[JIT] Make create autodiff subgraphs do in place updates to aliasDb #42141
Conversation
This file contains hidden or 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
[ghstack-poisoned]
facebook-github-bot
added
the
oncall: jit
eellison
changed the title
💊 CI failures summary and remediationsAs of commit 6ba44e5 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions on the GitHub issue tracker or post in the (internal) Dr. CI Users group. This comment has been revised 35 times. |
…tes to aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. [ghstack-poisoned]
eellison
changed the title
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. [ghstack-poisoned]
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
Krovatkin
approved these changes
Jul 30, 2020
| return a.user == b.user && a.offset == b.offset; | ||
| } | ||
|
|
||
| void copyAliasing(Node* merged_node, AliasDb& db) { |
There was a problem hiding this comment.
nitpick: should we just store AliasDb as a member of ValueMapper since we are already passing it in a c-tor which simplifies copyAliasing a little bit?
| if (shouldConsiderForMerge(consumer)) { | ||
| if (consumer->kind() != prim::DifferentiableGraph) { | ||
| // ValueMapper preserves the aliasing information of the node's outputs | ||
| ValueMapper vm(consumer, aliasDb_, 0); |
There was a problem hiding this comment.
nitpick: we could probably just remember the number of the outputs of a producer (which we can get off the producer rather than pass both producer and the offset) since we will need to update aliasing for all of them and then in a consumer we could just update the aliasing information for the last consumer.outputs().size() - producer.outputs().size() of the merged consumer?
There was a problem hiding this comment.
That's what I'm doing, see: https://github.com/pytorch/pytorch/pull/42141/files#diff-d35d4ac08adc5a54ab5b75564e9a1290R53
| auto g = n->owningGraph(); | ||
| // temporary node to put the aliasing properties of the node before its | ||
| // merged and destroyed | ||
| placeholder_node_ = g->insertNode(g->create(prim::Uninitialized, 0)); |
There was a problem hiding this comment.
nitpick: do we really need a node here? could we simply store all the values in a vector or will that confuse AA?
There was a problem hiding this comment.
Values are owned by the graph, if we store them in a vector but delete the node they will be dead references
| for (Value* v : new_outputs) { | ||
| auto maybe_last_use = firstOrLastUse(v, /*find_first*/ false); | ||
| // if it doesnt have a use it shouldnt have been added as output | ||
| TORCH_INTERNAL_ASSERT(maybe_last_use); |
There was a problem hiding this comment.
is this really true? could we have a node that produces two values but one of them isn't used?
There was a problem hiding this comment.
This is outputs of the autodiff subgraph, which are added individually. Even if there's a node with two outputs and one isn't used, the unused one will not be added an output to the autodiff subgraph
| // Inlining nodes may cause some subexpression to come back in the | ||
| // subgraphs (for example, copying constants in repeatedly will generate | ||
| // redundant prim::Constants). Run CSE to clean them up. | ||
| EliminateCommonSubexpression(curNode->g(attr::Subgraph)); |
There was a problem hiding this comment.
nitpick: should we move CSE into the next check? if we decide to inline it, the final CSE will clean all unmerged graphs up. In other words, we only need to clean a graph up if we aren't inlining it?
There was a problem hiding this comment.
This is existing logic in the file, I didn't change it. But yea, good point
|
|
||
| ValueMapper vm(producer, aliasDb_, consumer->outputs().size()); | ||
| SubgraphUtils::mergeNodeIntoSubgraph(producer, consumer); | ||
| vm.copyAliasing(consumer, aliasDb_); |
There was a problem hiding this comment.
nitpick: I wonder if it makes sense to do copyAliasing when vm goes out of scope?
There was a problem hiding this comment.
What do you mean ?
| TORCH_INTERNAL_ASSERT(i != last_uses_.size()); | ||
| db.replaceWithNewValue(placeholder_node_->outputs().at(i), v); | ||
| } | ||
| placeholder_node_->destroy(); |
There was a problem hiding this comment.
if copyAliasing is accidentally called twice destroy should throw an exception and not crash, right?
There was a problem hiding this comment.
If it's accidentally called twice we have other problems - this isn't an exposed API, this is only used within this file.
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
…o aliasDb" Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior. Description from #37106 holds pretty well: """ Recomputing the aliasdb on every fusion iteration + in every subblock is hugely expensive. Instead, update it in-place when doing fusion. The graph fuser pass operates by pushing nodes into a fusion group. So we start with `x, y = f(a, b, c)` and end with: ``` x_out, y_out = prim::fusionGroup(a, b, c) x_in, y_in = f(a_in, b_in, c_in) -> x_in, y_in ``` We destroy the x and y Value*s in the process. This operation is easy to express as an update to the aliasDb--x_out just takes on all the aliasing information x used to have. In particular, since we know f and prim::fusionGroup are purely functional, we don't have to mess with any write information. """ The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node. There are a few options: - attempt to make all subgraph utils & ir cloning logic update a map - mirror the subgraph utils implementation in create_autodiff_subgraph - uniquely map x, y and x_in, y_in so you can back out the correspondence. I went with the third option. This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated. Differential Revision: [D22798377](https://our.internmc.facebook.com/intern/diff/D22798377) [ghstack-poisoned]
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
Stack from ghstack:
Update alias db in-place instead of having to construct alias db from scratch on each change, causing O(n^2) behavior.
Description from #37106 holds pretty well:
"""
Recomputing the aliasdb on every fusion iteration + in every subblock
is hugely expensive. Instead, update it in-place when doing fusion.
The graph fuser pass operates by pushing nodes into a fusion group. So
we start with
x, y = f(a, b, c)and end with:
We destroy the x and y Value*s in the process. This operation is
easy to express as an update to the aliasDb--x_out just takes on all
the aliasing information x used to have. In particular, since we know
f and prim::fusionGroup are purely functional, we don't have to mess
with any write information.
"""
The one difficulty here is mapping x, y to x_out, y_out is not trivial in merging nodes into the autodiff subgraph node.
There are a few options:
I went with the third option.
This shouldn't affect the results of the pass at all. LMK if you think there's anything else I should be doing to test, I was thinking about maybe exposing an option to run create autodiff subgraphs without the post processor and check that the alias db was correctly updated.
Differential Revision: D22798377