Membership against weighting

inside answers which values are legal. dist answers how often each legal value appears, and values it leaves out become illegal. Neither can do the other's job.

Soft priority and ordering

Soft marks a default the solver may drop to satisfy hard rules; no ordering keyword is involved. solve...before orders the pick, which moves probability only, never legality.

The bench measures each answer: both weight forms over 2000 calls, a soft default beside a hard rule, and the skewed err bit with and without a dist.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

class weights;
  rand bit [4:0] a, b;
  constraint c_a { a dist { [1:10] :/ 50, 20 :/ 50 }; }    // :/ shares 50 across ten values
  constraint c_b { b dist { [1:10] := 50, 20 := 50 }; }    // := gives 50 to each value
endclass

class soft_order;
  rand bit [7:0] x;
  rand bit       mode;
  constraint c_soft { soft x inside {[10:20]}; }
  constraint c_hard { mode -> x > 100; }
endclass

class skew;
  rand bit       err;
  rand bit [7:0] code;
  constraint c_skew { err -> code == 255; }
endclass

class skew_fixed extends skew;
  constraint c_fix { err dist { 0 := 1, 1 := 1 }; }       // state the share outright
endclass

P(x == 5): with :/ 4.5% (5% predicted), with := 9.3% (9.1% predicted)
soft default kept whenever mode == 0 (400 calls); dropped for the 0 calls with mode == 1
err == 1 in 1000: plain 2 (uniform over pairs: about 4), with a dist 502
PASS: :/ and := matched their arithmetic, hard beat soft, and a dist fixed the skew

Notice the soft line: in 400 calls mode never came out 1. The hard rule mode -> x > 100 competes with the soft default, and Verilator kept the soft default every time by keeping mode at 0. Legal, but a test that needs mode == 1 has to ask for it.

Related topics

• Q&A: Controlling Randomization