portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Clean up but no major changes

Browse Source
This commit is contained in:
Peter Boyle 2015-04-03 22:54:13 +01:00
parent d081715504
commit d5eee231e0
10 changed files with 74 additions and 205 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
Grid_Cartesian.h
View File

@ -201,10 +201,8 @@ public:
block = block*_rdimensions[d];
}
if ( _isites != vComplex::Nsimd()) {
printf("bad layout for grid isites %d Nsimd %d\n",_isites,vComplex::Nsimd());
exit(0);
}
assert( _isites == vComplex::Nsimd());
};
};
@ -235,10 +233,8 @@ public:
int ocb=CheckerBoardFromOsite(osite);
if ( (source_cb+ocb)&1 ) {
printf("Checkerboard shift %d\n",(shift)/2);
return (shift)/2;
} else {
printf("Checkerboard shift %d\n",(shift+1)/2);
return (shift+1)/2;
}
}
@ -314,10 +310,7 @@ public:
block = block*_rdimensions[d];
}
if ( _isites != vComplex::Nsimd()) {
printf("bad layout for grid isites %d Nsimd %d\n",_isites,vComplex::Nsimd());
exit(0);
}
assert ( _isites == vComplex::Nsimd());
};
protected:
virtual int oIndex(std::vector<int> &coor)

70
Grid_Lattice.h
View File

@ -30,9 +30,7 @@ public:
Lattice(SimdGrid *grid) : _grid(grid) {
_odata.reserve(_grid->oSites());
if ( ((uint64_t)&_odata[0])&0xF) {
exit(-1);
}
assert((((uint64_t)&_odata[0])&0xF) ==0);
checkerboard=0;
}
@ -97,23 +95,22 @@ public:
template<class sobj>
friend void pokeSite(const sobj &s,Lattice<vobj> &l,std::vector<int> &site){
if ( l.checkerboard != l._grid->CheckerBoard(site)){
printf("Poking wrong checkerboard\n");
exit(EXIT_FAILURE);
}
typedef typename vobj::scalar_type stype;
typedef typename vobj::vector_type vtype;
assert( l.checkerboard == l._grid->CheckerBoard(site));
int o_index = l._grid->oIndex(site);
int i_index = l._grid->iIndex(site);
// BUGGY. This assumes complex real
Real *v_ptr = (Real *)&l._odata[o_index];
Real *s_ptr = (Real *)&s;
stype *v_ptr = (stype *)&l._odata[o_index];
stype *s_ptr = (stype *)&s;
v_ptr = v_ptr + 2*i_index;
for(int i=0;i<sizeof(sobj);i+=2*sizeof(Real)){
for(int i=0;i<sizeof(sobj);i+=2*sizeof(stype)){
v_ptr[0] = s_ptr[0];
v_ptr[1] = s_ptr[1];
v_ptr+=2*vComplex::Nsimd();
v_ptr+=2*vtype::Nsimd();
s_ptr+=2;
}
return;
@ -124,22 +121,22 @@ public:
template<class sobj>
friend void peekSite(sobj &s,const Lattice<vobj> &l,std::vector<int> &site){
// FIXME : define exceptions set and throw up.
if ( l.checkerboard != l._grid->CheckerBoard(site)){
printf("Peeking wrong checkerboard\n");
exit(EXIT_FAILURE);
}
typedef typename vobj::scalar_type stype;
typedef typename vobj::vector_type vtype;
assert( l.checkerboard== l._grid->CheckerBoard(site));
int o_index = l._grid->oIndex(site);
int i_index = l._grid->iIndex(site);
Real *v_ptr = (Real *)&l._odata[o_index];
Real *s_ptr = (Real *)&s;
stype *v_ptr = (stype *)&l._odata[o_index];
stype *s_ptr = (stype *)&s;
v_ptr = v_ptr + 2*i_index;
for(int i=0;i<sizeof(sobj);i+=2*sizeof(Real)){
for(int i=0;i<sizeof(sobj);i+=2*sizeof(stype)){
s_ptr[0] = v_ptr[0];
s_ptr[1] = v_ptr[1];
v_ptr+=2*vComplex::Nsimd();
v_ptr+=2*vtype::Nsimd();
s_ptr+=2;
}
return;
@ -291,15 +288,6 @@ public:
}
}; // class Lattice
/* Need to implement the multiplication return type matching S S -> S, S M -> M, M S -> M through
all nested possibilities.
template<template<class> class lhs,template<class> class rhs>
class MultTypeSelector {
template<typename vtype> using ltype = lhs
typedef lhs type;
};
*/
template<class obj1,class obj2>
void conformable(const Lattice<obj1> &lhs,const Lattice<obj2> &rhs)
{
@ -313,28 +301,6 @@ public:
uint32_t vec_len = lhs._grid->oSites();
#pragma omp parallel for
for(int ss=0;ss<vec_len;ss++){
const char * ptr =(const char*)&lhs._odata[ss];
#ifdef PREFETCH
v_prefetch0(sizeof(obj2), ptr);
#endif
for(int i=0;i<sizeof(obj2);i+=64){
_mm_prefetch(ptr+i+4096,_MM_HINT_T1);
_mm_prefetch(ptr+i+256,_MM_HINT_T0);
}
ptr =(const char*)&rhs._odata[ss];
#ifdef PREFETCH
v_prefetch0(sizeof(obj3), ptr);
#endif
for(int i=0;i<sizeof(obj3);i+=64){
_mm_prefetch(ptr+i+4096,_MM_HINT_T1);
_mm_prefetch(ptr+i+256,_MM_HINT_T0);
}
mult(&ret._odata[ss],&lhs._odata[ss],&rhs._odata[ss]);
}
}

27
Grid_cshift_common.h
View File

@ -1,5 +1,17 @@
#ifndef _GRID_CSHIFT_COMMON_H_
#define _GRID_CSHIFT_COMMON_H_
//////////////////////////////////////////////////////////////////////////////////////////
// Must not lose sight that goal is to be able to construct really efficient
// gather to a point stencil code. CSHIFT is not the best way, so probably need
// additional stencil support.
//
// Stencil based code could pre-exchange haloes and use a table lookup for neighbours
//
// Lattice <foo> could also allocate haloes which get used for stencil code.
//
// Grid could create a neighbour index table for a given stencil.
// Could also implement CovariantCshift.
//////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////////////////////
// Gather for when there is no need to SIMD split
@ -8,8 +20,6 @@ friend void Gather_plane_simple (Lattice<vobj> &rhs,std::vector<vobj,alignedAllo
{
int rd = rhs._grid->_rdimensions[dimension];
// printf("Gather plane _simple mask %d\n",cbmask);
if ( !rhs._grid->CheckerBoarded(dimension) ) {
int so = plane*rhs._grid->_ostride[dimension]; // base offset for start of plane
@ -31,7 +41,6 @@ friend void Gather_plane_simple (Lattice<vobj> &rhs,std::vector<vobj,alignedAllo
int o = 0; // relative offset to base within plane
int bo = 0; // offset in buffer
// int jjj=0;
#pragma omp parallel for collapse(2)
for(int n=0;n<rhs._grid->_slice_nblock[dimension];n++){
for(int b=0;b<rhs._grid->_slice_block[dimension];b++){
@ -39,11 +48,6 @@ friend void Gather_plane_simple (Lattice<vobj> &rhs,std::vector<vobj,alignedAllo
int ocb=1<<rhs._grid->CheckerBoardFromOsite(o+b);// Could easily be a table lookup
if ( ocb &cbmask ) {
buffer[bo]=rhs._odata[so+o+b];
// float * ptr = (float *)& rhs._odata[so+o+b];
// if( (cbmask!=3)&&(jjj<8)){
// printf("Gather_plane_simple %d %le bo %d\n",so+o+b,*ptr,bo);
// jjj++;
// }
bo++;
}
@ -215,7 +219,7 @@ friend void Copy_plane(Lattice<vobj>& lhs,Lattice<vobj> &rhs, int dimension,int
int ro = rplane*rhs._grid->_ostride[dimension]; // base offset for start of plane
int lo = lplane*lhs._grid->_ostride[dimension]; // base offset for start of plane
int o = 0; // relative offset to base within plane
// int jjj=0;
#pragma omp parallel for collapse(2)
for(int n=0;n<rhs._grid->_slice_nblock[dimension];n++){
for(int b=0;b<rhs._grid->_slice_block[dimension];b++){
@ -224,11 +228,6 @@ friend void Copy_plane(Lattice<vobj>& lhs,Lattice<vobj> &rhs, int dimension,int
if ( ocb&cbmask ) {
lhs._odata[lo+o+b]=rhs._odata[ro+o+b];
// float *ptr =(float *) &rhs._odata[ro+o+b];
// if((cbmask!=0x3)&&jjj<8) {
// printf("Copy_plane %d %le n,b=%d,%d mask %d ocb %d\n",ro+o+b,*ptr,n,b,cbmask,ocb);
// jjj++;
// }
}
}

82
Grid_cshift_mpi.h
View File

@ -1,36 +1,10 @@
#ifndef _GRID_MPI_CSHIFT_H_
#define _GRID_MPI_CSHIFT_H_
#ifndef MAX
#define MAX(x,y) ((x)>(y)?(x):(y))
#define MIN(x,y) ((x)>(y)?(y):(x))
//////////////////////////////////////////////////////////////////////////////////////////
// Must not lose sight that goal is to be able to construct really efficient
// gather to a point stencil code. CSHIFT is not the best way, so probably need
// additional stencil support.
//
// Could still do a templated syntax tree and make CSHIFT return lattice vector.
//
// Stencil based code could pre-exchange haloes and use a table lookup for neighbours
//
// Lattice <foo> could also allocate haloes which get used for stencil code.
//
// Grid could create a neighbour index table for a given stencil.
// Could also implement CovariantCshift.
//////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////
// Q. Further split this into separate sub functions?
/////////////////////////////////////////////////////////////
// CshiftCB_local
// CshiftCB_local_permute
// Cshift_comms_splice
// Cshift_comms
// Cshift_local
// Cshift_local_permute
#endif
friend Lattice<vobj> Cshift(Lattice<vobj> &rhs,int dimension,int shift)
{
@ -71,16 +45,10 @@ friend void Cshift_comms(Lattice<vobj>& ret,Lattice<vobj> &rhs,int dimension,int
sshift[1] = rhs._grid->CheckerBoardShift(rhs.checkerboard,dimension,shift,1);
if ( sshift[0] == sshift[1] ) {
// printf("Cshift_comms : single pass\n");
Cshift_comms(ret,rhs,dimension,shift,0x3);
} else {
// printf("Cshift_comms : two pass\n");
// printf("call1\n");
Cshift_comms(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
// printf("call2\n");
Cshift_comms(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
// printf("done\n");
}
}
@ -94,11 +62,8 @@ friend void Cshift_comms_simd(Lattice<vobj>& ret,Lattice<vobj> &rhs,int dimensio
if ( sshift[0] == sshift[1] ) {
Cshift_comms_simd(ret,rhs,dimension,shift,0x3);
} else {
// printf("call1 0x1 cb=even\n");
Cshift_comms_simd(ret,rhs,dimension,shift,0x1);// if checkerboard is unfavourable take two passes
// printf("call2 0x2 cb=odd\n");
Cshift_comms_simd(ret,rhs,dimension,shift,0x2);// both with block stride loop iteration
// printf("done\n");
}
}
@ -120,13 +85,10 @@ friend void Cshift_comms(Lattice<vobj> &ret,Lattice<vobj> &rhs,int dimension,int
assert(shift>=0);
assert(shift<fd);
// Packed gather sequence is clean
int buffer_size = rhs._grid->_slice_nblock[dimension]*rhs._grid->_slice_block[dimension];
std::vector<vobj,alignedAllocator<vobj> > send_buf(buffer_size);
std::vector<vobj,alignedAllocator<vobj> > recv_buf(buffer_size);
// This code could be simplified by multiple calls to single routine with extra params to
// encapsulate the difference in the code paths.
int cb= (cbmask==0x2)? 1 : 0;
int sshift= rhs._grid->CheckerBoardShift(rhs.checkerboard,dimension,shift,cb);
@ -137,8 +99,9 @@ friend void Cshift_comms(Lattice<vobj> &ret,Lattice<vobj> &rhs,int dimension,int
int comm_proc = (x+sshift)/rd;
if (!offnode) {
// printf("local x %d sshift %d offnode %d rd %d cb %d\n",x,sshift,offnode,rd,cb);
Copy_plane(ret,rhs,dimension,x,sx,cbmask);
} else {
int words = send_buf.size();
@ -146,29 +109,13 @@ friend void Cshift_comms(Lattice<vobj> &ret,Lattice<vobj> &rhs,int dimension,int
int bytes = words * sizeof(vobj);
// printf("nonlocal x %d sx %d sshift %d offnode %d rd %d cb %d cbmask %d rhscb %d comm_proc %d\n",
// x,sx,sshift,offnode,rd,cb,cbmask,rhs.checkerboard,comm_proc);
// Copy_plane(temp,rhs,dimension,x,sx,cbmask);
// Bug found; cbmask may differ between sx plan and rx plane.
Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask);
// for(int i=0;i<MIN(words,8);i++){
// float *ptr = (float *)&send_buf[i];
// printf("send buf shift %d cbmask %d i %d %le\n",sshift,cbmask,i,*ptr);
// }
// Gather_plane_simple (rhs,send_buf,dimension,sx,cbmask^0x3);
// for(int i=0;i<MIN(words,8);i++){
// float *ptr = (float *)&send_buf[i];
// printf("send buf shift %d cbmask %d i %d %le\n",sshift,cbmask,i,*ptr);
// }
// recv_buf=send_buf;
int rank = grid->_processor;
int recv_from_rank;
int xmit_to_rank;
grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
// printf("bytes %d node %d sending to %d receiving from %d\n",bytes,rank,xmit_to_rank,recv_from_rank );
grid->SendToRecvFrom((void *)&send_buf[0],
xmit_to_rank,
(void *)&recv_buf[0],
@ -224,46 +171,29 @@ friend void Cshift_comms_simd(Lattice<vobj> &ret,Lattice<vobj> &rhs,int dimensi
int cb = (cbmask==0x2)? 1 : 0;
int sshift= grid->CheckerBoardShift(rhs.checkerboard,dimension,shift,cb);
// printf("cshift-comms-simd: shift = %d ; sshift = %d ; cbmask %d ; simd_layout %d\n",shift,sshift,cbmask,simd_layout);
std::vector<int> comm_offnode(simd_layout);
std::vector<int> comm_proc (simd_layout); //relative processor coord in dim=dimension
// Strategy
//
//* Loop over source planes
//* if any communication needed extract and send
//* if communication needed extract and send
for(int x=0;x<rd;x++){
int comm_any = 0;
for(int s=0;s<simd_layout;s++) {
// does shift to "neighbour" takes us off node?
// coordinates (reduce plane, simd_lane) of neighbour?
// how many nodes away is this shift?
// where we should send to?
// where we should receive from?
int shifted_x = x+s*rd+sshift;
comm_offnode[s] = shifted_x >= ld;
comm_any = comm_any | comm_offnode[s];
comm_proc[s] = shifted_x/ld;
// printf("rd %d x %d shifted %d s=%d comm_any %d\n",rd, x,shifted_x,s,comm_any);
}
int o = 0;
int bo = x*grid->_ostride[dimension];
int sx = (x+sshift)%rd;
// Need Convenience function in _grid. Move this in
if ( comm_any ) {
for(int i=0;i<Nsimd;i++){
pointers[i] = (scalar_type *)&send_buf_extract[i][0];
}
Gather_plane_extract(rhs,pointers,dimension,sx,cbmask);
// for(int i=0;i<Nsimd;i++){
// printf("extracted %d %le\n",i,real(send_buf_extract[i][0]));
// }
for(int i=0;i<Nsimd;i++){
@ -283,14 +213,11 @@ friend void Cshift_comms_simd(Lattice<vobj> &ret,Lattice<vobj> &rhs,int dimensi
recv_from_rank,
bytes);
// printf("Cshift_simd comms %d %le %le\n",i,real(recv_buf_extract[i][0]),real(send_buf_extract[i][0]));
rpointers[i] = (scalar_type *)&recv_buf_extract[i][0];
} else {
rpointers[i] = (scalar_type *)&send_buf_extract[i][0];
// printf("Cshift_simd local %d %le \n",i,real(send_buf_extract[i][0]));
}
@ -311,7 +238,6 @@ friend void Cshift_comms_simd(Lattice<vobj> &ret,Lattice<vobj> &rhs,int dimensi
} else {
pointers[i] = rpointers[i];
}
// printf("Cshift_simd perm %d num %d wrap %d swiz %d %le unswiz %le\n",permute_slice,num,wrap,i,real(pointers[i][0]),real(rpointers[i][0]));
}
Scatter_plane_merge(ret,pointers,dimension,x,cbmask);

2
Grid_math_types.h
View File

@ -830,7 +830,7 @@ template<class vtype,int N> inline iMatrix<vtype,N> adj(const iMatrix<vtype,N> &
/////////////////////////////////////////////////////////////////
// Can only take the real/imag part of scalar objects, since
// lattice objects of different complexity are non-conformable.
// lattice objects of different complex nature are non-conformable.
/////////////////////////////////////////////////////////////////
template<class itype> inline auto real(const iScalar<itype> &z) -> iScalar<decltype(real(z._internal))>
{

22
Grid_mpi.cc
View File

@ -17,13 +17,15 @@ CartesianCommunicator::CartesianCommunicator(std::vector<int> &processors)
MPI_Cart_create(MPI_COMM_WORLD, _ndimension,&_processors[0],&periodic[0],1,&communicator);
MPI_Comm_rank(communicator,&_processor);
MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
printf("Hello world from processor [");
for(int i=0;i<_ndimension;i++){
printf("%d ",_processor_coor[i]);
_Nprocessors*=_processors[i];
}
printf("]\n");
fflush(stdout);
int Size;
MPI_Comm_size(communicator,&Size);
assert(Size==_Nprocessors);
}
void CartesianCommunicator::GlobalSumF(float &f){
@ -71,15 +73,3 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
}
#if 0
// Could possibly do a direct block strided send?
int MPI_Type_vector(
int count,
int blocklength,
int stride,
MPI_Datatype old_type,
MPI_Datatype *newtype_p
);
#endif

8
Grid_vComplexD.h
View File

@ -48,7 +48,6 @@ namespace Grid {
#endif
#ifdef AVX512
ret.v = _mm512_add_pd(a.v,b.v);
//printf("%s %f\n",__func__,_mm512_reduce_mul_pd(ret.v));
#endif
#ifdef QPX
ret.v = vec_add(a.v,b.v);
@ -210,7 +209,7 @@ namespace Grid {
#ifdef QPX
#error // Not implemented yet
#endif
default: exit(EXIT_FAILURE); break;
default: assert(0); break;
}
};
void vload(zvec& a){
@ -265,8 +264,7 @@ friend inline void vstore(vComplexD &ret, ComplexD *a){
//Note v has a3 a2 a1 a0
#endif
#ifdef QPX
printf("%s Not implemented\n",__func__);
exit(-1);
assert(0);
#endif
}
friend inline void vprefetch(const vComplexD &v)
@ -294,7 +292,7 @@ friend inline void vstore(vComplexD &ret, ComplexD *a){
#endif
#ifdef QPX
exit(0); // not implemented
assert(0);
#endif
return ret;
}

7
Grid_vComplexF.h
View File

@ -193,7 +193,7 @@ namespace Grid {
#ifdef QPX
#error
#endif
default: exit(EXIT_FAILURE); break;
default: assert(0); break;
}
};
@ -235,8 +235,7 @@ friend inline void vstore(vComplexF &ret, ComplexF *a){
//Note v has a3 a2 a1 a0
#endif
#ifdef QPX
printf("%s Not implemented\n",__func__);
exit(-1);
assert(0);
#endif
}
friend inline void vprefetch(const vComplexF &v)
@ -333,7 +332,7 @@ exit(-1);
ret.v = _mm512_mask_sub_ps(in.v,0xaaaa,ret.v,in.v); // Zero out 0+real 0-imag
#endif
#ifdef QPX
exit(0); // not implemented
assert(0);
#endif
return ret;
}

5
Grid_vRealD.h
View File

@ -161,7 +161,7 @@ namespace Grid {
#ifdef QPX
#error
#endif
default: exit(EXIT_FAILURE); break;
default: assert(0);break;
}
};
// gona be bye bye
@ -214,8 +214,7 @@ namespace Grid {
// Note v has a7 a6 a5ba4 a3 a2 a1 a0
#endif
#ifdef QPX
printf("%s Not implemented\n",__func__);
exit(-1);
assert(0);
#endif
}
friend inline void vprefetch(const vRealD &v)

5
Grid_vRealF.h
View File

@ -185,7 +185,7 @@ namespace Grid {
#ifdef QPX
#error not implemented
#endif
default: exit(EXIT_FAILURE); break;
default: assert(0); break;
}
};
@ -236,8 +236,7 @@ friend inline void vstore(vRealF &ret, float *a){
// Note v has a7 a6 a5ba4 a3 a2 a1 a0
#endif
#ifdef QPX
printf("%s Not implemented\n",__func__);
exit(-1);
assert(0);
#endif
}