Split an IEEE 754 half-precision word into its class, sign, true exponent and significand.
Hardware context: The weight debug unit of a half-precision accelerator and the reference model of a half-precision adder. Decode a half-precision word into class, sign, true exponent and significand.
A machine-learning accelerator stores its weights in half precision. The debug unit must show each weight as a class, a sign, an exponent and a significand. The first version printed nonsense for every very small weight.
This is an Implement problem. Write half_decode. It reads one 16-bit word: a sign bit, a 5-bit exponent field with bias 15, and a 10-bit fraction.
For a finite word, the value is (-1)^sign x sig x 2^(exp - 10). The table gives the outputs for each class.
The check applies three words, one per nanosecond, and records all four outputs after each.
Signals in the Timeline
| Signal | Width | Role |
|---|---|---|
script | 48 | input of each check |
Full specification
module half_decode (
input logic [15:0] h, // an IEEE 754 half-precision word
output logic [2:0] cls, // 0 zero, 1 subnormal, 2 normal, 3 infinity, 4 NaN
output logic sign, // the sign bit
output logic signed [5:0] exp, // the true exponent: -14 to 15, or 0
output logic [10:0] sig // the significand with its leading bit, or 0
);
// your design
endmodule
| Exponent field e | Fraction f | cls | exp | sig |
|---|---|---|---|---|
| 0 | 0 | 0 (zero) | 0 | 0 |
| 0 | not 0 | 1 (subnormal) | -14 | f with a 0 in front |
| 1 to 30 | any | 2 (normal) | e - 15 | f with a 1 in front |
| 31 | 0 | 3 (infinity) | 0 | 0 |
| 31 | not 0 | 4 (NaN) | 0 | 0 |
sign is bit 15 of the word for every class, zeros and NaNs included.
The result fields are named cls_step0, sign_step0, exp_step0 and sig_step0 for the first word, up to the same names with step2.
Examples
Example 1
Normal Numbers Near One- Input
h = 3c00; h = c000; h = 3555
- Output
cls_step0 = normal (2), sign_step0 = no, exp_step0 = 0, sig_step0 = 0x400, cls_step1 = normal (2), sign_step1 = yes, exp_step1 = 1, sig_step1 = 0x400, cls_step2 = normal (2), sign_step2 = no, exp_step2 = -2, sig_step2 = 0x555
3c00 has e = 15, so exp is 0 and sig is 400 in hex: the hidden 1 and ten zeros. That is 1.0. c000 is -2.0. The third word is a normal number a little below 0.5.
Example 2
Two Zeros and a Normal Number- Input
h = 0000; h = 8000; h = 4248
- Output
cls_step0 = zero (0), sign_step0 = no, exp_step0 = 0, sig_step0 = 0x0, cls_step1 = zero (0), sign_step1 = yes, exp_step1 = 0, sig_step1 = 0x0, cls_step2 = normal (2), sign_step2 = no, exp_step2 = 1, sig_step2 = 0x648
Both zeros have cls 0, exp 0 and sig 0, but their sign bits differ. The third word is a normal number with exponent 2.
Example 3
The Largest and Smallest Normal Numbers- Input
h = 7bff; h = 0400; h = fbff
- Output
cls_step0 = normal (2), sign_step0 = no, exp_step0 = 15, sig_step0 = 0x7ff, cls_step1 = normal (2), sign_step1 = no, exp_step1 = -14, sig_step1 = 0x400, cls_step2 = normal (2), sign_step2 = yes, exp_step2 = 15, sig_step2 = 0x7ff
7bff has e = 30, so exp is 15 and every fraction bit is 1: 65504. 0400 is the smallest normal number, with exp -14. fbff is -65504.
Constraints and Contract
- Inputs are known 0/1 values. Each word is held for one nanosecond and the outputs are sampled at its end.
- Keep the module name and the ports. The design is combinational.
Hints
Hint 1
First classify. Only two exponent fields are special: all zeros and all ones. The fraction then tells the two classes of each apart.
Hint 2
The exponent field is unsigned. In a wider or real-valued expression, e - 15 wraps to a huge number for a small field. Put a 0 bit in front and make it signed first.
Hint 3
A subnormal has no hidden 1, and its exponent is the same as that of the smallest normal number.
What the grader checks: Ten directed and twelve seeded cases of three words each run on the block. All four outputs are compared after every word with a model written in JavaScript. The model itself is checked against the half-precision type of Node.
22 checks on Icarus · browser samples / server checks. Run the examples first, then check the full regression. Full checks use the server judge. Downloads contain public samples only. A passing result establishes the tested cases, not a proof for all possible inputs.
See all 2 steps and the final capstone