Depth scale fixes

[hrydgard]

Changing depth scale during runtime had some side effects. Fixes #16941.

We can't bake scale factors into shaders anymore, so remove an instance of that. Also, when initializing depth buffer using DrawPixels we didn't obey the depth scale. This was fine for the testing done by Burnout Dominator's effect, but wrong.

Hm, noticed that Ratchet & Clank's depth trickery is a bit broken again on master too, and seems to have been for a while. This makes it even worse, so should probably look at that before merging.. Yeah we're not handling depth scale properly in either GenerateDraw2D565ToDepthFs or GenerateDraw2D565ToDepthDeswizzleFs.

Will fix Ratchet tomorrow.

[hrydgard]

Actually, I'm gonna peel off the depth clearing and commit that first. Then I plan to clean up our depth conversion functions (#16948), and then rebase this on top of that.

[hrydgard]

This now contains the commits from #16949 , and on top of that, fixes various issues with depth buffer conversion. For example, we can't use the same shader both for copying depth and converting color to depth.

This fixes the depth regression in Ratchet & Clank.

[unknownbrackets]

Would be a good idea to double check the Phantasy Star text dialog issue in D3D9 and etc. with this, as I think that was a good example of using mixed modes (transform and through) with equal depth that needs to round correctly.

-[Unknown]

[unknownbrackets]

Also just to note, originally when I wrote the depth download shader I simply had a pspautotest write specific depth values and then tested that a dmac copy + CPU read read back the exact values both on desktop and mobile. That motivated the two paths, for example.

-[Unknown]

[hrydgard]

Right, I'll do a little extra testing. I'm hoping that we can match it with a single path now.

Is that pspautotest that you used around, and if so can we enable it to run by default?

[hrydgard]

After some brief testing, I'm not seeing any issues whatsoever.

But yeah, ideally we should do another round of bit exact comparisons.. and as I said above, turn on such a test by default.

[unknownbrackets]

I'll look, but it's of course difficult to do mobile GPU tests, our CI doesn't even run with any GPU.

-[Unknown]

[hrydgard]

Right, i keep forgetting that little issue...

[hrydgard]

Though on Linux, we could theoretically leverage llvmpipe - while not a real gpu test, could catch issues like mismatched depth.