34 	 * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.  in fp16_ieee_to_fp32_bits()
47 	 * Extract mantissa and biased exponent of the input number into the bits 0-30 of the 32-bit word:  in fp16_ieee_to_fp32_bits()
57 …* If the initial number is normalized, some of its high 6 bits (sign == 0 and 5-bit exponent) equa…  in fp16_ieee_to_fp32_bits()
59 …* denormalized nonsign by renorm_shift, the unit bit of mantissa will shift into exponent, turning…  in fp16_ieee_to_fp32_bits()
60 	 * biased exponent into 1, and making mantissa normalized (i.e. without leading 1).  in fp16_ieee_to_fp32_bits()
71 	 * Iff half-precision number has exponent of 15, the addition overflows it into bit 31,  in fp16_ieee_to_fp32_bits()
74 …*                   0x7F800000 if the half-precision number had exponent of 15 (i.e. was NaN or in…  in fp16_ieee_to_fp32_bits()
88 …* 2. Shift nonsign right by 3 so the exponent (5 bits originally) becomes an 8-bit field and 10-bi…  in fp16_ieee_to_fp32_bits()
90 	 * 3. Add 0x70 to the exponent (starting at bit 23) to compensate the different in exponent bias  in fp16_ieee_to_fp32_bits()
92 …* 4. Subtract renorm_shift from the exponent (starting at bit 23) to account for renormalization. …  in fp16_ieee_to_fp32_bits()
94 	 * 5. Binary OR with inf_nan_mask to turn the exponent into 0xFF if the input was NaN or infinity.  in fp16_ieee_to_fp32_bits()
95 …* 6. Binary ANDNOT with zero_mask to turn the mantissa and exponent into zero if the input was zer…  in fp16_ieee_to_fp32_bits()
116 	 * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.  in fp16_ieee_to_fp32_value()
129 	 * Extract mantissa and biased exponent of the input number into the high bits of the 32-bit word:  in fp16_ieee_to_fp32_value()
139 	 * Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become mantissa and exponent  in fp16_ieee_to_fp32_value()
142 	 *       S|Exponent |          Mantissa  in fp16_ieee_to_fp32_value()
148 	 * Next, there are some adjustments to the exponent:  in fp16_ieee_to_fp32_value()
149 …* - The exponent needs to be corrected by the difference in exponent bias between single-precision…  in fp16_ieee_to_fp32_value()
152 …*   Therefore, if the biased exponent of the half-precision input was 0x1F (max possible value), t…  in fp16_ieee_to_fp32_value()
154 …*   - First, we adjust the exponent by (0xFF - 0x1F) = 0xE0 (see exp_offset below) rather than by …  in fp16_ieee_to_fp32_value()
155 	 *     by the difference in the exponent bias (see above).  in fp16_ieee_to_fp32_value()
156 …*   - Then we multiply the single-precision result of exponent adjustment by 2**(-112) to reverse …  in fp16_ieee_to_fp32_value()
157 …*     exponent adjustment by 0xE0 less the necessary exponent adjustment by 0x70 due to difference…  in fp16_ieee_to_fp32_value()
161 …* Note that the above operations do not handle denormal inputs (where biased exponent == 0). Howev…  in fp16_ieee_to_fp32_value()
176 	 * In a denormalized number the biased exponent is zero, and mantissa has on-zero bits.  in fp16_ieee_to_fp32_value()
188 	 * and with an exponent which would scale the corresponding mantissa bits to 2**(-24).  in fp16_ieee_to_fp32_value()
190 	 *    FP32 = (1 + mantissa * 2**(-23)) * 2**(exponent - 127)  in fp16_ieee_to_fp32_value()
191 …* Therefore, when the biased exponent is 126, a unit change in the mantissa of the input denormali…  in fp16_ieee_to_fp32_value()
206 …*   input exponent. The variable two_w contains input exponent in bits 27-31, therefore if its sma…  in fp16_ieee_to_fp32_value()
208 	 * - Combine the result of conversion of exponent and mantissa with the sign of the input number.  in fp16_ieee_to_fp32_value()
263 	 * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.  in fp16_alt_to_fp32_bits()
276 	 * Extract mantissa and biased exponent of the input number into the bits 0-30 of the 32-bit word:  in fp16_alt_to_fp32_bits()
286 …* If the initial number is normalized, some of its high 6 bits (sign == 0 and 5-bit exponent) equa…  in fp16_alt_to_fp32_bits()
288 …* denormalized nonsign by renorm_shift, the unit bit of mantissa will shift into exponent, turning…  in fp16_alt_to_fp32_bits()
289 	 * biased exponent into 1, and making mantissa normalized (i.e. without leading 1).  in fp16_alt_to_fp32_bits()
309 …* 2. Shift nonsign right by 3 so the exponent (5 bits originally) becomes an 8-bit field and 10-bi…  in fp16_alt_to_fp32_bits()
311 	 * 3. Add 0x70 to the exponent (starting at bit 23) to compensate the different in exponent bias  in fp16_alt_to_fp32_bits()
313 …* 4. Subtract renorm_shift from the exponent (starting at bit 23) to account for renormalization. …  in fp16_alt_to_fp32_bits()
315 …* 5. Binary ANDNOT with zero_mask to turn the mantissa and exponent into zero if the input was zer…  in fp16_alt_to_fp32_bits()
336 	 * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 - zero bits.  in fp16_alt_to_fp32_value()
349 	 * Extract mantissa and biased exponent of the input number into the high bits of the 32-bit word:  in fp16_alt_to_fp32_value()
359 	 * Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become mantissa and exponent  in fp16_alt_to_fp32_value()
362 	 *       S|Exponent |          Mantissa  in fp16_alt_to_fp32_value()
368 …* Next, the exponent is adjusted for the difference in exponent bias between single-precision and …  in fp16_alt_to_fp32_value()
370 …* half-precision exponent is 0x1F and after the adjustment is can not exceed 0x8F < 0xFE (largest …  in fp16_alt_to_fp32_value()
371 	 * exponent for non-finite values).  in fp16_alt_to_fp32_value()
373 …* Note that this operation does not handle denormal inputs (where biased exponent == 0). However, …  in fp16_alt_to_fp32_value()
383 	 * In a denormalized number the biased exponent is zero, and mantissa has on-zero bits.  in fp16_alt_to_fp32_value()
395 	 * and with an exponent which would scale the corresponding mantissa bits to 2**(-24).  in fp16_alt_to_fp32_value()
397 	 *    FP32 = (1 + mantissa * 2**(-23)) * 2**(exponent - 127)  in fp16_alt_to_fp32_value()
398 …* Therefore, when the biased exponent is 126, a unit change in the mantissa of the input denormali…  in fp16_alt_to_fp32_value()
413 …*   input exponent. The variable two_w contains input exponent in bits 27-31, therefore if its sma…  in fp16_alt_to_fp32_value()
415 	 * - Combine the result of conversion of exponent and mantissa with the sign of the input number.  in fp16_alt_to_fp32_value()