Okay, as far as I understand, there is something for fractional bits / rounding in upper-left?
The actual result seems to be only 10bits wide, but after summing up 8 values, it might become 13bit wide (yet another 13bit value).
Or would it be possible that the result is bigger than 10bits, in case you haven't found all outputs yet?
For the inputs, I am not sure which is which:
On the right is 17bits with something special on a 13bit portion.
On the bottom is 13bits with something special on a 11bit portion.
So both right & bottom inputs have some "13bit" stuff in there.
Best guess might be that the bottom input can be 11bit from RLE, or 13bit from previous multiplication (eight 10bit values summed).
What are think pink symbols in lower-left? If they are allowing to force the lower 2bits to zero, then they could expand the 11bit RLE values to 13bit?
The input on the right side is strange. If it is the scaletable then it should be only 13bits wide. Are you sure that the 13bits on the right side are outputs, not inputs?
For MDEC, I've no clue why there could be a 17bit input. Ah, unless, it's a smaller value sign-expanded to 17bits. Would that make sense? I don't understand the multiplier logic & don't know if it needs sign-expansion.
Another idea would be that they might re-use the multiplier for the YUV-to-RGB conversion (which requires multiplication by some constants; though I doubt that it'd be required to 17bit resolution in that stage).
EDIT: If the multiplier is re-used for different purposes - maybe the 17bit input is intended for smaller values with optional shift-amounts?
Like "17bit = (13bit shl 4)" for scaletable calculation, and "17bit = whatever smaller value" for YUV calculation.
That "shl 4" might be useful for getting the desired amount of fractional bits in the result, just an idea.
EDIT2: What is the small blue box in lower right of the image? Does that allow to disable all inputs on right side? Or is it the sign and allows to negate (or invert) the right input? When trying to figure out the rounding, I was having the feeling that the MDEC might "negate" negative values by a "NOT" operation instead of a real "NEG" operation.