Method to_float

cpp/fory/util/float16.cc:120–178

Convert float16 to float32 (exact: every f16 value is representable in f32).

// Convert float16 to float32 (exact: every f16 value is representable in f32).
float float16_t::to_float() const noexcept {
  const uint32_t sign = static_cast<uint32_t>(bits & 0x8000u) << 16;
  const uint32_t exp = (bits >> 10) & 0x1Fu;
  const uint32_t mantissa = bits & 0x03FFu;

  // NaN or Infinity (f16 exp = 0x1F)
  if (exp == 0x1Fu) {
    if (mantissa == 0u) {
      // ±Inf
      const uint32_t f32_bits = sign | 0x7F800000u;
      float result;
      std::memcpy(&result, &f32_bits, sizeof(result));
      return result;
    }
    // NaN: expand 10-bit f16 fraction to 23-bit f32 fraction by shifting
    // left 13 bits.  The f16 quiet bit (bit 9) maps to the f32 quiet bit
    // (bit 22), preserving quiet/signaling status and the payload.
    const uint32_t nan_payload = mantissa << 13;
    const uint32_t f32_bits = sign | 0x7F800000u | nan_payload;
    float result;
    std::memcpy(&result, &f32_bits, sizeof(result));
    return result;
  }

  // ±0
  if (exp == 0u && mantissa == 0u) {
    float result;
    std::memcpy(&result, &sign, sizeof(result));
    return result;
  }

  // Subnormal f16: normalize into a f32 normal.
  // f16 subnormals have true exponent -14 and no implicit leading 1.
  if (exp == 0u) {
    uint32_t m = mantissa;
    int32_t e = -14;
    // Shift left until the implicit leading 1 reaches bit 10.
    while ((m & 0x0400u) == 0u) {
      m <<= 1;
      e -= 1;
    }
    m &= 0x03FFu; // strip implicit leading 1
    const uint32_t exp32 = static_cast<uint32_t>(e + 127);
    const uint32_t mantissa32 = m << 13;
    const uint32_t f32_bits = sign | (exp32 << 23) | mantissa32;
    float result;
    std::memcpy(&result, &f32_bits, sizeof(result));
    return result;
  }

  // Normal f16: remap exponent bias (15 → 127) and zero-extend mantissa
  // (10 → 23 bits).
  const uint32_t exp32 = exp - 15u + 127u;
  const uint32_t mantissa32 = mantissa << 13;
  const uint32_t f32_bits = sign | (exp32 << 23) | mantissa32;
  float result;
  std::memcpy(&result, &f32_bits, sizeof(result));
  return result;