<Range> Node

Discussions - ACESNext VWG – Common LUT Format (CLF)
Tags: #<Tag:0x00007fa12d0e6778>
#4

Thanks, @Thomas_Mansencal for offering your opinion on the issue and apologies for always having meetings at conflicting times.

I also don’t prefer the wording (or existence) of that note but it’s pulled directly out of the v2.0 spec (but put into a note in this draft because I was hoping to get rid of it or otherwise address it).
Historically, I think it originated when the original author’s intent for Range to serve as a scale operator needed to be reconciled with a different interpretation that Range was basically a clamp operation. Thus statement “The Range element is also used to clamp values” was born.

#5

There is a lot of history on this topic and many hours were spent by the committee during the writing of the v2 spec back in 2014. Here is my attempt at a brief summary of the relevant points:

  1. Fundamentally, the spec provides two ways of applying a shift and scale operation: Matrix and Range. The difference is that the Matrix does not clamp and the Range does.

  2. When the v2 spec was written, Jim objected to having the Range always clamp and so we added the noClamp style option. However, this is purely a convenience for CLF authors (for example, under the hood OCIO will convert a noClamp Range to a Matrix, functionally they are no different). Also, there need not be any performance hit to using a Matrix since implementors can easily optimize the case where the off-diagonal terms are 0 (as does OCIO).

  3. As much as we all strive to avoid improper clamping, clamping is needed in order to accurately implement certain transforms (including some ACES transforms). So CLF absolutely needs a way to do this.

  4. There was an assertion during the call that Range cannot map [-0.5, 222] into [0,1]. However, that is incorrect, it can do that and furthermore allows the option to clamp or not.

  5. The objections that have been raised seem to be not fundamental issues with the math but more with the “aesthetics” of the design. Please keep in mind that this was a committee effort going back over a decade and that it won’t be possible to please everyone on naming. If we were starting from scratch, I would absolutely support the idea of further discussion about naming and perhaps doing a poll with various options. However, we’re talking about v3 of the spec, and a number of companies have implemented it at this point (it’s also in OCIO now). At the end of the day we want more people to implement this and I don’t think it helps the cause to signal that the syntax is open to being redone at every revision of the spec. :wink:

  6. There was some discussion during the call about the formulas for the case where only the min bounds or only the max bounds were specified. We added those formulae in the v2 spec to clarify for implementors what to do if they get a file that does that. However, there is only one important case of that type, which is when one needs to clamp at the low end or high end but not both. If people want to remove those formulae, we could add a limitation which is to say that if a Range only provides one set of bounds (i.e., either min or max), that it cannot shift or scale.

with respect,

Doug

2 Likes
#6

Thank you @doug_walker, for taking time to put these points in writing. I’m sorry we have had to go back yet again to go through all this, especially this late in the revision process and because it’s been discussed before.

I did review the conversations from around the 2014 v2 publication timeframe re: this issue and came to the same conclusions as your bullet points.

I agree with #5 that as of right now the functionality is there to satisfy both parties, it’s hard to please everyone, and there’s no benefit to changing names (in fact it could be detrimental to adoption as it will break existing implementations). Other than for “clarity” there’s no reason to re-segment into separate nodes or otherwise rename the operator. Just keep as-is and make even clearer the functionality is already built in for both requirements - scale and clamp.

I am glad it was brought up so that nothing is overlooked but I think we’re still doing the “best” thing here in terms of servicing the need to offer a “convenience scale/offset” and also a “clamp”, as well as not unnecessarily changing the existing spec just for the sake of change. If we were building from the ground up, yes, we could consider restructuring, but it seems a step too far.

I am proceeding on to finishing the other sections. As of now, nothing has changed in terms of functionality of the <Range> node from v2 to this version of the spec. We had already reached compromise previously and unless there is some yet-unheard-extremely-compelling reason to change, I am in favor of leaving as-is.

#7

I’m not aware of what was said in the past, is it documented publicly? In any case clarity is important. Clamping functionality of the Range node could be deprecated, and a new node introduced without breaking anything.

1 Like
#8

Could a compromise be to require the Range’s style to be defined?

#9

When all of minInValue, minOutValue, maxInValue, and maxOutValue are present, if style is not specified, the default behavior is “noClamp”.

If a minInValue/minOutValue pair is present, the result shall be clamped at the low end.

If a maxInValue/maxOutValue pair is present, the result shall be clamped at the high end.

This would appear to be contradictory. Should it not read “If only a minInValue/minOutValue pair…”?

#10

I’m sorry, but this discussion on Clamping is off point. My dislike was of the use
of implicit clamping an one end but not the other based on whether a parameter was
missing or not. (This is not a good method) Doug brought up the same argument that
because the original spec was implemented differently than what it stated, that we should defer to
the only one implementation. I still disagree with that.

But all of this misses my main objection. The LUT format is intended to be able to create a ProcessLIST that works on full bit resolution files in their native float representation. But the calculation of the (if not present) scale factor ALWAYS set floats to a 0…1.0 assumption (scale=1) But this is only true if the LUT only works in a range of 0…1.0 on the input side and this is EXPLICITLY not so. A range of -0.005 to 222 is a valid float input.

My recommendation is to delete the following text, and always include both MIN and MAX values in the Range Node. There should be no missing float range assumption.

note: the full range available turns out to be a wrong statement (we are not using the full range of floats just the 0 to 1.0 subset. [This can be a poor processing result squeezing the whole float range into a 1.0 container range.)

is to use the full range available with the inBitDepth or outBitDepth attribute used in place of the missing input range or missing output range, respectively, as calculated with these equations:

In the formulae below, if the bit depth is integral, then rangescalar() is defined as:

rangescalar(bitDepthInteger) = 2bitDepth − 1 If the bit depth is specified as floating point, then:

rangescalar(float) = 1.0 If only minimum values are specified, the formula is:

scale = rangescalar(outBitDepth) rangescalar(inBitDepth)

RGBout = MAX(minOutValue, RGBin × scale + minOutValue − minInValue × scale) If only maximum values are specified, the formula is:

scale = rangescalar(outBitDepth) rangescalar(inBitDepth)

RGBout = MIN(maxOutValue, RGBin × scale + maxOutValue − maxInValue × scale)

#11

Scott: please review the spreadsheets I did at that time. They were about the calculations
of floats going wrong under the implicit missing item equations.

#12

Since Doug is not presenting both points of view, I am sorry I think I have to responsd in line.

On a ‘lot of history’ – this conversation was never concluded and we ran out of time (which is why the CLF 2.0 spec was never finished and is listed as DRAFT)

Yes, matrix allows scale and offset, but it doesn’t allow extraction of a portion of a float range
and only work on that range. (the RANGE node is for that purpose and it still works. It is the implicit clamping at one end that is the problem – the workaround is to always use all 4 metadata items but the spec should not allow an implicit operation that causes the math to go off.

Again my objection was to the implicit one-side clamping, not about the presence of clamping.

on point #6 and this previous sentence, another solution is to make explicit the clamping behavior
ClampHi, ClampLo when needed. Adding another node with a different behavior is also an understandable approach but it needs to get back to the original intent on handling the full float range (not assuming float=1.0 scale) I propose deleting the objectionable text in this thread. The focus should be on that issue not the general existing of clamping – and no it is not a naming disagreement, it is about the fundamental math.

There should be a note about use of the matrix 3x4 node to scale and offset floats in an unclamped fashion. I agree that Clamping is essential.

I am sorry to say I may not be available until late in the call and request that this discussion
be deferred until I can join.

Thanks,

Jim

#13

What if we enforced that when style is ‘noClamp’ (even when only min or max pairs are provided) then no clamp is applied. Would that work? Then nothing is implicit because you are explicitly saying not to clamp.

If we went further and listened to @Greg_Cotten, we could require that style be set, which would require explicitly stating the intended behavior.

Side note: What might be a practical use case for specifying just min or max pairs and not wanting it to clamp?

#14

Yes I did review this but I had difficulty making sense of it for the min/max only cases when mixed with in/out bit depth differences. Perhaps if I had a practical example to help wrap my head around what the intended behavior is rather than just looking at numbers - then I might be able to better grasp the numbers to see some meaning in the current behavior.

For those interested in taking a look, here is the spreadsheet Jim is referencing: range_node_eval_v4.xlsx (53.6 KB)

#15

It is certainly much preferable.

1 Like
#16

Hey Jim, it looks like the website messed up the formatting of your message. It’s not clear what you’re objecting to or what your proposal is. As Scott suggested, I think it would help to have a specific example to discuss.

#17

My proposal was to delete the text in the Range node that deals with the assumed scaling behavior if one piece is missing.

An example is for an input of 0 to 10,000 nits and an output of 0 to 1023, the scale factor that gets calculated is (1023/10000) = 0.1023 when all four parameters are present. This is the correct conversion
from a 10000.00 float to 1023. With a 5000n input, RGBout = 5000 x 0.1023 + 0 - 0 x 0.1023 = 511.5 the right answer.

Exactly per the spec, if only the max values are specified, the scale parameter becomes ( 1023/ 1.0 ) or 1023 X as a multiplier. This is 10000 Times OFF. This is the damage from assuming all floats operate in the 0 to 1.0 range.

With an RGBin of 5000n and following the spec exactly with the equations as specified for the max only case…
RGBout = MIN(1023, 5000n x 1023 + 1023 - 10000 x 1023) or MIN ( 1023, -5113977 ) gives the wrong
number in the full float input situation. (The 511 is there but just of completely the wrong magnitude).

My proposal is to delete rangescale(float)=1.0 and everything after it and require all 4 parameters with a specific style addition of ClampHI or ClampLO. This minimizes the changes to the node input but if it still seen as necessary to keep the damaged math of the RANGE node as is, create a new RANGEF node with a note about the assumption in the existing RANGE node.

Jim

#18

Thank you Jim. When I plug in your example values I get a different number. But in any case, I’m open to prohibiting the use of only one side (max or min) in a case that does not allow a simple clamp.

We’ll discuss at the meeting today …

Doug

CLF Spec Revision VWG Meeting #6 - Tuesday 9/17/19 9am PDT
#19

It took me a little while, but I now follow your example, @jim_houston, and get the same result you do. I did not realise initially that your scale=1023 was coming not from maxOutValue=1023 but rather from outBitDepth=10 which is implicit, but not explicitly stated, in your example.

It does seem to me that the maths in the current spec is just simply wrong (or at least counterintuitive) for the single ended clamping cases. I have (I believe) a simpler, float only, example which illustrates this.

maxOutValue and maxInValue are not included at all in the calculation of scale. So if inBitDepth and outBitDepth are both 'float' then:

scale = rangescalar(float) / rangescalar(float) = 1.0

If I then set maxInValue=2.0 and maxOutValue=10.0, I might expect the operation to simply multiply input values by 5, and clamp at 10.0. Therefore 0.5 input would produce 2.5. However, testing this in Python, I get:

>>> maxInValue = 2.0
>>> maxOutValue = 10.0
>>> scale = 1.0
>>> RGBin = 0.5
>>> min(maxOutValue, RGBin * scale + maxOutValue - maxInValue * scale)
8.5

Is this really the intended behaviour?

It is effectively using implicit values of minInValue=0.0 and minOutValue=8.0 which is not what I think most people would expect. Intuitively you might expect 0.0 to be used for both.

Indeed in @jim_houston’s example for the result in the second case to match that in the first, implicit minimum values of zero would have to be used. But in fact an implicit minInValue=9999.0 is used.

#20

On a separate subject, I am not particularly keen on this sentence in the spec:
“If the input and output bit depths are not the same, a conversion should take place using the range elements.”

Who “should”? Does it mean that people creating CLFs should add e.g. an explicit range mapping from 0-1 to 0-1023? Or does it mean that CLF implementers should apply bit-depth based range scaling automatically? It kind of means both, which is a bit confusing.

If a CLF creator includes a Range node with inBitDepth="32f" outBitDepth="10i" if they provide only, for example, minInValue=0.0 and minOutValue=0, a normalised float to 10-bit scaling will be automatically applied (with negatives clamped). If they provide all four range values, they need to be values which correctly apply that scaling, and inBitDepth and outBitDepth are ignored in the calculation of the mapping.

But maybe it’s only me who finds this confusing…

#21

Hi Nick,

The only scaling in the formula for the “single-sided” case is what is necessary for the bit-depth scaling. I think there is a logic to that since you would need “both sides” in order to infer both a non-default scaling and an offset. So the “single-sided” case is basically a simple offset that is corrected for any bit-depth difference. So in your example, maxInValue does get mapped to maxOutValue, which I think is the primary expectation. Your example input value of 0.5, which is 1.5 less than maxInValue, gets mapped to 8.5, which is 1.5 less than maxOutValue. So I think it is a plausible result.

But as I wrote farther up in this thread, I agree that the “single-sided” case is probably not very useful for anything other than a simple clamp. In fact, in the interest of moving this forward and trying to bring us to consensus, I’ll go ahead and make a concrete proposal:

In Scott’s PDF on OverLeaf referenced earlier in the thread, change equation 3 from:

out = MAX(minOutValue, in × scale + minOutValue − minInValue × scale)

to simply:

out = MAX(minOutValue, in × scale)

and change equation 4 to:

out = MIN(maxOutValue, in × scale)

And add a note such as

Note: If only the minimum values are specified, the values must be set such that minOutValue = minInValue × scale. Likewise, if only the maximum values are specified, the values must be set such that maxOutValue = maxInValue × scale.

(One could argue that only the minIn or minOut is necessary, but I think it is important for readability to include both.)

I think that would address the issue that you and Jim raised. (Yes?) It is a change from the v2 spec but it’s minor enough that I’m not aware of any real-world use-cases that would actually break due to that change.

Regarding your other post about scaling, I agree that sentence is confusing. Actually, I don’t think that sentence is even necessary so perhaps the best thing is just to delete it.

Doug

#22

I think it would address the issue, as long as it is clearly documented. Currently it is slightly unclear whether the document is user-facing or developer-facing. From a user perspective it would be easy to make assumptions about expected behaviour, such as implicit minimum values of zero when omitted. It would be better if the behaviour was clearly described in the text. At the moment, it is necessary to run various numbers through the provided equations (as @jim_houston and I did) to find out what the behaviour is.

Plausible, yes, once you analyse what the maths does. But I don’t think that a pure offset, based only on the difference between in and out values is what people would necessarily intuitively expect from an operator called “Range”.

out = min(in, maxInValue) * (maxOutValue / maxInValue)

is the maths I would intuitively expect, as only scaling seems more logical to me than only offsetting.

However there is no obvious equivalent for minimums only, particularly as those could often both be zero.

1 Like
#23

No for sure! To me and without looking at the doc I would expect it to be equivalent to those:

Cheers,

Thomas