func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexDepthToSpace", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>

// CustomOpNotWeightOnly-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> tensor<*xf32> attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w = arith.constant dense<127.0> : tensor<1024x1x1x1xf32>

func.func @testSpaceToDepthF32(%arg0: tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32> {
  // CHECK: %[[ARG:.*]]: tensor<1x2x2x1xf32>
  // CHECK: "tfl.space_to_depth"(%[[ARG]]) <{block_size = 2 : i32}> : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  %0 = "tfl.space_to_depth"(%arg0) {block_size = 2: i32} : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  func.return %0 : tensor<1x1x1x4xf32>
}

// -----

// CHECK-LABEL: @NotReorderReshapeAddIfNotTailingDimAfter
func.func @NotReorderReshapeAddIfNotTailingDimAfter(%arg0: tensor<1x30x1x96xf32>) -> tensor<1x30x96xf32> {
  %cst = arith.constant dense<2.0> : tensor<1x30x96xf32>
  %shape = arith.constant dense<[1, 30, 96]> : tensor<3xi32>
  %1 = "tfl.reshape"(%arg0, %shape) : (tensor<1x30x1x96xf32>, tensor<3xi32>) -> tensor<1x30x96xf32>

    } : (tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32>
  func.return %1 : tensor<1x1x1x16xf32>

// CHECK: %0 = "tfl.dequantize"(%arg0)
// CHECK: %1 = "tfl.average_pool_2d"(%0)
// CHECK: %2 = "tfl.quantize"(%1)
// CHECK: %3 = "tfl.dequantize"(%2)
// CHECK: return %3 : tensor<1x1x1x16xf32>
}

// CHECK-LABEL: QuantizeMaximum

  func.func @simple_folding(%arg0: tensor<1x1x1x1xi32>, %arg1: tensor<1x1x1x1xf32>) -> tensor<?x?x?x?xf32> {
    // CHECK: %[[SHAPE:.*]] = "tf.Shape"
    // CHECK: %[[CONV:.*]] = "tf.Conv2DBackpropInput"(%[[SHAPE]]
    // CHECK-SAME: (tensor<4xi32>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
    // CHECK: return %[[CONV]] : tensor<1x1x1x1xf32>
    %0 = "tf.Shape"(%arg0) : (tensor<1x1x1x1xi32>) -> tensor<4xi32>

  // Move binary op batched RHS before reshape:
  // binary(reshape(lhs), rhs) => reshape(binary(lhs, flatten(rhs)))
  // Pattern targetted here is as follows-
  // [input, lhr, rhs] == [<1x1024x128>, <1x1024x8x16>, <1x1x8x16xf32>]
  // This is valid only when the-
  // 1.last dimension of lhs is equal to the number of elements in constant rhs.
  // 2.Reduded shape of rhs, here <8x16> is equal to last dimensions of lhs.

// CHECK:           %[[VAL_2:.*]] = "tf.Reshape"(%[[VAL_0]], %[[VAL_1]]) : (tensor<3x1x16xf32>, tensor<4xi64>) -> tensor<3x1x1x16xf32>
// CHECK:           %[[VAL_3:.*]] = arith.constant dense<[3, 8, 8, 16]> : tensor<4xi64>
// CHECK:           %[[VAL_4:.*]] = "tf.BroadcastTo"(%[[VAL_2]], %[[VAL_3]]) : (tensor<3x1x1x16xf32>, tensor<4xi64>) -> tensor<3x8x8x16xf32>
// CHECK:           return %[[VAL_4]] : tensor<3x8x8x16xf32>
// CHECK:         }

}

// -----

// CHECK-LABEL: @conv2d_backprop_filter_grouped
func.func @conv2d_backprop_filter_grouped(
    %input: tensor<1x2x2x2xf32>,
    %out_backprop: tensor<1x1x1x2xf32>
  ) -> tensor<2x2x1x2xf32> {

  // CHECK: mhlo.convolution(%arg0, %arg1)
  // CHECK-SAME:  batch_group_count = 2 : i64
  // CHECK-SAME:  feature_group_count = 1 : i64