Questions About Lowering from the onnx dialect

[AkhilJ99]

How exactly does one convert the code in onnx dialect like the one given below to a lower dialect like affine and std ?. Do these conversions already exist ?.
`
func @main_graph(%arg0: tensor<2x3xf32>) -> tensor<2x3xf32> {
%0 = "onnx.Add"(%arg0, %arg0) : (tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
return %0 : tensor<2x3xf32>
}
"onnx.EntryPoint"() {func = @main_graph, numInputs = 1 : i32, numOutputs = 1 : i32} : () -> ()

`

[tjingrant]

Simply run ./onnx-mlir --EmitMLIR add.mlir and you should get:

#map0 = affine_map<() -> (0)>
#map1 = affine_map<() -> (3)>
#map2 = affine_map<() -> (2)>


module {
  %0 = "krnl.packed_const"() {file_name = "/var/folders/1n/n6cdfhkd1hndnfq93trnf0vm0000gn/T/packed_const-c3909e.tmp", is_le = true, size_in_bytes = 0 : i64} : () -> i64
  func @main_graph(%arg0: memref<2x3xf32>) -> memref<2x3xf32> {
    %1 = alloc() : memref<2x3xf32>
    affine.for %arg1 = 0 to 2 {
      affine.for %arg2 = 0 to 3 {
        %2 = load %arg0[%arg1, %arg2] : memref<2x3xf32>
        %3 = load %arg0[%arg1, %arg2] : memref<2x3xf32>
        %4 = addf %2, %3 : f32
        store %4, %1[%arg1, %arg2] : memref<2x3xf32>
      }
    }
    return %1 : memref<2x3xf32>
  }
  "krnl.entry_point"() {func = @main_graph, numInputs = 1 : i32, numOutputs = 1 : i32} : () -> ()
}

[AkhilJ99]

Simply run ./onnx-mlir --EmitMLIR add.mlir and you should get:

#map0 = affine_map<() -> (0)>
#map1 = affine_map<() -> (3)>
#map2 = affine_map<() -> (2)>


module {
  %0 = "krnl.packed_const"() {file_name = "/var/folders/1n/n6cdfhkd1hndnfq93trnf0vm0000gn/T/packed_const-c3909e.tmp", is_le = true, size_in_bytes = 0 : i64} : () -> i64
  func @main_graph(%arg0: memref<2x3xf32>) -> memref<2x3xf32> {
    %1 = alloc() : memref<2x3xf32>
    affine.for %arg1 = 0 to 2 {
      affine.for %arg2 = 0 to 3 {
        %2 = load %arg0[%arg1, %arg2] : memref<2x3xf32>
        %3 = load %arg0[%arg1, %arg2] : memref<2x3xf32>
        %4 = addf %2, %3 : f32
        store %4, %1[%arg1, %arg2] : memref<2x3xf32>
      }
    }
    return %1 : memref<2x3xf32>
  }
  "krnl.entry_point"() {func = @main_graph, numInputs = 1 : i32, numOutputs = 1 : i32} : () -> ()
}

Thank you that was helpful

[AkhilJ99]

Also I have tried to lower the onnx implementation of AlexNet (https://github.com/onnx/models/blob/master/vision/classification/alexnet/model/bvlcalexnet-3.onnx) to the onnx dialect but have got the shape inference error.

First I emmited the onnxbasic

./build/bin/onnx-mlir --EmitONNXBasic bvlcalexnet-3.onnx 
Full MLIR code written to: 
   bvlcalexnet-3.onnx.mlir

Constant-free MLIR Code written to: 
   bvlcalexnet-3.tmp

Use:
   bvlcalexnet-3.onnx.mlir
to continue lowering the code to other dialects.

And then tried to lower to MLIR but got shape inference error

./build/bin/onnx-mlir --EmitMLIR bvlcalexnet-3.onnx.mlir
loc("bvlcalexnet-3.onnx.mlir":10:10): error: unable to infer shape of operation without shape inference interface
loc("bvlcalexnet-3.onnx.mlir":11:10): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":11:10): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":14:10): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":14:10): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":16:11): error: unable to infer shape of operation without shape inference interface
loc("bvlcalexnet-3.onnx.mlir":17:11): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":17:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":20:11): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":20:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":24:11): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":24:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":28:11): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":28:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":30:11): error: Input tensor not ranked
loc("bvlcalexnet-3.onnx.mlir":30:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":33:11): error: Input tensor(s) not ranked
loc("bvlcalexnet-3.onnx.mlir":33:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":35:22): error: unable to infer shape of operation without shape inference interface
loc("bvlcalexnet-3.onnx.mlir":38:11): error: Input tensor(s) not ranked
loc("bvlcalexnet-3.onnx.mlir":38:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":40:26): error: unable to infer shape of operation without shape inference interface
loc("bvlcalexnet-3.onnx.mlir":43:11): error: Input tensor(s) not ranked
loc("bvlcalexnet-3.onnx.mlir":43:11): error: shape inference failed
loc("bvlcalexnet-3.onnx.mlir":4:3): error: Shape inference failed, 21 operations couldn't be inferred

I have also tried using onnx-mlir-opt for shape inference but it throws another error

./build/bin/onnx-mlir-opt --shape-inference bvlcalexnet-3.onnx.mlir
bvlcalexnet-3.onnx.mlir:10:10: error: unable to infer shape of operation without shape inference interface
    %4 = "onnx.LRN"(%3) {alpha = 9.99999974E-5 : f32, beta = 7.500000e-01 : f32, bias = 1.000000e+00 : f32, size = 5 : i64} : (tensor<*xf32>) -> tensor<*xf32>
         ^
bvlcalexnet-3.onnx.mlir:10:10: note: see current operation: %4 = "onnx.LRN"(%3) {alpha = 9.99999974E-5 : f32, beta = 7.500000e-01 : f32, bias = 1.000000e+00 : f32, size = 5 : i64} : (tensor<1x96x54x54xf32>) -> tensor<*xf32>
bvlcalexnet-3.onnx.mlir:11:10: error: Input tensor not ranked
    %5 = "onnx.MaxPoolSingleOut"(%4) {kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]} : (tensor<*xf32>) -> tensor<*xf32>
         ^
bvlcalexnet-3.onnx.mlir:11:10: note: see current operation: %5 = "onnx.MaxPoolSingleOut"(%4) {kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]} : (tensor<*xf32>) -> tensor<*xf32>
bvlcalexnet-3.onnx.mlir:11:10: error: shape inference failed
    %5 = "onnx.MaxPoolSingleOut"(%4) {kernel_shape = [3, 3], pads = [0, 0, 0, 0], strides = [2, 2]} : (tensor<*xf32>) -> tensor<*xf32>
.
.
.
.

[doru1004]

@AkhilJ99 any updates on this?