// =============================================================================

func.func @sigmoidGrad(%arg0: tensor<?x32xf32>, %arg1: tensor<?x32xf32>) -> tensor<?x32xf32> {
  %0 = "tf.SigmoidGrad"(%arg0, %arg1) : (tensor<?x32xf32>, tensor<?x32xf32>) -> tensor<?x32xf32>
  func.return %0 : tensor<?x32xf32>
// CHECK-LABEL: sigmoidGrad
// CHECK-NEXT: [[ONE:%.+]] = arith.constant dense<1.000000e+00> : tensor<f32>

}

func.func @QDQsFollowedByTranspose(tensor<1x2xf32>) -> (tensor<2x1xf32>) {
^bb0(%arg0: tensor<1x2xf32>):
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.quantize"(%arg0){qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>}: (tensor<1x2xf32>) -> (tensor<1x2x!quant.uniform<u8:f32, 1.0>>)
  %1 = "tfl.dequantize"(%0): (tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> (tensor<1x2xf32>)

func.func @fuseUnpackAndConcatToReshape(%arg0: tensor<1x3x2xf32>) -> tensor<1x6xf32> {
  %0:3 = "tfl.unpack"(%arg0) {axis = 1 : i32, num = 3 : i32} : (tensor<1x3x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>)
  %1 = "tfl.concatenation"(%0#0, %0#1, %0#2) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x6xf32>
  func.return %1 : tensor<1x6xf32>

}

// -----

func.func @fully_connected(%arg0: tensor<1x37xf32>, %arg1: tensor<40x37xf32>, %arg2: tensor<40xf32>) -> tensor<1x40xf32> {
  %0 = "tfl.fully_connected"(%arg0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> tensor<1x40xf32>
  func.return %0 : tensor<1x40xf32>
}

// -----

    return %2 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
  }
}

    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2

}

// CHECK-LABEL: QuantizeConcat
func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

"/job:tpu_host_worker/replica:0/task:3/device:CPU:0"}, %arg15: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "927", tf.device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0"}, %arg16: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "929", tf.device = "/job:tpu_host_worker/replica:0/task:1/device:CPU:0"}, %arg17: tensor<*x!tf_type.resource<tensor<6x32xf32>>> {tf._user_specified_name = "931", tf.device = "/job:tpu_host_worker/replica:0/task:2/device:CPU:0"},...

// CHECK:           %[[VAL_11:.*]] = "tf.ArgMax"(%[[VAL_0]], %[[VAL_9]]) : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xi32>
// CHECK:           return %[[VAL_10]], %[[VAL_11]] : tensor<1x1xf32>, tensor<1x1xi32>
// CHECK:         }
func.func @convert_argmax_with_reshaped_iota(%arg0: tensor<1x32x1xf32>) -> (tensor<1x1xf32>, tensor<1x1xi32>) {

      %cst = stablehlo.constant dense<1.0> : tensor<1x3xf32>
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      %1 = stablehlo.subtract %cst, %0 : tensor<1x3xf32>
      %2 = stablehlo.add %0, %cst : tensor<1x3xf32>
      return %2 : tensor<1x3xf32>
    }
  }