// Applies previously trained QDA classifier to new data matrix x
//
// Inputs:
// i_x - N x p data matrix of N samples in p dimensions
// i_means - K x p matrix of class centroids
// i_icovmats - K x p x p array of inverse covariance matrices
// i_bias - vector of length containing the bias term for each discriminant function
// i_N - number of samples in x
// i_p - dimensionality of x
// i_K - number of classes
//
// Outputs:
// o_labels - labels for all samples in i_x
//
// Memory allocation requirements for the outputs:
// o_labels must be of length N
void QDAtest(
double* i_x, double* i_means, double* i_icovmats, double* i_bias,
int* i_N, int* i_p, int* i_K, int* o_labels
)
{
// Create local storage
int t, k, i, j;
double temp;
double* x_c = make1D( *i_p );
double** x = make2D( *i_N, *i_p );
double** means = make2D( *i_K, *i_p );
double*** icovmats = make3D( *i_K, *i_p, *i_p );
double** delta = make2D( *i_N, *i_K );
// Copy and reshape the inputs to local storage
cp_vec_mat( *i_N, *i_p, i_x, x );
cp_vec_mat( *i_K, *i_p, i_means, means );
cp_vec_3D( *i_K, *i_p, *i_p, i_icovmats, icovmats );
// Compute the delta function values
for( t = 0; t < *i_N; t++ )
for( k = 0; k < *i_K; k++ )
{
add_vec_vec( *i_p, 1.0, -1.0, x[t], means[k], x_c );
delta[t][k] = 0.0;
for( i = 0; i < *i_p; i++ )
for( j = 0; j < *i_p; j++ )
delta[t][k] += icovmats[k][i][j] * x_c[i] * x_c[j];
delta[t][k] *= -0.5;
delta[t][k] += i_bias[k];
}
// Pick the highest discriminant function value for each point
for( t = 0; t < *i_N; t++ )
{
o_labels[t] = 1;
temp = delta[t][0];
for( k = 1; k < *i_K; k++ )
{
if( delta[t][k] > temp )
{
o_labels[t] = k + 1;
temp = delta[t][k];
}
}
}
// Free up memory
free( x_c );
free2D( x, *i_N );
free2D( means, *i_K );
free3D( icovmats, *i_K, *i_p );
free2D( delta, *i_N );
}
int main (int nargs, char** args)
{
int n; /* number of points in each direction */
double h; /* grid spacing, same in all the directions */
double ***u_old, ***u_new, ***rhs;
double factor, factor2, l2_norm;
int i,j,k;
int max_iters=100;
if (nargs>1)
n = atoi(args[1]);
else
n = 256;
h = 1.0/(n-1);
u_old = alloc3D(n+2, n+2, n+2);
u_new = alloc3D(n+2, n+2, n+2);
rhs = alloc3D(n+2, n+2, n+2);
/* fill the right-hand side vector */
factor = (1.0-h*h*M_PI*M_PI/4)*3.0*M_PI*M_PI; /* use deferred correction */
for (k=0; k<= n+1; k++)
for (j=0; j<= n+1; j++)
for (i=0; i<= n+1; i++)
rhs[k][j][i] = 6.*h*h*factor*sin(M_PI*i*h)*sin(M_PI*j*h)*sin(M_PI*k*h);
/* use initial zero guess */
for (k=0; k<= n+1; k++)
for (j=0; j<= n+1; j++)
for (i=0; i<= n+1; i++)
u_old[k][j][i] = u_new[k][j][i] = 0.;
/* Jacobi iterations */
l2_norm = 1e+12;
factor = 1.0/24; factor2 = 6.0*h*h;
printf("\n=====Timings (sec) for 19-Point Jacobi, Solving Poisson Eqn ");
if(sizeof(REAL) == 4)
printf(" (Single Precision) =====\n");
if(sizeof(REAL) == 8)
printf(" (Double Precision) =====\n");
printf("Kernel\t Time(sec)\tGflops \tBW-ideal(GB/s)\tBW-algorithm (N=(%d) iters=%d)\n",n, max_iters);
printf("------\t----------\t--------\t--------------\t------------\n");
int nIters=0;
double time_elapsed= getTime();
double Gflops =0.0;
#pragma mint copy ( u_old, toDevice, ( n+2 ), n+2, ( n+2 ) )
#pragma mint copy ( u_new, toDevice, ( n+2 ), n+2, ( n+2 ) )
#pragma mint copy ( rhs, toDevice, ( n+2 ), n+2, ( n+2 ) )
#pragma mint parallel
{
int iters = 0 ;
while (iters < max_iters && l2_norm > 1e-9) {
++iters;
/* update each interior point */
#pragma mint for nest(all) tile(16,16,1)
for (k=1; k<= n; k++){
for (j=1; j<= n; j++){
for (i=1; i<= n; i++)
u_new[k][j][i] = factor*(rhs[k][j][i]
+factor2*(u_old[k][j][i-1]+u_old[k][j][i+1]
+u_old[k][j-1][i]+u_old[k][j+1][i]
+u_old[k+1][j][i]+u_old[k-1][j][i])
+u_old[k-1][j-1][i]+u_old[k-1][j+1][i]
+u_old[k-1][j][i-1]+u_old[k-1][j][i+1]
+u_old[k][j-1][i-1]+u_old[k][j+1][i-1]
+u_old[k][j-1][i+1]+u_old[k][j+1][i+1]
+u_old[k+1][j-1][i]+u_old[k+1][j+1][i]
+u_old[k+1][j][i-1]+u_old[k+1][j][i+1]);
}}
/* pointer swap */
#pragma mint single
{
REAL*** tmp;
tmp = u_old; u_old= u_new; u_new = tmp;
nIters = iters;
}
}
}
#pragma mint copy ( u_old, fromDevice, ( n+2 ), ( n+2 ), ( n+2 ) )
time_elapsed = getTime() - time_elapsed;
Gflops = (double)(nIters * (n) * (n) * (n) * 1.0e-9 * FLOPS) / time_elapsed ;
l2_norm = 0;
for (k=0; k<= n+1; k++)
for (j=0; j<= n+1; j++)
for (i=0; i<= n+1; i++) {
factor = sin(M_PI*i*h)*sin(M_PI*j*h)*sin(M_PI*k*h);
l2_norm += (factor-u_old[k][j][i])*(factor-u_old[k][j][i]);
}
printf("%s%3.3f \t%5.3f\n", "Poisson19 ", time_elapsed, Gflops);
printf(":N %d M %d K %d , iteration %d\n", n, n, n , nIters);
printf(":max: %20.12e, l2norm: %20.12e\n",factor,sqrt(l2_norm*h*h*h));
//printf("Total iterations used: %d, l2-norm of error=%e\n",
// nIters,sqrt(l2_norm*h*h*h));
free3D(u_new);
free3D(u_old);
free3D(rhs);
return 0;
}
// Trains a QDA classifier
//
// Inputs:
// i_x - N x p data matrix of N samples in p dimensions
// i_y - N x 1 vector of labels
// i_N - number of samples
// i_p - data dimensionality
// i_K - number of classes
// i_debug_output - level of debug output
//
// Output:
// o_priors - K x 1 vector of Bayesian priors, computed as fraction of points from each class
// o_means - K x p matrix of means approximated from the data
// o_covmats - K x p x p array of covariance matrices approximated from the data
// o_icovmats - K x p x p array of inverse covariance matrices
// o_bias - K x 1 vector of bias terms for discriminant function computations
//
// Memory allocation requirements for outputs:
// o_priors must be of length K
// o_means must be of length K*p
// o_covmat must be of length K*p*p
// o_icovmat must be of length K*p*p
// o_bias must be of lenght K
// o_status must be of length 1
// o_info must be of length 1
//
// Meaning of o_status:
// 0 - Everything is OK
// 1 - The covariance matrix for class *o_info is singular
void QDAtrain(
double* i_x, int* i_y, int* i_N, int* i_p, int* i_K, double* i_cov_reg,
int* i_debug_output,
double* o_priors, double* o_means, double* o_covmats, double* o_icovmats,
double* o_bias, int* o_status, int* o_info
)
{
if( *i_debug_output > 0 )
printf( "Currently running C version of the QDA.train code\n" );
// Create local storage
int i, j, t, k;
int* Nk = (int*)malloc( *i_K * sizeof(int) );
double** x = make2D( *i_N, *i_p );
double** means = make2D( *i_K, *i_p );
double*** covmats = make3D( *i_K, *i_p, *i_p );
double*** icovmats = make3D( *i_K, *i_p, *i_p );
// Copy and reshape the inputs
cp_vec_mat( *i_N, *i_p, i_x, x );
// Init the counts and means to 0
for( k = 0; k < *i_K; k++ )
{
Nk[k] = 0;
for( i = 0; i < *i_p; i++ )
means[k][i] = 0.0;
}
// Init the covariance matrices to 0
for( k = 0; k < *i_K; k++ )
for( i = 0; i < *i_p; i++ )
for( j = 0; j < *i_p; j++ )
covmats[k][i][j] = 0.0;
// Traverse one sample at a time to compute mean contributions
for( t = 0; t < *i_N; t++ )
{
k = i_y[t] - 1; // account for 0-based index
// Count the sample
Nk[k]++;
// Add contribution to the mean
add_vec_vec( *i_p, 1.0, 1.0, means[k], x[t], means[k] );
}
// Compute the means and the priors
for( k = 0; k < *i_K; k++ )
{
o_priors[k] = ((double)Nk[k]) / ((double)*i_N);
mult_const_vec( *i_p, 1.0 / ((double)Nk[k]), means[k] );
}
// Subtract the means from the samples and compute the covariance matrices
for( t = 0; t < *i_N; t++ )
{
k = i_y[t] - 1; // account for 0-based index
add_vec_vec( *i_p, 1.0, -1.0, x[t], means[k], x[t] );
for( i = 0; i < *i_p; i++ )
for( j = 0; j < *i_p; j++ )
covmats[k][i][j] += x[t][i] * x[t][j];
}
for( k = 0; k < *i_K; k++ )
for( i = 0; i < *i_p; i++ )
for( j = 0; j < *i_p; j++ )
covmats[k][i][j] /= (Nk[k] - 1);
// Regularize the covariance matrices
if( *i_cov_reg > 0 )
{
for( k = 0; k < *i_K; k++ )
for( i = 0; i < *i_p; i++ )
for( j = 0; j < *i_p; j++ )
{
covmats[k][i][j] *= (1 - *i_cov_reg);
if( i == j )
covmats[k][i][j] += *i_cov_reg;
}
}
// Compute the bias terms
for( k = 0; k < *i_K; k++ )
o_bias[k] = log( o_priors[k] ) - 0.5 * covmat_logdet( *i_p, covmats[k] );
// Compute the covariance matrix inverses
for( k = 0; k < *i_K; k++ )
{
*o_status = covmat_inverse( *i_p, covmats[k], icovmats[k] );
if( *o_status != 0 )
{
free2D( x, *i_N );
free2D( means, *i_K );
free3D( covmats, *i_K, *i_p );
free3D( icovmats, *i_K, *i_p );
*o_info = k;
return;
}
}
// Copy the results from local storage to outputs
cp_mat_vec( *i_K, *i_p, means, o_means );
cp_3D_vec( *i_K, *i_p, *i_p, covmats, o_covmats );
cp_3D_vec( *i_K, *i_p, *i_p, icovmats, o_icovmats );
// Free memory
free2D( x, *i_N );
free2D( means, *i_K );
free3D( covmats, *i_K, *i_p );
free3D( icovmats, *i_K, *i_p );
}
