Using a fixed point FontUnit dimension for internal calculations.
cbreeden opened this issue · comments
On IRC we discussed the possibility of using a scale of internal font units (in xits it's 1/1000 em) for internal dimension in ReX. We concluded that it would probably be best if we use a 24.8 fixedpoint representation of the internal font unit. A cursory glance of the code showed that *2, /2, and +/- is by far the most common operation. However, *0.68 is used for scripts scaling, and it is unclear what the distribution of *0.68 use is to make a fully informed decision.
Pertinent numbers from the intel optimization manual:
| Latency | Throughput | |
|---|---|---|
| ADD | 1 | 0.5 |
| MUL | 7,8 | 6,7 |
| SHL | 1 | 1 |
| FADD | 5 | 4,1 |
| FMUL | 5 | 1 |
Considering the number of *2, /2 uses, it seems unlikely that f64 will afford the same bitshift optimizations that a u32 would. FDIV is not listed, but expensive.
sebk also pointed out that it could be possible to put a scaling call in both the layout and renderer, but depending on a feature flag, one would be a no op and the other wouldn't.
Using a scaled font unit would also rid of all the Pixels -> FontUnit conversions in the Layout (there seem to be quite a few).