thespydog / sdl_gpu_examples Goto Github PK

It seems a bit strange in this code snippet:

if (SDL_GpuGetDriver(device) == SDL_GPU_DRIVER_VULKAN)
{
	shader = SDL_GpuCreateShader(device, &shaderInfo);
}
else
{
	shader = SDL_CompileFromSPIRV(device, &shaderInfo, SDL_FALSE);
}

It seems the design intent of these examples is to load up the SPIR-V blobs and cross compile them. This would make shipping the SPIR-V blobs a cross-platform shader solution. Is this correct? If so, why should I, as a user, need to ever call SDL_GpuCreateShader, unless I write a platform-dependent shader? Even if I'm targeting Vulkan, can I not just also call SDL_CompileFromSPIRV?

There are now examples in the main SDL repo, so let's move to that system so it can get merged along with the API!

For example, ClearScreen.c could probably be 01-clearscreen.c, and we can organize the rest to either be good for a learning resource or for porting to a new target.

First of all, this is fantastic work.

On my machine, the ComputeSpriteBatch is the only one failing: nothing shows up, and instead it's a flow of error logs:

WARN: Failed to recreate swapchain!
ERROR: Device does not support requested present mode!

I am on Ubuntu Linux 24.04, with a AMD Radeon 5700XT, a decent GPU:

INFO: Welcome to the SDL_Gpu example suite!
INFO: Press A/D to move between examples!
INFO: STARTING EXAMPLE: ClearScreen
INFO: Validation layers enabled, expect debug level performance!
INFO: SDL_Gpu Driver: Vulkan
INFO: Vulkan Device: AMD Radeon RX 5700 XT (RADV NAVI10)
INFO: Vulkan Driver: radv Mesa 24.0.5-1ubuntu1
INFO: Vulkan Conformance: 1.2.1

All other examples work as expected.

ComputeSpriteBatch seems to be using SDL_GPU_PRESENTMODE_IMMEDIATE.

Many other Vulkan apps do work as expected on my workstation, not sure which presentation mode they do use.

I can try on a Windows machine with different GPU on Monday.

We're using stb for PNG but for official examples we can settle with BMP since the files are very tiny!

Cycling transfer buffers isn't as trivial as GPU buffers on the library side, but it should be easy for the app side - we should demonstrate this to avoid having to cobble together host-memory cycling in SDL.