// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range -tfl-quantize="enable-dynamic-range-quantization=true" | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range -tfl-quantize="enable-dynamic-range-quantization=true enable-weight-only-quantization=true" | FileCheck --check-prefix=PerChannelWeightOnly %s

// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="enable-dynamic-range-per-channel-quantization=false" | FileCheck --check-prefix=PerTensor %s
// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="enable-float16-quantization" | FileCheck --check-prefix=Float16 %s

// 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>

// RUN: tf-opt %s -tfl-prepare-quantize-dynamic-range="enable-custom-op-quantization=CustomTestOp=1" -tfl-quantize="enable-dynamic-range-quantization=true enable-custom-op-weight-only=CustomTestOp=false" -tfl-post-quantize="enable-no-side-effect=CustomTestOp=false" | FileCheck --check-prefix=NotPrune %s

  func.return %rst : tensor<8xf32>

// CHECK: %[[CONSTANT:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT]]) <{qtype = tensor<8x!quant.uniform<u8:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantFolded
func.func @fakeQuantFolded() -> (tensor<8xf32>) {

  func.return %rst : tensor<8xf32>

// CHECK: %[[CONSTANT:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT]]) <{qtype = tensor<8x!quant.uniform<u4:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantFolded
func.func @fakeQuantFolded() -> (tensor<8xf32>) {

             "bool", "false",
             "Whether to run weight-only for post-training dynamic range quantization pass">,
      Option<"enable_legacy_quantize_", "legacy-quantize",
             "bool", "false",
             "Use legacy quantize mode in test. Valid when `-legacy-quantize` is set.">,
      ListOption<"ops_blocklist_flag_", "ops-blocklist",
             "std::string", "Names of ops to blocklist from quantization">,

    // This is the argument used to refer to the pass in
    // the textual format (on the commandline for example).
    return "quant-quantize-weights";
  }

  StringRef getDescription() const final {
    // This is a brief description of the pass.
    return "Quantize weights used by quantizable ops.";
  }

  void getDependentDialects(DialectRegistry& registry) const override {

#include "tensorflow/compiler/mlir/tensorflow/ir/tf_dialect.h"

//===----------------------------------------------------------------------===//
// The prepare-quantize-drq Pass.
//
namespace mlir {
namespace quant {

namespace {

using QuantizationUnit = std::pair<Operation*, int>;
using QuantizationUnits = llvm::SetVector<QuantizationUnit>;

  // Inserts the Quantize ops for requantizing a block argument.
  void RequantizeArg(BlockArgument arg, RequantizeStates& states);

  // Inserts the Quantize and Dequantize ops to quantize the value and returns
  // the Quantize op.
  void RequantizeValue(Value value, RequantizeStates& states, Location loc);

  // Returns the quantization parameter satisfies the same scale