#include "tensorflow/compiler/mlir/quantization/common/quantization_lib/quantization_utils.h"

namespace mlir {
namespace TFL {

//===----------------------------------------------------------------------===//
// The actual Quantize Pass.
//===----------------------------------------------------------------------===//
namespace {
#define GEN_PASS_DEF_QUANTIZEPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true" | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true disable-per-channel=true" | FileCheck --check-prefix=PerTensor %s

// CHECK-LABEL: uint8_to_int8
func.func @uint8_to_int8(%arg0: tensor<2x2xf32>) -> tensor<2x2xf32> {
  %1 = "tfl.quantize"(%arg0) {qtype = tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>} : (tensor<2x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>

// RUN: tf-opt %s -tfl-prepare-quantize="post-training-quantize=true" -tfl-quantize="numeric-verify=true log-if-failed=true" | FileCheck --check-prefix=DEBUG %s
// RUN: tf-opt %s -tfl-prepare-quantize="post-training-quantize=true" -tfl-quantize="numeric-verify=true log-if-failed=true whole-model-verify=true" | FileCheck --check-prefix=MODEL-DEBUG %s

// DEBUG-LABEL: QuantizeConv2D

// 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">,