// CHECK:           %[[VAL_12:.*]] = "tfl.reshape"(%[[VAL_10]], %[[VAL_7]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x512xf32>, tensor<4xi32>) -> tensor<128x1x1x512xf32>

// object containing configuration parameters, builtins have a predetermined
// set of acceptable options.
// LINT.IfChange
enum BuiltinOperator : int32 {
  ADD = 0,
  AVERAGE_POOL_2D = 1,
  CONCATENATION = 2,
  CONV_2D = 3,
  DEPTHWISE_CONV_2D = 4,
  DEPTH_TO_SPACE = 5,
  DEQUANTIZE = 6,
  EMBEDDING_LOOKUP = 7,
  FLOOR = 8,
  FULLY_CONNECTED = 9,
  HASHTABLE_LOOKUP = 10,
  L2_NORMALIZATION = 11,
  L2_POOL_2D = 12,

// object containing configuration parameters, builtins have a predetermined
// set of acceptable options.
// LINT.IfChange
enum BuiltinOperator : byte {
  ADD = 0,
  AVERAGE_POOL_2D = 1,
  CONCATENATION = 2,
  CONV_2D = 3,
  DEPTHWISE_CONV_2D = 4,
  DEPTH_TO_SPACE = 5,
  DEQUANTIZE = 6,
  EMBEDDING_LOOKUP = 7,
  FLOOR = 8,
  FULLY_CONNECTED = 9,
  HASHTABLE_LOOKUP = 10,
  L2_NORMALIZATION = 11,
  L2_POOL_2D = 12,

  fused_activation_function:ActivationFunctionType;
  dilation_w_factor:int = 1;
  dilation_h_factor:int = 1;
  // Parameters for Conv2D version 8 or above.
  // When set, quantized_bias_type defines the dtype for both bias and accumulator.
  quantized_bias_type: TensorType;
}

// Options for both Conv3D and Conv3DTranspose.
table Conv3DOptions {
  padding:Padding;
  stride_d:int;
  stride_w:int;
  stride_h:int;

  %0 = "tf.Conv2D"(%arg0, %arg1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true
  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>
  %1 = "tf.Relu6"(%0) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>


  %3 = "tf.Conv2D"(%arg0, %arg1) {

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst_1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true

    } else if (function_name.contains("conv2d")) {
      // For Conv2D, the channel dimension must be static to calculate the
      // feature group count.
      if (!HasStaticShapeAtDims(call_op->getOperand(0), /*dims=*/3)) {
        return absl::InternalError(
            "The channel dimension of Conv2D is required to be static.");
      }
    } else if (function_name.contains("conv3d")) {

// CHECK: %[[CONV_1:.*]] = "tf.Conv2D"(%[[GATHER]], %[[DEQUANTIZED_1]]) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}> {device = ""} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1024xf32>) -> tensor<1x3x2x1024xf32>

// CHECK: %[[CONV2D:.*]] = "tf.Conv2D"(%arg0, %[[CONST]]) <{data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}> : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>

    }
  }

  // Handle Conv2D input, stride and filter.
  HandleConv2DInput(conv2d, block_size);
  HandleConv2DStride(conv2d);
  HandleConv2DFilter(conv2d, block_size);

  // Book keeping new filter shape for backprop filter rewrite.
  // Filter shape is defined in HandleConv2DFilter, thus it is RankedTensorType.
  filter_shape =
      mlir::cast<RankedTensorType>(conv2d.getFilter().getType()).getShape();