// csw_r = csw_t to recover the isotropic version
//////////////////////////////////////////////////////////////////
template <class Impl>
template<class Impl, class CloverHelpers>
class WilsonCloverFermion : public WilsonFermion<Impl>,
public WilsonCloverHelpers<Impl>
{

2
Grid/qcd/action/fermion/WilsonCloverHelpers.h
View File

@ -209,6 +209,8 @@ public:
};
////////////////////////////////////////////////////////
template<class Impl> class CompactWilsonCloverHelpers {
public:

2
Grid/qcd/action/fermion/WilsonCloverTypes.h
View File

@ -47,8 +47,6 @@ class CompactWilsonCloverTypes {
public:
INHERIT_IMPL_TYPES(Impl);
static_assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3, "Wrong dimensions");
static constexpr int Nred = Nc * Nhs; // 6
static constexpr int Nblock = Nhs; // 2
static constexpr int Ndiagonal = Nred; // 6

16
Grid/qcd/action/fermion/WilsonCompressor.h
View File

@ -117,19 +117,19 @@ public:
typedef decltype(coalescedRead(*in)) sobj;
typedef decltype(coalescedRead(*out0)) hsobj;
unsigned int Nsimd = vobj::Nsimd();
constexpr unsigned int Nsimd = vobj::Nsimd();
unsigned int mask = Nsimd >> (type + 1);
int lane = acceleratorSIMTlane(Nsimd);
int j0 = lane &(~mask); // inner coor zero
int j1 = lane |(mask) ; // inner coor one
const vobj *vp0 = &in[k];
const vobj *vp1 = &in[m];
const vobj *vp = (lane&mask) ? vp1:vp0;
auto sa = coalescedRead(*vp,j0);
auto sb = coalescedRead(*vp,j1);
const vobj *vp0 = &in[k]; // out0[j] = merge low bit of type from in[k] and in[m]
const vobj *vp1 = &in[m]; // out1[j] = merge hi bit of type from in[k] and in[m]
const vobj *vp = (lane&mask) ? vp1:vp0;// if my lane has high bit take vp1, low bit take vp0
auto sa = coalescedRead(*vp,j0); // lane to read for out 0, NB 50% read coalescing
auto sb = coalescedRead(*vp,j1); // lane to read for out 1
hsobj psa, psb;
projector::Proj(psa,sa,mu,dag);
projector::Proj(psb,sb,mu,dag);
projector::Proj(psa,sa,mu,dag); // spin project the result0
projector::Proj(psb,sb,mu,dag); // spin project the result1
coalescedWrite(out0[j],psa);
coalescedWrite(out1[j],psb);
#else

187
Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h
View File

@ -32,22 +32,23 @@
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
GridCartesian& Fgrid,
GridRedBlackCartesian& Hgrid,
const RealD _mass,
const RealD _csw_r,
const RealD _csw_t,
const RealD _cF,
const WilsonAnisotropyCoefficients& clover_anisotropy,
const ImplParams& impl_p)
template<class Impl, class CloverHelpers>
CompactWilsonCloverFermion<Impl, CloverHelpers>::CompactWilsonCloverFermion(GaugeField& _Umu,
GridCartesian& Fgrid,
GridRedBlackCartesian& Hgrid,
const RealD _mass,
const RealD _csw_r,
const RealD _csw_t,
const RealD _cF,
const WilsonAnisotropyCoefficients& clover_anisotropy,
const ImplParams& impl_p)
: WilsonBase(_Umu, Fgrid, Hgrid, _mass, impl_p, clover_anisotropy)
, csw_r(_csw_r)
, csw_t(_csw_t)
, cF(_cF)
, open_boundaries(impl_p.boundary_phases[Nd-1] == 0.0)
, fixedBoundaries(impl_p.boundary_phases[Nd-1] == 0.0)
, Diagonal(&Fgrid), Triangle(&Fgrid)
, DiagonalEven(&Hgrid), TriangleEven(&Hgrid)
, DiagonalOdd(&Hgrid), TriangleOdd(&Hgrid)
@ -58,83 +59,85 @@ CompactWilsonCloverFermion<Impl>::CompactWilsonCloverFermion(GaugeField& _Umu,
, BoundaryMask(&Fgrid)
, BoundaryMaskEven(&Hgrid), BoundaryMaskOdd(&Hgrid)
{
assert(Nd == 4 && Nc == 3 && Ns == 4 && Impl::Dimension == 3);
csw_r *= 0.5;
csw_t *= 0.5;
if (clover_anisotropy.isAnisotropic)
csw_r /= clover_anisotropy.xi_0;
ImportGauge(_Umu);
if (open_boundaries) {
if (fixedBoundaries) {
this->BoundaryMaskEven.Checkerboard() = Even;
this->BoundaryMaskOdd.Checkerboard() = Odd;
CompactHelpers::SetupMasks(this->BoundaryMask, this->BoundaryMaskEven, this->BoundaryMaskOdd);
}
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Dhop(const FermionField& in, FermionField& out, int dag) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Dhop(const FermionField& in, FermionField& out, int dag) {
WilsonBase::Dhop(in, out, dag);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopOE(const FermionField& in, FermionField& out, int dag) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopOE(const FermionField& in, FermionField& out, int dag) {
WilsonBase::DhopOE(in, out, dag);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopEO(const FermionField& in, FermionField& out, int dag) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopEO(const FermionField& in, FermionField& out, int dag) {
WilsonBase::DhopEO(in, out, dag);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDir(const FermionField& in, FermionField& out, int dir, int disp) {
WilsonBase::DhopDir(in, out, dir, disp);
if(this->open_boundaries) ApplyBoundaryMask(out);
if(this->fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::DhopDirAll(const FermionField& in, std::vector<FermionField>& out) {
WilsonBase::DhopDirAll(in, out);
if(this->open_boundaries) {
if(this->fixedBoundaries) {
for(auto& o : out) ApplyBoundaryMask(o);
}
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::M(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField& in, FermionField& out) {
out.Checkerboard() = in.Checkerboard();
WilsonBase::Dhop(in, out, DaggerNo); // call base to save applying bc
Mooee(in, Tmp);
axpy(out, 1.0, out, Tmp);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Mdag(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField& in, FermionField& out) {
out.Checkerboard() = in.Checkerboard();
WilsonBase::Dhop(in, out, DaggerYes); // call base to save applying bc
MooeeDag(in, Tmp);
axpy(out, 1.0, out, Tmp);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Meooe(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Meooe(const FermionField& in, FermionField& out) {
WilsonBase::Meooe(in, out);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MeooeDag(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeooeDag(const FermionField& in, FermionField& out) {
WilsonBase::MeooeDag(in, out);
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField& in, FermionField& out) {
if(in.Grid()->_isCheckerBoarded) {
if(in.Checkerboard() == Odd) {
MooeeInternal(in, out, DiagonalOdd, TriangleOdd);
@ -144,16 +147,16 @@ void CompactWilsonCloverFermion<Impl>::Mooee(const FermionField& in, FermionFiel
} else {
MooeeInternal(in, out, Diagonal, Triangle);
}
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeDag(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField& in, FermionField& out) {
Mooee(in, out); // blocks are hermitian
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField& in, FermionField& out) {
if(in.Grid()->_isCheckerBoarded) {
if(in.Checkerboard() == Odd) {
MooeeInternal(in, out, DiagonalInvOdd, TriangleInvOdd);
@ -163,27 +166,27 @@ void CompactWilsonCloverFermion<Impl>::MooeeInv(const FermionField& in, FermionF
} else {
MooeeInternal(in, out, DiagonalInv, TriangleInv);
}
if(open_boundaries) ApplyBoundaryMask(out);
if(fixedBoundaries) ApplyBoundaryMask(out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeInvDag(const FermionField& in, FermionField& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField& in, FermionField& out) {
MooeeInv(in, out); // blocks are hermitian
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::Mdir(const FermionField& in, FermionField& out, int dir, int disp) {
DhopDir(in, out, dir, disp);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MdirAll(const FermionField& in, std::vector<FermionField>& out) {
DhopDirAll(in, out);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
assert(!open_boundaries); // TODO check for changes required for open bc
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField& force, const FermionField& X, const FermionField& Y, int dag) {
assert(!fixedBoundaries); // TODO check for changes required for open bc
// NOTE: code copied from original clover term
conformable(X.Grid(), Y.Grid());
@ -254,7 +257,7 @@ void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionFi
}
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu); // checked
force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu); // checked
count++;
}
@ -264,18 +267,18 @@ void CompactWilsonCloverFermion<Impl>::MDeriv(GaugeField& force, const FermionFi
force += clover_force;
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
assert(0);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField& mat, const FermionField& U, const FermionField& V, int dag) {
assert(0);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField& in,
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField& in,
FermionField& out,
const CloverDiagonalField& diagonal,
const CloverTriangleField& triangle) {
@ -288,8 +291,8 @@ void CompactWilsonCloverFermion<Impl>::MooeeInternal(const FermionField&
CompactHelpers::MooeeKernel(diagonal.oSites(), 1, in, out, diagonal, triangle);
}
template<class Impl>
void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
template<class Impl, class CloverHelpers>
void CompactWilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField& _Umu) {
// NOTE: parts copied from original implementation
// Import gauge into base class
@ -302,6 +305,7 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
GridBase* grid = _Umu.Grid();
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
CloverField TmpOriginal(grid);
CloverField TmpInverse(grid);
// Compute the field strength terms mu>nu
double t2 = usecond();
@ -321,22 +325,30 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
TmpOriginal += Helpers::fillCloverXT(Ex) * csw_t;
TmpOriginal += Helpers::fillCloverYT(Ey) * csw_t;
TmpOriginal += Helpers::fillCloverZT(Ez) * csw_t;
TmpOriginal += this->diag_mass;
// Instantiate the clover term
// - In case of the standard clover the mass term is added
// - In case of the exponential clover the clover term is exponentiated
double t4 = usecond();
CloverHelpers::InstantiateClover(TmpOriginal, TmpInverse, csw_t, this->diag_mass);
// Convert the data layout of the clover term
double t4 = usecond();
double t5 = usecond();
CompactHelpers::ConvertLayout(TmpOriginal, Diagonal, Triangle);
// Possible modify the boundary values
double t5 = usecond();
if(open_boundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
// Invert the clover term in the improved layout
// Modify the clover term at the temporal boundaries in case of open boundary conditions
double t6 = usecond();
CompactHelpers::Invert(Diagonal, Triangle, DiagonalInv, TriangleInv);
if(fixedBoundaries) CompactHelpers::ModifyBoundaries(Diagonal, Triangle, csw_t, cF, this->diag_mass);
// Invert the Clover term
// In case of the exponential clover with (anti-)periodic boundary conditions exp(-Clover) saved
// in TmpInverse can be used. In all other cases the clover term has to be explictly inverted.
// TODO: For now this inversion is explictly done on the CPU
double t7 = usecond();
CloverHelpers::InvertClover(TmpInverse, Diagonal, Triangle, DiagonalInv, TriangleInv, fixedBoundaries);
// Fill the remaining clover fields
double t7 = usecond();
double t8 = usecond();
pickCheckerboard(Even, DiagonalEven, Diagonal);
pickCheckerboard(Even, TriangleEven, Triangle);
pickCheckerboard(Odd, DiagonalOdd, Diagonal);
@ -347,20 +359,19 @@ void CompactWilsonCloverFermion<Impl>::ImportGauge(const GaugeField& _Umu) {
pickCheckerboard(Odd, TriangleInvOdd, TriangleInv);
// Report timings
double t8 = usecond();
#if 0
std::cout << GridLogMessage << "CompactWilsonCloverFermion::ImportGauge timings:"
<< " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
<< ", allocations = " << (t2 - t1) / 1e6
<< ", field strength = " << (t3 - t2) / 1e6
<< ", fill clover = " << (t4 - t3) / 1e6
<< ", convert = " << (t5 - t4) / 1e6
<< ", boundaries = " << (t6 - t5) / 1e6
<< ", inversions = " << (t7 - t6) / 1e6
<< ", pick cbs = " << (t8 - t7) / 1e6
<< ", total = " << (t8 - t0) / 1e6
<< std::endl;
#endif
double t9 = usecond();
std::cout << GridLogDebug << "CompactWilsonCloverFermion::ImportGauge timings:" << std::endl;
std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
std::cout << GridLogDebug << "allocations = " << (t2 - t1) / 1e6 << std::endl;
std::cout << GridLogDebug << "field strength = " << (t3 - t2) / 1e6 << std::endl;
std::cout << GridLogDebug << "fill clover = " << (t4 - t3) / 1e6 << std::endl;
std::cout << GridLogDebug << "instantiate clover = " << (t5 - t4) / 1e6 << std::endl;
std::cout << GridLogDebug << "convert layout = " << (t6 - t5) / 1e6 << std::endl;
std::cout << GridLogDebug << "modify boundaries = " << (t7 - t6) / 1e6 << std::endl;
std::cout << GridLogDebug << "invert clover = " << (t8 - t7) / 1e6 << std::endl;
std::cout << GridLogDebug << "pick cbs = " << (t9 - t8) / 1e6 << std::endl;
std::cout << GridLogDebug << "total = " << (t9 - t0) / 1e6 << std::endl;
}
NAMESPACE_END(Grid);

112
Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
View File

@ -34,8 +34,8 @@
NAMESPACE_BEGIN(Grid);
template<class Impl>
WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField& _Umu,
template<class Impl, class CloverHelpers>
WilsonCloverFermion<Impl, CloverHelpers>::WilsonCloverFermion(GaugeField& _Umu,
GridCartesian& Fgrid,
GridRedBlackCartesian& Hgrid,
const RealD _mass,
@ -74,8 +74,8 @@ WilsonCloverFermion<Impl>::WilsonCloverFermion(GaugeField&
}
// *NOT* EO
template <class Impl>
void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::M(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
@ -89,8 +89,8 @@ void WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
out += temp;
}
template <class Impl>
void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::Mdag(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
@ -104,8 +104,8 @@ void WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
out += temp;
}
template <class Impl>
void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::ImportGauge(const GaugeField &_Umu)
{
double t0 = usecond();
WilsonFermion<Impl>::ImportGauge(_Umu);
@ -131,47 +131,11 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
CloverTerm += Helpers::fillCloverXT(Ex) * csw_t;
CloverTerm += Helpers::fillCloverYT(Ey) * csw_t;
CloverTerm += Helpers::fillCloverZT(Ez) * csw_t;
CloverTerm += diag_mass;
double t4 = usecond();
int lvol = _Umu.Grid()->lSites();
int DimRep = Impl::Dimension;
CloverHelpers::Instantiate(CloverTerm, CloverTermInv, csw_t, this->diag_mass);
double t5 = usecond();
{
autoView(CTv,CloverTerm,CpuRead);
autoView(CTIv,CloverTermInv,CpuWrite);
thread_for(site, lvol, {
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
typename SiteClover::scalar_object Qx = Zero(), Qxinv = Zero();
peekLocalSite(Qx, CTv, lcoor);
//if (csw!=0){
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++){
auto zz = Qx()(j, k)(a, b);
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
}
// if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
EigenInvCloverOp = EigenCloverOp.inverse();
//std::cout << EigenInvCloverOp << std::endl;
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
// }
pokeLocalSite(Qxinv, CTIv, lcoor);
});
}
double t6 = usecond();
// Separate the even and odd parts
pickCheckerboard(Even, CloverTermEven, CloverTerm);
pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
@ -184,48 +148,44 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
double t7 = usecond();
double t6 = usecond();
#if 0
std::cout << GridLogMessage << "WilsonCloverFermion::ImportGauge timings:"
<< " WilsonFermion::Importgauge = " << (t1 - t0) / 1e6
<< ", allocations = " << (t2 - t1) / 1e6
<< ", field strength = " << (t3 - t2) / 1e6
<< ", fill clover = " << (t4 - t3) / 1e6
<< ", misc = " << (t5 - t4) / 1e6
<< ", inversions = " << (t6 - t5) / 1e6
<< ", pick cbs = " << (t7 - t6) / 1e6
<< ", total = " << (t7 - t0) / 1e6
<< std::endl;
#endif
std::cout << GridLogDebug << "WilsonCloverFermion::ImportGauge timings:" << std::endl;
std::cout << GridLogDebug << "WilsonFermion::Importgauge = " << (t1 - t0) / 1e6 << std::endl;
std::cout << GridLogDebug << "allocations = " << (t2 - t1) / 1e6 << std::endl;
std::cout << GridLogDebug << "field strength = " << (t3 - t2) / 1e6 << std::endl;
std::cout << GridLogDebug << "fill clover = " << (t4 - t3) / 1e6 << std::endl;
std::cout << GridLogDebug << "instantiation = " << (t5 - t4) / 1e6 << std::endl;
std::cout << GridLogDebug << "pick cbs = " << (t6 - t5) / 1e6 << std::endl;
std::cout << GridLogDebug << "total = " << (t6 - t0) / 1e6 << std::endl;
}
template <class Impl>
void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::Mooee(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerNo, InverseNo);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeDag(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerYes, InverseNo);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInv(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerNo, InverseYes);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInvDag(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerYes, InverseYes);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
{
out.Checkerboard() = in.Checkerboard();
CloverField *Clover;
@ -278,8 +238,8 @@ void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionFie
} // MooeeInternal
// Derivative parts unpreconditioned pseudofermions
template <class Impl>
void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MDeriv(GaugeField &force, const FermionField &X, const FermionField &Y, int dag)
{
conformable(X.Grid(), Y.Grid());
conformable(X.Grid(), force.Grid());
@ -349,7 +309,7 @@ void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X,
}
PropagatorField Slambda = Gamma(sigma[count]) * Lambda; // sigma checked
Impl::TraceSpinImpl(lambda, Slambda); // traceSpin ok
force_mu -= factor*Helpers::Cmunu(U, lambda, mu, nu); // checked
force_mu -= factor*CloverHelpers::Cmunu(U, lambda, mu, nu); // checked
count++;
}
@ -360,15 +320,15 @@ void WilsonCloverFermion<Impl>::MDeriv(GaugeField &force, const FermionField &X,
}
// Derivative parts
template <class Impl>
void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
{
assert(0);
}
// Derivative parts
template <class Impl>
void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
template<class Impl, class CloverHelpers>
void WilsonCloverFermion<Impl, CloverHelpers>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
{
assert(0); // not implemented yet
}

85
Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
View File

@ -4,12 +4,13 @@ Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
Copyright (C) 2015
Copyright (C) 2022
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Fabian Joswig <fabian.joswig@ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -599,11 +600,47 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
Current curr_type,
unsigned int mu)
{
if(curr_type != Current::Vector)
{
std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
exit(1);
}
Gamma g5(Gamma::Algebra::Gamma5);
conformable(_grid, q_in_1.Grid());
conformable(_grid, q_in_2.Grid());
conformable(_grid, q_out.Grid());
assert(0);
auto UGrid= this->GaugeGrid();
PropagatorField tmp_shifted(UGrid);
PropagatorField g5Lg5(UGrid);
PropagatorField R(UGrid);
PropagatorField gmuR(UGrid);
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT,
};
Gamma gmu=Gamma(Gmu[mu]);
g5Lg5=g5*q_in_1*g5;
tmp_shifted=Cshift(q_in_2,mu,1);
Impl::multLinkField(R,this->Umu,tmp_shifted,mu);
gmuR=gmu*R;
q_out=adj(g5Lg5)*R;
q_out-=adj(g5Lg5)*gmuR;
tmp_shifted=Cshift(q_in_1,mu,1);
Impl::multLinkField(g5Lg5,this->Umu,tmp_shifted,mu);
g5Lg5=g5*g5Lg5*g5;
R=q_in_2;
gmuR=gmu*R;
q_out-=adj(g5Lg5)*R;
q_out-=adj(g5Lg5)*gmuR;
}
@ -617,9 +654,51 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
unsigned int tmax,
ComplexField &lattice_cmplx)
{
if(curr_type != Current::Vector)
{
std::cout << GridLogError << "Only the conserved vector current is implemented so far." << std::endl;
exit(1);
}
int tshift = (mu == Nd-1) ? 1 : 0;
unsigned int LLt = GridDefaultLatt()[Tp];
conformable(_grid, q_in.Grid());
conformable(_grid, q_out.Grid());
assert(0);
auto UGrid= this->GaugeGrid();
PropagatorField tmp(UGrid);
PropagatorField Utmp(UGrid);
PropagatorField L(UGrid);
PropagatorField zz (UGrid);
zz=Zero();
LatticeInteger lcoor(UGrid); LatticeCoordinate(lcoor,Nd-1);
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT,
};
Gamma gmu=Gamma(Gmu[mu]);
tmp = Cshift(q_in,mu,1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu);
tmp = ( Utmp*lattice_cmplx - gmu*Utmp*lattice_cmplx ); // Forward hop
tmp = where((lcoor>=tmin),tmp,zz); // Mask the time
q_out = where((lcoor<=tmax),tmp,zz); // Position of current complicated
tmp = q_in *lattice_cmplx;
tmp = Cshift(tmp,mu,-1);
Impl::multLinkField(Utmp,this->Umu,tmp,mu+Nd); // Adjoint link
tmp = -( Utmp + gmu*Utmp );
// Mask the time
if (tmax == LLt - 1 && tshift == 1){ // quick fix to include timeslice 0 if tmax + tshift is over the last timeslice
unsigned int t0 = 0;
tmp = where(((lcoor==t0) || (lcoor>=tmin+tshift)),tmp,zz);
} else {
tmp = where((lcoor>=tmin+tshift),tmp,zz);
}
q_out+= where((lcoor<=tmax+tshift),tmp,zz); // Position of current complicated
}
NAMESPACE_END(Grid);

3
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -498,6 +498,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDag); return;}
#endif
acceleratorFenceComputeStream();
} else if( interior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt); return;}
@ -505,11 +506,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagInt); return;}
#endif
} else if( exterior ) {
acceleratorFenceComputeStream();
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagExt); return;}
#endif
acceleratorFenceComputeStream();
}
assert(0 && " Kernel optimisation case not covered ");
}

5
Grid/qcd/action/fermion/instantiation/CompactWilsonCloverFermionInstantiation.cc.master
View File

@ -9,6 +9,7 @@
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: Daniel Richtmann <daniel.richtmann@gmail.com>
Author: Mattia Bruno <mattia.bruno@cern.ch>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -32,10 +33,12 @@
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/CompactWilsonCloverFermion.h>
#include <Grid/qcd/action/fermion/implementation/CompactWilsonCloverFermionImplementation.h>
#include <Grid/qcd/action/fermion/CloverHelpers.h>
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class CompactWilsonCloverFermion<IMPLEMENTATION>;
template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactCloverHelpers<IMPLEMENTATION>>;
template class CompactWilsonCloverFermion<IMPLEMENTATION, CompactExpCloverHelpers<IMPLEMENTATION>>;
NAMESPACE_END(Grid);

51
Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonAdjImplD/WilsonKernelsInstantiationWilsonAdjImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonAdjImplF/WilsonKernelsInstantiationWilsonAdjImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

7
Grid/qcd/action/fermion/instantiation/WilsonCloverFermionInstantiation.cc.master
View File

@ -8,7 +8,8 @@
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
Author: Mattia Bruno <mattia.bruno@cern.ch>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
@ -31,10 +32,12 @@
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
#include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
#include <Grid/qcd/action/fermion/CloverHelpers.h>
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonCloverFermion<IMPLEMENTATION>;
template class WilsonCloverFermion<IMPLEMENTATION, CloverHelpers<IMPLEMENTATION>>;
template class WilsonCloverFermion<IMPLEMENTATION, ExpCloverHelpers<IMPLEMENTATION>>;
NAMESPACE_END(Grid);

51
Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonImplD/WilsonKernelsInstantiationWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonImplF/WilsonKernelsInstantiationWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexAntiSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexAntiSymmetricImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplD/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/WilsonTwoIndexSymmetricImplF/WilsonKernelsInstantiationWilsonTwoIndexSymmetricImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplD/WilsonKernelsInstantiationZWilsonImplD.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

51
Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc
View File

@ -1,51 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015, 2020
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#ifndef AVX512
#ifndef QPX
#ifndef A64FX
#ifndef A64FXFIXEDSIZE
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#endif
#endif
#endif
#endif
NAMESPACE_BEGIN(Grid);
#include "impl.h"
template class WilsonKernels<IMPLEMENTATION>;
NAMESPACE_END(Grid);

1
Grid/qcd/action/fermion/instantiation/ZWilsonImplF/WilsonKernelsInstantiationZWilsonImplF.cc Symbolic link
View File

@ -0,0 +1 @@
../WilsonKernelsInstantiation.cc.master

24
Grid/qcd/action/fermion/instantiation/generate_instantiations.sh
View File

@ -18,6 +18,10 @@ WILSON_IMPL_LIST=" \
GparityWilsonImplF \
GparityWilsonImplD "
COMPACT_WILSON_IMPL_LIST=" \
WilsonImplF \
WilsonImplD "
DWF_IMPL_LIST=" \
WilsonImplF \
WilsonImplD \
@ -40,13 +44,23 @@ EOF
done
CC_LIST="WilsonCloverFermionInstantiation CompactWilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
CC_LIST="WilsonCloverFermionInstantiation WilsonFermionInstantiation WilsonKernelsInstantiation WilsonTMFermionInstantiation"
for impl in $WILSON_IMPL_LIST
do
for f in $CC_LIST
do
ln -f -s ../$f.cc.master $impl/$f$impl.cc
ln -f -s ../$f.cc.master $impl/$f$impl.cc
done
done
CC_LIST="CompactWilsonCloverFermionInstantiation"
for impl in $COMPACT_WILSON_IMPL_LIST
do
for f in $CC_LIST
do
ln -f -s ../$f.cc.master $impl/$f$impl.cc
done
done
@ -63,14 +77,14 @@ for impl in $DWF_IMPL_LIST $GDWF_IMPL_LIST
do
for f in $CC_LIST
do
ln -f -s ../$f.cc.master $impl/$f$impl.cc
ln -f -s ../$f.cc.master $impl/$f$impl.cc
done
done
# overwrite the .cc file in Gparity directories
for impl in $GDWF_IMPL_LIST
do
ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
ln -f -s ../WilsonKernelsInstantiationGparity.cc.master $impl/WilsonKernelsInstantiation$impl.cc
done
@ -84,7 +98,7 @@ for impl in $STAG_IMPL_LIST
do
for f in $CC_LIST
do
ln -f -s ../$f.cc.master $impl/$f$impl.cc
ln -f -s ../$f.cc.master $impl/$f$impl.cc
done
done

2
Grid/qcd/action/gauge/Photon.h
View File

@ -49,7 +49,7 @@ NAMESPACE_BEGIN(Grid);
typedef Lattice<SiteLink> LinkField;
typedef Lattice<SiteField> Field;
typedef Field ComplexField;
typedef LinkField ComplexField;
};
typedef QedGImpl<vComplex> QedGImplR;

8
Grid/qcd/utils/GaugeFix.h
View File

@ -55,12 +55,12 @@ public:
}
}
static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
GridBase *grid = Umu.Grid();
GaugeMat xform(grid);
SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog);
SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
}
static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1) {
static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
GridBase *grid = Umu.Grid();
@ -122,6 +122,8 @@ public:
}
}
std::cout << GridLogError << "Gauge fixing did not converge in " << maxiter << " iterations." << std::endl;
if (err_on_no_converge) assert(0);
};
static Real SteepestDescentStep(std::vector<GaugeMat> &U,GaugeMat &xform,Real & alpha, GaugeMat & dmuAmu,int orthog) {
GridBase *grid = U[0].Grid();

3
Grid/qcd/utils/WilsonLoops.h
View File

@ -125,7 +125,6 @@ public:
return sumplaq / vol / faces / Nc; // Nd , Nc dependent... FIXME
}
//////////////////////////////////////////////////
// average over all x,y,z the temporal loop
//////////////////////////////////////////////////
@ -165,7 +164,7 @@ public:
double vol = Umu.Grid()->gSites();
return p.real() / vol / 4.0 / 3.0;
return p.real() / vol / (4.0 * Nc ) ;
};
//////////////////////////////////////////////////

4
Grid/serialisation/JSON_IO.cc
View File

@ -26,7 +26,7 @@
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
#ifndef GRID_HIP
NAMESPACE_BEGIN(Grid);
@ -82,7 +82,7 @@ void JSONWriter::writeDefault(const std::string &s, const std::string &x)
if (s.size())
ss_ << "\""<< s << "\" : \"" << os.str() << "\" ," ;
else
ss_ << os.str() << " ," ;
ss_ << "\""<< os.str() << "\" ," ;
}
// Reader implementation ///////////////////////////////////////////////////////

4
Grid/serialisation/JSON_IO.h
View File

@ -54,7 +54,7 @@ namespace Grid
void pop(void);
template <typename U>
void writeDefault(const std::string &s, const U &x);
#ifdef __NVCC__
#if defined(GRID_CUDA) || defined(GRID_HIP)
void writeDefault(const std::string &s, const Grid::ComplexD &x)
{
std::complex<double> z(real(x),imag(x));
@ -101,7 +101,7 @@ namespace Grid
void readDefault(const std::string &s, std::vector<U> &output);
template <typename U, typename P>
void readDefault(const std::string &s, std::pair<U,P> &output);
#ifdef __NVCC__
#if defined(GRID_CUDA) || defined(GRID_HIP)
void readDefault(const std::string &s, ComplexD &output)
{
std::complex<double> z;

2
Grid/serialisation/Serialisation.h
View File

@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include "BinaryIO.h"
#include "TextIO.h"
#include "XmlIO.h"
#if (!defined(GRID_CUDA)) && (!defined(GRID_HIP))
#ifndef GRID_HIP
#include "JSON_IO.h"
#endif

32
Grid/stencil/Stencil.h
View File

@ -80,11 +80,14 @@ void Gather_plane_simple_table (commVector<std::pair<int,int> >& table,const Lat
///////////////////////////////////////////////////////////////////
template<class cobj,class vobj,class compressor>
void Gather_plane_exchange_table(const Lattice<vobj> &rhs,
commVector<cobj *> pointers,int dimension,int plane,int cbmask,compressor &compress,int type) __attribute__((noinline));
commVector<cobj *> pointers,
int dimension,int plane,
int cbmask,compressor &compress,int type) __attribute__((noinline));
template<class cobj,class vobj,class compressor>
void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const Lattice<vobj> &rhs,
Vector<cobj *> pointers,int dimension,int plane,int cbmask,
void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,
const Lattice<vobj> &rhs,
std::vector<cobj *> &pointers,int dimension,int plane,int cbmask,
compressor &compress,int type)
{
assert( (table.size()&0x1)==0);
@ -92,14 +95,15 @@ void Gather_plane_exchange_table(commVector<std::pair<int,int> >& table,const La
int so = plane*rhs.Grid()->_ostride[dimension]; // base offset for start of plane
auto rhs_v = rhs.View(AcceleratorRead);
auto rhs_p = &rhs_v[0];
auto p0=&pointers[0][0];
auto p1=&pointers[1][0];
auto tp=&table[0];
accelerator_forNB(j, num, vobj::Nsimd(), {
compress.CompressExchange(p0,p1, &rhs_v[0], j,
so+tp[2*j ].second,
so+tp[2*j+1].second,
type);
compress.CompressExchange(p0,p1, rhs_p, j,
so+tp[2*j ].second,
so+tp[2*j+1].second,
type);
});
rhs_v.ViewClose();
}
@ -230,8 +234,8 @@ public:
};
struct Merge {
cobj * mpointer;
Vector<scalar_object *> rpointers;
Vector<cobj *> vpointers;
// std::vector<scalar_object *> rpointers;
std::vector<cobj *> vpointers;
Integer buffer_size;
Integer type;
};
@ -406,6 +410,7 @@ public:
comms_bytes+=bytes;
shm_bytes +=2*Packets[i].bytes-bytes;
}
_grid->StencilBarrier();// Synch shared memory on a single nodes
}
void CommunicateComplete(std::vector<std::vector<CommsRequest_t> > &reqs)
@ -420,7 +425,7 @@ public:
////////////////////////////////////////////////////////////////////////
void Communicate(void)
{
if ( CartesianCommunicator::CommunicatorPolicy == CartesianCommunicator::CommunicatorPolicySequential ){
if ( 0 ){
thread_region {
// must be called in parallel region
int mythread = thread_num();
@ -569,7 +574,7 @@ public:
d.buffer_size = buffer_size;
dv.push_back(d);
}
void AddMerge(cobj *merge_p,Vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
void AddMerge(cobj *merge_p,std::vector<cobj *> &rpointers,Integer buffer_size,Integer type,std::vector<Merge> &mv) {
Merge m;
m.type = type;
m.mpointer = merge_p;
@ -582,6 +587,7 @@ public:
}
template<class decompressor> void CommsMergeSHM(decompressor decompress) {
mpi3synctime-=usecond();
accelerator_barrier();
_grid->StencilBarrier();// Synch shared memory on a single nodes
mpi3synctime+=usecond();
shmmergetime-=usecond();
@ -1114,8 +1120,8 @@ public:
int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
Vector<cobj *> rpointers(maxl);
Vector<cobj *> spointers(maxl);
std::vector<cobj *> rpointers(maxl);
std::vector<cobj *> spointers(maxl);
///////////////////////////////////////////
// Work out what to send where

3
Grid/threads/Accelerator.cc
View File

@ -195,12 +195,15 @@ void acceleratorInit(void)
#ifdef GRID_SYCL
cl::sycl::queue *theGridAccelerator;
cl::sycl::queue *theCopyAccelerator;
void acceleratorInit(void)
{
int nDevices = 1;
cl::sycl::gpu_selector selector;
cl::sycl::device selectedDevice { selector };
theGridAccelerator = new sycl::queue (selectedDevice);
// theCopyAccelerator = new sycl::queue (selectedDevice);
theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
zeInit(0);

36
Grid/threads/Accelerator.h
View File

@ -247,7 +247,6 @@ inline int acceleratorIsCommunicable(void *ptr)
//////////////////////////////////////////////
// SyCL acceleration
//////////////////////////////////////////////
#ifdef GRID_SYCL
NAMESPACE_END(Grid);
#include <CL/sycl.hpp>
@ -262,6 +261,7 @@ NAMESPACE_END(Grid);
NAMESPACE_BEGIN(Grid);
extern cl::sycl::queue *theGridAccelerator;
extern cl::sycl::queue *theCopyAccelerator;
#ifdef __SYCL_DEVICE_ONLY__
#define GRID_SIMT
@ -289,7 +289,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
cgh.parallel_for( \
cl::sycl::nd_range<3>(global,local), \
[=] (cl::sycl::nd_item<3> item) /*mutable*/ \
[[intel::reqd_sub_group_size(8)]] \
[[intel::reqd_sub_group_size(16)]] \
{ \
auto iter1 = item.get_global_id(0); \
auto iter2 = item.get_global_id(1); \
@ -298,19 +298,19 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
}); \
});
#define accelerator_barrier(dummy) theGridAccelerator->wait();
#define accelerator_barrier(dummy) { theGridAccelerator->wait(); }
inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) {
theGridAccelerator->memcpy(to,from,bytes);
}
inline void acceleratorCopySynchronise(void) { theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
inline void acceleratorCopySynchronise(void) { theCopyAccelerator->wait(); }
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes);}
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
inline int acceleratorIsCommunicable(void *ptr)
{
#if 0
@ -481,9 +481,10 @@ inline void acceleratorCopySynchronise(void) { hipStreamSynchronize(copyStream);
#define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) thread_for2d(iter1,num1,iter2,num2,{ __VA_ARGS__ });
accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { memcpy(to,from,bytes);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { memcpy(to,from,bytes);}
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { thread_bcopy(from,to,bytes); }
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { thread_bcopy(from,to,bytes);}
inline void acceleratorCopySynchronise(void) {};
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
@ -510,7 +511,16 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
inline void acceleratorFreeCpu (void *ptr){free(ptr);};
#endif
//////////////////////////////////////////////
// Fencing needed ONLY for SYCL
//////////////////////////////////////////////
#ifdef GRID_SYCL
inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
#else
// Ordering within a stream guaranteed on Nvidia & AMD
inline void acceleratorFenceComputeStream(void){ };
#endif
///////////////////////////////////////////////////
// Synchronise across local threads for divergence resynch

17
Grid/threads/Threads.h
View File

@ -72,3 +72,20 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define thread_region DO_PRAGMA(omp parallel)
#define thread_critical DO_PRAGMA(omp critical)
#ifdef GRID_OMP
inline void thread_bcopy(void *from, void *to,size_t bytes)
{
uint64_t *ufrom = (uint64_t *)from;
uint64_t *uto = (uint64_t *)to;
assert(bytes%8==0);
uint64_t words=bytes/8;
thread_for(w,words,{
uto[w] = ufrom[w];
});
}
#else
inline void thread_bcopy(void *from, void *to,size_t bytes)
{
bcopy(from,to,bytes);
}
#endif

7
Grid/util/Sha.h
View File

@ -27,6 +27,7 @@
/* END LEGAL */
extern "C" {
#include <openssl/sha.h>
#include <openssl/evp.h>
}
#ifdef USE_IPP
#include "ipp.h"
@ -70,10 +71,8 @@ public:
static inline std::vector<unsigned char> sha256(const void *data,size_t bytes)
{
std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
SHA256_CTX sha256;
SHA256_Init (&sha256);
SHA256_Update(&sha256, data,bytes);
SHA256_Final (&hash[0], &sha256);
auto digest = EVP_get_digestbyname("SHA256");
EVP_Digest(data, bytes, &hash[0], NULL, digest, NULL);
return hash;
}
static inline std::vector<int> sha256_seeds(const std::string &s)