// CHECK-LABEL: conv2DBiasAdd_noActivation
func.func @conv2DBiasAdd_noActivation(%arg0: tensor<128xf32>, %arg1: tensor<1x1x3x128xf32>, %arg2: tensor<8x32x32x3xf32>) -> (tensor<*xf32>) {

//   <Contraction> + BiasAdd + <Activation> -> <FusedContraction>
//
// Note that fusion with activation is preferred, but a contraction and BiasAdd
// can also be replaced by a _FusedConv2D if there is no other activation
// function.
// i.e., this class also supports the following fusion:
//   <Contraction> + BiasAdd -> <FusedContraction>
//

// CHECK-NEXT: [[o6:%.*]] = tfrt_fallback_async.executeop key({{[0-9]+}}) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf._FusedConv2D"([[arg1]], [[o3]], [[o2]])
// CHECK-NEXT: [[o7:%.*]] = tfrt_fallback_async.executeop key({{[0-9]+}}) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AvgPool"([[o6]])

  %input_scale = "tf.Const"() {value = dense<1.0> : tensor<f32>} : () -> tensor<f32>
  %side_input_scale = "tf.Const"() {value = dense<2.0> : tensor<f32>} : () -> tensor<f32>
  %conv2d = "tf._FusedConv2D"(%input, %filter, %bias, %act, %input_scale, %side_input_scale) {

           {"FusedBatchNorm", "FusedBatchNormV2", "FusedBatchNormV3",
            "_FusedBatchNormEx", "FusedBatchNormGrad", "FusedBatchNormGradV2",
            "FusedBatchNormGradV3"}},
          {"Conv", {"_FusedConv2D"}},
          {"SORT", {"TopKV2"}},  // XLA version much faster then TF version.
          {"MISC",
           // clang-format off
     {"ApproxTopK", "BroadcastTo", "ExpandDims", "Fill", "NoOp",

are specified by `args`.
* If there is an op A specified, the output of op X is the input to op A, and op
A produces the _FusedConv2D output. Otherwise, op X produces the _FusedConv2D
output.

*NOTE*: Do not invoke this operator directly in Python. Grappler is expected to
create these operators.
  }];

  let arguments = (ins