Closed tfarago closed 2 years ago
Looks good, maybe make the g_warning
a g_error
however unlikely it is to occur in reality. Do you have performance numbers already? And why the heck have we never noticed that low-hanging fruit?
Right, it's changed. However, there is no performance bump because on HW side (at least on NVIDIA) nothing changes. To populate all channels with one value takes no time, or they even fake it and don't do it actually, who knows... I got 10 s for CL_INTENSITY and CL_R and 100 sinograms of size 2024x3000.
So this is a PR about compatibility, not performance optimization.
Related also to #86
Now we will finally be able to dynamically handle various channel orders supported for
CL_FLOAT
data type. I have implemented it in the buffer in a very dynamic way which will create a tiny overhead but the code is way shorter and we don't need to mess with the python ufo.numpy.