imops 0.8.1__cp38-cp38-win32.whl → 0.8.3__cp38-cp38-win32.whl

imops/src/_fast_morphology.pyx

@@ -11,135 +11,319 @@ cimport numpy as np
 
 from cython.parallel import prange
 
+from libc.stdlib cimport abort, free, malloc
 
-# TODO: Move generic functions like this to the separate file
-cdef inline np.uint8_t get_pixel3d(np.uint8_t* input,
-                                   Py_ssize_t rows, Py_ssize_t cols, Py_ssize_t dims,
-                                   Py_ssize_t r, Py_ssize_t c, Py_ssize_t d,
-                                   np.uint8_t cval, layout = b'C') nogil:
-    if (r < 0) or (r >= rows) or (c < 0) or (c >= cols) or (d < 0) or (d >= dims):
-        return cval
-
-    if layout == b'C':
-        return input[r * cols * dims + c * dims + d]
-
-    if layout == b'F':
-        return input[rows * cols * d + rows * c + r]
-
-
-cdef inline np.uint8_t max_in_footprint(np.uint8_t* input, np.uint8_t* footprint,
-                                        Py_ssize_t r, Py_ssize_t c, Py_ssize_t d,
-                                        Py_ssize_t rows, Py_ssize_t cols, Py_ssize_t dims,
-                                        Py_ssize_t f_rows, Py_ssize_t f_cols, Py_ssize_t f_dims) nogil:
-    cdef Py_ssize_t i, j, k
-    cdef Py_ssize_t i_r, i_c, i_d
-
-    cdef int rows_shift = r + f_rows // 2
-    cdef int cols_shift = c + f_cols // 2
-    cdef int dims_shift = d + f_dims // 2
-
-    for i in range(f_rows):
-        i_r = rows_shift - i
-        for j in range(f_cols):
-            i_c = cols_shift - j
-            for k in range(f_dims):
-                i_d = dims_shift - k
-
-                if (
-                    get_pixel3d(
-                        footprint,
-                        f_rows, f_cols, f_dims,
-                        i, j, k,
-                        False,
-                    ) and
-                    get_pixel3d(
-                        input,
-                        rows, cols, dims,
-                        i_r, i_c, i_d,
-                        False,
-                    )
-                ):
-                    return True
-
-    return False
-
-
-cdef inline np.uint8_t min_in_footprint(np.uint8_t* input, np.uint8_t* footprint,
-                                        Py_ssize_t r, Py_ssize_t c, Py_ssize_t d,
-                                        Py_ssize_t rows, Py_ssize_t cols, Py_ssize_t dims,
-                                        Py_ssize_t f_rows, Py_ssize_t f_cols, Py_ssize_t f_dims) nogil:
-    cdef Py_ssize_t i, j, k
-    cdef Py_ssize_t i_r, i_c, i_d
-
-    cdef int rows_shift = r - f_rows // 2
-    cdef int cols_shift = c - f_cols // 2
-    cdef int dims_shift = d - f_dims // 2
-
-    for i in range(f_rows):
-        i_r = rows_shift + i
-        for j in range(f_cols):
-            i_c = cols_shift + j
-            for k in range(f_dims):
-                i_d = dims_shift + k
-
-                if (
-                    get_pixel3d(
-                        footprint,
-                        f_rows, f_cols, f_dims,
-                        i, j, k,
-                        False,
-                    ) and not
-                    get_pixel3d(
-                        input,
-                        rows, cols, dims,
-                        i_r, i_c, i_d,
-                        True,
-                    )
-                ):
-                    return False
-
-    return True
-
-
-def _binary_dilation(np.uint8_t[:, :, :] input, np.uint8_t[:, :, :] footprint, Py_ssize_t num_threads):
-    cdef np.uint8_t[:, :, ::1] contiguous_input = np.ascontiguousarray(input)
-    cdef np.uint8_t[:, :, ::1] contiguous_footprint = np.ascontiguousarray(footprint)
-
-    cdef np.uint8_t[:, :, ::1] dilated = np.zeros_like(input, dtype=np.uint8)
-
-    cdef Py_ssize_t rows = input.shape[0], cols = input.shape[1], dims = input.shape[2]
-    cdef Py_ssize_t f_rows = footprint.shape[0], f_cols = footprint.shape[1], f_dims = footprint.shape[2]
-    cdef Py_ssize_t i, j, k
-
-    for i in prange(rows, nogil=True, num_threads=num_threads):
-        for j in prange(cols):
-            for k in prange(dims):
-                dilated[i, j, k] = max_in_footprint(
-                    &contiguous_input[0, 0, 0],
-                    &contiguous_footprint[0, 0, 0],
-                    i, j, k, rows, cols, dims, f_rows, f_cols, f_dims
-                )
-
-    return np.asarray(dilated)
-
-
-def _binary_erosion(np.uint8_t[:, :, :] input, np.uint8_t[:, :, :] footprint, Py_ssize_t num_threads):
-    cdef np.uint8_t[:, :, ::1] contiguous_input = np.ascontiguousarray(input)
-    cdef np.uint8_t[:, :, ::1] contiguous_footprint = np.ascontiguousarray(footprint)
-
-    cdef np.uint8_t[:, :, ::1] eroded = np.zeros_like(input, dtype=np.uint8)
-
-    cdef Py_ssize_t rows = input.shape[0], cols = input.shape[1], dims = input.shape[2]
-    cdef Py_ssize_t f_rows = footprint.shape[0], f_cols = footprint.shape[1], f_dims = footprint.shape[2]
-    cdef Py_ssize_t i, j, k
-
-    for i in prange(rows, nogil=True, num_threads=num_threads):
-        for j in prange(cols):
-            for k in prange(dims):
-                eroded[i, j, k] = min_in_footprint(
-                    &contiguous_input[0, 0, 0],
-                    &contiguous_footprint[0, 0, 0],
-                    i, j, k, rows, cols, dims, f_rows, f_cols, f_dims
-                )
-
-    return np.asarray(eroded)
+
+cdef struct array_iterator:
+    int rank_m1
+    int dimensions[3]
+    int coordinates[3]
+    int strides[3]
+    int backstrides[3]
+
+
+cdef int init_array_iterator(array_iterator *arr_iter, int *shape, int *strides) nogil:
+    cdef int i
+
+    arr_iter.rank_m1 = 2
+
+    for i in range(3):
+        arr_iter[0].dimensions[i] = shape[i] - 1
+        arr_iter[0].coordinates[i] = 0
+        arr_iter[0].strides[i] = strides[i]
+        arr_iter[0].backstrides[i] = strides[i] * arr_iter[0].dimensions[i]
+
+    return 0
+
+
+cdef struct filter_iterator:
+    int strides[3]
+    int backstrides[3]
+    int bound1[3]
+    int bound2[3]
+
+
+cdef int init_filter_iterator(
+    filter_iterator *filter_iter,
+    int *a_shape,
+    int *f_shape,
+    int filter_size,
+    np.uint8_t is_dilation
+) nogil:
+    cdef int i, rank = 3, step, orgn
+
+    filter_iter[0].strides[rank - 1] = filter_size
+    for i in range(rank - 2, -1, -1):
+        step = a_shape[i + 1] if a_shape[i + 1] < f_shape[i + 1] else f_shape[i + 1]
+        filter_iter[0].strides[i] = filter_iter[0].strides[i + 1] * step
+
+    for i in range(rank):
+        step = a_shape[i] if a_shape[i] < f_shape[i] else f_shape[i]
+        orgn = f_shape[i] // 2
+        if is_dilation and f_shape[i] % 2 == 0:
+            orgn -= 1
+
+        filter_iter[0].backstrides[i] = (step - 1) * filter_iter[0].strides[i]
+        filter_iter[0].bound1[i] = orgn
+        filter_iter[0].bound2[i] = a_shape[i] - f_shape[i] + orgn
+
+    return 0
+
+
+cdef int init_filter_offsets(
+    int *a_shape, int *a_strides,
+    np.uint8_t *footprint, int *f_shape,
+    int footprint_size,
+    int **offsets, int *border_flag_value,
+    np.uint8_t is_dilation
+) nogil:
+    cdef int i, j, k
+    cdef int filter_size = 1, offsets_size = 1
+    cdef int max_size = 0, max_stride = 0
+    cdef int coordinates[3]
+    cdef int position[3]
+    cdef int *po
+    cdef int stride, offset, orgn, cc
+
+    for i in range(3):
+        filter_size *= f_shape[i]
+
+    for i in range(3):
+        offsets_size *= (a_shape[i] if a_shape[i] < f_shape[i] else f_shape[i])
+
+    offsets[0] = <int*>malloc(offsets_size * footprint_size * sizeof(int))
+    if offsets[0] == NULL:
+        abort()
+
+    for i in range(3):
+        if a_shape[i] > max_size:
+            max_size = a_shape[i]
+
+        stride = -a_strides[i] if a_strides[i] < 0 else a_strides[i]
+
+        if stride > max_stride:
+            max_stride = stride
+
+        coordinates[i] = 0
+        position[i] = 0
+
+    border_flag_value[0] = max_size * max_stride + 1
+    po = offsets[0]
+
+    for j in range(offsets_size):
+        for k in range(filter_size):
+            offset = 0
+
+            # only calculate an offset if the footprint is 1:
+            if footprint[k]:
+                # find offsets along all axes:
+                for i in range(3):
+                    orgn = f_shape[i] // 2
+                    if is_dilation and f_shape[i] % 2 == 0:
+                        orgn -= 1
+                    cc = coordinates[i] - orgn + position[i]
+
+                    if cc < 0 or cc >= a_shape[i]:
+                        offset = border_flag_value[0]
+                        break
+                    else:
+                        # use an offset that is possibly mapped from outside the border:
+                        cc -= position[i]
+                        offset += a_strides[i] * cc
+
+                # store the offset
+                po[0] = offset
+                po += 1
+
+            # next point in the filter:
+            for i in range(3 - 1, -1, -1):
+                if coordinates[i] < f_shape[i] - 1:
+                    coordinates[i] += 1
+                    break
+                else:
+                    coordinates[i] = 0
+
+        # move to the next array region:
+        for i in range(3 - 1, -1, -1):
+            orgn = f_shape[i] // 2
+            if is_dilation and f_shape[i] % 2 == 0:
+                orgn -= 1
+
+            if position[i] == orgn:
+                position[i] += a_shape[i] - f_shape[i] + 1
+
+                if position[i] <= orgn:
+                    position[i] = orgn + 1
+            else:
+                position[i] += 1
+
+            if position[i] < a_shape[i]:
+                break
+            else:
+                position[i] = 0
+
+    return 0
+
+
+cdef inline int filter_iterator_offset(filter_iterator *filter_iter, int *coordinates) nogil:
+    cdef int i, position, offset = 0
+
+    for i in range(3):
+        if filter_iter[0].bound1[i] > filter_iter[0].bound2[i]:
+            position = coordinates[i]
+        elif coordinates[i] < filter_iter[0].bound1[i]:
+            position = coordinates[i]
+        elif coordinates[i] >= filter_iter[0].bound2[i]:
+            position = filter_iter[0].bound1[i] + coordinates[i] - filter_iter[0].bound2[i]
+        else:
+            position = filter_iter[0].bound1[i]
+
+        offset += filter_iter[0].strides[i] * position
+
+    return offset
+
+
+cdef inline int worker(
+    np.uint8_t *input, np.uint8_t *footprint, np.uint8_t *output,
+    int *a_shape, int *a_strides, int *f_shape,
+    int footprint_size,
+    int *offsets, int border_flag_value,
+    int start, int end,
+    np.uint8_t border_value,
+    np.uint8_t is_dilation,
+) nogil:
+    cdef np.uint8_t _true = not is_dilation
+    cdef np.uint8_t _false = not _true
+
+    cdef array_iterator input_iter
+    cdef filter_iterator filter_iter
+
+    init_array_iterator(&input_iter, a_shape, a_strides)
+
+    init_filter_iterator(&filter_iter, a_shape, f_shape, footprint_size, is_dilation)
+
+    cdef int temp = start
+    for i in range(3):
+        input_iter.coordinates[i] = temp // input_iter.strides[i]
+        temp %= input_iter.strides[i]
+
+    cdef np.uint8_t *pi = input + start
+    cdef np.uint8_t *po = output + start
+    cdef int *oo = offsets + filter_iterator_offset(&filter_iter, <int*>input_iter.coordinates)
+
+    cdef np.uint8_t out
+    cdef int _oo, _pp
+
+    for j in range(start, end):
+        out = _true
+
+        for i in range(footprint_size):
+            _oo = oo[i]
+
+            if _oo == border_flag_value:
+                if border_value == _false:
+                    out = _false
+                    break
+            elif pi[_oo] == _false:
+                out = _false
+                break
+
+        po[0] = out
+
+        # pointers and filter next
+        for i in range(input_iter.rank_m1, -1, -1):
+            _pp = input_iter.coordinates[i]
+
+            if _pp < input_iter.dimensions[i]:
+                if _pp < filter_iter.bound1[i] or _pp >= filter_iter.bound2[i]:
+                    oo += filter_iter.strides[i]
+
+                input_iter.coordinates[i] += 1
+                break
+            else:
+                input_iter.coordinates[i] = 0
+
+                oo -= filter_iter.backstrides[i]
+
+        pi += 1
+        po += 1
+
+    return 0
+
+
+def _binary_operation(
+    np.uint8_t[:, :, :] input,
+    np.uint8_t[:, :, :] footprint,
+    np.uint8_t[:, :, ::1] out,
+    Py_ssize_t num_threads,
+    np.uint8_t border_value,
+    np.uint8_t is_dilation,
+) -> np.ndarray:
+    cdef np.uint8_t[:, :, ::1] c_input = np.ascontiguousarray(input)
+    cdef np.uint8_t[:, :, ::1] c_footprint = np.ascontiguousarray(footprint)
+
+    cdef int f_shape[3]
+    cdef int a_shape[3]
+    cdef int a_strides[3]
+    for i in range(3):
+        f_shape[i] = c_footprint.shape[i]
+        a_shape[i] = c_input.shape[i]
+        a_strides[i] = c_input.strides[i]
+
+    cdef int footprint_size = np.sum(c_footprint)
+    cdef int size = np.size(c_input)
+
+    cdef int *offsets
+    cdef int border_flag_value
+
+    init_filter_offsets(
+        a_shape, a_strides,
+        &c_footprint[0, 0, 0], f_shape, footprint_size,
+        &offsets, &border_flag_value,
+        is_dilation
+    )
+
+    cdef int task, start, chunk
+
+    cdef int _mod = size % num_threads, _div = size // num_threads
+
+    for task in prange(num_threads, nogil=True, num_threads=num_threads):
+        if task < _mod:
+            chunk = _div + 1
+            start = chunk * task
+        else:
+            chunk = _div
+            start = chunk * task + _mod
+
+        worker(
+            &c_input[0, 0, 0], &c_footprint[0, 0, 0], &out[0, 0, 0],
+            a_shape, a_strides, f_shape,
+            footprint_size,
+            offsets, border_flag_value,
+            start, start + chunk,
+            border_value,
+            is_dilation
+        )
+
+    free(offsets)
+
+    return np.asarray(out)
+
+
+def _binary_erosion(
+    np.uint8_t[:, :, :] input,
+    np.uint8_t[:, :, :] footprint,
+    np.uint8_t[:, :, ::1] out,
+    Py_ssize_t num_threads,
+) -> np.ndarray:
+    return _binary_operation(input, footprint, out, num_threads, True, False)
+
+
+def _binary_dilation(
+    np.uint8_t[:, :, :] input,
+    np.uint8_t[:, :, :] footprint,
+    np.uint8_t[:, :, ::1] out,
+    Py_ssize_t num_threads,
+) -> np.ndarray:
+    inverted_footprint = np.array(footprint[::-1, ::-1, ::-1])
+    return _binary_operation(input, inverted_footprint, out, num_threads, False, True)