STORAGE_TYPES.float is same as STORAGE_TYPES.double

[ansstuff]

Is this intended?

When reading processed image data from Wand back to Numpy array, I've noticed the following:

• outputArray = numpy.array(img, dtype=numpy.float32) treats image data as uint8 bytes, so it's useless when image has HDR float32 data
• okay, so instead I use export_pixels() which has storage arg
• but outputArray = img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "float") outputs data as float64, exactly like outputArray = img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "double")
• so I do post-conversion outputArray = numpy.array(img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "float"), dtype=numpy.float32)

Wouldn't it be better to treat STORAGE_TYPES.float as float32 instead of it being redundant copy of STORAGE_TYPES.double?

[emcconville]

Is this intended

Yes.

outputArray = numpy.array(img, dtype=numpy.float32) treats image data as uint8 bytes, so it's useless when image has HDR float32 data

Wand's __array_interface__ defaults Q8 for compatibility with other utilities that use numpy as a proxy.

okay, so instead I use export_pixels() which has storage arg

That would be the correct way to export raw pixel data.

but outputArray = img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "float") outputs data as float64, exactly like outputArray = img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "double")

Very possible on modern hardware. Each export will have a value between 0.0 & 1.0 with different memory allocations. If I remember correctly, the values are calculated by something like...

floor( Value / QuantumRange + 0.5 ) / 1.0

so I do post-conversion outputArray = numpy.array(img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "float"), dtype=numpy.float32)

Don't waste time with floats. If HDRI is important, ensure ImageMagick is compiled with HDRI support, and export to the quantum storage type. That's ImageMagick's internal storage. Then once exported, you should be able to convert as needed. I think it would look something like...

outputArray = img.export_pixels(0, 0, imgWidth, imgHeight, "RGBA", "quantum")
outputArray *= wand.version.QUANTUM_SCALE  # Convert values between 0.0 & 1.0

Wouldn't it be better to treat STORAGE_TYPES.float as float32 instead of it being redundant copy of STORAGE_TYPES.double?

Not all hardware architecture may support this. As long as Python's ctypes.c_float matches C's float, and etc, things should work. Trying to be smart with float32, and you'll have overflow & faults on ppc, arm, and all those exotic IBM mainframes.

[ansstuff]

I see.

Don't waste time with floats.

My input and output are in float32 arrays, so I don't think I'm losing any precision by exporting pixels to float64 and relying on built-in normalization.