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

Working simple OpenMP offloading with cudaMallocManaged; cshift not working

Browse Source
This commit is contained in:
Meifeng Lin
2021-09-29 15:23:13 -07:00
parent 229ce57fef
commit c2f8ba194e
5 changed files with 47 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
Grid/qcd/action/fermion/WilsonCloverFermion.h
View File

@@ -266,7 +266,8 @@ protected:
T = Zero();
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
int size=CloverTerm.Grid()->oSites();
accelerator_for(i, size,1,
{
T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
@@ -284,7 +285,8 @@ protected:
autoView(T_v, T,AcceleratorWrite);
autoView(F_v, F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
int size=CloverTerm.Grid()->oSites();
accelerator_for(i, size,1,
{
T_v[i]()(0, 1) = -F_v[i]()();
T_v[i]()(1, 0) = F_v[i]()();
@@ -302,7 +304,8 @@ protected:
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
int size=CloverTerm.Grid()->oSites();
accelerator_for(i, size,1,
{
T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
T_v[i]()(1, 1) = timesI(F_v[i]()());
@@ -320,7 +323,8 @@ protected:
autoView( T_v , T, AcceleratorWrite);
autoView( F_v , F, AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
int size=CloverTerm.Grid()->oSites();
accelerator_for(i, size,1,
{
T_v[i]()(0, 1) = timesI(F_v[i]()());
T_v[i]()(1, 0) = timesI(F_v[i]()());
@@ -338,7 +342,8 @@ protected:
autoView( T_v ,T,AcceleratorWrite);
autoView( F_v ,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
int size=CloverTerm.Grid()->oSites();
accelerator_for(i, size,1,
{
T_v[i]()(0, 1) = -(F_v[i]()());
T_v[i]()(1, 0) = (F_v[i]()());
@@ -357,7 +362,8 @@ protected:
autoView( T_v , T,AcceleratorWrite);
autoView( F_v , F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
int size=CloverTerm.Grid()->oSites();
accelerator_for(i, size,1,
{
T_v[i]()(0, 0) = timesI(F_v[i]()());
T_v[i]()(1, 1) = timesMinusI(F_v[i]()());

10
Grid/qcd/action/fermion/WilsonImpl.h
View File

@@ -110,8 +110,10 @@ public:
autoView( out_v, out, AcceleratorWrite);
autoView( phi_v, phi, AcceleratorRead);
autoView( Umu_v, Umu, AcceleratorRead);
int size=out.Grid()->oSites();
typedef decltype(coalescedRead(out_v[0])) calcSpinor;
accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
accelerator_for(sss,size,Nsimd,{
calcSpinor tmp;
multLink(tmp,Umu_v[sss],phi_v(sss),mu);
coalescedWrite(out_v[sss],tmp);
@@ -203,7 +205,8 @@ public:
autoView( tmp_v , tmp, AcceleratorWrite);
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
accelerator_for(sss,tmp.Grid()->oSites(),1,{
int size=tmp.Grid()->oSites();
accelerator_for(sss,size,1,{
int sU=sss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;
@@ -217,7 +220,8 @@ public:
const int Nsimd = SiteSpinor::Nsimd();
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
accelerator_for(sss,mat.Grid()->oSites(),Nsimd,{
int size=mat.Grid()->oSites();
accelerator_for(sss,size,Nsimd,{
int sU=sss;
typedef decltype(coalescedRead(mat_v[sU](mu)() )) ColorMatrixType;
ColorMatrixType sum;

6
Grid/qcd/action/gauge/GaugeImplTypes.h
View File

@@ -88,7 +88,8 @@ public:
static inline void AddLink(Field &U, LinkField &W, int mu) { // U[mu] += W
autoView(U_v,U,AcceleratorWrite);
autoView(W_v,W,AcceleratorRead);
accelerator_for( ss, U.Grid()->oSites(), 1, {
int size=U.Grid()->oSites();
accelerator_for( ss, size, 1, {
U_v[ss](mu) = U_v[ss](mu) + W_v[ss]();
});
}
@@ -133,7 +134,8 @@ public:
//auto start = std::chrono::high_resolution_clock::now();
autoView(U_v,U,AcceleratorWrite);
autoView(P_v,P,AcceleratorRead);
accelerator_for(ss, P.Grid()->oSites(),1,{
int size=P.Grid()->oSites();
accelerator_for(ss, size,1,{
for (int mu = 0; mu < Nd; mu++) {
U_v[ss](mu) = ProjectOnGroup(Exponentiate(P_v[ss](mu), ep, Nexp) * U_v[ss](mu));
}

34
Grid/threads/Accelerator.h
View File

@@ -28,6 +28,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
#pragma once
//#ifndef ACCELERATOR_H
//#define ACCELERATOR_H
#include <string.h>
#ifdef HAVE_MALLOC_MALLOC_H
@@ -434,30 +437,33 @@ inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(bas
//OpenMP Target Offloading
#ifdef OMPTARGET
uint32_t nt=acceleratorThreads();
//uint32_t gpu_threads=acceleratorThreads();
#define accelerator
#define accelerator_inline strong_inline
#define accelerator_for(iterator,num,nsimd, ... ) \
DO_PRAGMA(omp target teams distribute parallel for thread_limit(nt)) \
#define accelerator_for(i,num,nsimd, ... ) \
_Pragma("omp target teams distribute parallel for") \
for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
#define accelerator_forNB(iterator,num,nsimd, ... ) \
DO_PRAGMA(omp target teams distribute parallel for thread_limit(nt) nowait) \
#define accelerator_forNB(i,num,nsimd, ... ) \
_Pragma("omp target teams distribute parallel for nowait") \
for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
#define accelerator_barrier(dummy) DO_PRAGMA(omp barrier)
#define accelerator_barrier(dummy) _Pragma("omp barrier")
#define accelerator_for2d(iter1, num1, iter2, num2, nsimd, ... ) \
DO_PRAGMA(omp target teams distribute parallel for thread_limit(nt) collapse(2)) \
for ( uint64_t i=0;i<num;i++) { __VA_ARGS__ } ;
_Pragma("omp target teams distribute parallel for collapse(2)") \
for ( uint64_t iter1=0;iter1<num1;iter1++) \
for ( uint64_t iter2=0;iter2<num2;iter2++) { __VA_ARGS__ } ;
accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { DO_PRAGMA(omp target enter data map(to:from[0:bytes])}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ DO_PRAGMA(omp target exit data map(from:from[0:bytes])}
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) {;}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){;}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { memcpy(to,from,bytes);}
inline void acceleratorCopySynchronize(void) {};
inline void acceleratorCopySynchronize(void) {;};
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
inline void acceleratorMemSet(void *base,int value,size_t bytes) { memset(base,value,bytes);}
#ifdef OMPTARGET_MANAGED
#include <cuda_runtime_api.h>
inline void *acceleratorAllocShared(size_t bytes)
{
void *ptr=NULL;
@@ -469,11 +475,14 @@ inline void *acceleratorAllocShared(size_t bytes)
return ptr;
};
inline void acceleratorFreeShared(void *ptr){cudaFree(ptr);};
inline void *acceleratorAllocDevice(size_t bytes){return memalign(GRID_ALLOC_ALIGN,bytes);};
inline void acceleratorFreeDevice(void *ptr){free(ptr);};
#else
inline void *acceleratorAllocShared(size_t bytes){return memalign(GRID_ALLOC_ALIGN,bytes);};
inline void *acceleratorAllocDevice(size_t bytes){return memalign(GRID_ALLOC_ALIGN,bytes);};
inline void acceleratorFreeShared(void *ptr){free(ptr);};
inline void acceleratorFreeDevice(void *ptr){free(ptr);};
#endif
//OpenMP CPU threads
#else
@@ -504,6 +513,7 @@ inline void *acceleratorAllocDevice(size_t bytes){return memalign(GRID_ALLOC_ALI
inline void acceleratorFreeShared(void *ptr){free(ptr);};
inline void acceleratorFreeDevice(void *ptr){free(ptr);};
#endif
#endif
#endif // CPU target
@@ -567,3 +577,5 @@ accelerator_inline void acceleratorFence(void)
}
NAMESPACE_END(Grid);
//#endif

2
Grid/threads/Threads.h
View File

@@ -46,7 +46,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#endif
#ifdef GRID_OMP
#define DO_PRAGMA_(x) _Pragma (#x)
#define DO_PRAGMA_(x) _Pragma ("x")
#define DO_PRAGMA(x) DO_PRAGMA_(x)
#define thread_num(a) omp_get_thread_num()
#define thread_max(a) omp_get_max_threads()