func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

  func.return %24 : tensor<1x4xf32>
}

    // CHECK-NEXT: %[[LEAKYGRAD:.*]] = mhlo.multiply %[[GRADIENT:.*]], %[[ALPHA]] : tensor<1x4x4xf32>
    // CHECK-NEXT: %[[CMP:.*]] = mhlo.compare GT, %[[INP:.*]], %[[ZERO]], NOTYPE : (tensor<1x4x4xf32>, tensor<1x4x4xf32>) -> tensor<1x4x4xi1>
    // CHECK-NEXT: %[[RES:.*]] = mhlo.select %[[CMP]], %[[GRADIENT]], %[[LEAKYGRAD]] : tensor<1x4x4xi1>, tensor<1x4x4xf32>
    // CHECK-NEXT: return %[[RES]] : tensor<1x4x4xf32>

    %14 = stablehlo.multiply %12, %13 : tensor<1x4x3xf32>  // s1 * s2
    %15 = call @uniform_quantize_1(%14, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>
    %16 = call @uniform_dequantize_0(%15, %5, %6) : (tensor<1x4x3xi8>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xf32>
    return %16 : tensor<1x4x3xf32>
  }

func.func @batchmatmulv2_basic(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
// CHECK-LABEL:   func @batchmatmulv2_basic
// CHECK-SAME:        ([[LHS:%.*]]: tensor<1x4x2xf32>, [[RHS:%.*]]: tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
// CHECK:           [[LHSSHAPE:%.*]] = shape.shape_of [[LHS]] : tensor<1x4x2xf32>
// CHECK:           [[RHSSHAPE:%.*]] = shape.shape_of [[RHS]] : tensor<3x2x4xf32>

func.func @reshape_removeIdentity(tensor<4x4x4xf32>) -> tensor<4x4x4xf32> {
^bb0(%arg0: tensor<4x4x4xf32>) :
  %cst = arith.constant dense<[4, 4, 4]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<4x4x4xf32>, tensor<3xi32>) -> tensor<4x4x4xf32>
  func.return %0 : tensor<4x4x4xf32>

// CHECK-LABEL: func @reshape_removeIdentity
// CHECK:  return %arg0 : tensor<4x4x4xf32>
}

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<[[0.1]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

  func.func @batchmatmulv2(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
    // CHECK: mhlo.reduce
    // CHECK: mhlo.dot_general
    // CHECK: mhlo.transpose
    %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
    func.return %0 : tensor<3x4x4xf32>
  }

  %2 = "tfl.fully_connected" (%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x40x37xf32>, tensor<4x37xf32>, none) -> (tensor<1x40x4xf32>)
  %3 = "tfl.add"(%2, %cst1) {fused_activation_function = "NONE"} : (tensor<1x40x4xf32>, tensor<1x1x4xf32>) -> tensor<1x40x4xf32>
  func.return %1, %3 : tensor<1x40x4xf32>, tensor<1x40x4xf32>