// 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-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-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

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true" -cse | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true legacy-float-scale=true" -cse| FileCheck --check-prefix=Legacy %s

// CHECK-LABEL: QuantizeLstmCellInput
func.func @QuantizeLstmCellInput(%arg0: tensor<1x28x28xf32>) -> tensor<1x28x20xf32> {
    %cst_2 = "tfl.no_value"() {value = unit} : () -> none

  // CHECK: %[[DQ_1:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_5:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

// 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>

A. Unique TensorFlower <******@****.***> 1713119208 -0700

    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    Value value = tf_op.getOutputs();
    auto quantize = rewriter.create<TFL::QuantizeOp>(
        tf_op.getLoc(), qtype.getValue(), value, qtype);
    auto dequantize = rewriter.create<TFL::DequantizeOp>(
        tf_op.getLoc(), res_type, quantize.getOutput());

// CHECK: %[[q0:.*]] = "tfl.quantize"(%arg1) <{qtype = tensor<2x1x!quant.uniform<u8:f32, 0.0078431372549019607:128>>}>
// CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) <{qtype = tensor<2x2x!quant.uniform<u8:f32, 1.000000e+00:128>>}>
// CHECK: %[[add:.*]] = tfl.add(%[[q1]], %[[q0]]) <{fused_activation_function = "NONE"}> : (tensor<2x2x!quant.uniform<u8:f32, 1.000000e+00:128>>

  }
};

// Inserts a "tfl.quantize" and "tfl.dequantize" op pair (QDQs) after the
// "tf.FakeQuantWithMinMaxVarsOp" to be constant folded. Since the constant
// folding logic will use a "arith.constant" op to replace the
// "tf.FakeQuantWithMinMaxVarsOp", the "tfl.quantize" op is used to preserve
// the quantization parameters as a TypeAttr and "tfl.dequantize" op used to