GPU Vulkan: Only modify surface on window claim or release

[thatcosmonaut]

According to the Vulkan spec, we have to destroy swapchains before any associated surfaces.

I can't actually repro the validation error on my dev environment, so let's have the issue reporter verify the fix.

Resolves #14648

[zopsicle]

I still get the validation error. With some debug printing I narrowed it down to this call:

SDL/src/gpu/vulkan/SDL_gpu_vulkan.c

Lines 4607 to 4614 in 9092ddf

	if (swapchainSupportDetails.capabilities.currentExtent.width == 0 ||
	swapchainSupportDetails.capabilities.currentExtent.height == 0) {
	renderer->vkDestroySurfaceKHR(
	renderer->instance,
	windowData->surface,
	NULL);
	windowData->surface = VK_NULL_HANDLE;
	if (swapchainSupportDetails.formatsLength > 0) {

I think the old swapchain still exists here, because it is used later in swapchainCreateInfo.oldSwapchain = windowData->swapchain;. If I replace this call to vkDestroySurfaceKHR with a call to VULKAN_INTERNAL_DestroySwapchain, the validation error disappears. There are some other calls to vkDestroySurfaceKHR nearby, they may have this same issue but I don't know how to trigger them.

---

Side note: there is a comment that says "NVIDIA + Win32 can return 0 extent when the window is minimized." I think this is always the case on Windows, not just with Nvidia, quoting the Vulkan spec on VkWin32SurfaceCreateInfoKHR:

The window size may become (0, 0) on this platform (e.g. when the window is minimized), and so a swapchain cannot be created until the size changes.

[thatcosmonaut]

I see... I'm trying to remember why exactly we're destroying the surface here.

[thatcosmonaut]

This problem was introduced by this patch: #14113

Before, we would always destroy the old swapchain, but now we hold on to it.

[zopsicle]

IIUC the surface can just be kept alive the entire time between claiming the window and unclaiming the window, and reused with each swapchain recreation. It would only have to be destroyed when claiming the window fails, and when unclaiming the window.

[thatcosmonaut]

I believe you are correct. I'll revise this patch.

[thatcosmonaut]

This is much cleaner now.

Tagging @zopsicle and @slouken for re-review.

[zopsicle]

The new version does not trigger the validation error.