Originally Posted by erkq
The LUT simply holds the transformation data generated by the calibration process. All that math still has to be done in real time by the LUT holder. It may even be more math, given the quantity of data in the LUT compared to the settings of the Panny/Fury.
I'm not sure you fully understand the math that is required to do either 3DLUT processing or tone mapping.
For 3DLUT processing you read the 8 nearest 3DLUT data points and use trilinear interpolation between them. That's very simple math, requires no "pow" instructions, and besides would be needed for high quality calibration, anyway. So baking the tone mapping into the 3DLUT consumes no extra power over just using a 3DLUT calibration without tone mapping.
When talking about tone mapping algorithms, if you simply clip in YCbCr that's ultra easy and should consume less power than applying a 3DLUT. But who wants simple YCbCr clipping? It looks *awful*.
For high quality tone mapping, you should convert YCbCr to ICtCp, then apply the tone mapping in ICtCp, then reduce either saturation or luminance (or both) where colors are out of gamut after tone mapping. Finally, the resulting ICtCp data needs to be converted back to YCbCr. This whole processing chain consumes quite a lot of power, involving many matrix multiplications and "pow" instructions, which I think are expensive for a FPGA.
To give you some real world data: A Skylake GPU can do 3DLUT processing for 4Kp24 without any problems (even when using a massive 256^3 3DLUT, which is much bigger than the 65^3 3DLUT used by Lumagen). But the same Skylake GPU can't run highest quality tone mapping algorithms in real time for 4Kp24.
Now of course I don't know which exact tone mapping algorithms the Pany uses, but I highly doubt it can compete in quality with what Calman would be able to bake into a 3DLUT.