Function halfToFloat

numeric_types/half.hpp:229–319

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

 * Based on Mike Acton's half.c implementation.
 */
float halfToFloat(half h)
{
  // Constants for bit masks, shifts, and biases
  const uint16_t ONE = 0x0001;
  const uint16_t TWO = 0x0002;
  const uint32_t FLOAT_EXP_MASK = 0x7f800000;
  const uint32_t FLOAT_MANTISSA_MASK = 0x007fffff;
  const uint16_t FLOAT_EXP_BIAS = 0x007f;
  const uint16_t HALF_EXP_BIAS = 0x000f;
  const uint16_t HALF_SIGN_MASK = 0x8000;
  const uint16_t HALF_EXP_MASK = 0x7c00;
  const uint16_t HALF_MANTISSA_MASK = 0x03ff;
  const uint16_t HALF_EXP_POS = 0x000a;
  const uint16_t FLOAT_EXP_POS = 0x0017;
  const uint16_t FLOAT_SIGN_POS = 0x001f;
  const uint16_t HALF_SIGN_POS = 0x000f;
  const uint16_t HALF_FLOAT_DENORM_SA_OFFSET = 0x000a;
  const uint32_t HALF_FLOAT_BIAS_OFFSET = HALF_EXP_BIAS - FLOAT_EXP_BIAS;
  const uint16_t HALF_FLOAT_SIGN_POS_OFFSET = FLOAT_SIGN_POS - HALF_SIGN_POS;
  const uint16_t HALF_FLOAT_MANTISSA_POS_OFFSET = FLOAT_EXP_POS - HALF_EXP_POS;

  // Extracting the sign, exponent, and mantissa from the 16-bit float
  const uint32_t halfSignMasked = h.data & HALF_SIGN_MASK;
  const uint32_t halfExpMasked = h.data & HALF_EXP_MASK;
  const uint16_t halfMantissa = h.data & HALF_MANTISSA_MASK;

  // Shifting the sign bit to the correct position for the 32-bit float
  const uint32_t floatSign = halfSignMasked << HALF_FLOAT_SIGN_POS_OFFSET;

  // Adjusting the exponent
  const uint16_t halfExpHalfBias = halfExpMasked >> HALF_EXP_POS;
  const uint32_t floatExp = halfExpHalfBias - HALF_FLOAT_BIAS_OFFSET;

  // Shifting the mantissa to the correct position for the 32-bit float
  const uint32_t floatMantissa = halfMantissa << HALF_FLOAT_MANTISSA_POS_OFFSET;

  // Checking conditions for zero, denormalized, infinity, and NaN
  const uint32_t isExpNonZero = halfExpMasked != 0;
  const uint32_t isMantissaNonZero = halfMantissa != 0;
  const uint32_t isZero = !(isExpNonZero || isMantissaNonZero);
  const uint32_t isDenorm = !isZero && !isExpNonZero;
  const uint32_t isExpFlagged = halfExpMasked == HALF_EXP_MASK;
  const uint32_t isInf = isExpFlagged && !isMantissaNonZero;
  const uint32_t isNan = isExpFlagged && isMantissaNonZero;

  // Handling denormalized numbers
  const uint16_t halfMantissaLeadingZeros = __builtin_clz(halfMantissa) - 16;
  const uint16_t halfDenormShiftAmount =
      halfMantissaLeadingZeros + HALF_FLOAT_DENORM_SA_OFFSET;
  const uint32_t halfFloatDenormMantissaShiftAmount =
      halfDenormShiftAmount - TWO;
  const uint32_t halfFloatDenormMantissa =
      halfMantissa << halfFloatDenormMantissaShiftAmount;
  const uint32_t floatDenormMantissa =
      halfFloatDenormMantissa & FLOAT_MANTISSA_MASK;
  const uint32_t halfFloatDenormShiftAmount = ONE - halfDenormShiftAmount;
  const uint32_t floatDenormExp = halfFloatDenormShiftAmount + FLOAT_EXP_BIAS;
  const uint32_t floatDenormExpPacked = floatDenormExp << FLOAT_EXP_POS;