// CHECK: %[[X_CENTER:.+]] = mhlo.subtract %[[X]], %[[MEAN_BCAST]] : tensor<4x256xf32>
  // CHECK: %[[X_SCALED:.+]] = mhlo.multiply %[[X_CENTER]], %[[SCALE_BCAST]] : tensor<4x256xf32>
  // CHECK: %[[X_NORMED:.+]] = mhlo.divide %[[X_SCALED]], %[[STDDEV_BCAST]] : tensor<4x256xf32>
  // CHECK: %[[RESULT:.+]] = mhlo.add %[[X_NORMED]], %[[OFFSET_BCAST]] : tensor<4x256xf32>
  %0 = "mhlo.batch_norm_inference"(%x, %scale, %offset, %mean, %variance)

  // CHECK-DAG: %[[LHS_1:.*]] = mhlo.convert %arg0 : tensor<3x2xi32>
  // CHECK-DAG: %[[LHS_2:.*]] = mhlo.bitcast_convert %[[LHS_1]] : (tensor<3x2xi32>) -> tensor<3x2x!quant.uniform<i32:f32, 2.000000e+00:4>>
  // CHECK-DAG: %[[RHS_1:.*]] = mhlo.convert %arg1 : tensor<2xi32>
  // CHECK-DAG: %[[RHS_2:.*]] = mhlo.bitcast_convert %[[RHS_1]] : (tensor<2xi32>) -> tensor<2x!quant.uniform<i32:f32, 2.000000e+00:4>>

#include "xla/mlir_hlo/mhlo/IR/hlo_ops.h"  // IWYU pragma: keep

namespace mlir {
namespace odml {

// Given a Composite op that wraps a core.aten.avg_pool2d, returns the padding
// configuration required for the `tfl.pad` if the padding part of the op is
// to be done before average pooling.
DenseIntElementsAttr GetPadOpAttr(Builder& builder, mhlo::CompositeOp op);

    if (failed(output_type)) {
      return failure();
    }

    auto input_quant = rewriter.create<mhlo::BitcastConvertOp>(
        op->getLoc(), *input_quant_type, input);
    auto result = rewriter.create<mhlo::UniformQuantizeOp>(
        op->getLoc(), *output_type, input_quant);
    rewriter.replaceOpWithNewOp<mhlo::BitcastConvertOp>(
        op,
        output_type->clone(

#include "tensorflow/compiler/mlir/tensorflow/ir/tf_saved_model.h"
#include "tensorflow/compiler/mlir/tensorflow/transforms/passes.h"
#include "xla/mlir_hlo/mhlo/IR/register.h"
#include "xla/mlir_hlo/mhlo/transforms/passes.h"
#include "tensorflow/core/ir/types/dialect.h"

int main(int argc, char** argv) {
  tensorflow::InitMlir y(&argc, &argv);

  mlir::registerAllPasses();

  }
};

// Converts TensorFlow DiagPartOp to HLO ops using reduction on masked matrix.
// For a Rank-2 input, it creates the following ops:
//   %1 = "mhlo.iota"() {iota_dimension = 0 : i64}
//   %2 = "mhlo.iota"() {iota_dimension = 1 : i64}
//   %3 = "mhlo.compare"(%1, %2) {comparison_direction = "EQ"}
//   %4 = mhlo.constant dense<0.000000e+00> : tensor<f32>
//   %5 = "mhlo.broadcast"(%4)
//   %6 = "mhlo.select"(%3, %input, %5)

    lhs_flattend = rewriter.create<mhlo::ReshapeOp>(
        loc,
        RankedTensorType::get(lhs_flattened_shape, lhs_type.getElementType()),
        lhs_transposed.getResult());
  } else {
    auto lhs_flattend_shape_op = BuildDotOperandFlattenedShapeOp(
        lhs, lhs_dot_dimensions_info, builder, /*is_lhs=*/true);
    lhs_flattend = rewriter.create<mhlo::DynamicReshapeOp>(
        loc,

// Converts a mhlo.reduce op with a mlho binary operation into a tensorflow
// reduction operation. If the initial value can be ignored, then convert it
// into a single TfReduceOp. Otherwise, convert it into a TfReduceOp followed by
// a TfBinaryOp.
// For example:
//   1) A mhlo::ReduceOp on value `x` with a mhlo::AndOp and a constant initial
// value `true` is converted to a TF::Any on value `x`.

#include "xla/client/sharding_builder.h"
#include "xla/mlir_hlo/mhlo/IR/hlo_ops.h"
#include "xla/primitive_util.h"
#include "xla/side_effect_util.h"
#include "xla/translate/mhlo_to_hlo/type_to_shape.h"

namespace mlir {

using func::FuncOp;

namespace mhlo {

namespace {
constexpr char kShardingAttr[] = "mhlo.sharding";
constexpr char kFrontendAttributesAttr[] = "mhlo.frontend_attributes";

      rewriter.replaceOpWithNewOp<mhlo::TransposeOp>(
          scatter_op, scatter_op.getResult(0).getType(), tf_scatter_op,
          inverse_permutation);

      return success();
    }
  }
};

using ConvertScatterAddOp =
    ConvertScatterOp<mhlo::AddOp, TF::TensorScatterAddOp>;
using ConvertScatterMaxOp =
    ConvertScatterOp<mhlo::MaxOp, TF::TensorScatterMaxOp>;
using ConvertScatterMinOp =