Composing legalizations about mlir-emitc HOT 6 CLOSED

This is indeed not possible with the passes we have right now. Is this relevant for the work on your side?

from mlir-emitc.

This would allow using the argmax from TOSA with a primarily MHLO model.

Is it a requirement that the lowering to EmitC be a one-shot?

from mlir-emitc.

I see. This is not a strict requirement, it just how we currently have implemented this. So far there wasn't a use case for mixed models on our side. Anyway, if it is helping we can think about refactoring this.

from mlir-emitc.

This is indeed not possible with the passes we have right now. Is this relevant for the work on your side?

I must correct my statement. The conversion passes are implemented as partial conversions and it is indeed possible to pass mixed input files. For example

func @mixed_dialects(%arg0: tensor<2xi32>) -> i32 {
  %0 = "mhlo.constant"() {value = dense<1> : tensor<2xi32>} : () -> tensor<2xi32>
  %1 = "tosa.const"() {value = dense<[3, 0, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %2 = emitc.call "func_a" (%0, %1) : (tensor<2xi32>, tensor<4xi32>) -> i32
  return %2 : i32
}

can be converted by passing --convert-tosa-to-emitc --convert-mhlo-to-emit to emitc-opt to

func @mixed_dialects(%arg0: tensor<2xi32>) -> i32 {
  %0 = "emitc.constant"() {value = dense<1> : tensor<2xi32>} : () -> tensor<2xi32>
  %1 = "emitc.constant"() {value = dense<[3, 0, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %2 = emitc.call "func_a"(%0, %1) : (tensor<2xi32>, tensor<4xi32>) -> i32
  return %2 : i32
}

Using emitc-translate or mlir-translate (from MLIR upstream) this can be translated to

int32_t mixed_dialects(Tensor<int32_t, 2> v1) {
  Tensor<int32_t, 2> v2 = {1, 1};
  Tensor<int32_t, 4> v3 = {3, 0, 1, 2};
  int32_t v4 = func_a(v2, v3);
  return v4;
}

from mlir-emitc.

This would allow using the argmax from TOSA with a primarily MHLO model.

The problem you were facing is that there is no conversion for tosa.argmax. I've created #279 to track this. Please re-open or fill a new issue if I got things wrong.

from mlir-emitc.

Ah I hadn't checked argmax support there. Thanks!

from mlir-emitc.