[TorchToLinalg] Issue Tracking for Convolution Support

[zjgarvey]

There is some missing support for lowering aten.convolution to linalg:

• Grouped convolution for spatial_dims != 2. E.g. Failed to lower "torch.aten.convolution" to linalg #3429
• "Fake" depthwise convolution, where input and output channels differ. Failed to lower "torch.aten.convolution" to linalg #3429
• 1-D (groupsize = 1) quantized convolution see this code block
• 3-D depthwise convolution see this code
• 1-D and 3-D quantized depthwise convolution see this code

Please feel free to file issues for individual items and link them here.

[pdhirajkumarprasad]

Here are the issues that we are seeing in models

1>

module {
  func.func @main_graph( %arg1: !torch.vtensor<[1,96,1,1],f32>) -> !torch.vtensor<[1,32,1,1],f32>  attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.opset_versions = {ai.onnx.contrib = 1 : si64}, torch.onnx_meta.producer_name = "vai_q_onnx", torch.onnx_meta.producer_version = "1.17.0+5e1d84e"} {
    %54 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<32x24x1x1xf32>} : () -> !torch.vtensor<[32,24,1,1],f32> 
    %57 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<32xf32>} : () -> !torch.vtensor<[32],f32> 
    %1070 = torch.operator "onnx.Conv"(%arg1, %54, %57) {torch.onnx.dilations = [1 : si64, 1 : si64], torch.onnx.group = 4 : si64, torch.onnx.kernel_shape = [1 : si64, 1 : si64], torch.onnx.pads = [0 : si64, 0 : si64, 0 : si64, 0 : si64], torch.onnx.strides = [1 : si64, 1 : si64]} : (!torch.vtensor<[1,96,1,1],f32>, !torch.vtensor<[32,24,1,1],f32>, !torch.vtensor<[32],f32>) -> !torch.vtensor<[1,32,1,1],f32> 
    return %1070 : !torch.vtensor<[1,32,1,1],f32>
  }
}

error

model.torch_onnx.mlir:5:13: error: 'linalg.generic' op write affecting operations on global resources are restricted to workgroup distributed contexts.
    %1070 = torch.operator "onnx.Conv"(%arg1, %54, %57) {torch.onnx.dilations = [1 : si64, 1 : si64], torch.onnx.group = 4 : si64, torch.onnx.kernel_shape = [1 : si64, 1 : si64], torch.onnx.pads = [0 : si64, 0 : si64, 0 : si64, 0 : si64], torch.onnx.strides = [1 : si64, 1 : si64]} : (!torch.vtensor<[1,96,1,1],f32>, !torch.vtensor<[32,24,1,1],f32>, !torch.vtensor<[32],f32>) -> !torch.vtensor<[1,32,1,1],f32> 
            ^
model.torch_onnx.mlir:5:13: note: see current operation: 

2>

module {
  func.func @main_graph( %arg3: !torch.vtensor<[?,768,?],f32>) -> (!torch.vtensor<[?,768,?],f32> ) attributes {torch.onnx_meta.ir_version = 6 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.4.0"} {
    %5 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<768x192x1xf32>} : () -> !torch.vtensor<[768,192,1],f32> 
    %6 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1.0> : tensor<768xf32>} : () -> !torch.vtensor<[768],f32> 
    %237 = torch.operator "onnx.Conv"(%arg3, %5, %6) {torch.onnx.dilations = [1 : si64], torch.onnx.group = 4 : si64, torch.onnx.kernel_shape = [1 : si64], torch.onnx.pads = [0 : si64, 0 : si64], torch.onnx.strides = [1 : si64]} : (!torch.vtensor<[?,768,?],f32>, !torch.vtensor<[768,192,1],f32>, !torch.vtensor<[768],f32>) -> !torch.vtensor<[?,768,?],f32> 
    return %237 : !torch.vtensor<[?,768,?],f32>
  }
}

error:

model.torch_onnx.mlir:5:12: error: failed to legalize operation 'torch.aten.convolution' that was explicitly marked illegal
    %237 = torch.operator "onnx.Conv"(%arg3, %5, %6) {torch.onnx.dilations = [1 : si64], torch.onnx.group = 4 : si64, torch.onnx.kernel_shape = [1 : si64], torch.onnx.pads = [0 : si64, 0 : si64], torch.onnx.strides = [1 : si64]} : (!torch.vtensor<[?,768,?],f32>, !torch.vtensor<[768,192,1],f32>, !torch.vtensor<[768],f32>) -> !torch.vtensor<[?,768,?],f32> 
           ^
model.torch_onnx.mlir:5:12: note: see current operation: %9 = "torch.aten.convolution"(%arg3, %3, %4, %7, %6, %7, %1, %8, %0) : (!torch.vtensor<[?,768,?],f32>, !torch.vtensor<[768,192,1],f32>, !torch.vtensor<[768],f32>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int) -> !torch.vtensor<[?,768,?],f32>