void joint_histogram(double* H,
unsigned int clampI,
unsigned int clampJ,
PyArrayIterObject* iterI,
const PyArrayObject* imJ_padded,
const double* Tvox,
int affine,
int interp)
{
const signed short* J=(signed short*)imJ_padded->data;
size_t dimJX=imJ_padded->dimensions[0]-2;
size_t dimJY=imJ_padded->dimensions[1]-2;
size_t dimJZ=imJ_padded->dimensions[2]-2;
signed short Jnn[8];
double W[8];
signed short *bufI, *bufJnn;
double *bufW;
signed short i, j;
size_t off;
size_t u2 = imJ_padded->dimensions[2];
size_t u3 = u2+1;
size_t u4 = imJ_padded->dimensions[1]*u2;
size_t u5 = u4+1;
size_t u6 = u4+u2;
size_t u7 = u6+1;
double wx, wy, wz, wxwy, wxwz, wywz;
double W0, W2, W3, W4;
size_t x, y, z;
int nn, nx, ny, nz;
double Tx, Ty, Tz;
double *bufTvox = (double*)Tvox;
void (*interpolate)(unsigned int, double*, unsigned int, const signed short*, const double*, int, void*);
void* interp_params = NULL;
rk_state rng;
/* Reset the source image iterator */
PyArray_ITER_RESET(iterI);
/* Make sure the iterator the iterator will update coordinate values */
UPDATE_ITERATOR_COORDS(iterI);
/* Set interpolation method */
if (interp==0)
interpolate = &_pv_interpolation;
else if (interp>0)
interpolate = &_tri_interpolation;
else { /* interp < 0 */
interpolate = &_rand_interpolation;
rk_seed(-interp, &rng);
interp_params = (void*)(&rng);
}
/* Re-initialize joint histogram */
memset((void*)H, 0, clampI*clampJ*sizeof(double));
/* Looop over source voxels */
while(iterI->index < iterI->size) {
/* Source voxel intensity */
bufI = (signed short*)PyArray_ITER_DATA(iterI);
i = bufI[0];
/* Compute the transformed grid coordinates of current voxel */
if (affine) {
/* Get voxel coordinates and apply transformation on-the-fly*/
x = iterI->coordinates[0];
y = iterI->coordinates[1];
z = iterI->coordinates[2];
_affine_transform(&Tx, &Ty, &Tz, Tvox, x, y, z);
}
else
/* Use precomputed transformed coordinates */
bufTvox = _precomputed_transform(&Tx, &Ty, &Tz, (const double*)bufTvox);
/* Test whether the current voxel is below the intensity
threshold, or the transformed point is completly outside
the reference grid */
if ((i>=0) &&
(Tx>-1) && (Tx<dimJX) &&
(Ty>-1) && (Ty<dimJY) &&
(Tz>-1) && (Tz<dimJZ)) {
/*
Nearest neighbor (floor coordinates in the padded
image, hence +1).
Notice that using the floor function doubles excetution time.
FIXME: see if we can replace this with assembler instructions.
*/
nx = FLOOR(Tx) + 1;
ny = FLOOR(Ty) + 1;
nz = FLOOR(Tz) + 1;
/* The convention for neighbor indexing is as follows:
*
* Floor slice Ceil slice
*
* 2----6 3----7 y
* | | | | ^
* | | | | |
* 0----4 1----5 ---> x
*/
/*** Trilinear interpolation weights.
Note: wx = nnx + 1 - Tx, where nnx is the location in
the NON-PADDED grid */
wx = nx - Tx;
wy = ny - Ty;
wz = nz - Tz;
wxwy = wx*wy;
wxwz = wx*wz;
wywz = wy*wz;
/*** Prepare buffers */
bufJnn = Jnn;
bufW = W;
/*** Initialize neighbor list */
off = nx*u4 + ny*u2 + nz;
nn = 0;
/*** Neighbor 0: (0,0,0) */
W0 = wxwy*wz;
APPEND_NEIGHBOR(off, W0);
/*** Neighbor 1: (0,0,1) */
APPEND_NEIGHBOR(off+1, wxwy-W0);
/*** Neighbor 2: (0,1,0) */
W2 = wxwz-W0;
APPEND_NEIGHBOR(off+u2, W2);
/*** Neightbor 3: (0,1,1) */
W3 = wx-wxwy-W2;
APPEND_NEIGHBOR(off+u3, W3);
/*** Neighbor 4: (1,0,0) */
W4 = wywz-W0;
APPEND_NEIGHBOR(off+u4, W4);
/*** Neighbor 5: (1,0,1) */
APPEND_NEIGHBOR(off+u5, wy-wxwy-W4);
/*** Neighbor 6: (1,1,0) */
APPEND_NEIGHBOR(off+u6, wz-wxwz-W4);
/*** Neighbor 7: (1,1,1) */
APPEND_NEIGHBOR(off+u7, 1-W3-wy-wz+wywz);
/* Update the joint histogram using the desired interpolation technique */
interpolate(i, H, clampJ, Jnn, W, nn, interp_params);
} /* End of IF TRANSFORMS INSIDE */
/* Update source index */
PyArray_ITER_NEXT(iterI);
} /* End of loop over voxels */
return;
}
int joint_histogram(PyArrayObject* JH,
unsigned int clampI,
unsigned int clampJ,
PyArrayIterObject* iterI,
const PyArrayObject* imJ_padded,
const PyArrayObject* Tvox,
long interp)
{
const signed short* J=(signed short*)imJ_padded->data;
size_t dimJX=imJ_padded->dimensions[0]-2;
size_t dimJY=imJ_padded->dimensions[1]-2;
size_t dimJZ=imJ_padded->dimensions[2]-2;
signed short Jnn[8];
double W[8];
signed short *bufI, *bufJnn;
double *bufW;
signed short i, j;
size_t off;
size_t u2 = imJ_padded->dimensions[2];
size_t u3 = u2+1;
size_t u4 = imJ_padded->dimensions[1]*u2;
size_t u5 = u4+1;
size_t u6 = u4+u2;
size_t u7 = u6+1;
double wx, wy, wz, wxwy, wxwz, wywz;
double W0, W2, W3, W4;
int nn, nx, ny, nz;
double *H = (double*)PyArray_DATA(JH);
double Tx, Ty, Tz;
double *tvox = (double*)PyArray_DATA(Tvox);
void (*interpolate)(unsigned int, double*, unsigned int, const signed short*, const double*, int, void*);
void* interp_params = NULL;
prng_state rng;
/*
Check assumptions regarding input arrays. If it fails, the
function will return -1 without doing anything else.
iterI : assumed to iterate over a signed short encoded, possibly
non-contiguous array.
imJ_padded : assumed C-contiguous (last index varies faster) & signed
short encoded.
H : assumed C-contiguous.
Tvox : assumed C-contiguous:
either a 3x4=12-sized array (or bigger) for an affine transformation
or a 3xN array for a pre-computed transformation, with N equal
to the size of the array corresponding to iterI (no checking
done)
*/
if (PyArray_TYPE(iterI->ao) != NPY_SHORT) {
fprintf(stderr, "Invalid type for the array iterator\n");
return -1;
}
if ( (!PyArray_ISCONTIGUOUS(imJ_padded)) ||
(!PyArray_ISCONTIGUOUS(JH)) ||
(!PyArray_ISCONTIGUOUS(Tvox)) ) {
fprintf(stderr, "Some non-contiguous arrays\n");
return -1;
}
/* Reset the source image iterator */
PyArray_ITER_RESET(iterI);
/* Set interpolation method */
if (interp==0)
interpolate = &_pv_interpolation;
else if (interp>0)
interpolate = &_tri_interpolation;
else { /* interp < 0 */
interpolate = &_rand_interpolation;
prng_seed(-interp, &rng);
interp_params = (void*)(&rng);
}
/* Re-initialize joint histogram */
memset((void*)H, 0, clampI*clampJ*sizeof(double));
/* Looop over source voxels */
while(iterI->index < iterI->size) {
/* Source voxel intensity */
bufI = (signed short*)PyArray_ITER_DATA(iterI);
i = bufI[0];
/* Compute the transformed grid coordinates of current voxel */
Tx = *tvox; tvox++;
Ty = *tvox; tvox++;
Tz = *tvox; tvox++;
/* Test whether the current voxel is below the intensity
threshold, or the transformed point is completly outside
the reference grid */
if ((i>=0) &&
(Tx>-1) && (Tx<dimJX) &&
(Ty>-1) && (Ty<dimJY) &&
(Tz>-1) && (Tz<dimJZ)) {
/*
Nearest neighbor (floor coordinates in the padded
image, hence +1).
Notice that using the floor function doubles excetution time.
FIXME: see if we can replace this with assembler instructions.
*/
nx = FLOOR(Tx) + 1;
ny = FLOOR(Ty) + 1;
nz = FLOOR(Tz) + 1;
/* The convention for neighbor indexing is as follows:
*
* Floor slice Ceil slice
*
* 2----6 3----7 y
* | | | | ^
* | | | | |
* 0----4 1----5 ---> x
*/
/*** Trilinear interpolation weights.
Note: wx = nnx + 1 - Tx, where nnx is the location in
the NON-PADDED grid */
wx = nx - Tx;
wy = ny - Ty;
wz = nz - Tz;
wxwy = wx*wy;
wxwz = wx*wz;
wywz = wy*wz;
/*** Prepare buffers */
bufJnn = Jnn;
bufW = W;
/*** Initialize neighbor list */
off = nx*u4 + ny*u2 + nz;
nn = 0;
/*** Neighbor 0: (0,0,0) */
W0 = wxwy*wz;
APPEND_NEIGHBOR(off, W0);
/*** Neighbor 1: (0,0,1) */
APPEND_NEIGHBOR(off+1, wxwy-W0);
/*** Neighbor 2: (0,1,0) */
W2 = wxwz-W0;
APPEND_NEIGHBOR(off+u2, W2);
/*** Neightbor 3: (0,1,1) */
W3 = wx-wxwy-W2;
APPEND_NEIGHBOR(off+u3, W3);
/*** Neighbor 4: (1,0,0) */
W4 = wywz-W0;
APPEND_NEIGHBOR(off+u4, W4);
/*** Neighbor 5: (1,0,1) */
APPEND_NEIGHBOR(off+u5, wy-wxwy-W4);
/*** Neighbor 6: (1,1,0) */
APPEND_NEIGHBOR(off+u6, wz-wxwz-W4);
/*** Neighbor 7: (1,1,1) */
APPEND_NEIGHBOR(off+u7, 1-W3-wy-wz+wywz);
/* Update the joint histogram using the desired interpolation technique */
interpolate(i, H, clampJ, Jnn, W, nn, interp_params);
} /* End of IF TRANSFORMS INSIDE */
/* Update source index */
PyArray_ITER_NEXT(iterI);
} /* End of loop over voxels */
return 0;
}