Textures are offset by one pixel

It looks like textures are being drawn one pixel up and to the left of where they should be, which is causing lines to appear.

[Zer0xFF]

I've been look at this and I've narrowed down the issue to bilinear filtering (forced & none forced).

when ever bilinear filtering is applied to a texture it seems to cause it to drift (even on native (1x) resolution).

offset fix force bilinear filtering off:

offset fix force bilinear filtering on:

no fix and force bilinear filtering on:

offset fix force bilinear filtering off (force on has no effect, since its on for this texture):

no offset fix:

diff --git a/Source/gs/GSH_OpenGL/GSH_OpenGL_Shader.cpp b/Source/gs/GSH_OpenGL/GSH_OpenGL_Shader.cpp
index 86c2e450..66dff8cc 100644
--- a/Source/gs/GSH_OpenGL/GSH_OpenGL_Shader.cpp
+++ b/Source/gs/GSH_OpenGL/GSH_OpenGL_Shader.cpp
@@ -229,6 +229,7 @@ Framework::OpenGl::CShader CGSH_OpenGL::GenerateFragmentShader(const SHADERCAPS&
                }
                else
                {
+                       shaderBuilder << "      texCoord.st -= g_texelSize.st * 0.5;" << std::endl;
                        shaderBuilder << "      float tlIdx = texture(g_texture, texCoord.st                                     ).r * 255.0;" << std::endl;
                        shaderBuilder << "      float trIdx = texture(g_texture, texCoord.st + vec2(g_texelSize.x, 0)            ).r * 255.0;" << std::endl;
                        shaderBuilder << "      float blIdx = texture(g_texture, texCoord.st + vec2(0, g_texelSize.y)            ).r * 255.0;" << std::endl;

or

+                       shaderBuilder << "      texCoord.st -= vec2(0.5 / 256.0);" << std::endl;

^^ the 2nd one seems to work better (in KH) with forced bilinear filter, not sure why (since texture is treated as a square?), even with forced bilinear filter the intro loads correctly.

offset fix with forced bilinear filter off

no offset fix and forced bilinear filter off (note the text it has few missing pixels)

offset fix with forced bilinear (text is fine, but character break up, note character would break up without fix as well)

[jpd002]

That offset makes sense. I think you need to remove 0.5 from the texture size when computing factors for bilinear filtering.

The force bilinear option shouldn't be considered. It was made for debugging purposes, I don't think
there's another valid use case for that.

[jpd002]

Another thing to consider that I just thought of while working on bilinear filtering on that experimental branch. In the current GS handler, when we do the bilinear filtering in the fragment shader, we don't differentiate between minification and magnification.

In the case of KH, it might be possible that it uses point sampling for minification and bilinear for magnification, but that the emulator does bilinear whatever the case is.

[Zer0xFF]

In the case of KH, it might be possible that it uses point sampling for minification and bilinear for magnification, but that the emulator does bilinear whatever the case is.

(my knowledge is rather rouge in this area)
but I assume this isn't the case since this worked (correctly with offset fix) so I assume its bilinear magnification? however, I couldn't find anything about point sampling and pixel offsetting.

[jpd002]

I made a few KH frame dumps and the game uses bilinear filtering for both min and mag, so, what I said doesn't apply here. The offset fix would probably be appropriate here.

I don't know about Nights, I'll have to look into it.

FFX also has problems with its text rendering, but it uses point sampling for min and bilinear for mag. I don't know if the LOD factor would make it magnify, but it would be interesting to see what happens if we apply the offset.

[Zer0xFF]

I've opened a PR #780 with the change, @ToppestOfDogs can report on that game or any game once build is done. Download would appear here in about 10 mins https://ci.appveyor.com/project/jpd002/play/builds/26153236/job/tj2bjdnnyhfq0h4u/artifacts

[Zer0xFF]

It seems to have had an effect.

But I suppose the question is, are you using software or hardware render in PCSX2, since that might explain the difference in sharpness.

[unknownbrackets]

Can ST have a fraction here? I assume that's why there's still issues in #780.

With bilinear filtering, it's important that you snap the filtered texels to the grid. It might even be better to use texelFetch().

I'd recommend:

• When doing an indexed texture (isIndexedTextureSource) with manual filtering, turn off bilinear filtering on the texture. The indexes definitely shouldn't be filtered and need NEAREST.
• Either round down st (floor) first, or use textureOffset/textureGather. You don't want the addition of g_texelSize to round differently/down and cause you to filter incorrectly.
• For a graphical improvement, if you can detect that a palette (esp 16 color) is a linear gradient, turn ON linear filtering for both the palette and texture, and then just do a linear filtered index lookup. Not sure how common this is in PS2 games, it may at least be when drawing some text?

-[Unknown]