func.return %0: tensor<2x3x7xf32>
}

// CHECK-LABEL: testBatchMatMulToMatMul
func.func @testBatchMatMulToMatMul(%arg0: tensor<2x3xf32>, %arg1: tensor<3x2xf32>) -> tensor<2x2xf32> {
  // CHECK: %0 = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<2x3xf32>, tensor<3x2xf32>) -> tensor<2x2xf32>

      type: DT_INT32
    }
  }
  attr {
    key: "keep_dims"
    value {
      b: false
    }
  }
}
node {
  name: "fc1000_1/MatMul"
  op: "MatMul"
  input: "avg_pool_1/Mean"
  input: "fc1000_1/MatMul/ReadVariableOp"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }

      "IRFFT3D",
      "If",
      "InTopKV2",
      "L2Loss",
      "LeakyRelu",
      "LinSpace",
      "ListDiff",
      "LogMatrixDeterminant",
      "LowerBound",
      "MatMul",
      "MatrixBandPart",
      "MatrixDiag",
      "MatrixDiagPart",
      "MatrixDiagPartV2",
      "MatrixDiagPartV3",
      "MatrixDiagV2",
      "MatrixDiagV3",
      "MatrixInverse",

  // - rhs: [RHSBATCHDIMS..., RHSROWS, RHSCOLS]
  // - result: [broadcast(LHSBATCHDIMS, RHSBATCHDIMS)..., LHSROWS, RHSCOLS]
  // To perform the matmul, we need to first broadcast lhs and rhs to a common
  // set of leading dimensions before doing the actual matmul.
  // That's what the code below does.
  // In particular, we populate out_lhs and out_rhs to have dimension structure:

}

def TF__FusedMatMulOp : TF_Op<"_FusedMatMul", [Pure, TF_SameOperandsAndResultElementTypeResolveRef]> {
  let summary = [{
Performs a MatMul followed by a specified series of operations.
  }];

  let description = [{
The inputs to the MatMul are specified by `a` and `b`. The series of operations
that follows is specified by the `fused_ops` attribute, which is a list of TF op

  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "fc1000_1/MatMul"
  op: "MatMul"
  input: "avg_pool_1/Mean"
  input: "fc1000_1/MatMul/Cast"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_HALF
    }
  }
  attr {

* `tf.config.experimental.enable_tensor_float_32_execution`