return %7 : tensor<1x3xf32>
  }
// Test that the inputs and output of the tf.XlaCallModule op has been replaced
// by quantized types, and the corresponding quantfork.dcast ops that turned
// those quantized types back to float types are removed.
// CHECK: %[[CONST_0:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<4x3xf32>

// 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>

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

  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: %[[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>

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

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.