While improving the numba performance for the float with nans case, we also found out how to improve the numba performance for the float with no nans case. (24ms --> 12ms)

The gap between numpy vs numba has widened; and the gap between numba and cupy has narrowed.

testable with:

VideoSpeedTest.py --dtype float --size 6000x4000 --levels 0,2

On a particular machine, improvement was from 38 fps to 57 fps. (50% improvement)

Revisiting the "optimized" rescaleData for uint16, it has now turned into a pessimization.
Remove the code as it is difficult to maintain for no benefit.

Surprisingly, for an uint16 image the same size as the float32 image above, the float32 image is running (slightly) faster.

import pyqtgraph as pg
import numpy as np

pg.setConfigOptions(imageAxisOrder='row-major')
# pg.setConfigOptions(useNumba=True)

Y = np.sinc(np.linspace(-6, 10, 4000))[:, np.newaxis]
X = np.sinc(np.linspace(-15, 9, 6000))
data = np.abs(Y * X)

data *= 2000
data = data.astype(np.uint16)

pg.mkQApp()
imv = pg.ImageView()
imv.show()
imv.setImage(data, levels=(0.0, 1000.0))
hlut = imv.getHistogramWidget()
hlut.gradient.setColorMap(pg.colormap.get("CET-C1"))
hlut.gradient.showTicks(False)
pg.exec()
print(imv.imageItem.qimage.format())

Change uint16 mono images to use functions_numba.rescale_only instead of functions_numba.rescaleData.

VideoSpeedTest.py --dtype uint16 --size 6000x4000 --levels 0,10000

On a particular machine, improvement was from 46 fps to 83 fps. (80% improvement)

Added support for float RGB images with nans.
The supported data types can thus be summarized as (uint8, uint16, float32, float64) x (1 chan, 3 chans) provided no multi-channel levels are used.

import pyqtgraph as pg
import numpy as np
pg.setConfigOptions(imageAxisOrder='row-major')
# pg.setConfigOptions(useNumba=True)

bias = 50

Y = np.sinc(np.linspace(-6, 10, 4000))[:, np.newaxis]
X = np.sinc(np.linspace(-15, 9, 6000))
data = np.abs(Y * X)
mask1 = data > 0.95
mask2 = (0.5 < data) & (data < 0.6)
data = 10*np.log10(data)
data += bias
data = np.dstack((data, data * 0.5, data * 0.25))
data[..., 1][mask1] = np.nan
data[..., 2][mask2] = np.nan
data = data.astype(np.float32)

pg.mkQApp()
imv = pg.ImageView(levelMode='mono')
imv.show()
imv.setImage(data, levels=(-50 + bias, 0 + bias))
pg.exec()
print(imv.imageItem.qimage.format())

that numba rescale_and_clip function seems to be exactly what's needed haha, wow, that's really fast.

Every time I think you've gotten as much out of ImageItem performance as we can get, you prove me wrong 😆 Thanks for the PR @pijyoi Merging.