// RUN: stablehlo-quant-opt %s -split-input-file -stablehlo-quantize -verify-each=false | FileCheck %s

// Tests for PopulateFusedGemmStylePatterns are handled in
// quantize_composite_functions for module-level evaluation of functions.

module attributes {tf_saved_model.semantics} {
// CHECK: quantize_simple_xla_call_module(%[[ARG_0:.+]]: tensor<1x4xf32>)
  func.func private @quantize_simple_xla_call_module(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {

// RUN: tf-quant-opt %s -split-input-file -quant-lift-quantizable-spots-as-functions -quant-quantize -verify-each=false | FileCheck %s

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>

  patterns.add<StableHloQuantization, StableHloQuantizationReverse>(&ctx);

  PopulateCommonQuantizationPatterns(ctx, patterns,
                                     enable_per_channel_quantized_weight_);

  // Quantize all quantizable ops, including ops that are not compute-heavy.
  PopulateAllQuantizablePatterns(ctx, patterns);

  if (failed(applyPatternsAndFoldGreedily(module_op, std::move(patterns)))) {

      auto &requantize = states_manager_.GetOperandRequantizeState(op, i);
      if (state.IsEmpty() && requantize.pos == RequantizeState::NO_REQUANTIZE) {
        input_specs.push_back(original_input_specs[i]);
      } else if (requantize.pos == RequantizeState::ON_OUTPUT) {
        input_specs.push_back(TypeAttr::get(requantize.params));
      } else {

// CHECK: %[[IDENTITY:.*]] = "tf.Identity"(%cst_0) : (tensor<200x100x300xi8>) -> tensor<200x100x300xi8>
// CHECK: %[[DEQUANTIZED:.*]] = "tf.PartitionedCall"(%[[IDENTITY]]) <{config = "", config_proto = "", executor_type = "", f = @composite_dequantize_uniform}> : (tensor<200x100x300xi8>) -> tensor<200x100x300xf32>

      "DataFormatDimMap",
      "DataFormatVecPermute",
      "DepthToSpace",
      "DepthwiseConv2dNative",
      "DepthwiseConv2dNativeBackpropFilter",
      "DepthwiseConv2dNativeBackpropInput",
      "Dequantize",
      "Diag",
      "DynamicInfeedEnqueueTupleOp",
      "DynamicInfeedDequeueTupleOp",
      "DynamicStitch",
      "DynamicPartition",
      "Einsum",
      "EmptyTensorList",

// Quantization ops.
//===----------------------------------------------------------------------===//
def TFL_DequantizeOp: TFL_Op<"dequantize", [NoMemoryEffect]> {
  let summary = "Dequantize operator";

  let description = [{
    Converts quantized array of integers to floating-points according to the
    quantization parameters.
  }];

Jake Harmon <******@****.***> 1694027275 -0700

  // ranges.
  bool SetInputNodesQuantizationParams(func::FuncOp func);

  // The function might contain more stats ops than required, and it will
  // introduce requantize if the calibration stats have conflicts. This method
  // tries to remove all the redundant stats ops.
  bool RemoveRedundantStats(func::FuncOp func);

  // Whether the func contains Quantize ops. This is used to determine whether
  // to use the quantization parameters from the fixed output range property.
  bool ContainsQuantizeOps(func::FuncOp func);

  QuantizationSpecs quant_specs_;

  Option<bool> enable_post_training_quantize_{
      *this, "post-training-quantize", llvm::cl::init(false),