在同一个子通道中将图像用作输入附件和颜色输出附件是否合法？

Question

I'm trying to optimize my code with a move that I'm not confident with. I'm experimenting with deferred rendering, so I have built a render pass with two subpasses. The first one output the albedo/normal/depth images, and second one uses them as input attachments to create the final image. Until now, I was using an image to write the albedo, and another one to ouput the final calculations using the G-buffer.

I was wondering if it was possible to use the same image for the albedo and the final calculations. So I tried to use just one image, as color output attachment in the first subpass, and both as color output attachment and input attachment in the second subpass. Given that an input attachment can be either SHADER_READ_ONLY_OPTIMAL or GENERAL, I used the general layout since I need to write in the image.

It works almost perfectly, but first, I'm not sure not being lucky, and second, I get this message from the validation layers :

"Layout for input attachment is GENERAL but should be READ_ONLY_OPTIMAL.

Layout for color attachment is GENERAL but should be COLOR_ATTACHMENT_OPTIMAL."

Where the image refered by the input attachment and the color attachment is the image I used for reads and writes. I don't know if this message is a driver bug or simply a way to tell me that I'm doing wrong. Moreover, even if it's a legal move, I think that the use of GENERAL layout instead of COLOR_ATTACHMENT_OPTIMAL is likely to make performances worse than simply having two images.

Answer 1

The answer to the question in your title is yes, you may use an attachment as both an input attachment and a color attachment in the same subpass. The validation layer should not be outputting a warning. It is not a violation of the validation rules (and indeed, only the general layout allows the use of an attachment as both input and color). The validation layers can tell when you're using the same attachment for both in a subpass, so it shouldn't give a warning for using the general layout in this case.

Feel free to submit an issue to the validation layer project on GitHub .

However, that's not entirely what your main text asks. The answer to this question:

if it was possible to use the same image for the albedo and the final calculations

Is "no".

In the lighting pass of a deferred renderer, you will generally need to cover the same pixels several times. You may have more lights than you could reasonably compute in a single pass, or you may need to do special shadows or other things that don't easily combine well, or various other things. So you will typically need to render the same area of the screen multiple times, using additive blending for each rendering with the output color.

If you're writing to the same image that stored the albedo, then the later read (assuming you did the proper barrier gymnastics to allow reading from a value written to the attachment within the same subpass) will be reading the light intensity , not an albedo value. Therefore, it's going to be doing the wrong math.

So yes, you're going to need another image.