void update(par* p, gsl_vector** dbiases, gsl_matrix** dweights, double step){
regularisation(p, dbiases, dweights);
momentum_update(p, dbiases, dweights, step);
}
double pewarpcost(float *x, float *bX, float *bY, float *dbY, float *bZ,
mwSignedIndex *bdims, uint8_T *target, float *source, uint8_T *mask,
mwSignedIndex *ddims, float *skrn, int skrnl, float xscale,
float regstrength, float *H, uint8_T *warped, float *def)
{
int nvox = ddims[0] * ddims[1] * ddims[2];
/* float *def = mxMalloc(sizeof(float) * nvox); */
float *jac = mxMalloc(sizeof(float) * nvox);
/* uint8_T *warped = mxMalloc(sizeof(uint8_T) * nvox); */
/* float *H = mxMalloc(sizeof(float) * 256 * 256); */
float *Htmp = mxMalloc(sizeof(float) * 256 * 256);
double *s1 = mxMalloc(sizeof(double) * 256);
double *s2 = mxMalloc(sizeof(double) * 256);
double s, sL, s1L, s2L;
double cost;
/* mexPrintf(">>> Start -> %lu (CLOCKS_PER_SEC: %lu)\n", clock(), CLOCKS_PER_SEC); */
getDef(def, ddims, bdims[0], bX, bdims[1], bY, bdims[2], bZ, x, xscale);
/* The actual Jacobian is 1 + jac, but it is easier to add 1 later. */
getDef(jac, ddims, bdims[0], bX, bdims[1], dbY, bdims[2], bZ, x, xscale);
/* mexPrintf("getDef done -> %lu\n", clock()); */
/* For some reason, execution is single-threaded without the num_threads clause.
* TODO: Find out why and change. */
#pragma omp parallel num_threads(4) shared(s, sL, s1L, s2L)
{
peResampleAndApplyJacobian(warped, source, def, /* jac, */ mask, ddims);
/* if (omp_get_thread_num() == 1)
mexPrintf("peResampleAndApplyJacobian done -> %lu\n", clock()); */
hist2(H, target, warped, mask, nvox);
/* if (omp_get_thread_num() == 1)
mexPrintf("hist2 done -> %lu\n", clock()); */
smoothRows(Htmp, H, 256, 256, skrn, skrnl);
smoothCols(H, Htmp, 256, 256, skrn, skrnl);
/* if (omp_get_thread_num() == 1)
mexPrintf("smooth... done -> %lu\n", clock()); */
#pragma omp single
{
s = histSumAll(H, 65536);
}
#pragma omp sections
{
#pragma omp section
{
sL = histLogAll(H, s, 65536);
}
#pragma omp section
{
s1L = histLogVert(H, s1, s, 256, 256);
}
#pragma omp section
{
s2L = histLogHorz(H, s2, s, 256, 256);
}
}
/* if (omp_get_thread_num() == 1)
mexPrintf("hist... done -> %lu\n", clock()); */
}
cost = - (s1L + s2L) / sL + regstrength * regularisation(jac, nvox);
/* mexPrintf("regularisation done -> %lu\n", clock());
mexPrintf("<<<\n"); */
/* mexPrintf("cost: %f sL: %f s1L: %f s2L: %f #voxels: %d\n", cost, sL, s1L, s2L, nvox); */
/* mxFree(def); */
mxFree(jac);
/* mxFree(warped); */
/* mxFree(H); */
mxFree(Htmp);
mxFree(s1);
mxFree(s2);
return cost;
}
