maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[mlir][linalg] ComprehensiveBufferize: Do not copy InitTensorOp results 

E.g.:

```
%2 = memref.alloc() {alignment = 128 : i64} : memref<256x256xf32>
%3 = memref.alloc() {alignment = 128 : i64} : memref<256x256xf32>

// ... (%3 is not written to)

linalg.copy(%3, %2) : memref<256x256xf32>, memref<256x256xf32>
vector.transfer_write %11, %2[%c0, %c0] {in_bounds = [true, true]} : vector<256x256xf32>, memref<256x256xf32>
```

Avoid copies of %3 if %3 came directly from an InitTensorOp.

Differential Revision: https://reviews.llvm.org/D109742
This commit is contained in:
Matthias Springer 2021-09-15 17:26:29 +09:00
parent e90d55e1c9
commit 934e2f695e
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
mlir/lib/Dialect/Linalg/Transforms/ComprehensiveBufferize.cpp
View File

@ -2190,7 +2190,8 @@ static LogicalResult bufferize(OpBuilder &b, VectorTransferOpInterface op,
newInputBuffer = createNewAllocDeallocPairForShapedValue(
b, loc, writeOp.source(), aliasInfo);
Value v = lookup(bvm, writeOp.source());
b.create<CopyOp>(loc, v, newInputBuffer);
if (!isInitTensorOp(writeOp.source()))
b.create<CopyOp>(loc, v, newInputBuffer);
} else {
// InPlace write will result in memref.tensor_load(x) which must
// canonicalize away with one of it uses.