// CHECK: return %[[RESULT]] : tensor<2x3x4x6xf32>
}

// -----

func.func @batchMatMulOneDim(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x5x6xf32>) -> tensor<3x4x6xf32> {
  %0 = "tf.BatchMatMul"(%arg0, %arg1) : (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>

  // CHECK-LABEL: batchMatMulOneDim

  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,ikm->ijm"}: (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>
  // CHECK-LABEL: einsum_basic
  // CHECK: "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = false}> : (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
}

func.func @lower_einsum(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x5x6xf32>) -> tensor<3x4x6xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,ikm->ijm"}: (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>
}
// CHECK-LABEL: lower_einsum
// CHECK: "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = false}> : (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>

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>

  %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>
}

    %13 = stablehlo.subtract %10, %12 : tensor<1x4x3xf32>  // Precalculated zp_neg.
    %14 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2] : (tensor<1x1x3xf32>) -> tensor<1x4x3xf32>  // Optional
    %15 = stablehlo.multiply %13, %14 : tensor<1x4x3xf32>  // s1 * s2
    %16 = call @uniform_quantize_1(%15, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>

        version = 5 : i64
      } : (tensor<?x3x4x3xf32>, tensor<2x3x3x2xf32>, tensor<1x1x1x2xf32>) -> tensor<?x3x4x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<?x3x4x2xf32>) -> tensor<?x3x4x2xf32>
    return %2 : tensor<?x3x4x2xf32>
  }

    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<3x4x2xf32>) -> tensor<3x4x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<3x4x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x4x2xf32>
    %7 = "stablehlo.gather"(%6, %arg2) {
    dimension_numbers = #stablehlo.gather<
      offset_dims = [2, 3],
      collapsed_slice_dims = [0],

// CHECK-SAME: (%[[ARG0:.*]]: tensor<3x4xf32>)
func.func @expm1(%arg0: tensor<3x4xf32>) -> tensor<3x4xf32> {
  %0 = "tf.Expm1"(%arg0) : (tensor<3x4xf32>) -> tensor<3x4xf32>
  func.return %0 : tensor<3x4xf32>
  // CHECK: %[[ONE:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<f32>}> : () -> tensor<f32>
  // CHECK: %[[EXP:.*]] = "tf.Exp"(%[[ARG0]]) : (tensor<3x4xf32>) -> tensor<3x4xf32>

    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>