Faster vars used tracking in simplify let visitor #8205
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Deletes a bunch of code and speeds up lowering time of local laplacian with 20 pyramid levels by ~2.5%
It was O(n) for n facts. This makes it O(log(n)) This was particularly bad for pipelines with lots of inputs or outputs, because those pipelines have lots of asserts, which make for lots of facts to substitute in. Speeds up lowering of local laplacian with 20 pyramid levels (which has only one input and one output) by 1.09x Speeds up lowering of the adams 2019 cost model training pipeline (lots of weight inputs and lots outputs due to derivatives) by 1.5x Speeds up resnet50 (tons of weight inputs) lowering by 7.3x!
…o abadams/faster_substitute_facts
…o abadams/faster_substitute_facts
…o abadams/faster_substitute_facts
Interval::is_single_point() used to only compare expressions by shallow equality to see if they are the same Expr object. However, bounds_of_expr_in_scope is really improved if it uses deep equality instead, so it has a prepass that goes over the provided scope, calls equal(min, max) on everything, and fixes up anything where deep equality is true but shallow equality. This prepass costs O(n) for n things in scope, regardless of how complex the expression being analyzed is. So if you ask for the bounds of '4' say in a context where there are lots of things in the scope, it's absurdly slow. We were doing this! BoxTouched calls bounds_of_expr_in_scope lots of times on small index Exprs within the same very large scope. It's better to just make Interval::is_single_point() check deep equality. This speeds up local laplacian lowering by 1.1x, and resnet50 lowering by 1.5x. There were also places where intervals that were a single point were diverging due to carelessly written code. E.g. the interval [40*8, 40*8], where both of those 40*8s are the same Mul node, was being simplified like this: interval.min = simplify(interval.min); interval.max = simplify(interval.max); Not only does this do double the simplification work it should, but it also caused something that was a single point to diverge into not being a single point, because the repeated constant-folding creates a new Expr. With the new is_single_point this matters a lot less, but even so, I centralized simplification of intervals into a single helper that doesn't do the pointless double-simplification for single points. Some of these shallowly-unequal but deeply-equal Intervals were being created in bounds inference itself after the prepass, which may have been generating suboptimal bounds. This change should fix that in addition to the compile-time benefits. Also added a simplify call in SkipStages because I noticed when it processed specializations it was creating things like (condition) || (!condition).
This visitor shows up as the main cost of lowering in very large pipelines. This visitor is for tracking which lets are actually used for real inside the body of a let block (as opposed to the tracking we do when mutating, which is approximate, because we could construct and Expr that uses a Var and then discard it in a later mutation). The old implementation made a map of all variables referenced, and then checked each let name against that map one by one. If there are a small number of lets outside a huge Stmt, this is bad, because the data structure has to hold a number of names proportional to the stmt size instead of proportional to the number of lets. This new implementation instead makes a hash set of the let names, and than traverses the Stmt, removing names from the set as they are encountered. This is a big speed-up. We then make the speed-up larger by about the same factor again doing the following: 1) Only add names to the map that might be used based on the recursive mutate call. These are very very likely to be used, because we saw them at least once, and mutations that remove *all* uses of a Var are rare. 2) The visitor should early out when the map becomes empty. The let variables are often all used immediately, so this is frequent. Speeds up lowering of local laplacian by 1.44x, 2.6x, and 4.8x respectively for 20, 50, and 100 pyramid levels. Speeds up lowering of resnet50 by 1.04x. Speeds up lowering of lens blur by 1.06x
|
It looks like this PR also contains #8202 ? Might need to merge with main to clean it up |
rootjalex
reviewed
Apr 22, 2024
…sed_in_simplify_let
|
Merge with main done |
|
With the last merge, when combined with #8204, this is 6.7x faster for local laplacian with 100 pyramid levels. But Amdahl's law has kicked in here. For just the pass that was slow (the simplification pass just after vectorization, because it did the first simplification after unify_duplicate_lets), this is actually over 400x faster. |
rootjalex
reviewed
Apr 24, 2024
src/Simplify_Let.cpp
Outdated
| std::map<std::string, int> vars_used; | ||
| count_var_uses(result, vars_used); | ||
|
|
||
| std::unordered_set<std::string> unused_vars(frames.size() * 2); |
There was a problem hiding this comment.
Where does frames.size() * 2 come from?
There was a problem hiding this comment.
Oh, each frame can maybe insert two names. Maybe add a comment above this definition, for explanation?
There was a problem hiding this comment.
Actually I think old_uses and new_uses might be mutually exclusive. Will test.
|
Is it reasonable to write tests for this new functionality, or do existing tests already stress it enough? |
|
Everything breaks if you remove a let that you should not have. I also confirmed the existing tests fail if you leave too many lets in (it messes up some instruction selection tests). |
|
Just needs an approval |
1 task
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This visitor shows up as the main cost of lowering in very large
pipelines.
This visitor is for tracking which lets are actually used for real
inside the body of a let block (as opposed to the tracking we do when
mutating, which is approximate, because we could construct an Expr that
uses a Var and then discard it in a later mutation).
The old implementation made a map of all variables referenced, and then
checked each let name against that map one by one. If there are a small
number of lets outside a huge Stmt, this is bad, because the data
structure has to hold a number of names proportional to the Stmt size
instead of proportional to the number of lets.
This new implementation instead makes a hash set of the let names, and
than traverses the Stmt, removing names from the set as they are
encountered. This is a big speed-up.
We then make the speed-up larger by about the same factor again doing
the following:
Only add names to the map that might be used based on the recursive
mutate call. These are very very likely to be used, because we saw them
at least once, and mutations that remove all uses of a Var are rare.
The visitor should early out when the map becomes empty. The let
variables are often all used immediately, so this is frequent.
Speeds up lowering of local laplacian by 1.44x, 2.6x, and 4.8x
for 20, 50, and 100 pyramid levels respectively.
Speeds up lowering of resnet50 by 1.04x.
Speeds up lowering of lens blur by 1.06x
Built on #8202
Complementary with #8204 though the total speed-up from merging both is not going to be the product of the two speed-ups. The two PRs solve the same local laplacian pathological lowering time problem in two different ways.