def _text_contains_case_sensitive_numba(
length: int,
valid_bits: np.ndarray,
valid_offset: int,
offsets: np.ndarray,
data: np.ndarray,
pat: bytes,
) -> np.ndarray:
failure_function = compute_kmp_failure_function(pat)
# Initialise boolean (bit-packaed) output array.
output_size = length // 8
if length % 8 > 0:
output_size += 1
output = np.empty(output_size, dtype=np.uint8)
if length % 8 > 0:
# Zero trailing bits
output[-1] = 0
has_nulls = valid_bits.size > 0
for row_idx in range(length):
if has_nulls and not _check_valid_row(row_idx, valid_bits,
valid_offset):
continue
matched_len = 0
contains = False
for str_idx in range(offsets[row_idx], offsets[row_idx + 1]):
if matched_len == len(pat):
contains = True
break
# Manually inlined utils.kmp.append_to_kmp_matching for
# performance
while matched_len > -1 and pat[matched_len] != data[str_idx]:
matched_len = failure_function[matched_len]
matched_len = matched_len + 1
if matched_len == len(pat):
contains = True
# Write out the result into the bit-mask
byte_offset_result = row_idx // 8
bit_offset_result = row_idx % 8
mask_result = np.uint8(1 << bit_offset_result)
current = output[byte_offset_result]
if contains: # must be logical, not bit-wise as different bits may be flagged
output[byte_offset_result] = current | mask_result
else:
output[byte_offset_result] = current & ~mask_result
return output
def _text_count_case_sensitive_numba(
length: int,
valid_bits: np.ndarray,
valid_offset: int,
offsets: np.ndarray,
data: np.ndarray,
pat: bytes,
) -> np.ndarray:
failure_function = compute_kmp_failure_function(pat)
output = np.empty(length, dtype=np.int64)
has_nulls = valid_bits.size > 0
for row_idx in range(length):
if has_nulls and not _check_valid_row(row_idx, valid_bits,
valid_offset):
continue
matched_len = 0
output[row_idx] = 0
if len(pat) == 0:
output[row_idx] = offsets[row_idx + 1] - offsets[row_idx] + 1
continue
for str_idx in range(offsets[row_idx], offsets[row_idx + 1]):
# Manually inlined utils.kmp.append_to_kmp_matching for performance
while matched_len > -1 and pat[matched_len] != data[str_idx]:
matched_len = failure_function[matched_len]
matched_len = matched_len + 1
if matched_len == len(pat):
output[row_idx] += 1
# `matched_len=0` ensures overlapping matches are not counted.
# This matches the behavior of Python's builtin `count`
# function.
matched_len = 0
return output
def _text_replace_case_sensitive_numba(
length: int,
valid_bits: np.ndarray,
valid_offset: int,
offsets: np.ndarray,
data: np.ndarray,
pat: bytes,
repl: bytes,
max_repl: int,
) -> Tuple[np.ndarray, np.ndarray]:
failure_function = compute_kmp_failure_function(pat)
# Computes output buffer offsets
output_offsets = np.empty(length + 1, dtype=np.int32)
cumulative_offset = 0
has_nulls = valid_bits.size > 0
match_len_change = len(repl) - len(pat)
for row_idx in range(length):
output_offsets[row_idx] = cumulative_offset
if has_nulls and not _check_valid_row(row_idx, valid_bits,
valid_offset):
continue
row_len = offsets[row_idx + 1] - offsets[row_idx]
cumulative_offset += row_len
matched_len = 0
matches_done = 0
for str_idx in range(offsets[row_idx], offsets[row_idx + 1]):
# Manually inlined utils.kmp.append_to_kmp_matching for performance
while matched_len > -1 and pat[matched_len] != data[str_idx]:
matched_len = failure_function[matched_len]
matched_len = matched_len + 1
if matched_len == len(pat):
matches_done += 1
matched_len = 0
if matches_done == max_repl:
break
cumulative_offset += match_len_change * matches_done
output_offsets[length] = cumulative_offset
output_buffer = np.empty(cumulative_offset, dtype=np.uint8)
output_pos = 0
for row_idx in range(length):
if has_nulls and not _check_valid_row(row_idx, valid_bits,
valid_offset):
continue
matched_len = 0
matches_done = 0
write_idx = offsets[row_idx]
for read_idx in range(offsets[row_idx], offsets[row_idx + 1]):
# A modified version of utils.kmp.append_to_kmp_matching
while matched_len > -1 and pat[matched_len] != data[read_idx]:
matched_len = failure_function[matched_len]
matched_len = matched_len + 1
if read_idx - write_idx == len(pat):
output_buffer[output_pos] = data[write_idx]
output_pos += 1
write_idx += 1
if matched_len == len(pat):
matched_len = 0
if matches_done != max_repl:
matches_done += 1
write_idx = read_idx + 1
for char in repl:
output_buffer[output_pos] = char
output_pos += 1
while write_idx < offsets[row_idx + 1]:
output_buffer[output_pos] = data[write_idx]
output_pos += 1
write_idx += 1
return output_offsets, output_buffer