Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Cleanups having to do with fidelity of sets of instructions #916

Open
wants to merge 1 commit into
base: master
Choose a base branch
from

Conversation

macrologist
Copy link
Contributor

Move defs near to use sites; refactor instrs fidelity

Fidelity calculations used by the compressor and by the fidelity addresser were needlessly constructing logical schedules.

Peeking at the logical schedule fidelity calculations, they too have been cleaned up and refatored for clarity and performance.

Copy link
Member

@stylewarning stylewarning left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I did a once-over to check for obvious things. Looks pretty good.

In my initial review though I was a bit confused about using the minimum. It seemed the old code didn't actually calculate minima, but it mentioned it in a doc string?

I'd personally like some clarity on that.

Not approving or requesting changes until I understand it better.

src/addresser/fidelity-addresser.lisp Outdated Show resolved Hide resolved
(defun calculate-instructions-log-fidelity (instructions chip-specification)
"Calculates the fidelity of a sequence of native INSTRUCTIONS on a chip with architecture governed by CHIP-SPECIFICATION (and with assumed perfect parallelization across resources)."
(reduce #'+ instructions
:key (lambda (instr)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

might be silly but maybe we should rewrite the reduce as a loop to avoid the closure, or you should FLET it (it would have a nice name) and declare it dynamic-extent

SBCL is maybe smart enough to do this already; I don't know

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I could use loop - I was just too lazy to determine the container type of instructions.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Opted for flet instead: citing above laziness.

(let ((running-fidelity 0d0))
(map-lschedule-in-topological-order
lschedule
(lambda (instr)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

consider flet + DX

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

what do you mean by + DX? Are you asking that I not use incf ? I can't tell

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I refactored to use flet.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

DX is SBCLian slang for "dynamic extent (declaration)"

(let ((min-fidelity 1.0d0))
(map-lschedule-in-topological-order
lschedule
(lambda (instr)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

consider flet + DX

src/chip/chip-specification.lisp Outdated Show resolved Hide resolved
(let (fidelity)
(a:when-let* ((obj (lookup-hardware-object chip-spec instr))
(specs-hash (hardware-object-gate-information obj))
(binding (and (< 0 (hash-table-count specs-hash))
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

plusp


(defun calculate-instructions-fidelity (instructions chip-specification)
"Calculates the fidelity of a sequence of native INSTRUCTIONS on a chip with architecture governed by CHIP-SPECIFICATION (and with assumed perfect parallelization across resources)."
(exp (- (calculate-instructions-log-fidelity instructions chip-specification))))
(reduce #'min instructions
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this correct? why the minimum across instructions?

Copy link
Contributor Author

@macrologist macrologist Feb 8, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes so, the old function here depended upon a mathematical trick that approximated the minimum fidelity by passing the negative sum of log squared of individual fidelities to an exponential function. What you got was not the minimum fidelity but was "close".

This change cuts out the middle man, gives you the minimum, and doesn't have to calculate the sum of log squared terms.

The function as it had appeared in compressor was more-or-less copied from its similarly named function in logical-schedule. The trick was explained there in the (old) docstring to lschedule-calculate-fidelity. I.e.

min(x,y) ~ exp(-(sqrt (ln(x)^2 + ln(y)^2))).

I assume this trick was used in order to recycle the lschedule-calculate-log-fidelity function. But that recycling isn't really worth it, in my opinion: it is much faster to just get the actual minimum.

Copy link
Contributor Author

@macrologist macrologist Feb 8, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh, and I assume it wants the minimum because the doc string of the function that this function is directly imitating literally states that's what it is trying to do.

Copy link
Contributor

@ecpeterson ecpeterson Feb 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Huh. I think I'm originally responsible for this goofy expression, and I don't think I was looking for an approximate minimum — if x = y, for instance, you won't get min(x, y) = x but x^{-sqrt 2}. I also don't think I'm responsible for that old docstring; there was a refactor in 2019–2020 where someone else did their best to read the tea leaves. That said, I don't think there's any true justification to what I wrote, as fidelities do not easily compose, neither horizontally nor vertically. If you want something with at least some thin mathematical justification, you could consider using sum everywhere, since infinitesimal infidelities do compose OK and we're all paid to be very optimistic about long-term hardware performance.

Copy link
Contributor Author

@macrologist macrologist Feb 9, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I suppose this is a way to use the composition-as-a-sum-of-log-squares and normalize it to [0,1] so that it lives in the same set as the fidelity scores?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That is nice that it keeps to [0, 1]; I'm not sure if that's accidental or intentional. On the other hand, the minimum fidelity of an nQ gate is 1/(1 + 2^n), so maybe not that nice. And sums of infinitesimals won't escape [0, 1] either. I dunno.

@macrologist
Copy link
Contributor Author

macrologist commented Feb 9, 2024

Some of the changes I have made here have to do with removing needless logical schedule construction. The assessment of needlessness was based on looking at the code as it actually is (and determining that extra work was being done), but not necessarily what it was intended to be.

Specifically, all of these fidelity calculations first construct a logical schedule, but then use that schedule to iterate over every instruction, adding up the (log squared) fidelities of each, and then either using that sum or using the (exp -sum) in some kind of comparison. In the calculations as they exist presently, the ordering of operations in a logical schedule doesn't seem to be relevant.

But! maybe they're meant to be. Maybe it was decided that the logical schedule is the place to calculate fidelities b/c there may at some point be a better model of whole-program fidelity calculation that takes into account the logical ordering of operations.

I now have some doubts about the eventual emergence of such a model based on what @ecpeterson mentioned above: that fidelities don't generally compose well.

So, I don't know if this PR should remove the instantiation of logical schedules from the pipeline of instructions -> fidelity score because doing so may or may not be a violation of intended design. @stylewarning

@ecpeterson
Copy link
Contributor

In the calculations as they exist presently, the ordering of operations in a logical schedule doesn't seem to be relevant.
...

Something else to bear in mind is that the addresser has a cross-competing concern about instruction ordering: it wants to make good decisions so that the total resulting schedule has good fidelity, but it also needs to make a decision about what SWAPs to immediately insert, and so it can't get too wistful about instructions down the line. We dealt with this by adding a decay factor to logically-temporally distant instructions, which is fine for this scheduling application but less fine for calculating the total fidelity of a schedule.

I only bring this up to say (1) that there is a place where we jam in ordering info into these values + (2) something like that jamming is unavoidable in greedy-ish scheduling, so it's always going to distort the compiler's internal perspective on these values, and so maybe it isn't so important to come up with a really nice (/ computationally expensive) formula. (Users' external perspectives on these values are another matter! They probably do want meaningful values.)

@macrologist
Copy link
Contributor Author

Sorry I don't know if I made my concern clear. My question has to do with a very specific case: Several of these fidelity calculations start with a sequence of instructions, build a logical schedule out of them, calculate fidelity, then throw the recently built logical schedule away. In these situations, the logical schedule is only being built so that it can be passed to a fidelity calculation routine. (This is the situation in some of the compressor code as well as a sub-step in the fidelity-addresser code. )

Maybe this is fine if what you want to do is to take into account the ordering of operations in your calculation of fidelity. Right now that is not happening - the actual computation doesn't take order into account. But perhaps one might in the future expect order to be significant - if that is the case, it makes sense to create disposable logical schedules.

I just don't know if there is a meaningful way to think about fidelity on a logical schedule that doesn't also apply to simple sequences of instructions.

@ecpeterson
Copy link
Contributor

I myself think it is fine not to bother calculating these small logical schedules and to use a fairly dumb fidelity estimate which need not rely on logical ordering.

One last comment: Certain Places care foremost about coherence limitations and so only really care about the “longest chain” of instructions. I guess this is comparable to what we’re doing with the temporal strategy? It would be nice to have a similarly clear objective for the “fidelity strategy” if it’s not, in fact, computing fidelities.

Move defs near to use sites; refactor instrs fidelity

Fidelity calculations used by the compressor and by the fidelity
addresser were needlessly constructing logical schedules.

Peeking at the logical schedule fidelity calculations, they too have
been cleaned up and refatored for clarity and performance.

Generally opting for treating fidelity of a circuit as the minimum
fidelity of its constituent gates. Prior to this commit, when a composite
fidelity needed to fall in the range (0, 1] the calculation

(exp (- (sqrt (reduce #'+ fidelities
                 :key (lambda (f) (* (log f) (log f)))))))

was being used. For fidelities f in (0, 1] this produced a
non-negative value f0 that was at most the minimal f.
@macrologist macrologist force-pushed the log-sched-fidelity-cleanup branch from a1a3cdd to 3f281cf Compare February 13, 2024 23:50
@macrologist
Copy link
Contributor Author

macrologist commented Feb 14, 2024

I opted to replace most of the

(exp (- (sqrt (reduce #'+ FIDELITIES :key (lambda (f) (* (log f) (log f))))))) 

calculations with a simpler

(reduce #'min FIDELITIES) 

expression. Both will yield values in the interval (0, 1] when each member of FIDELITIES is also in (0, 1].

The first has the added benefit of getting worse as more "bad" fidelities enter and so is perhaps more reflective of compounding bad-upon-bad. The downside is that more computation is required for empirically uncertain benefit.

Today I have pushed the most recent version of these changes as a kind of indication of life. I am considering what next to do.

Some thoughts that have occurred to me:

  • re-refactor to make the (presently excised) construction of logical-schedules significant by considering, perhaps, fidelity on longest paths
  • generalize the selection of the fidelity-composition-routine into a PRAGMA for, e.g., compression strategies.
  • OR instead of foisting this choice onto users, do the work to produce evidence on the effect of choices to compilation output, and based on those, select one over the other.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants