// Allocate sufficient memory to tensor buffer.
  int32_t num_strings = offset_.size() - 1;
  // Total bytes include:
  //   * size of content (data_.size)
  //   * offset of each tensor (sizeof(int32_t) * num_strings)
  //   * length of whole buffer (int32_t)
  //   * num of strings (int32_t).
  int32_t bytes = data_.size()                            // size of content
                  + sizeof(int32_t) * (num_strings + 2);  // size of header

	src := C.CString(input)

	cn := C.int32_t(0)

	switch caser {
	case "fold":
		cn = C.ucasemap_utf8FoldCase(cm,
			dst, C.int32_t(len(buf)),
			src, C.int32_t(len(input)),
			&err)
	case "lower":
		cn = C.ucasemap_utf8ToLower(cm,
			dst, C.int32_t(len(buf)),
			src, C.int32_t(len(input)),
			&err)
	case "upper":
		cn = C.ucasemap_utf8ToUpper(cm,
			dst, C.int32_t(len(buf)),
			src, C.int32_t(len(input)),

#include <cstdint>
#include <optional>
#include <utility>

#include "absl/types/optional.h"

namespace mlir {
namespace quant {

using QuantizedMultiplier = std::pair<int32_t, int32_t>;
using QuantizedRange = std::pair<int32_t, int32_t>;

// Decompose double precision multiplier to integer multiplier and exponent.
//    double_multiplier = int_multiplier * 2 ^ (-31 + exponent)
// int_multiplier will be range of (2^31, 2^30].

                          Value &padding_high) {
  Value zero = CreateScalarConstValue<int32_t>(builder, loc, 0);
  Value one = CreateScalarConstValue<int32_t>(builder, loc, 1);
  Value two = CreateScalarConstValue<int32_t>(builder, loc, 2);
  Value filter_size = CreateScalarConstValue<int32_t>(builder, loc, filter_sz);
  Type int32_scalar_type = zero.getType();

  auto scalar_add = [&](Value lhs, Value rhs) {

}

TEST(CalibrationParametersTest, CalculateNumBinsDivisible) {
  int32_t num_bins = CalculateActualNumBins(
      /*min_value=*/0.0, /*max_value=*/4.0, /*bin_width=*/2.0);

  // Expect 2 bins: [0, 2), [2, 4].
  EXPECT_EQ(num_bins, 2);
}

TEST(CalibrationParametersTest, CalculateNumBinsNotDivisible) {
  int32_t num_bins = CalculateActualNumBins(
      /*min_value=*/0.0, /*max_value=*/5.0, /*bin_width=*/2.0);

#include <gtest/gtest.h>
#include "mlir/Support/LogicalResult.h"  // from @llvm-project

namespace mlir::quant::stablehlo {
namespace {

TEST(UtilsTest, QuantizeMultiplierNormalMultipliers) {
  int32_t quantized_fraction;
  int32_t shift;

  EXPECT_TRUE(succeeded(QuantizeMultiplier(1.2, quantized_fraction, shift)));
  EXPECT_EQ(quantized_fraction, 19661);
  EXPECT_EQ(shift, 1);

bool DenseI64AttrToI32Vector(const DenseIntElementsAttr& dense_attr,
                             std::vector<int32_t>* out_vec) {
  std::vector<int32_t> ret(dense_attr.getNumElements());
  auto range = dense_attr.getValues<int64_t>();
  std::transform(range.begin(), range.end(), ret.begin(),
                 [](int64_t attr) { return static_cast<int32_t>(attr); });
  *out_vec = std::move(ret);
  return true;
}

QuantizedRange CalculateQuantizedRange(double scale, int32_t zero_point,
                                       std::optional<double> rmin,
                                       std::optional<double> rmax, int32_t qmin,
                                       int32_t qmax) {
  auto quantize = [scale, zero_point](float f) {
    return zero_point + static_cast<int32_t>(std::round(f / scale));
  };

}

// Calculates the bin index of the current value.
inline int32_t CalculateBinIndex(const float value, const float lower_bound,
                                 const float bin_width) {
  return std::floor((value - lower_bound) / bin_width);
}

// Same as `CalculateBinIndex` but clamps to avoid out-of-bound.
inline int32_t CalculateBinIndexSafe(const float value, const float lower_bound,

llvm::SmallVector<int32_t> InvertPermutation(llvm::SmallVector<int32_t> arr) {
  llvm::SmallVector<int32_t> inverse_arr(arr.size());
  for (int32_t i = 0; i < arr.size(); ++i) {
    inverse_arr[arr[i]] = i;
  }
  return inverse_arr;
}

llvm::SmallVector<int64_t> PermuteShape(llvm::ArrayRef<int64_t> shape,
                                        llvm::ArrayRef<int32_t> perm) {