TFDynamicRangeQuantization>(ctx, quant_params) {}
};

// Removes quantize-dequantize pairs that are not used in the quantization.
// The benefit of this pattern is set to lower value than other patterns, so
// that the other patterns can work on quantize/dequantize ops first.
class RemoveUnusedQdqPattern
    : public OpRewritePattern<quantfork::DequantizeCastOp> {
 public:

    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-opt %s -tfl-prepare-quantize -tfl-quantize  | FileCheck %s
// RUN: tf-opt %s -tfl-quantize="legacy-quantize=true" | FileCheck --check-prefix=LEGACY %s
// RUN: tf-opt %s -tfl-prepare-quantize -tfl-quantize="ops-blocklist=tfl.fully_connected,tfl.softmax locs-blocklist=Block,NullBlock" | FileCheck --check-prefix=BLOCK %s

// CHECK-LABEL: QuantizeFloatConst
func.func @QuantizeFloatConst() -> tensor<2x2x!quant.uniform<u8:f32, 7.8431372549019615E-4:128>> {

// 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>

};

class QuantizeConstPattern : public OpRewritePattern<QuantizeOp> {
 public:
  explicit QuantizeConstPattern(MLIRContext* context, bool legacy_float_scale)
      : OpRewritePattern<QuantizeOp>(context),
        legacy_float_scale_(legacy_float_scale) {}
  LogicalResult matchAndRewrite(QuantizeOp op,
                                PatternRewriter& rewriter) const override {

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

// MixedPrecision-NEXT: %[[q:.*]] = "tfl.quantize"(%arg0)
// MixedPrecision-NEXT: %[[dq:.*]] = "tfl.dequantize"(%[[q]])
// MixedPrecision-NEXT: %[[q_0:.*]] = "tfl.quantize"(%arg1)
// MixedPrecision-NEXT: %[[dq_0:.*]] = "tfl.dequantize"(%[[q_0]])
// MixedPrecision-NEXT: %[[c:.*]] = "tfl.concatenation"(%[[dq]], %[[dq_0]])
// MixedPrecision-NEXT: %[[q_1:.*]] = "tfl.quantize"(%[[c]])

  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)))) {

// RUN: tf-opt %s -tfl-quantize-variables | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize -tfl-quantize -tfl-post-quantize -tfl-quantize-variables -tfl-quantize -tfl-post-quantize | FileCheck --check-prefix=WHOLE-PASSES %s

// CHECK-LABEL: QuantizeReadVariable
func.func @QuantizeReadVariable() -> (tensor<1x2x1x3x!quant.uniform<i8:f32, 1.0>>) {
  %1 = "tfl.var_handle"() : () -> tensor<!tf_type.resource>

  %2 = "tfl.dequantize"(%1#0) : (tensor<2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2xf32>
  %3 = "tfl.dequantize"(%1#1) : (tensor<2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2xf32>

  // unused quantization ops should be removed as well.
  %4 = "tfl.dequantize"(%1#2) : (tensor<2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2xf32>