New "colour" for drawPxel

[Myndale]

Well why not but how do you flip a single bit in a byte then ?

XOR it with the bit you want to flip:

Code: Select all

if (color == INVERT)
        _displayBuffer[x + (y / 8) * LCDWIDTH_NOROT] ^= _BV(y % 8);


Means you can also flip multiple bits in a byte with the one operation.

[rodot]

I tried your method, it's like 4% faster... seems that the compiler already does a pretty good optimization.
If you manage to implement it for it to be way faster, then I'd be glad to accept your pull request

[Myndale]

Yeah, the performance increase of that trick is pretty much eclipsed by the overhead of the division and multiplication in calculating the screen address. Also the penalty of doing a branch is minor on Arduino due to its simplistic 2-stage single-level instruction pipeline. Conditional branching is particularly expensive on architectures with multiple overlapping pipelines because they all have to be stalled while the CPU figures out which branch it needs to take. Good branch prediction helps but the CPU doesn't always get it right, so on those architectures you really need to do everything you can to avoid "if" statements in your performance-critical code sections.

[rodot]

Thanks for this clarification, I'll have to read some documentation about CPU architecture and branch prediction to have a deeper understanding of what you said

[DFX2KX]

Yeah, the performance increase of that trick is pretty much eclipsed by the overhead of the division and multiplication in calculating the screen address. Also the penalty of doing a branch is minor on Arduino due to its simplistic 2-stage single-level instruction pipeline. Conditional branching is particularly expensive on architectures with multiple overlapping pipelines because they all have to be stalled while the CPU figures out which branch it needs to take. Good branch prediction helps but the CPU doesn't always get it right, so on those architectures you really need to do everything you can to avoid "if" statements in your performance-critical code sections.

doesn't that affect pretty much any conditional? If so, that would suck trying to avoid for and while, too.

[Myndale]

doesn't that affect pretty much any conditional? If so, that would suck trying to avoid for and while, too.

Absolutely! Which is why you use things like unrolled loops to minimize branching and the overhead of loop mechanics etc. In general though it's difficult or impossible to eliminate a for or while loop unless you know in advance how many iterations you need to make, whereas with if statements it's a lot easier. For example a boolean is typically stored as either a 0 or 1, so a statement like this...

Code: Select all

if (my_boolean)
   some_variable = value1;
else
   some_variable = value2;


...can be replaced with a linear interpolation, eliminating the conditional altogether:

Code: Select all

   some_variable = my_boolean*value1 + (1-my_boolean) * value2;


It's much slower on a device like Arduino but it can be used to get huge speed-ups on some other architectures.

[DFX2KX]

I didn't know you could do a conditional test with just math. Myndale. You're officially the forum computer sage.

[TheTurnipKing]

This is a good thread. We have learned a lot.

[Myndale]

You're officially the forum computer sage.

Lol, I dunno about that, but thanks anyway One of my favourite resources for this kind of thing is the Bit Twiddling Hacks page, has some really clever hacks:

https://graphics.stanford.edu/~seander/bithacks.html