// Ternary ops patterns.
//===----------------------------------------------------------------------===//
// Multi-pattern consisting of matching stand-alone convolution op followed by
// activation op.
multiclass FuseActFnIntoConvOpPat<Op ActFnOp, ConstantStrAttr ActFnAttr> {
  def FuseActivationFuncWithConv#ActFnOp#ActFnAttr : Pat<
    (ActFnOp (TFL_Conv2DOp:$conv_out $input, $filter, $bias, $h_factor,

// dynamic batch dimension is properly quantized.

// Note that this checks for identical condition as
// quantize_conv_with_bias_dynamic_fn, omitting stablehlo.maximum.
// This is because activation clipping which includes 0.0f can be simply
// omitted from the graph as the lifted function's out_scale and out_zp are
// already calculated based on the clipped distribution.

    // optimizations in the pipeline.
    METHOD_NO_QUANTIZE = 1;

    // Static range quantization. Quantized tensor values' ranges are statically
    // determined. The activation and weight are quantized to INT8 while bias is
    // quantized to INT32.
    METHOD_STATIC_RANGE_INT8 = 2;

    // Dynamic range quantization. Quantized tensor values' ranges are

  QuantizationSpecs specs = 4;

  // Configures the quantization debugger.
  DebuggerConfig debugger_config = 5;

  // Defines calibration options for quantization. This option is only used for
  // activation of static range quantization (SRQ). Quantization calibration
  // method is set to MIN_MAX by default.
  CalibrationOptions calibration_options = 6;

  // Path to file to save the quantization report, which is essentially a

// CHECK: %[[QCONST_0:.+]] =  "tfl.pseudo_qconst"()
// CHECK: "tfl.batch_matmul"(%[[ARG]], %[[QCONST_0]]) <{adj_x = false, adj_y = false}>

// -----

// Tests static range quantized dot_general with activation as RHS

*   Add `UnifiedGRU` as the new GRU implementation for tf2.0. Change the default
    recurrent activation function for GRU from `hard_sigmoid` to `sigmoid`, and
    `reset_after` to True in 2.0. Historically recurrent activation is
    `hard_sigmoid` since it is fast than 'sigmoid'. With new unified backend
    between CPU and GPU mode, since the CuDNN kernel is using sigmoid, we change

    QuantizableResult,
    PredOpTrait<"input and output must have same element type",
      TFL_TCresVTEtIsSameAsOp<0, 0>>]> {
  let summary = "Hardswish activation function.";
  let description = [{
    Computes hard-swish activation function
      f(x) -> (x * relu6(x+3))/6
    element-wise.
  }];

  let arguments = (ins TFL_TensorOf<[F32, QUI8, QI8]>:$input);

// RUN: tf-opt %s -tf-fused-kernel-matcher | FileCheck %s

//===----------------------------------------------------------------------===//
// Conv2D + BiasAdd + <Activation> fusions.
//===----------------------------------------------------------------------===//

// CHECK-LABEL: conv2DBiasAdd_noActivation
func.func @conv2DBiasAdd_noActivation(%arg0: tensor<128xf32>, %arg1: tensor<1x1x3x128xf32>, %arg2: tensor<8x32x32x3xf32>) -> (tensor<*xf32>) {

      Option<"quantize_signed_", "quantize-signed", "bool", "false",
             "signed inference type. Only used in tests">,
      Option<"activation_number_of_bits_", "activation-number-of-bits", "int", "8",
             "number of bits for inference type. Only used in tests">,
      Option<"post_training_quantize_", "post-training-quantize", "bool", "false",

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true activation-number-of-bits=16" -cse | FileCheck %s

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>