92 lines
2.1 KiB
Plaintext
92 lines
2.1 KiB
Plaintext
#include "typeT.h"
|
|
|
|
static inline
|
|
uniform int gcd(uniform int a, uniform int b)
|
|
{
|
|
while ( a != 0 )
|
|
{
|
|
uniform int c = a;
|
|
a = b%a;
|
|
b = c;
|
|
}
|
|
return b;
|
|
}
|
|
|
|
static inline
|
|
int __rj(const int i, const uniform int joverb, const uniform int m, const uniform int b)
|
|
{
|
|
return (i + joverb) % m;
|
|
}
|
|
|
|
static inline
|
|
int __di(const int i, const uniform int j, const uniform int m, const uniform int n, const uniform int joverb)
|
|
{
|
|
return (((i+joverb) % m) + j*m) % n;
|
|
}
|
|
|
|
static inline
|
|
int __sj(const int i, const uniform int j, const uniform int m, const uniform int n, const int iovera)
|
|
{
|
|
return (j + i*n - iovera) % m;
|
|
}
|
|
|
|
static inline
|
|
void transpose_serial(uniform T A[], const uniform int m, const uniform int n)
|
|
{
|
|
const uniform int tmpSize = max(m,n) * programCount;
|
|
uniform T * uniform tmp = uniform new uniform T [tmpSize];
|
|
uniform int * uniform joverb = uniform new uniform int[n];
|
|
uniform int * uniform iovera = uniform new uniform int[m];
|
|
|
|
uniform T (*uniform tmp2D)[programCount] = (uniform T (*uniform)[programCount])tmp;
|
|
|
|
const uniform int c = gcd(m,n);
|
|
const uniform int a = m/c;
|
|
const uniform int b = n/c;
|
|
foreach (j = 0 ... n)
|
|
joverb[j] = j/b;
|
|
foreach (i = 0 ... m)
|
|
iovera[i] = i/a;
|
|
|
|
if (c > 1)
|
|
{
|
|
for (uniform int j = 0; j < n; j++)
|
|
{
|
|
const uniform int base = j*m;
|
|
const uniform int __joverb = joverb[j];
|
|
foreach (i = 0 ... m)
|
|
tmp[i] = A[base + __rj(i,__joverb,m,b)];
|
|
foreach (i = 0 ... m)
|
|
A[base + i] = tmp[i];
|
|
}
|
|
}
|
|
|
|
foreach (i = 0 ... m)
|
|
{
|
|
for (uniform int j = 0; j < n; j++)
|
|
tmp2D[__di(i,j,m,n,joverb[j])][programIndex] = A[j*m + i];
|
|
for (uniform int j = 0; j < n; j++)
|
|
A[j*m + i] = tmp2D[j][programIndex];
|
|
}
|
|
|
|
for (uniform int j = 0; j < n; j++)
|
|
{
|
|
const uniform int base = j*m;
|
|
foreach (i = 0 ... m)
|
|
tmp[i] = A[base + __sj(i,j,m,n,iovera[i])];
|
|
foreach (i = 0 ... m)
|
|
A[base + i] = tmp[i];
|
|
}
|
|
|
|
delete iovera;
|
|
delete joverb;
|
|
delete tmp;
|
|
}
|
|
|
|
export
|
|
void transpose(uniform T A[], const uniform int m, const uniform int n)
|
|
{
|
|
transpose_serial(A, m, n);
|
|
}
|
|
|