// func.func func_name_${key1}_fn (...) {
//   ...${key2}...
// }
// ```
// The above template with generate two functions by substituting `key1` and
// `key2` with given values.

module {

  for main_op in ["Conv2D", "DepthwiseConv2D", "MatMul"] {
    parameters[
      {"quantized_ops": ["${main_op}", "BiasAdd"], "act_func": "internal_requantize_no_activation_fn", "output_type": "!tf_type.qint8"},

  // Canonicalization includes const folding, which is utilized here to optimize
  // away ops that can't get constant folded after PrepareTF pass. For example,
  // tf.Conv2D is split into tf.Transpose and tfl.Conv2D.
  pass_manager->addNestedPass<mlir::func::FuncOp>(
      mlir::createCanonicalizerPass());
  pass_manager->addNestedPass<mlir::func::FuncOp>(mlir::createCSEPass());

    dilation_width_factor = rewriter.getI32IntegerAttr(1);
  }

  StringAttr padding;
  if (!TFPaddingIsSameOrValid(op, &padding)) return failure();

  // TensorFlow Conv3D has no bias, optimization patterns will fuse Conv3D
  // with other ops can fill the bias.
  Value none = rewriter.create<TFL::NoValueOp>(
      op->getLoc(), rewriter.getNoneType(), rewriter.getUnitAttr());

  %0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device...

			rows.raw = rows.raw[:0]
			test(tt.name, tt.in, tt.want)
		}
	})

	// The numbers below are only valid for 64-bit interface word sizes,
	// and gc. With 32-bit words there are more convT2E allocs, and
	// with gccgo, only pointers currently go in interface data.
	// So only care on amd64 gc for now.
	measureAllocs := false
	switch runtime.GOARCH {
	case "amd64", "arm64":

  // TFR_ConstantTensorOp (
  //   ConstantOp (ConstAttr<F32Attr (in_scale[0] * in_scale[1] /
  //   out_scale))
  // )
  // Currently, all decompositions using this pattern (Conv2D, FC) have the
  // following preconditions:
  // * out_scale: float scalar attribute
  // * in_scale[0] (input scale): float scalar, given by tf.Const -> tfr.cast
  // * in_scale[1] (filter scale): float scalar/vector

    quantization_axis = -1 : i64, quantization_min_val = -128 : i64,
    quantization_max_val = 127 : i64
  } : (
    tensor<3x3x10x20x!tf_type.qint8>, tensor<f32>, tensor<i32>
  ) -> tensor<3x3x10x20xf32>
  %0 = "tf.Conv2D"(%input, %filter_new) {
    Tin = "tfdtype$DT_FLOAT", Tout = "tfdtype$DT_FLOAT",
    attr_map = "", batch_group_count = 1 : i64,
    explicit_padding = [], feature_group_count = 1 : i64, lhs_dilation = [1, 1],