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

Massive clean up extract merge.

Browse Source 
Simpler and GPU friendly
This commit is contained in:
paboyle 2018-02-24 22:21:08 +00:00
parent 36ea5f6b77
commit 7574c18cef
1 changed files with 103 additions and 246 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

349
lib/tensors/Tensor_extract_merge.h
View File

@ -29,260 +29,117 @@ Author: Christopher Kelly <ckelly@phys.columbia.edu>
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_EXTRACT_H
#define GRID_EXTRACT_H
#pragma once
NAMESPACE_BEGIN(Grid);
/////////////////////////////////////////////////////////////////
// Generic extract/merge/permute
/////////////////////////////////////////////////////////////////
namespace Grid{
template<class __T> using ExtractPointerArray = AcceleratorVector<__T *,GRID_MAX_SIMD>;
template<class __T> using ExtractBuffer = AcceleratorVector<__T ,GRID_MAX_SIMD>;
////////////////////////////////////////////////////////////////////////////////////////////////
// Extract/merge a fundamental vector type, to pointer array with offset
////////////////////////////////////////////////////////////////////////////////////////////////
template<class vsimd,class scalar>
strong_inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const vsimd >::type * y,
std::vector<scalar *> &extracted,int offset){
// FIXME: bounce off memory is painful
static const int Nsimd=sizeof(vsimd)/sizeof(scalar);
int Nextr=extracted.size();
int s=Nsimd/Nextr;
//void extract(const vobj &vec,ExtractBuffer<typename vobj::scalar_object> &extracted);
//void extract(const vobj &vec,ExtractPointerArray<sobj> &extracted, int offset);
//void merge(vobj &vec,ExtractBuffer<typename vobj::scalar_object> &extracted)
//void merge(vobj &vec,ExtractPointerArray<typename vobj::scalar_object> &extracted)
//
scalar*buf = (scalar *)y;
////////////////////////////////////////////////////////////////////////
// Extract to contiguous array scalar object
////////////////////////////////////////////////////////////////////////
template<class vobj,class sobj> accelerator_inline
void extract(const vobj &vec,ExtractBuffer<sobj> &extracted)
{
typedef typename GridTypeMapper<vobj>::scalar_type scalar_type;
typedef typename GridTypeMapper<vobj>::vector_type vector_type;
const int words=sizeof(vobj)/sizeof(vector_type);
const int Nsimd=vector_type::Nsimd();
const int Nextr=extracted.size();
const int s=Nsimd/Nextr;
scalar_type *sp = (scalar_type *)&extracted[0];
scalar_type *vp = (scalar_type *)&vec;
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
extracted[i][offset] = buf[i*s];
}
};
////////////////////////////////////////////////////////////////////////
// Merge simd vector from array of scalars to pointer array with offset
////////////////////////////////////////////////////////////////////////
template<class vsimd,class scalar>
strong_inline void merge(typename std::enable_if<!isGridTensor<vsimd>::value, vsimd >::type * y,
std::vector<scalar *> &extracted,int offset){
static const int Nsimd=sizeof(vsimd)/sizeof(scalar);
int Nextr=extracted.size();
int s=Nsimd/Nextr; // can have sparse occupation of simd vector if simd_layout does not fill it
// replicate n-fold. Use to allow Integer masks to
// predicate floating point of various width assignments and maintain conformable.
scalar *buf =(scalar *) y;
for(int i=0;i<Nextr;i++){
for(int ii=0;ii<s;ii++){
buf[i*s+ii]=extracted[i][offset];
}
}
};
////////////////////////////////////////////////////////////////////////////////////////////////
// Extract a fundamental vector type to scalar array
////////////////////////////////////////////////////////////////////////////////////////////////
template<class vsimd,class scalar>
strong_inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const vsimd >::type &y,std::vector<scalar> &extracted){
int Nextr=extracted.size();
int Nsimd=vsimd::Nsimd();
int s=Nsimd/Nextr;
scalar *buf = (scalar *)&y;
for(int i=0;i<Nextr;i++){
extracted[i]=buf[i*s];
#ifdef PARANOID
for(int ii=1;ii<s;ii++){
if ( buf[i*s]!=buf[i*s+ii] ){
std::cout<<GridLogMessage << " SIMD extract failure splat = "<<s<<" ii "<<ii<<" " <<Nextr<<" "<< Nsimd<<" "<<std::endl;
for(int vv=0;vv<Nsimd;vv++) {
std::cout<<GridLogMessage<< buf[vv]<<" ";
}
std::cout<<GridLogMessage<<std::endl;
assert(0);
}
assert(buf[i*s]==buf[i*s+ii]);
}
#endif
}
};
////////////////////////////////////////////////////////////////////////
// Merge simd vector from array of scalars
////////////////////////////////////////////////////////////////////////
template<class vsimd,class scalar>
strong_inline void merge(typename std::enable_if<!isGridTensor<vsimd>::value, vsimd >::type &y,std::vector<scalar> &extracted){
int Nextr=extracted.size();
static const int Nsimd=vsimd::Nsimd();
int s=Nsimd/Nextr;
scalar *buf = (scalar *)&y;
for(int i=0;i<Nextr;i++){
for(int ii=0;ii<s;ii++){
buf[i*s+ii]=extracted[i]; // replicates value
}
}
};
////////////////////////////////////////////////////////////////////////
// Extract to contiguous array scalar object
////////////////////////////////////////////////////////////////////////
template<class vobj> strong_inline void extract(const vobj &vec,std::vector<typename vobj::scalar_object> &extracted)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
static const int words=sizeof(vobj)/sizeof(vector_type);
int Nextr=extracted.size();
std::vector<scalar_type *> pointers(Nextr);
for(int i=0;i<Nextr;i++)
pointers[i] =(scalar_type *)& extracted[i];
vector_type *vp = (vector_type *)&vec;
for(int w=0;w<words;w++){
extract<vector_type,scalar_type>(&vp[w],pointers,w);
sp[i*words+w] = vp[w*Nsimd+i*s] ;
}
}
////////////////////////////////////////////////////////////////////////
// Extract to a bunch of scalar object pointers, with offset
////////////////////////////////////////////////////////////////////////
template<class vobj> strong_inline
void extract(const vobj &vec,std::vector<typename vobj::scalar_object *> &extracted, int offset)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
static const int words=sizeof(vobj)/sizeof(vector_type);
static const int Nsimd=vobj::vector_type::Nsimd();
int Nextr=extracted.size();
int s = Nsimd/Nextr;
scalar_type * vp = (scalar_type *)&vec;
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
scalar_type * pointer = (scalar_type *)& extracted[i][offset];
pointer[w] = vp[i*s+w*Nsimd];
}
}
}
////////////////////////////////////////////////////////////////////////
// Extract to a bunch of scalar object pointers of different scalar type, with offset. Useful for precision change
////////////////////////////////////////////////////////////////////////
template<class vobj, class sobj> strong_inline
void extract1(const vobj &vec,std::vector<sobj*> &extracted, int offset)
{
typedef typename vobj::scalar_type vobj_scalar_type ;
typedef typename vobj::vector_type vobj_vector_type ;
typedef typename sobj::scalar_type sobj_scalar_type ;
static const int words=sizeof(vobj)/sizeof(vobj_vector_type);
static const int Nsimd=vobj_vector_type::Nsimd();
int Nextr=extracted.size();
int s = Nsimd/Nextr;
vobj_scalar_type * vp = (vobj_scalar_type *)&vec;
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
sobj_scalar_type * pointer = (sobj_scalar_type *)& extracted[i][offset];
pointer[w] = vp[i*s+w*Nsimd];
}
}
}
////////////////////////////////////////////////////////////////////////
// Merge a contiguous array of scalar objects
////////////////////////////////////////////////////////////////////////
template<class vobj> strong_inline
void merge(vobj &vec,std::vector<typename vobj::scalar_object> &extracted)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
static const int words=sizeof(vobj)/sizeof(vector_type);
int Nextr = extracted.size();
std::vector<scalar_type *> pointers(Nextr);
for(int i=0;i<Nextr;i++)
pointers[i] =(scalar_type *)& extracted[i];
vector_type *vp = (vector_type *)&vec;
for(int w=0;w<words;w++){
merge<vector_type,scalar_type>(&vp[w],pointers,w);
}
}
////////////////////////////////////////////////////////////////////////
// Merge a bunch of different scalar object pointers, with offset
////////////////////////////////////////////////////////////////////////
template<class vobj> strong_inline
void merge(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int offset)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=sizeof(vector_type)/sizeof(scalar_type);
const int words=sizeof(vobj)/sizeof(vector_type);
int Nextr=extracted.size();
int s=Nsimd/Nextr;
scalar_type *pointer;
scalar_type *vp = (scalar_type *)&vec;
// assert( (((uint64_t)vp)&(sizeof(scalar_type)-1)) == 0);
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
for(int ii=0;ii<s;ii++){
pointer=(scalar_type *)&extracted[i][offset];
vp[w*Nsimd+i*s+ii] = pointer[w];
}
}
}
}
template<class vobj> strong_inline void merge1(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int offset)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
static const int Nsimd=vobj::vector_type::Nsimd();
static const int words=sizeof(vobj)/sizeof(vector_type);
scalar_type *vp = (scalar_type *)&vec;
// assert( (((uint64_t)vp)&(sizeof(scalar_type)-1)) == 0);
for(int w=0;w<words;w++){
for(int i=0;i<Nsimd;i++){
vp[w*Nsimd+i] = ((scalar_type *)&extracted[i][offset])[w];
}}
}
template<class vobj> strong_inline void merge2(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int offset)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=vobj::vector_type::Nsimd();
const int words=sizeof(vobj)/sizeof(vector_type);
scalar_type *pointer;
scalar_type *vp = (scalar_type *)&vec;
// assert( (((uint64_t)vp)&(sizeof(scalar_type)-1)) == 0);
for(int w=0;w<words;w++){
for(int i=0;i<Nsimd;i++){
pointer=(scalar_type *)&extracted[i][offset];
vp[w*Nsimd+i] =pointer[w];
}
}
}
}
#endif
////////////////////////////////////////////////////////////////////////
// Merge a contiguous array of scalar objects
////////////////////////////////////////////////////////////////////////
template<class vobj,class sobj> accelerator_inline
void merge(vobj &vec,ExtractBuffer<sobj> &extracted)
{
typedef typename GridTypeMapper<vobj>::scalar_type scalar_type;
typedef typename GridTypeMapper<vobj>::vector_type vector_type;
const int words=sizeof(vobj)/sizeof(vector_type);
const int Nsimd=vector_type::Nsimd();
const int Nextr = extracted.size();
const int s=Nsimd/Nextr;
scalar_type *sp = (scalar_type *)&extracted[0];
scalar_type *vp = (scalar_type *)&vec;
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
vp[w*Nsimd+i*s] = sp[i*words+w] ;
}
}
}
////////////////////////////////////////////////////////////////////////
// Extract to a bunch of scalar object pointers of different scalar type, with offset. Useful for precision change
////////////////////////////////////////////////////////////////////////
template<class vobj, class sobj> accelerator_inline
void extract(const vobj &vec,ExtractPointerArray<sobj> &extracted, int offset)
{
typedef typename GridTypeMapper<vobj>::scalar_type scalar_type;
typedef typename GridTypeMapper<vobj>::vector_type vector_type;
const int words=sizeof(vobj)/sizeof(vector_type);
const int Nsimd=vector_type::Nsimd();
const int Nextr=extracted.size();
const int s = Nsimd/Nextr;
scalar_type * vp = (scalar_type *)&vec;
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
scalar_type * pointer = (scalar_type *)& extracted[i][offset];
pointer[w] = vp[w*Nsimd+i*s];
}
}
}
////////////////////////////////////////////////////////////////////////
// Merge bunch of scalar object pointers of different scalar type, with offset. Useful for precision change
////////////////////////////////////////////////////////////////////////
template<class vobj, class sobj> accelerator_inline
void merge(vobj &vec,ExtractPointerArray<sobj> &extracted, int offset)
{
typedef typename GridTypeMapper<vobj>::scalar_type scalar_type;
typedef typename GridTypeMapper<vobj>::vector_type vector_type;
const int words=sizeof(vobj)/sizeof(vector_type);
const int Nsimd=vector_type::Nsimd();
const int Nextr=extracted.size();
const int s = Nsimd/Nextr;
scalar_type * vp = (scalar_type *)&vec;
for(int w=0;w<words;w++){
for(int i=0;i<Nextr;i++){
scalar_type * pointer = (scalar_type *)& extracted[i][offset];
vp[w*Nsimd+i*s] = pointer[w] ;
}
}
}
NAMESPACE_END(Grid);