include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"

// Quantize attribute $0 by using quantization parameter from %1.
def QuantizeByQuantizedType : NativeCodeCall<"quant::Quantize($0, $1.getValue())">;
def F32ElementsAttr : ElementsAttrBase<
  CPred<"$_self.cast<ElementsAttr>().getShapedType().getElementType().isF32()">, "float constant tensor">;

// Squash tfl.dequantize and tfl.quantize pairs.

    auto func_op =
        dyn_cast_or_null<func::FuncOp>(symbol_table.lookup(func_name));
    if (!func_op) return failure();
    // The quantized fusion should have requantize and return ops at the end.
    auto return_op = dyn_cast_or_null<func::ReturnOp>(
        func_op.getRegion().getBlocks().front().getTerminator());
    if (!return_op) return failure();

  // 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>

    // 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());

    %0 = "tfl.dequantize"(%arg0) : (tensor<1x224x224x3x!quant.uniform<u8:f32, 1.0>>) -> tensor<1x224x224x3xf32>
    %1 = "tfl.dequantize"(%arg1) : (tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 1.0>>) -> tensor<32x3x3x3xf32>
    %2 = "tfl.dequantize"(%arg2) : (tensor<32x!quant.uniform<i32:f32, 1.0>>) -> tensor<32xf32>

  }
};

// 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

// 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)))) {

// CHECK:   %{{.*}} = "tfl.dequantize"(%{{.*}}) : (tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>) -> tensor<1x401408xf32>

  %cst = arith.constant dense<[1, 401408]> : tensor<2xi32>