/* Main computational kernel. The whole function will be timed,
including the call and return. */
static
void kernel_correlation(int m, int n,
DATA_TYPE float_n,
DATA_TYPE POLYBENCH_2D(data,M,N,m,n),
DATA_TYPE POLYBENCH_2D(symmat,M,M,m,m),
DATA_TYPE POLYBENCH_1D(mean,M,m),
DATA_TYPE POLYBENCH_1D(stddev,M,m))
{
int i, j, j1, j2;
DATA_TYPE eps = 0.1f;
#define sqrt_of_array_cell(x,j) sqrt(x[j])
#pragma acc data \
copyin (data), \
copyout (mean, stddev, symmat)
{
/* Determine mean of column vectors of input data matrix */
#pragma acc parallel
{
#pragma acc loop
for (j = 0; j < _PB_M; j++)
{
mean[j] = 0.0;
#pragma acc loop
for (i = 0; i < _PB_N; i++)
mean[j] += data[i][j];
mean[j] /= float_n;
}
/* Determine standard deviations of column vectors of data matrix. */
#pragma acc loop
for (j = 0; j < _PB_M; j++)
{
stddev[j] = 0.0;
#pragma acc loop
for (i = 0; i < _PB_N; i++)
stddev[j] += (data[i][j] - mean[j]) * (data[i][j] - mean[j]);
stddev[j] /= float_n;
stddev[j] = sqrt_of_array_cell(stddev, j);
/* The following in an inelegant but usual way to handle
near-zero std. dev. values, which below would cause a zero-
divide. */
stddev[j] = stddev[j] <= eps ? 1.0 : stddev[j];
}
/* Center and reduce the column vectors. */
#pragma acc loop
for (i = 0; i < _PB_N; i++)
#pragma acc loop
for (j = 0; j < _PB_M; j++)
{
data[i][j] -= mean[j];
data[i][j] /= sqrt(float_n) * stddev[j];
}
/* Calculate the m * m correlation matrix. */
#pragma acc loop
for (j1 = 0; j1 < _PB_M-1; j1++)
{
symmat[j1][j1] = 1.0;
#pragma acc loop
for (j2 = j1+1; j2 < _PB_M; j2++)
{
symmat[j1][j2] = 0.0;
#pragma acc loop
for (i = 0; i < _PB_N; i++)
symmat[j1][j2] += (data[i][j1] * data[i][j2]);
symmat[j2][j1] = symmat[j1][j2];
}
}
}
}
symmat[_PB_M-1][_PB_M-1] = 1.0;
}
void correlation(DATA_TYPE* data, DATA_TYPE* mean, DATA_TYPE* stddev, DATA_TYPE* symmat)
{
int i, j, j1, j2;
// Determine mean of column vectors of input data matrix
for (j = 1; j <= M; j++)
{
mean[j] = 0.0;
for (i = 1; i <= N; i++)
{
mean[j] += data[i*(M+1) + j];
}
mean[j] /= (DATA_TYPE)FLOAT_N;
}
// Determine standard deviations of column vectors of data matrix.
for (j = 1; j <= M; j++)
{
stddev[j] = 0.0;
for (i = 1; i <= N; i++)
{
stddev[j] += (data[i*(M+1) + j] - mean[j]) * (data[i*(M+1) + j] - mean[j]);
}
stddev[j] /= FLOAT_N;
stddev[j] = sqrt_of_array_cell(stddev, j);
stddev[j] = stddev[j] <= EPS ? 1.0 : stddev[j];
}
// Center and reduce the column vectors.
for (i = 1; i <= N; i++)
{
for (j = 1; j <= M; j++)
{
data[i*(M+1) + j] -= mean[j];
data[i*(M+1) + j] /= sqrt(FLOAT_N) ;
data[i*(M+1) + j] /= stddev[j];
}
}
// Calculate the m * m correlation matrix.
for (j1 = 1; j1 <= M-1; j1++)
{
symmat[j1*(M+1) + j1] = 1.0;
for (j2 = j1+1; j2 <= M; j2++)
{
symmat[j1*(M+1) + j2] = 0.0;
for (i = 1; i <= N; i++)
{
symmat[j1*(M+1) + j2] += (data[i*(M+1) + j1] * data[i*(M+1) + j2]);
}
symmat[j2*(M+1) + j1] = symmat[j1*(M+1) + j2];
}
}
symmat[M*(M+1) + M] = 1.0;
}
/* Main computational kernel. The whole function will be timed,
including the call and return. */
static
void kernel_correlation(int m, int n,
DATA_TYPE float_n,
DATA_TYPE POLYBENCH_2D(data,M,N,m,n),
DATA_TYPE POLYBENCH_2D(symmat,M,M,m,m),
DATA_TYPE POLYBENCH_1D(mean,M,m),
DATA_TYPE POLYBENCH_1D(stddev,M,m))
{
int i, j, j1, j2;
DATA_TYPE eps = 0.1f;
#define sqrt_of_array_cell(x,j) sqrt(x[j])
/* Determine mean of column vectors of input data matrix */
#pragma omp parallel
{
#ifdef PRINT_ENV
printf("- Thread[%d] Working\n", omp_get_thread_num());
#endif
#pragma omp for private (i)
for (j = 0; j < _PB_M; j++)
{
mean[j] = 0.0;
for (i = 0; i < _PB_N; i++)
mean[j] += data[i][j];
mean[j] /= float_n;
}
/* Determine standard deviations of column vectors of data matrix. */
#pragma omp for private (i)
for (j = 0; j < _PB_M; j++)
{
stddev[j] = 0.0;
for (i = 0; i < _PB_N; i++)
stddev[j] += (data[i][j] - mean[j]) * (data[i][j] - mean[j]);
stddev[j] /= float_n;
stddev[j] = sqrt_of_array_cell(stddev, j);
/* The following in an inelegant but usual way to handle
near-zero std. dev. values, which below would cause a zero-
divide. */
stddev[j] = stddev[j] <= eps ? 1.0 : stddev[j];
}
/* Center and reduce the column vectors. */
#pragma omp for private (j)
for (i = 0; i < _PB_N; i++)
for (j = 0; j < _PB_M; j++)
{
data[i][j] -= mean[j];
data[i][j] /= sqrt(float_n) * stddev[j];
}
/* Calculate the m * m correlation matrix. */
#pragma omp for private (j2, i)
for (j1 = 0; j1 < _PB_M-1; j1++)
{
symmat[j1][j1] = 1.0;
for (j2 = j1+1; j2 < _PB_M; j2++)
{
symmat[j1][j2] = 0.0;
for (i = 0; i < _PB_N; i++)
symmat[j1][j2] += (data[i][j1] * data[i][j2]);
symmat[j2][j1] = symmat[j1][j2];
}
}
symmat[_PB_M-1][_PB_M-1] = 1.0;
}
}
void runCorr(DATA_TYPE pdata[M+1][N+1], DATA_TYPE psymmat[M+1][M+1], DATA_TYPE pstddev[M+1], DATA_TYPE pmean[M+1], DATA_TYPE pfloat_n, DATA_TYPE peps)
{
int i, j, j1, j2;
#define sqrt_of_array_cell(x,j) sqrt(x[j])
/* Determine mean of column vectors of input data matrix */
#pragma hmppcg grid blocksize 32 X 8
for (j = 1; j <= M; j++)
{
pmean[j] = 0.0;
for (i = 1; i <= N; i++)
{
pmean[j] += pdata[i][j];
}
pmean[j] /= pfloat_n;
}
/* Determine standard deviations of column vectors of data matrix. */
#pragma hmppcg grid blocksize 32 X 8
for (j = 1; j <= M; j++)
{
pstddev[j] = 0.0;
for (i = 1; i <= N; i++)
{
pstddev[j] += (pdata[i][j] - pmean[j]) * (pdata[i][j] - pmean[j]);
}
pstddev[j] /= pfloat_n;
pstddev[j] = sqrt_of_array_cell(pstddev, j);
/* The following in an inelegant but usual way to handle
near-zero std. dev. values, which below would cause a zero-
divide. */
pstddev[j] = pstddev[j] <= peps ? 1.0 : pstddev[j];
}
/* Center and reduce the column vectors. */
#pragma hmppcg grid blocksize 32 X 8
for (i = 1; i <= N; i++)
{
for (j = 1; j <= M; j++)
{
pdata[i][j] -= pmean[j];
pdata[i][j] /= sqrt(pfloat_n) * pstddev[j];
}
}
/* Calculate the m * m correlation matrix. */
#pragma hmppcg grid blocksize 32 X 8
for (j1 = 1; j1 <= M-1; j1++)
{
psymmat[j1][j1] = 1.0;
for (j2 = j1+1; j2 <= M; j2++)
{
psymmat[j1][j2] = 0.0;
for (i = 1; i <= N; i++)
{
psymmat[j1][j2] += (pdata[i][j1] * pdata[i][j2]);
}
psymmat[j2][j1] = psymmat[j1][j2];
}
}
psymmat[M][M] = 1.0;
}
