// CHECK-SAME:   slice_sizes = dense<[1, 4, 128]>
  // CHECK-SAME: (tensor<2x4x128xf32>, tensor<2x1xi32>) -> tensor<2x4x128xf32>
  %0 =  "tf.GatherNd"(%arg0, %arg1) {Tindices = i32, Tparams = i32, device = ""} : (tensor<2x4x128xf32>, tensor<2x1xi32>) -> tensor<2x4x128xf32>
  func.return %0 : tensor<2x4x128xf32>
}

//===----------------------------------------------------------------------===//

// CHECK: return %[[v5]] : tensor<?x?x8xf32>
}

func.func @einsum_no_reshape(%arg0 : tensor<?x?x8x128xf32>, %arg1 : tensor<1x?x8x128xf32>) -> tensor<?x?x8x?xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {device = "", equation = "BQNH,BTNH->BQNT"} : (tensor<?x?x8x128xf32>, tensor<1x?x8x128xf32>) -> tensor<?x?x8x?xf32>
  func.return %0 : tensor<?x?x8x?xf32>
// CHECK-LABEL: einsum_no_reshape

  // CHECK-LABEL: test_batch_matmul_broadcast_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  func.return %0: tensor<2x2x2xf32>
}

}

// CHECK-LABEL: func @move_across_broadcastable_op
func.func @move_across_broadcastable_op(%arg0: tensor<1x4x1x8xf32>, %arg1: tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32> {

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}>
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg0, %arg1) : (tensor<1x4x1x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>

// CHECK-NOT: "tfl.batch_matmul"
func.func @Batchmatmul2FullyconnectedAdjx(%arg0: tensor<4x2x128xf32>) -> (tensor<4x128x1xf32>) {
  %0 = arith.constant dense<[[1.0], [2.0]]> : tensor<2x1xf32>
  %1 = "tfl.batch_matmul"(%arg0, %0) {adj_x = true, adj_y = false, asymmetric_quantize_inputs = false} : (tensor<4x2x128xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>
  func.return %1 : tensor<4x128x1xf32>

// CHECK:           %[[VAL_15:.*]] = "tfl.reshape"(%[[VAL_14]], %[[VAL_5]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x128xf32>, tensor<3xi32>) -> tensor<1x384x128xf32>

  %2 = "tf.Conv2D"(%1, %arg1) {padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<2x68x1x3xf32>, tensor<5x1x3x8xf32>) -> tensor<2x64x1x8xf32>
  %3 = "tf.Squeeze"(%2) {squeeze_dims = [-2]} : (tensor<2x64x1x8xf32>) -> tensor<2x64x8xf32>
  %4 = "tf.BatchToSpaceND"(%3, %cst_1, %cst) : (tensor<2x64x8xf32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<1x128x8xf32>
  func.return %4 : tensor<1x128x8xf32>

// CHECK:         return %0 : tensor<1x?x4x128xf32>
func.func @real_dynamic_slice_strides_equal_to_1_signed(%arg0: tensor<1x?x4x256xf32>, %arg1: tensor<4xi32>, %arg2: tensor<4xi32>) -> tensor<1x?x4x128xf32> {
%cst = mhlo.constant dense<1> : tensor<4xi32>

      // CHECK-SAME:     scatter_dimension = 0
      //
      %1 = "tf.XlaReduceScatter"(%arg0, %cst_0, %cst) {reduce_op = "Add"} : (tensor<128x128xf32>, tensor<4x2xi32>, tensor<i32>) -> tensor<64x128xf32>
      func.return %1 : tensor<64x128xf32>
  }

  // CHECK-LABEL: func @tf_mod
  func.func @tf_mod(%arg1: tensor<2x2xf32>) -> tensor<2x2xf32> {
    %cst = "tf.Const"() {value = dense<7.000000e+00> : tensor<f32>} : () -> tensor<f32>

  } : (tensor<6x7x8xf32>, tensor<4xi32>, tensor<4xi32>, tensor<4xi32>) -> tensor<1x4x1x8xf32>
  func.return %0 : tensor<1x4x1x8xf32>

  // CHECK-DAG: %[[NEW_DIMS:.*]] = arith.constant dense<[6, 1, 7, 1, 8]> : tensor<5xi32>
  // CHECK-DAG: %[[BEGIN:.*]] = arith.constant dense<0> : tensor<5xi32>