[Opt] Dive into container statements to find local loads/stores for optimization, and optimize loads of new allocas to 0 #662

xumingkuan · 2020-03-26T20:14:45Z

Related issue = #656

xumingkuan · 2020-03-26T20:18:38Z

I don't see a trivial way to check the statements for each kind of container statement. Shall we add a function like virtual std::vector<Stmt *> get_contained_statements() in addition to virtual bool is_container_statement()?

taichi/transforms/simplify.cpp

yuanming-hu · 2020-03-26T20:22:29Z

I don't see a trivial way to check the statements for each kind of container statement. Shall we add a function like virtual std::vector<Stmt *> get_contained_statements() in addition to virtual bool is_container_statement()?

This can be a very useful analysis pass! Note that we use the visitor pattern to avoid adding too many methods in each statement. You can declare a class class GetContainedStatements : public IRVisitor and implement the contained statements under void visit(XStmt * stmt) for each container statement (e.g. body for WhileStmt).

yuanming-hu · 2020-03-26T20:23:18Z

(Maybe I reviewed too early - feel free to re-request my review when you need next time.)

xumingkuan · 2020-03-26T20:53:52Z

I found BasicStmtVisitor already dives into container statements. I think this looks much clearer now, ignoring if (!advanced_optimization) clauses :)

The number of statements of test_ad_if_mutable reduced from 92 to 70.

yuanming-hu · 2020-03-26T20:54:58Z

The number of statements of test_ad_if_mutable reduced from 92 to 70.

AWESOME!!

xumingkuan · 2020-03-26T21:57:35Z

Diving into container statements does nothing to test_ad_if_mutable, but optimizing

$1 = alloca
$2 = local load [ [$1[0]]]

to

$1 = alloca
$2 = const [0]

further reduces the number of statements to 58.

yuanming-hu · 2020-03-26T22:00:13Z

Cool! At some point in the future, you may want to expand your test set and write an automatic tool to summarize how many instructions are generated in each test case :-)

taichi/transforms/simplify.cpp

yuanming-hu · 2020-03-26T22:28:20Z

taichi/transforms/simplify.cpp

 #include "taichi/ir/ir.h"

 TLANG_NAMESPACE_BEGIN

+// Find if there is a load following a store in a basic block


It seems to me that LocalLoadSearcher is just finding loads, since LocalStore is not mentioned here. I'm confused...

Oh yes, LocalLoadSearcher is finding loads all over the root, with root actually be a statement after LocalStore, passed in as an argument. I'd better change the comment to simply Find if there is a load.

yuanming-hu · 2020-03-26T22:29:42Z

taichi/transforms/simplify.cpp

+  }
+};
+
+// Find if there is a store preceding a load in a basic block


Again, isn't this simply Find if there is a store?

yuanming-hu · 2020-03-26T22:31:01Z

taichi/transforms/simplify.cpp

+  }
+};
+
+// Find the **last** store preceding a load in a basic block


Again, how does this ensure preceding?

xumingkuan · 2020-03-26T23:06:17Z

Oh no I'm breaking tests. Go debugging.

xumingkuan · 2020-03-26T23:28:23Z

I see what the bug is: sometimes I find a store that is not the last one in LocalStoreForwarder.
There is one thing I want to make sure: can we optimize

$30 = atomic add($1, $8)
$31 = local load [ [$1[0]]]

to (replace all usages of $31 with $30)? Or does AtomicOpStmt provide a return value? And what about $31 = local load [ [$1[1]]]?

yuanming-hu · 2020-03-26T23:37:07Z

Yes, AtomicOpStmt returns the old value

xumingkuan · 2020-03-26T23:45:52Z

Yes, AtomicOpStmt returns the old value

I see. So if the last statement is an AtomicOpStmt, there's no way we can directly optimize local load.

xumingkuan · 2020-03-27T00:25:29Z

Should be working now. The good news is number 58 here doesn't change.

Diving into container statements does nothing to test_ad_if_mutable, but optimizing
$1 = alloca
$2 = local load [ [$1[0]]]
to
$1 = alloca
$2 = const [0]
further reduces the number of statements to 58.

yuanming-hu

Great job!

taichi/transforms/simplify.cpp

…ptimization, and optimize loads of new allocas to 0 (taichi-dev#662) * refactor BasicBlockSimplify::visit(LocalStoreStmt*) * refactor again, making use of visitors * [skip ci] same for visit(LocalLoadStmt *) * [skip ci] LocalStoreForwarder * Optimize loads of new allocas to const[0] * [skip ci] enforce code format * Add assertions and update comments * fixed LocalStoreForwarder * [skip ci] enforce code format * [skip ci] minor: update comment * [skip ci] minor: update comment * Update simplify.cpp Co-authored-by: Taichi Gardener <taichigardener@gmail.com> Co-authored-by: Yuanming Hu <yuanming-hu@users.noreply.github.com>

refactor BasicBlockSimplify::visit(LocalStoreStmt*)

2511b39

yuanming-hu reviewed Mar 26, 2020

View reviewed changes

taichi/transforms/simplify.cpp Outdated Show resolved Hide resolved

taichi/transforms/simplify.cpp Outdated Show resolved Hide resolved

refactor again, making use of visitors

c0760bf

xumingkuan added 3 commits March 26, 2020 17:09

[skip ci] same for visit(LocalLoadStmt *)

d26f36d

[skip ci] LocalStoreForwarder

43ba503

Optimize loads of new allocas to const[0]

55a61e4

xumingkuan changed the title ~~[opt] Dive into container statements to find local loads/stores for optimization~~ [Opt] Dive into container statements to find local loads/stores for optimization, and optimize loads of new allocas to 0 Mar 26, 2020

[skip ci] enforce code format

9665cf7

xumingkuan marked this pull request as ready for review March 26, 2020 21:55

xumingkuan requested a review from yuanming-hu March 26, 2020 21:58

xumingkuan mentioned this pull request Mar 26, 2020

Advanced optimization #656

Closed

18 tasks

yuanming-hu reviewed Mar 26, 2020

View reviewed changes

Add assertions and update comments

bca4610

xumingkuan and others added 2 commits March 26, 2020 20:19

fixed LocalStoreForwarder

9c1b114

[skip ci] enforce code format

c3a362f

xumingkuan requested a review from yuanming-hu March 27, 2020 00:25

xumingkuan added 2 commits March 26, 2020 20:27

[skip ci] minor: update comment

d95c748

[skip ci] minor: update comment

079b473

yuanming-hu approved these changes Mar 27, 2020

View reviewed changes

taichi/transforms/simplify.cpp Outdated Show resolved Hide resolved

taichi/transforms/simplify.cpp Show resolved Hide resolved

Update simplify.cpp

f61a4f3

yuanming-hu merged commit d6c96a8 into taichi-dev:master Mar 27, 2020

xumingkuan deleted the load branch March 27, 2020 18:46

xumingkuan mentioned this pull request Apr 27, 2020

[Opt] [Refactor] [Bug] Use state machines for optimization on variables #859

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Opt] Dive into container statements to find local loads/stores for optimization, and optimize loads of new allocas to 0 #662

[Opt] Dive into container statements to find local loads/stores for optimization, and optimize loads of new allocas to 0 #662

xumingkuan commented Mar 26, 2020

xumingkuan commented Mar 26, 2020

yuanming-hu commented Mar 26, 2020

yuanming-hu commented Mar 26, 2020

xumingkuan commented Mar 26, 2020

yuanming-hu commented Mar 26, 2020

xumingkuan commented Mar 26, 2020 •

edited

yuanming-hu commented Mar 26, 2020

yuanming-hu Mar 26, 2020

xumingkuan Mar 26, 2020 •

edited

yuanming-hu Mar 26, 2020

yuanming-hu Mar 26, 2020

xumingkuan commented Mar 26, 2020

xumingkuan commented Mar 26, 2020 •

edited

yuanming-hu commented Mar 26, 2020

xumingkuan commented Mar 26, 2020

xumingkuan commented Mar 27, 2020

yuanming-hu left a comment

[Opt] Dive into container statements to find local loads/stores for optimization, and optimize loads of new allocas to 0 #662

[Opt] Dive into container statements to find local loads/stores for optimization, and optimize loads of new allocas to 0 #662

Conversation

xumingkuan commented Mar 26, 2020

xumingkuan commented Mar 26, 2020

yuanming-hu commented Mar 26, 2020

yuanming-hu commented Mar 26, 2020

xumingkuan commented Mar 26, 2020

yuanming-hu commented Mar 26, 2020

xumingkuan commented Mar 26, 2020 • edited

yuanming-hu commented Mar 26, 2020

yuanming-hu Mar 26, 2020

Choose a reason for hiding this comment

xumingkuan Mar 26, 2020 • edited

Choose a reason for hiding this comment

yuanming-hu Mar 26, 2020

Choose a reason for hiding this comment

yuanming-hu Mar 26, 2020

Choose a reason for hiding this comment

xumingkuan commented Mar 26, 2020

xumingkuan commented Mar 26, 2020 • edited

yuanming-hu commented Mar 26, 2020

xumingkuan commented Mar 26, 2020

xumingkuan commented Mar 27, 2020

yuanming-hu left a comment

Choose a reason for hiding this comment

xumingkuan commented Mar 26, 2020 •

edited

xumingkuan Mar 26, 2020 •

edited

xumingkuan commented Mar 26, 2020 •

edited