SDL_SoftStretch compatibility

[Starbuck5]

Hello, I worked a bit on testing this against pygame-ce (pygame-community/pygame-ce#3287). I feel that pygame-ce is a pretty bad scenario for this library because we do plenty of weird stuff and also try to expose almost all SDL APIs.

I ran into some crashes that I have yet to debug, but one thing I do see straight away is differences in SDL_SoftStretch.

In our implementation of pygame.transform.scale we lock the surface and then call SDL_SoftStretch. Now that softstretch is actually a blit, this yields a "surface can't be blit while locked" error.

On another note, I don't understand how softstretch can be replaced with a scaled blit. If you're trying to scale a surface with alpha wouldn't the blit scaled blend with whatever is in the destination surface?

[slouken]

Can you provide the parameters for the call, including formats of the src and dst surfaces? SDL_SoftStretch() isn't a scaled blit, but some format conversion paths through it might involve some pixel conversion or blitting, so it would be helpful to know exactly what you're running into.

Also, I've been looking forward to the pygame test results. You guys put the API through a torture test, and if sdl2-compat can pass that, I'm pretty confident it'll handle almost anything a real game can toss at us.

[Starbuck5]

If I'm reading the source of SDL2 compat properly the functions are implemented as a scaled blit: https://github.com/libsdl-org/sdl2-compat/blob/main/src/sdl2_compat.c#L3370-L3380

[slouken]

Ah, you are correct. I was looking at the SDL3 implementation. I'm not sure why they differ. I'll have to take a look.

[slouken]

So SDL_SoftStretch() is intended to be equivalent to blitting with blending turned off, and sdl2-compat assumes that blending is turned off for surfaces that it's stretching, which is obviously not correct. Do you have a minimal repro of incorrect behavior we can use to validate any fixes?

[Starbuck5]

I haven’t made a reproducer in C yet and the reproducer in Python fails because of the locking issue.

Interestingly, looking at the SDL source the softstretch function still exists, it’s just not publicly exposed.

[Starbuck5]

Our goal in pygame-ce is to get all our modules compiling with both SDL2 and SDL3, so I think the way we’ll do it is by vendoring the function into our own source.

[slouken]

I'm considering re-adding it to the API. It was removed as a simplification, but if there's a legitimate use case for it still, we can put it back.

[Starbuck5]

I've generated some examples to help test.

#define SDL_MAIN_HANDLED
#include <SDL.h>

int main() {
    SDL_Surface* surf = SDL_LoadBMP("alien.bmp");

    SDL_Surface* larger_surf = 
        SDL_CreateRGBSurfaceWithFormat(0, 100, 100, surf->format->BytesPerPixel, surf->format->format);

    SDL_FillRect(larger_surf, NULL, SDL_MapRGB(surf->format, 0, 0, 0));

    SDL_SoftStretch(surf, NULL, larger_surf, NULL);

    SDL_SaveBMP(larger_surf, "output.bmp");

    return 0;
}

Source file, zipped because github doesn't like bmp uploads. alien.zip

SDL2 output (reencoded as png)

SDL2 compat output (reencoded as png)

#define SDL_MAIN_HANDLED
#include <SDL.h>
#include <stdio.h>

int main() {
    SDL_Surface* surf = SDL_LoadBMP("alien.bmp");

    SDL_Surface* larger_surf = 
        SDL_CreateRGBSurfaceWithFormat(0, 100, 100, surf->format->BytesPerPixel, surf->format->format);

    SDL_FillRect(larger_surf, NULL, SDL_MapRGB(surf->format, 0, 0, 0));

    SDL_LockSurface(surf);
    if (SDL_SoftStretch(surf, NULL, larger_surf, NULL) < 0) {
        printf("Error: %s", SDL_GetError());
        return -1;
    }
    SDL_UnlockSurface(surf);

    SDL_SaveBMP(larger_surf, "output.bmp");

    return 0;
}

SDL2 produces image as expected still, SDL3 produces error: "Surfaces must not be locked during blit"

[slouken]

So I'm less concerned about the locked error, we can fix that. I'm more concerned about incorrect output. Is the output correct if you comment out that check in SDL?

[Starbuck5]

In my second code, if SDL_SoftStretch is not error checked it outputs a fully black surface.

[slouken]

Okay, thank you for the excellent test case. I re-added SDL_SoftStretch() to the API, and confirmed that fixes this issue.

[Starbuck5]

Thank you!

[slouken]

I'm going to reopen this, as it relies on a version of SDL that isn't released yet. Once SDL is released with the new function in a couple of weeks, I'll re-add this.