implement rescaleData as a blocked iterator using np.nditer (#1648)

* implement rescaleData_blocked

clip limits should be int if data is int

* add test for rescaleData_blocked

* dispatch to different versions depending on numpy or cupy

* make rescaleData() the only entry-point
This commit is contained in:
pijyoi 2021-04-06 14:02:52 +08:00 committed by GitHub
parent 758c038411
commit aa57c7a685
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -13,6 +13,7 @@ import re
import struct
import sys
import warnings
import math
import numpy as np
from .util.cupy_helper import getCupy
@ -1032,6 +1033,49 @@ def clip_array(arr, vmin, vmax, out=None):
return np.core.umath.clip(arr, vmin, vmax, out=out)
def _rescaleData_nditer(data_in, scale, offset, work_dtype, out_dtype, clip):
"""Refer to documentation for rescaleData()"""
data_out = np.empty_like(data_in, dtype=out_dtype)
# integer clip operations are faster than float clip operations
# so test to see if we can perform integer clipping
fits_int32 = False
if data_in.dtype.kind in 'ui' and out_dtype.kind in 'ui':
# estimate whether data range after rescale will fit within an int32.
# this means that the input dtype should be an 8-bit or 16-bit integer type.
# casting to an int32 will lose the fractional part, therefore the
# output dtype must be an integer kind.
lim_in = np.iinfo(data_in.dtype)
dst_bounds = scale * (lim_in.min - offset), scale * (lim_in.max - offset)
if dst_bounds[1] < dst_bounds[0]:
dst_bounds = dst_bounds[1], dst_bounds[0]
lim32 = np.iinfo(np.int32)
fits_int32 = lim32.min < dst_bounds[0] and dst_bounds[1] < lim32.max
it = np.nditer([data_in, data_out],
flags=['external_loop', 'buffered'],
op_flags=[['readonly'], ['writeonly', 'no_broadcast']],
op_dtypes=[None, work_dtype],
casting='unsafe',
buffersize=32768)
with it:
for x, y in it:
y[...] = x
y -= offset
y *= scale
# Clip before converting dtype to avoid overflow
if clip is not None:
if fits_int32:
# converts to int32, clips back to float32
np.core.umath.clip(y.astype(np.int32), clip[0], clip[1], out=y)
else:
clip_array(y, clip[0], clip[1], out=y)
return data_out
def rescaleData(data, scale, offset, dtype=None, clip=None):
"""Return data rescaled and optionally cast to a new dtype.
@ -1040,32 +1084,43 @@ def rescaleData(data, scale, offset, dtype=None, clip=None):
data => (data-offset) * scale
"""
if dtype is None:
dtype = data.dtype
out_dtype = data.dtype
else:
dtype = np.dtype(dtype)
out_dtype = np.dtype(dtype)
if dtype.kind in 'ui':
lim = np.iinfo(dtype)
if out_dtype.kind in 'ui':
lim = np.iinfo(out_dtype)
if clip is None:
# don't let rescale cause integer overflow
clip = lim.min, lim.max
clip = max(clip[0], lim.min), min(clip[1], lim.max)
# make clip limits integer-valued (no need to cast to int)
# this improves performance, especially on Windows
clip = [math.trunc(x) for x in clip]
if np.can_cast(data, np.float32):
work_dtype = np.float32
else:
work_dtype = np.float64
d2 = data.astype(work_dtype, copy=True)
d2 -= offset
d2 *= scale
# Clip before converting dtype to avoid overflow
if clip is not None:
clip_array(d2, clip[0], clip[1], out=d2)
cp = getCupy()
if cp and cp.get_array_module(data) == cp:
# Cupy does not support nditer
# https://github.com/cupy/cupy/issues/5021
# don't copy if no change in dtype
data = d2.astype(dtype, copy=False)
return data
data_out = data.astype(work_dtype, copy=True)
data_out -= offset
data_out *= scale
# Clip before converting dtype to avoid overflow
if clip is not None:
clip_array(data_out, clip[0], clip[1], out=data_out)
# don't copy if no change in dtype
return data_out.astype(out_dtype, copy=False)
else:
return _rescaleData_nditer(data, scale, offset, work_dtype, out_dtype, clip)
def applyLookupTable(data, lut):