History

Luke Hutton 6bb2043859

[mlir][tosa] Add pass to assign static input shape to TOSA functions (#171156 )

This commit introduces the `--tosa-experimental-input-shape` pass, which
allows a user to convert dynamically shaped input arguments of TOSA
functions to a user defined static shape. Here is a simple example:
```bash
func.func @test(%arg0: tensor<2x?xi32>, %arg1: tensor<?x256xf32>, %arg2: tensor<?x9xf32>) -> (tensor<2x?xi32>, tensor<?x256xf32>, tensor<?x9xf32>) {
    %0 = tosa.add %arg0, %arg0 : (tensor<2x?xi32>, tensor<2x?xi32>) -> tensor<2x?xi32>
    %1 = tosa.reciprocal %arg1 : (tensor<?x256xf32>) -> tensor<?x256xf32>
    %2 = tosa.sub %arg2, %arg2 : (tensor<?x9xf32>, tensor<?x9xf32>) -> tensor<?x9xf32>
    return %0, %1, %2 : tensor<2x?xi32>, tensor<?x256xf32>, tensor<?x9xf32>
}

$ mlir-opt --tosa-experimental-input-shape="args=arg0:2x16,arg2:64x9" test.mlir
func.func @test(%arg0: tensor<2x16xi32>, %arg1: tensor<?x256xf32>, %arg2: tensor<64x9xf32>) -> (tensor<2x?xi32>, tensor<?x256xf32>, tensor<?x9xf32>) {
    %0 = tosa.add %arg0, %arg0 : (tensor<2x16xi32>, tensor<2x16xi32>) -> tensor<2x?xi32>
    %1 = tosa.reciprocal %arg1 : (tensor<?x256xf32>) -> tensor<?x256xf32>
    %2 = tosa.sub %arg2, %arg2 : (tensor<64x9xf32>, tensor<64x9xf32>) -> tensor<?x9xf32>
    return %0, %1, %2 : tensor<2x?xi32>, tensor<?x256xf32>, tensor<?x9xf32>
}
```

When used in conjunction with the `--tosa-infer-shapes` pass, it can be
used to resolve simple TOSA functions (those that don't include TOSA
shape operations) to propagate static shape information. Continuing from
the example above:
```bash
$ mlir-opt --tosa-infer-shapes test2.mlir
func.func @test(%arg0: tensor<2x16xi32>, %arg1: tensor<?x256xf32>, %arg2: tensor<64x9xf32>) -> (tensor<2x?xi32>, tensor<?x256xf32>, tensor<?x9xf32>) {
    %0 = tosa.add %arg0, %arg0 : (tensor<2x16xi32>, tensor<2x16xi32>) -> tensor<2x16xi32>
    %cast = tensor.cast %0 : tensor<2x16xi32> to tensor<2x?xi32>
    %1 = tosa.reciprocal %arg1 : (tensor<?x256xf32>) -> tensor<?x256xf32>
    %2 = tosa.sub %arg2, %arg2 : (tensor<64x9xf32>, tensor<64x9xf32>) -> tensor<64x9xf32>
    %cast_0 = tensor.cast %2 : tensor<64x9xf32> to tensor<?x9xf32>
    return %cast, %1, %cast_0 : tensor<2x?xi32>, tensor<?x256xf32>, tensor<?x9xf32>
}
```
Note: tosa-infer-shapes currently doesn't have an option to update the
function signature.

Co-authored-by: Kaushik Varadharajan <kaushik.varadharajan@arm.com>

2026-01-15 17:31:34 +00:00

benchmark/python

…

cmake/modules

[mlir][Python] don't build libnanobind if module only has "pure" sources (#175090 )

2026-01-12 17:09:16 +00:00

docs

[mlir][Python] add docs about downstream type/attr implementation (#175259 )

2026-01-10 05:25:03 +00:00

examples

[MLIR] Add missing dialects to C API (#82190 )

2026-01-07 12:51:33 -08:00