// CHECK{LITERAL}: %cst_1 = arith.constant dense<[[[[1.000000e+00], [2.000000e+00]], [[2.000000e+00], [4.000000e+00]]]]> : tensor<1x2x2x1xf32>
// CHECK: %3 = tfl.mul %2, %cst_1 {fused_activation_function = "NONE"} : tensor<1x2x2x1xf32>
// CHECK: %cst_2 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>

// CHECK: return %6 : tensor<1x5x5x3xf32>
}

// CHECK-LABEL: bias_adjust_duplicate_filter
func.func @bias_adjust_duplicate_filter(%arg0: tensor<1x5x5x2xf32>) -> (tensor<1x5x5x3xf32>, tensor<1x5x5x3xf32>) {
  %0 = "quantfork.stats"(%arg0) {
    layerStats = dense<[-1.28e-5, 1.27e-5]> : tensor<2xf32>
  } : (tensor<1x5x5x2xf32>) -> tensor<1x5x5x2xf32>

  %4 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x2x5x5xf32>, tensor<4x2x3x3xf32>) -> tensor<1x4x5x5xf32>
  %5 = stablehlo.add %4, %3 : tensor<1x4x5x5xf32>
  %6 = stablehlo.maximum %5, %2 : tensor<1x4x5x5xf32>
  return %6 : tensor<1x4x5x5xf32>
}

// CHECK: %[[CUSTOM_AGGREGATOR_1:.+]], {{.*}}, {{.*}}, {{.*}} = "tf.CustomAggregator"(%[[XLA_CALL_MODULE]]) {{.*}} : (tensor<1x4x4x8xf32>) -> (tensor<1x4x4x8xf32>, tensor<f32>, tensor<f32>, tensor<0xi64>)

// CHECK: %[[TRANSPOSE_2:.+]] = stablehlo.transpose %[[CUSTOM_AGGREGATOR_1]], dims = [0, 3, 1, 2] : (tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32>

// CHECK: %[[TRANSPOSE_2:.+]] = stablehlo.transpose %[[CONV]], dims = [0, 3, 1, 2] : (tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32>

// -----

// Tests that the conversion doesn't happen when the input dimension numbers
// are not [b, f, 0, 1].

  %custom_2 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  %custom_3 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>

func.func @broadcast_to_to_reshape(%arg0: tensor<4x4x4xf32>, %arg1 : tensor<4xi32>) -> tensor<1x4x4x4xf32> {
  %0 = "tfl.broadcast_to"(%arg0, %arg1) : (tensor<4x4x4xf32>, tensor<4xi32>) -> tensor<1x4x4x4xf32>
  // CHECK: "tfl.reshape"
  // CHECK-SAME: (tensor<4x4x4xf32>, tensor<4xi32>) -> tensor<1x4x4x4xf32>
  func.return %0 : tensor<1x4x4x4xf32>
}

// Converts tfl.broadcast_to to tfl.reshape if input and output have the same

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

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>
// CHECK: return %[[CUSTOM_0]] : tensor<1x2x2x1xf32>
}

    // CHECK-NEXT: %[[CMP:.*]] = mhlo.compare GT, %[[INP]], %[[ZERO]], NOTYPE : (tensor<1x4x4x3xf32>, tensor<1x4x4x3xf32>) -> tensor<1x4x4x3xi1>
    // CHECK-NEXT: %[[RES:.*]] = mhlo.select %[[CMP]], %[[INP]], %[[LEAKY]] : tensor<1x4x4x3xi1>, tensor<1x4x4x3xf32>
    // CHECK-NEXT: return %[[RES]] : tensor<1x4x4x3xf32>
    %0 = "tf.LeakyRelu"(%arg0) {alpha = 2.000000e-01 : f32, device = ""} : (tensor<1x4x4x3xf32>) -> tensor<1x4x4x3xf32>