[Opt] Global variable optimizations #857

xumingkuan · 2020-04-23T19:19:28Z

Related issue = #656
Related PR = #859

Also dive into container statements to eliminate identical global loads.

Benchmark:
Before this PR:

After this PR:

After this PR, before 17fec6a:

After this PR, before fc14b95:

After this PR, before 5837637:

mpm99.py: 2900 -> 2875 statements. Before fc14b95, it's 2898 statements.

[Click here for the format server]

xumingkuan · 2020-04-23T19:20:29Z

What is this for?

taichi/taichi/transforms/lower_access.cpp

Line 155 in df1c2fa

auto merge = Stmt::make<ElementShuffleStmt>(lanes, true);

yuanming-hu · 2020-04-23T19:22:56Z

What is this for?

taichi/taichi/transforms/lower_access.cpp

Line 155 in df1c2fa

auto merge = Stmt::make<ElementShuffleStmt>(lanes, true);

In vectorized cases we still have to load the vector lanes one by one. Then we need this line to merge the lanes into a vector.

xumingkuan · 2020-04-23T19:23:39Z

It seems to me that there are 3 classes of statements referring to a global variable: GlobalTemporaryStmt, GlobalPtrStmt, GetChStmt. Is this correct?

Which classes are loading global variables, and which classes are modifying them?

yuanming-hu · 2020-04-23T19:26:42Z

We need to do the optimization before lower_access, where pointers still have forms like x[i, j, k]. This will make the aliasing analysis much easier.

Then we only have to treat GlobalTempraryStmt and GlobalPtrStmt, since GetChStmt will only be generated after lower_access.

xumingkuan · 2020-04-23T19:30:03Z

So are there only GlobalLoadStmt, GlobalStoreStmt, AtomicOpStmt that load or store GlobalTemporaryStmt, GlobalPtrStmt?

yuanming-hu · 2020-04-23T19:57:28Z

Yes

xumingkuan · 2020-04-23T20:08:25Z

Is GlobalPtrStmt accessing several SNodes with the same indices?

Maybe I should TI_ASSERT(width() == 1)...

yuanming-hu · 2020-04-23T20:16:08Z

please assert for now

# Conflicts: # taichi/ir/ir.h

xumingkuan · 2020-04-28T02:08:07Z

I found that in mpm99.py, lower_access is actually emitting many statements that seems optimizable, but we can't do the optimization without aliasing analysis...

xumingkuan · 2020-04-28T02:09:13Z

taichi/ir/ir.cpp

+    for (int i = 0; i < (int)ptr1->indices.size(); i++) {
+      if (!irpass::analysis::same_statements(ptr1->indices[i],
+                                             ptr2->indices[i])) {
+        if (ptr1->indices[i]->is<ConstStmt>() &&
+            ptr2->indices[i]->is<ConstStmt>()) {
+          // different constants
+          return false;
+        }
+      }
+    }
+    return true;


How to use value_diff to make this function analyze better?

Let's do that in a future PR. value_diff has not been touched for at least 6 months...

yuanming-hu

Thanks! I won't have time to go through every detail in variable_optimization.cpp before the group meeting, but I trust your impl. Please fix the nits.

yuanming-hu · 2020-04-28T02:53:50Z

taichi/ir/ir.cpp

@@ -18,6 +18,41 @@ IRBuilder &current_ast_builder() {
  return context->builder();
 }

+bool maybe_same_address(Stmt *var1, Stmt *var2) {


maybe_same_global_address?

I included the Alloca case (i.e. this function can check if two allocas are the same by directly checking if var1 == var2) although it is not used in that way.

If both Stmt* are allocas, they have the same address iff var1 == var2.

If only one of them is an alloca, they can never share the same address.

If both of them are global temps, they have the same address iff they have the same offset. (Is checking ret_type unnecessary?)

If only one of them is a global temp, they can never share the same address.

If both of them are global ptrs (or GetChStmts), we can check by SNode::id.

In other cases (probably after lower_access), we don't know if the two statements share the same address.

Is checking ret_type unnecessary?

A well-formed IR should not have cases with different ret_type but it's good to be safe here.

The logic here sounds good. Maybe we should copy-paste these as comments to the code :-)

I just realized that offset is a sufficient condition of "might be the same address".

Right. In case you are interested, there's the LLVM's design of alias analysis: https://llvm.org/docs/AliasAnalysis.html

taichi/ir/ir.cpp

yuanming-hu · 2020-04-28T02:57:26Z

taichi/ir/ir.cpp

+    for (int i = 0; i < (int)ptr1->indices.size(); i++) {
+      if (!irpass::analysis::same_statements(ptr1->indices[i],
+                                             ptr2->indices[i])) {
+        if (ptr1->indices[i]->is<ConstStmt>() &&
+            ptr2->indices[i]->is<ConstStmt>()) {
+          // different constants
+          return false;
+        }
+      }
+    }
+    return true;


Let's do that in a future PR. value_diff has not been touched for at least 6 months...

yuanming-hu · 2020-04-28T02:58:16Z

taichi/ir/ir.cpp

+            ptr2->indices[i]->is<ConstStmt>()) {
+          // different constants
+          return false;
+        }


I'm confused here: shouldn't we just return false here whenever two statements are not the same?

same_statements is checking if two statements are guaranteed to be the same.

For here, however, we want to check if two statements are possible to be the same. For example, arg[0] and arg[1] can be the same while same_statements tell us that they're different.

Now I understand. Thanks. Let's briefly document: return true when two statements might be the same address; false when two statements cannot be the same address.

yuanming-hu · 2020-04-28T03:00:45Z

taichi/ir/state_machine.cpp

@@ -117,12 +117,16 @@ void StateMachine::store(Stmt *store_stmt) {
 }

 void StateMachine::load(Stmt *load_stmt) {


Let's briefly document the behavior when load_stmt == nullptr.

xumingkuan · 2020-04-28T03:24:05Z

taichi/transforms/simplify.cpp

@@ -1162,7 +1162,6 @@ void full_simplify(IRNode *root, const CompileConfig &config, Kernel *kernel) {
    alg_simp(root, config);
    die(root);
    whole_kernel_cse(root);
-    variable_optimization(root);


Interestingly, avoid doing some variable optimizations leads to final optimization on some tests.

Right, now you need something like https://ieeexplore.ieee.org/document/1611550

xumingkuan · 2020-04-28T03:57:29Z

May I merge this into master now?

yuanming-hu · 2020-04-28T03:58:22Z

Yes, please. I trust your impl.

xumingkuan · 2020-04-28T04:06:33Z

Oh, I forgot to use the advanced_optimization variable for variable_optimization...

[skip ci] draft

e4df53d

xumingkuan mentioned this pull request Apr 24, 2020

[opt] Eliminate useless local stores and atomics #858

Merged

xumingkuan and others added 9 commits April 23, 2020 21:55

Merge remote-tracking branch 'origin/master' into opt-global

ce68dab

Merge branch 'master' into opt-global

7075833

# Conflicts: # taichi/ir/ir.h

refactor variable_optimization.cpp

4274db3

GlobalTemporaryStmt optimization

69c7783

[skip ci] GlobalPtrOptimize

428b8f1

remove unused code

8ce63cb

OtherVariableOptimize

5837637

Dive into container statements to eliminate identical global loads

fc14b95

[skip ci] enforce code format

81f4c4c

xumingkuan commented Apr 28, 2020

View reviewed changes

xumingkuan requested a review from yuanming-hu April 28, 2020 02:12

xumingkuan marked this pull request as ready for review April 28, 2020 02:29

yuanming-hu requested changes Apr 28, 2020

View reviewed changes

xumingkuan added 2 commits April 27, 2020 23:09

Reduce compilation time!

17fec6a

[skip ci] documentation about load(nullptr)

3e380ac

xumingkuan commented Apr 28, 2020

View reviewed changes

xumingkuan and others added 3 commits April 27, 2020 23:35

[skip ci] documentation of maybe_same_address

90d586d

[skip ci] more documentation about maybe_same_address

3ec767e

[skip ci] enforce code format

9daa183

xumingkuan requested a review from yuanming-hu April 28, 2020 03:57

yuanming-hu approved these changes Apr 28, 2020

View reviewed changes

xumingkuan mentioned this pull request Apr 28, 2020

Advanced optimization #656

Closed

18 tasks

xumingkuan merged commit c3168f9 into taichi-dev:master Apr 28, 2020

xumingkuan added a commit to xumingkuan/taichi that referenced this pull request Apr 28, 2020

Use advanced_optimization for taichi-dev#857

b476c4d

xumingkuan added a commit that referenced this pull request Apr 28, 2020

[opt] Use advanced_optimization for #857 (#894)

d7b9c4d

xumingkuan deleted the opt-global branch May 9, 2020 21:33

xumingkuan mentioned this pull request May 11, 2020

[IR] Use JIT compilation/evaluation for systematic constant folding #839

Merged

xumingkuan mentioned this pull request Jun 15, 2020

[Opt] Store-to-load forwarding with Control-flow graph #1248

Merged

xumingkuan mentioned this pull request Jun 25, 2020

[Opt] Dead store and stack-related operation elimination by control-flow graph #1324

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Opt] Global variable optimizations #857

[Opt] Global variable optimizations #857

xumingkuan commented Apr 23, 2020 •

edited

Loading

xumingkuan commented Apr 23, 2020 •

edited

Loading

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 23, 2020

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 23, 2020

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 23, 2020

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 28, 2020

xumingkuan Apr 28, 2020

yuanming-hu Apr 28, 2020

yuanming-hu left a comment

yuanming-hu Apr 28, 2020

xumingkuan Apr 28, 2020

xumingkuan Apr 28, 2020

yuanming-hu Apr 28, 2020

xumingkuan Apr 28, 2020

yuanming-hu Apr 28, 2020

yuanming-hu Apr 28, 2020

yuanming-hu Apr 28, 2020

xumingkuan Apr 28, 2020

yuanming-hu Apr 28, 2020

yuanming-hu Apr 28, 2020

xumingkuan Apr 28, 2020

yuanming-hu Apr 28, 2020

xumingkuan commented Apr 28, 2020

yuanming-hu commented Apr 28, 2020

xumingkuan commented Apr 28, 2020 •

edited

Loading

		@@ -117,12 +117,16 @@ void StateMachine::store(Stmt *store_stmt) {
		}

		void StateMachine::load(Stmt *load_stmt) {

[Opt] Global variable optimizations #857

[Opt] Global variable optimizations #857

Conversation

xumingkuan commented Apr 23, 2020 • edited Loading

xumingkuan commented Apr 23, 2020 • edited Loading

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 23, 2020

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 23, 2020

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 23, 2020

yuanming-hu commented Apr 23, 2020

xumingkuan commented Apr 28, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yuanming-hu left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

xumingkuan commented Apr 28, 2020

yuanming-hu commented Apr 28, 2020

xumingkuan commented Apr 28, 2020 • edited Loading

xumingkuan commented Apr 23, 2020 •

edited

Loading

xumingkuan commented Apr 23, 2020 •

edited

Loading

xumingkuan commented Apr 28, 2020 •

edited

Loading