Merge pull request #1794 from pijyoi/numba_lookup

This commit is contained in:
Ogi Moore 2021-06-15 16:42:54 -07:00 committed by GitHub
commit 61dedcb4a5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 66 additions and 18 deletions

View File

@ -20,3 +20,24 @@ def rescaleData(data, scale, offset, dtype, clip):
rescale_functions[key] = func
func(data, scale, offset, clip[0], clip[1], out=data_out)
return data_out
@numba.jit(nopython=True)
def _rescale_and_lookup1d_function(data, scale, offset, lut, out):
vmin, vmax = 0, lut.shape[0] - 1
for r in range(data.shape[0]):
for c in range(data.shape[1]):
val = (data[r, c] - offset) * scale
val = min(max(val, vmin), vmax)
out[r, c] = lut[int(val)]
def rescale_and_lookup1d(data, scale, offset, lut):
# data should be floating point and 2d
# lut is 1d
data_out = np.empty_like(data, dtype=lut.dtype)
_rescale_and_lookup1d_function(data, float(scale), float(offset), lut, data_out)
return data_out
@numba.jit(nopython=True)
def numba_take(lut, data):
# numba supports only the 1st two arguments of np.take
return np.take(lut, data)

View File

@ -603,16 +603,25 @@ class ImageItem(GraphicsObject):
maxVal = xp.nextafter(maxVal, 2*maxVal)
rng = maxVal - minVal
rng = 1 if rng == 0 else rng
image = fn.rescaleData(image, scale/rng, offset=minVal, dtype=dtype, clip=(0, num_colors-1))
levels = None
fn_numba = fn.getNumbaFunctions()
if xp == numpy and image.flags.c_contiguous and dtype == xp.uint16 and fn_numba is not None:
lut, augmented_alpha = self._convert_2dlut_to_1dlut(lut)
image = fn_numba.rescale_and_lookup1d(image, scale/rng, minVal, lut)
if image.dtype == xp.uint32:
image = image[..., xp.newaxis].view(xp.uint8)
return image, None, None, augmented_alpha
else:
image = fn.rescaleData(image, scale/rng, offset=minVal, dtype=dtype, clip=(0, num_colors-1))
if image.dtype == xp.uint16 and image.ndim == 2:
image, augmented_alpha = self._apply_lut_for_uint16_mono(image, lut)
lut = None
levels = None
# image is now of type uint8
return image, levels, lut, augmented_alpha
if image.dtype == xp.uint16 and image.ndim == 2:
image, augmented_alpha = self._apply_lut_for_uint16_mono(image, lut)
lut = None
# image is now of type uint8
return image, levels, lut, augmented_alpha
def _try_combine_lut(self, image, levels, lut):
augmented_alpha = False
@ -722,22 +731,40 @@ class ImageItem(GraphicsObject):
# if we are contiguous, we can take a faster codepath where we
# ensure that the lut is 1d
if lut.ndim == 2:
if lut.shape[1] == 3: # rgb
# convert rgb lut to rgba so that it is 32-bits
lut = xp.column_stack([lut, xp.full(lut.shape[0], 255, dtype=xp.uint8)])
augmented_alpha = True
if lut.shape[1] == 4: # rgba
lut = lut.view(xp.uint32)
lut, augmented_alpha = self._convert_2dlut_to_1dlut(lut)
fn_numba = fn.getNumbaFunctions()
if xp == numpy and fn_numba is not None:
image = fn_numba.numba_take(lut, image)
else:
image = lut[image]
image = lut.ravel()[image]
lut = None
# now both levels and lut are None
if image.dtype == xp.uint32:
image = image.view(xp.uint8).reshape(image.shape + (4,))
image = image[..., xp.newaxis].view(xp.uint8)
return image, augmented_alpha
def _convert_2dlut_to_1dlut(self, lut):
# converts:
# - uint8 (N, 1) to uint8 (N,)
# - uint8 (N, 3) or (N, 4) to uint32 (N,)
# this allows faster lookup as 1d lookup is faster
xp = self._xp
augmented_alpha = False
if lut.ndim == 1:
return lut, augmented_alpha
if lut.shape[1] == 3: # rgb
# convert rgb lut to rgba so that it is 32-bits
lut = xp.column_stack([lut, xp.full(lut.shape[0], 255, dtype=xp.uint8)])
augmented_alpha = True
if lut.shape[1] == 4: # rgba
lut = lut.view(xp.uint32)
lut = lut.ravel()
return lut, augmented_alpha
def _try_make_qimage(self, image, levels, lut, augmented_alpha):
xp = self._xp