Function halfFromFloat

numeric_types/half.hpp:106–222

* @brief Converts a 32-bit float to a 16-bit half-precision float. * * Based on Mike Acton's half.c implementation. */

 * Based on Mike Acton's half.c implementation.
 */
half halfFromFloat(float f)
{
  union
  {
    float f;
    uint32_t u;
  } floatUnion = {f};

  uint32_t float32 = floatUnion.u;

  // Constants for bit masks, shifts, and biases
  const uint16_t ONE = 0x0001;
  const uint32_t FLOAT_SIGN_MASK = 0x80000000;
  const uint32_t FLOAT_EXP_MASK = 0x7f800000;
  const uint32_t FLOAT_MANTISSA_MASK = 0x007fffff;
  const uint32_t FLOAT_HIDDEN_BIT = 0x00800000;
  const uint32_t FLOAT_ROUND_BIT = 0x00001000;
  const uint16_t FLOAT_EXP_BIAS = 0x007f;
  const uint16_t HALF_EXP_BIAS = 0x000f;
  const uint16_t FLOAT_SIGN_POS = 0x001f;
  const uint16_t HALF_SIGN_POS = 0x000f;
  const uint16_t FLOAT_EXP_POS = 0x0017;
  const uint16_t HALF_EXP_POS = 0x000a;
  const uint16_t HALF_EXP_MASK = 0x7c00;
  const uint16_t FLOAT_EXP_FLAGGED_VALUE = 0x00ff;
  const uint16_t HALF_EXP_MASK_VALUE = HALF_EXP_MASK >> HALF_EXP_POS;
  const uint16_t HALF_EXP_MAX_VALUE = HALF_EXP_MASK_VALUE - ONE;
  const uint16_t FLOAT_HALF_SIGN_POS_OFFSET = FLOAT_SIGN_POS - HALF_SIGN_POS;
  const uint16_t FLOAT_HALF_BIAS_OFFSET = FLOAT_EXP_BIAS - HALF_EXP_BIAS;
  const uint16_t FLOAT_HALF_MANTISSA_POS_OFFSET = FLOAT_EXP_POS - HALF_EXP_POS;
  const uint16_t HALF_NAN_MIN = HALF_EXP_MASK | ONE;

  // Extracting the sign, exponent, and mantissa from the 32-bit float
  const uint32_t floatSignMasked = float32 & FLOAT_SIGN_MASK;
  const uint32_t floatExpMasked = float32 & FLOAT_EXP_MASK;
  const uint16_t halfSign =
      static_cast<uint16_t>(floatSignMasked >> FLOAT_HALF_SIGN_POS_OFFSET);
  const uint16_t floatExp =
      static_cast<uint16_t>(floatExpMasked >> FLOAT_EXP_POS);
  const uint32_t floatMantissa = float32 & FLOAT_MANTISSA_MASK;

  // Check for NaN
  if ((floatExpMasked == FLOAT_EXP_MASK) && (floatMantissa != 0))
  {
    half result;
    result.data =
        HALF_EXP_MASK | (floatMantissa >> FLOAT_HALF_MANTISSA_POS_OFFSET);
    return result;
  }

  // Adjusting the exponent and rounding the mantissa
  const uint16_t floatExpHalfBias = floatExp - FLOAT_HALF_BIAS_OFFSET;
  const uint32_t floatMantissaRoundMask = floatMantissa & FLOAT_ROUND_BIT;
  const uint32_t floatMantissaRoundOffset = floatMantissaRoundMask << ONE;
  const uint32_t floatMantissaRounded =
      floatMantissa + floatMantissaRoundOffset;

  // Handling denormalized numbers