jiri/LA_library
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

*** empty log message ***

Browse Source
This commit is contained in:
jiri
2010-06-25 15:28:19 +00:00
parent eb0aaf9adf
commit 074c943862
13 changed files with 1938 additions and 464 deletions
Download Patch File Download Diff File Expand all files Collapse all files

276
mat.h
View File

@@ -33,12 +33,20 @@ protected:
T *v;
#endif
int *count;
#ifdef CUDALA
GPUID location;
#endif
public:
friend class NRVec<T>;
friend class NRSMat<T>;
inline NRMat() : nn(0), mm(0), v(0), count(0) {};
inline NRMat(const int n, const int m);
inline NRMat() : nn(0), mm(0), v(0), count(0)
{
#ifdef CUDALA
location = DEFAULT_LOC;
#endif
};
inline NRMat(const int n, const int m ,const GPUID loc= undefined);
inline NRMat(const T &a, const int n, const int m);
NRMat(const T *a, const int n, const int m);
inline NRMat(const NRMat &rhs);
@@ -57,6 +65,13 @@ public:
#endif
const bool operator==(const NRMat &rhs) const {return !(*this != rhs);};
inline int getcount() const {return count?*count:0;}
#ifdef CUDALA
inline GPUID getlocation() const {return location;}
void moveto(const GPUID dest);
#else
inline GPUID getlocation() const {return cpu;}
void moveto(const GPUID dest) {};
#endif
NRMat & operator=(const NRMat &rhs); //assignment
void randomize(const typename LA_traits<T>::normtype &x); //fill with random numbers
NRMat & operator=(const T &a); //assign a to diagonal
@@ -88,7 +103,7 @@ public:
const NRMat operator*(const NRSMat<T> &rhs) const; // Mat * Smat
const NRMat operator&(const NRMat &rhs) const; // direct sum
const NRMat operator|(const NRMat<T> &rhs) const; // direct product
const NRVec<T> operator*(const NRVec<T> &rhs) const {NRVec<T> result(nn); result.gemv((T)0,*this,'n',(T)1,rhs); return result;}; // Mat * Vec
const NRVec<T> operator*(const NRVec<T> &rhs) const {NRVec<T> result(nn,rhs.getlocation()); result.gemv((T)0,*this,'n',(T)1,rhs); return result;}; // Mat * Vec
const NRVec<complex<T> > operator*(const NRVec<complex<T> > &rhs) const {NRVec<complex<T> > result(nn); result.gemv((T)0,*this,'n',(T)1,rhs); return result;}; // Mat * Vec
const NRVec<T> rsum() const; //sum of rows
const NRVec<T> csum() const; //sum of columns
@@ -157,9 +172,14 @@ public:
namespace LA {
// ctors
template <typename T>
NRMat<T>::NRMat(const int n, const int m) : nn(n), mm(m), count(new int)
NRMat<T>::NRMat(const int n, const int m, const GPUID loc) : nn(n), mm(m), count(new int)
{
*count = 1;
#ifdef CUDALA
location= (loc==undefined?DEFAULT_LOC:loc);
if(location==cpu)
{
#endif
#ifdef MATPTR
v = new T*[n];
v[0] = new T[m*n];
@@ -167,14 +187,29 @@ NRMat<T>::NRMat(const int n, const int m) : nn(n), mm(m), count(new int)
#else
v = new T[m*n];
#endif
#ifdef CUDALA
}
else
{
v= (T*) gpualloc(n*m*sizeof(T));
}
#endif
}
template <typename T>
NRMat<T>::NRMat(const T &a, const int n, const int m) : nn(n), mm(m), count(new int)
{
#ifdef CUDALA
location=DEFAULT_LOC;
#endif
int i;
T *p;
*count = 1;
#ifdef CUDALA
if(location==cpu)
{
#endif
#ifdef MATPTR
v = new T*[n];
p = v[0] = new T[m*n];
@@ -186,12 +221,29 @@ NRMat<T>::NRMat(const T &a, const int n, const int m) : nn(n), mm(m), count(new
for (i=0; i< n*m; i++) *p++ = a;
else
memset(p, 0, n*m*sizeof(T));
#ifdef CUDALA
}
else
{
v= (T*) gpualloc(n*m*sizeof(T));
cublasSetVector(n*m,sizeof(T),&a,0,v,1);
}
#endif
}
template <typename T>
NRMat<T>::NRMat(const T *a, const int n, const int m) : nn(n), mm(m), count(new int)
{
#ifdef CUDALA
location=DEFAULT_LOC;
#endif
*count = 1;
#ifdef CUDALA
if(location==cpu)
{
#endif
#ifdef MATPTR
v = new T*[n];
v[0] = new T[m*n];
@@ -201,11 +253,25 @@ NRMat<T>::NRMat(const T *a, const int n, const int m) : nn(n), mm(m), count(new
v = new T[m*n];
memcpy(v, a, n*m*sizeof(T));
#endif
#ifdef CUDALA
}
else
{
v= (T*) gpualloc(n*m*sizeof(T));
cublasSetVector(n*m,sizeof(T),a,1,v,1);
}
#endif
}
//copy constructor
template <typename T>
NRMat<T>::NRMat(const NRMat &rhs)
{
#ifdef CUDALA
location=rhs.location;
#endif
nn = rhs.nn;
mm = rhs.mm;
count = rhs.count;
@@ -213,9 +279,16 @@ NRMat<T>::NRMat(const NRMat &rhs)
if (count) ++(*count);
}
template <typename T>
NRMat<T>::NRMat(const NRSMat<T> &rhs)
{
NOT_GPU(rhs);
#ifdef CUDALA
location=rhs.location;
#endif
int i;
nn = mm = rhs.nrows();
count = new int;
@@ -244,6 +317,10 @@ NRMat<T>::NRMat(const NRVec<T> &rhs, const int n, const int m, const int offset)
{
if (offset < 0 || n*m + offset > rhs.nn) laerror("matrix dimensions and offset incompatible with vector length");
#ifdef CUDALA
location=rhs.location;
#endif
nn = n;
mm = m;
count = rhs.count;
@@ -303,6 +380,7 @@ inline T & NRMat<T>::operator()(const int i, const int j)
if (_LA_count_check && *count != 1) laerror("Mat lval use of (,) with count > 1");
if (i<0 || i>=nn &&nn>0 || j<0 || j>=mm && mm>0) laerror("Mat (,) out of range");
if (!v) laerror("(,) for unallocated Mat");
NOT_GPU(*this);
#endif
#ifdef MATPTR
return v[i][j];
@@ -310,12 +388,14 @@ inline T & NRMat<T>::operator()(const int i, const int j)
return v[i*mm+j];
#endif
}
template <typename T>
inline const T & NRMat<T>::operator()(const int i, const int j) const
{
#ifdef DEBUG
if (i<0 || i>=nn&&nn>0 || j<0 || j>=mm&& mm>0) laerror("Mat (,) out of range");
if (!v) laerror("(,) for unallocated Mat");
NOT_GPU(*this); //in principle we could copy the element to CPU memory, yielding, however, a highly inneficient contruct
#endif
#ifdef MATPTR
return v[i][j];
@@ -391,7 +471,7 @@ inline const complex<double> NRMat< complex<double> >::amax() const
}
//basi stuff to be available for any type ... must be in .h
//basic stuff to be available for any type ... must be in .h
// dtor
template <typename T>
NRMat<T>::~NRMat()
@@ -399,10 +479,21 @@ NRMat<T>::~NRMat()
if (!count) return;
if (--(*count) <= 0) {
if (v) {
#ifdef CUDALA
if(location==cpu)
#endif
{
#ifdef MATPTR
delete[] (v[0]);
#endif
delete[] v;
}
#ifdef CUDALA
else
{
gpufree(v);
}
#endif
}
delete count;
}
@@ -415,14 +506,27 @@ NRMat<T> & NRMat<T>::operator=(const NRMat<T> &rhs)
if (this !=&rhs)
{
if (count)
if (--(*count) ==0 ) {
if (--(*count) ==0 )
{
#ifdef CUDALA
if(location==cpu)
{
#endif
#ifdef MATPTR
delete[] (v[0]);
#endif
delete[] v;
#ifdef CUDALA
}
else gpufree(v);
#endif
delete count;
}
v = rhs.v;
#ifdef CUDALA
location=rhs.location;
#endif
nn = rhs.nn;
mm = rhs.mm;
count = rhs.count;
@@ -437,46 +541,8 @@ template <typename T>
NRMat<T> & NRMat<T>::operator|=(const NRMat<T> &rhs)
{
if (this == &rhs) return *this;
#ifdef DEBUG
if (!rhs.v) laerror("unallocated rhs in Mat operator |=");
#endif
if (count)
if (*count > 1) {
--(*count);
nn = 0;
mm = 0;
count = 0;
v = 0;
}
if (nn != rhs.nn || mm != rhs.mm) {
if (v) {
#ifdef MATPTR
delete[] (v[0]);
#endif
delete[] (v);
v = 0;
}
nn = rhs.nn;
mm = rhs.mm;
}
if (!v) {
#ifdef MATPTR
v = new T*[nn];
v[0] = new T[mm*nn];
#else
v = new T[mm*nn];
#endif
}
#ifdef MATPTR
for (int i=1; i< nn; i++) v[i] = v[i-1] + mm;
memcpy(v[0], rhs.v[0], nn*mm*sizeof(T));
#else
memcpy(v, rhs.v, nn*mm*sizeof(T));
#endif
if (!count) count = new int;
*count = 1;
*this = rhs;
this->copyonwrite();
return *this;
}
@@ -486,9 +552,13 @@ void NRMat<T>::copyonwrite()
{
if (!count) laerror("Mat::copyonwrite of undefined matrix");
if (*count > 1) {
(*count)--;
count = new int;
*count = 1;
(*count)--;
count = new int;
*count = 1;
#ifdef CUDALA
if(location==cpu) //matrix is in CPU memory
{
#endif
#ifdef MATPTR
T **newv = new T*[nn];
newv[0] = new T[mm*nn];
@@ -499,10 +569,21 @@ void NRMat<T>::copyonwrite()
T *newv = new T[mm*nn];
memcpy(newv, v, mm*nn*sizeof(T));
v = newv;
#endif
#ifdef CUDALA
}
else //matrix is in GPU memory
{
T *newv = (T *) gpualloc(mm*nn*sizeof(T));
if(sizeof(T)%sizeof(float)!=0) laerror("cpu memcpy alignment problem");
cublasScopy(nn*mm*sizeof(T)/sizeof(float),(const float *) v,1,(float *)newv,1);
v = newv;
}
#endif
}
}
template <typename T>
void NRMat<T>::resize(int n, int m)
{
@@ -519,10 +600,18 @@ if(m==0) n=0;
if(n==0 && m==0)
{
if(--(*count) <= 0) {
#ifdef CUDALA
if(location==cpu)
{
#endif
#ifdef MATPTR
if(v) delete[] (v[0]);
#endif
if(v) delete[] v;
#ifdef CUDALA
}
else gpufree(v);
#endif
delete count;
}
count=0;
@@ -543,6 +632,10 @@ if(m==0) n=0;
*count = 1;
nn = n;
mm = m;
#ifdef CUDALA
if(location==cpu)
{
#endif
#ifdef MATPTR
v = new T*[nn];
v[0] = new T[m*n];
@@ -550,12 +643,22 @@ if(m==0) n=0;
#else
v = new T[m*n];
#endif
#ifdef CUDALA
}
else
v = (T *) gpualloc(n*m*sizeof(T));
#endif
return;
}
// At this point *count = 1, check if resize is necessary
if (n!=nn || m!=mm) {
nn = n;
mm = m;
#ifdef CUDALA
if(location==cpu)
{
#endif
#ifdef MATPTR
delete[] (v[0]);
#endif
@@ -566,6 +669,14 @@ if(m==0) n=0;
for (int i=1; i< n; i++) v[i] = v[i-1] + m;
#else
v = new T[m*n];
#endif
#ifdef CUDALA
}
else
{
gpufree(v);
v=(T *) gpualloc(n*m*sizeof(T));
}
#endif
}
}
@@ -587,7 +698,11 @@ return r;
// I/O
template <typename T>
std::ostream& operator<<(std::ostream &s, const NRMat<T> &x)
{
{
#ifdef CUDALA
if(x.getlocation()==cpu)
{
#endif
int i,j,n,m;
n=x.nrows();
m=x.ncols();
@@ -597,18 +712,43 @@ std::ostream& operator<<(std::ostream &s, const NRMat<T> &x)
for(j=0; j<m;j++) s << (typename LA_traits_io<T>::IOtype) x[i][j] << (j==m-1 ? '\n' : ' '); // endl cannot be used in the conditional expression, since it is an overloaded function
}
return s;
}
#ifdef CUDALA
}
else
{
NRMat<T> tmp=x;
tmp.moveto(cpu);
return s<<tmp;
}
#endif
}
template <typename T>
std::istream& operator>>(std::istream &s, NRMat<T> &x)
{
#ifdef CUDALA
if(x.getlocation()==cpu)
{
#endif
int i,j,n,m;
s >> n >> m;
x.resize(n,m);
typename LA_traits_io<T>::IOtype tmp;
for(i=0;i<n;i++) for(j=0; j<m;j++) { s>>tmp; x[i][j]=tmp;}
return s;
}
#ifdef CUDALA
}
else
{
NRMat<T> tmp;
tmp.moveto(cpu);
s >> tmp;
tmp.moveto(x.getlocation());
x=tmp;
return s;
}
#endif
}
//optional indexing from 1
@@ -671,6 +811,38 @@ NRMat<T> & NRMat<T>::operator^=(const NRMat<T> &rhs){
}
#ifdef CUDALA
template<typename T>
void NRMat<T>::moveto(const GPUID dest)
{
if(location==dest) return;
location=dest;
if(v && !count) laerror("internal inconsistency of reference counting 1");
if (!count) return;
if(v && *count==0) laerror("internal inconsistency of reference counting 2");
if(!v) return;
T *vold = v;
if(dest == cpu) //moving from GPU to CPU
{
v = new T[nn*mm];
gpuget(nn*mm,sizeof(T),vold,v);
if(*count == 1) gpufree(vold);
else {--(*count); count = new int(1);}
}
else //moving from CPU to GPU
{
v=(T *) gpualloc(nn*mm*sizeof(T));
gpuput(nn*mm,sizeof(T),vold,v);
if(*count == 1) delete[] vold;
else {--(*count); count = new int(1);}
}
}
#endif
//end CUDALA