llvmbot
commented
11 months ago @llvm/issue-subscribers-mlir-linalg
Author: None (MaheshRavishankar)
The pattern [RemoveOutsDependency](https://github.com/llvm/llvm-project/blob/f73844d92b36cb6801ac50ea721f4ba29b35d7a9/mlir/lib/Dialect/Linalg/Transforms/ElementwiseOpFusion.cpp#L1802) is meant to do the following
```
%0 = <some operation>
%1 = linalg.generic {
iterator_types = ["parallel", "parallel"], indexing_maps = [...]}
ins(....) outs(%0 : tensor<?x?xf32>) {...}
```
The `linalg.generic` is all parallel. So the use of `%0 -> %1` is a "false" dependency cause the actual values of `%0` arent used in `%1`. They are only really present to represent the shape of the output. So it is presumably legal to convert this to
```
%0 = <some operation>
%empty = tensor.empty(...)
%1 = linalg.generic {
iterator_types = ["parallel", "parallel"], indexing_maps = [...]}
ins(....) outs(%empty : tensor<?x?xf32>) {...}
```
where the dependency is now broken. This has advantages in terms of ability to DCE code and also helps fusion.
This pattern though was added before `DestinationPassingStyleOpInterface`. One thing the above pattern does is break destination passing style. A better approach is to no replace `outs` with `tensor.empty()`, instead restrict this pattern to do the following. If `%0` instead of "some operation" is "some destination passing style operation)
```
%0 = <some destination passing op> init(%dest : tensor<?x?xf32>) ...
%1 = linalg.generic {
iterator_types = ["parallel", "parallel"], indexing_maps = [...]}
ins(....) outs(%0 : tensor<?x?xf32>) {...}
```
and `%dest` is the operand that is "tied" to `%0`, the pattern above should convert this to
```
%0 = <some destination passing op> init(%dest : tensor<?x?xf32>) ...
%1 = linalg.generic {
iterator_types = ["parallel", "parallel"], indexing_maps = [...]}
ins(....) outs(%dest : tensor<?x?xf32>) {...}
```
This still breaks the false `%0` -> `%1` dependency. It does introduce a `%dest` to `%1` dependency, but if that is also defined by a destination passing style op, then the pattern can further apply, propagating as far up the use-def chain as possible.
joker-eph
MaheshRavishankar
okkwon
The pattern RemoveOutsDependency is meant to do the following
The
linalg.genericis all parallel. So the use of%0 -> %1is a "false" dependency cause the actual values of%0arent used in%1. They are only really present to represent the shape of the output. So it is presumably legal to convert this towhere the dependency is now broken. This has advantages in terms of ability to DCE code and also helps fusion.
This pattern though was added before
DestinationPassingStyleOpInterface. One thing the above pattern does is break destination passing style. A better approach is to no replaceoutswithtensor.empty(), instead restrict this pattern to do the following. If%0instead of "some operation" is "some destination passing style operation)and
%destis the operand that is "tied" to%0, the pattern above should convert this toThis still breaks the false
%0->%1dependency. It does introduce a%destto%1dependency, but if that is also defined by a destination passing style op, then the pattern can further apply, propagating as far up the use-def chain as possible.