Optimized mask.from_surface() when converting an alpha surface.

[itzpr3d4t0r]

This PR massively speeds up mask.from_surface() when dealing with the threshold argument (for 24 and 32 bit surfaces, about a 12% improvement for 16 and lower).

It achieves so by directly accessing the pixel's alpha channels instead of using SDL's functionality which is not only slow but actually did more work that we actually need for this operation by also returning the rgb channels.

With this new strategy we just jump from pixel's alpha to pixel's alpha and check that against the threshold.

I've also added a fast path for 24-bit surfaces bringing an astonishing 737 times improved runtime since we now use bitmask_fill.

I've also added a validity check for the threshold argument which wasn't there before.

here are the results:

made with this test program:

from data_utils import plot_tests, run_tests
import pygame
from pygame.mask import from_surface

pygame.init()

# ===========| CONFIG |===========

DO_TEST = 0
MAX_SIZE = 1000
REPETITIONS = 100
NUM_CALLS = 1

TITLE = "from_surface"
X_LABEL = "Surface size"
DO_SCATTER = True
MODE = "MIN"
LIMIT_TO_RANGE = MAX_SIZE
COMPARE_LIST = [(-1, 0)]
kwargs_dict = {"from_surface": from_surface}

def test_setup(curr_size: int, g: dict):
    s = pygame.Surface((curr_size, curr_size), pygame.SRCALPHA, 32)
    pygame.draw.circle(s, (255, 255, 255, 128), (curr_size // 2, curr_size // 2), curr_size // 2)
    g["s"] = s

tests = [
    ("mask from_surface new", "from_surface(s)"),
]

files = [
    ("mask from_surface new", "red"),
    ("mask from_surface old", "white")
]

if DO_TEST:
    run_tests(tests, test_setup, MAX_SIZE, REPETITIONS, NUM_CALLS, **kwargs_dict)

pygame.quit()

plot_tests(TITLE, files, MODE, LIMIT_TO_RANGE, DO_SCATTER, False, COMPARE_LIST,
           X_LABEL)

The data utils file (as txt):
data_utils.txt

[andrewhong04]

The work you do never ceases to amaze me. This works on my machine just fine.

[damusss]

:pg_chad:

It had no effect when I posted it and I didn't test it locally to be sure it worked, even tho I trust that it does

[damusss]

I've tested locally with different sizes and I get a 65+% improvement for 32 bit surfaces and 10% with 16 bit surfaces, which is really impressive, good job!

[ankith26]

Also thinking more about it, I realised this probably breaks support for the more obscure kind of pixel formats (like say, SDL_PIXELFORMAT_ARGB2101010 or FOURCC style formats like YUV) but the pygame codebase probably doesn't support these in other places to begin with, so we can let that slide for now

[ankith26]

Okay I am now satisfied that this PR should work for the common/normal cases.

Thanks for contributing! 🥳