C/C++ generated code sample about mlir-emitc HOT 2 CLOSED

Hi,

Asking here is perfectly fine. As the EmitC dialect is already upstreamed to MLIR core, it is also fine to ask question on https://llvm.discourse.group/c/mlir/31 or to discuss these on discord in the #mlir channel.

Regarding your question: Right now, we don't have any math implemented in EmitC itself. The only operators that can be applied with emitc.apply are & and *. Everything else need to be done via emitc.call ops. Thus you do more than a single line of C++ code. Instead of adding an example to the repo, let me to give on and comment on it here. For example the snippet

builtin.module  {
  builtin.func @test_add(%arg0: tensor<13x21x3xf32>, %arg1: tensor<13x21x3xf32>) -> tensor<13x21x3xf32> {
    %0 = "tosa.add"(%arg0, %arg1) : (tensor<13x21x3xf32>, tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
    return %0 : tensor<13x21x3xf32>
  }
}

can be converted to

builtin.module  {
  emitc.include "emitc_tosa.h"
  builtin.func @test_add(%arg0: tensor<13x21x3xf32>, %arg1: tensor<13x21x3xf32>) -> tensor<13x21x3xf32> {
    %0 = emitc.call "emitc::tosa::add"(%arg0, %arg1) : (tensor<13x21x3xf32>, tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
    return %0 : tensor<13x21x3xf32>
  }
}

using emitc-opt --tosa-to-emitc-pipeline. This can be translated with emitc-translate --mlir-to-cpp to the C++ source code below:

#include "emitc_tosa.h"
Tensor<float, 13, 21, 3> test_add(Tensor<float, 13, 21, 3> v1, Tensor<float, 13, 21, 3> v2) {
  Tensor<float, 13, 21, 3> v3 = emitc::tosa::add(v1, v2);
  return v3;
}

The called C++ function is a templated one and refers to the standard library:

It is not the aim of EmitC to be an MLIR dialect that represents C++ as complete as possible. However, we are of course willing to evolve the dialect. You might also want to take a look into this discussion.

I hope this answers your question?

from mlir-emitc.

Yes, I have a clearer picture now. Thanks!

from mlir-emitc.