Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Vulkan bandwidth optimizations (configure renderpass load/store better) #15917

Merged
merged 2 commits into from
Aug 28, 2022

Conversation

hrydgard
Copy link
Owner

@hrydgard hrydgard commented Aug 28, 2022

This makes a few straightforward optimizations to when we load/store from framebuffers (in depal, for example, can ignore depth/stencil entirely when drawing to temp FBO, and limit the drawn rectangle in the renderpass).

In reinterpret, by adding a check for full coverage in BlitUsingRaster we know that previous contents of the framebuffer will be irrelevant so can safely set the current render pass to loadop = DONT_CARE.

There's more to be done here later, mainly we should be able to create color-only renderpasses when possible, so we don't even have to consider the depth channel.

@hrydgard hrydgard added Vulkan GE emulation Backend-independent GPU issues labels Aug 28, 2022
@hrydgard hrydgard added this to the v1.14.0 milestone Aug 28, 2022
switch (action) {
case VKRRenderPassLoadAction::CLEAR: return VK_ATTACHMENT_LOAD_OP_CLEAR;
case VKRRenderPassLoadAction::KEEP: return VK_ATTACHMENT_LOAD_OP_LOAD;
default: return VK_ATTACHMENT_LOAD_OP_DONT_CARE;
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: I feel like you've switched to doing this recently, and it's very anti-defensive-programming.

Let's say we have CLEAR, KEEP, DONT_CARE, but Vulkan 2.0 comes out and we add KEEP_OR_CLEAR_IF_CHEAPER (I know this is unlikely, but things happen.)

The intent of compiler warnings is that they'd now notify you, "oh no, this switch doesn't include a case for KEEP_OR_CLEAR_IF_CHEAPER, and the two parts of code that are in different places are misaligned." Including a default says you don't want this help from the compiler, which I already disagree with in usual cases.

Replacing one case with default just seems even worse. Now a grep or search for VKRRenderPassLoadAction::DONT_CARE won't even find it, plus confusion is likely if a new value is added. It may even become unclear, later, whether the default intended to cover both the values or not until someone checks dates. Like if (storeOp == STORE) { ... } else { ...} assumptions, the same.

Personally I think the better pattern is:

switch (foo) {
case A: return 1;
case B: return 2;
case C: return 3;
}
return 1;

Probably load/store actions will never change, I've just been seeing it more often so saying something.

-[Unknown]

Copy link
Owner Author

@hrydgard hrydgard Aug 28, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No I'm very aware of the benefits of fully specifying all the cases and then letting the compiler warn, it's just that in C++ it's too loose.

I've gotten used to the Rust compiler forcing me to do it and allowing me to safely not specifying a default, while back in C++ land where everything is looser and the compiler will warn if I don't have a default case, I've started resorting to this.

I agree that it's bad, and having a separate return after the switch is a pretty good solution that I didn't consider.

@unknownbrackets unknownbrackets merged commit 3e865ab into master Aug 28, 2022
@unknownbrackets unknownbrackets deleted the vulkan-bandwidth-opt branch August 28, 2022 21:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
GE emulation Backend-independent GPU issues Vulkan
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants