Pixel Shaders for RetroArch 3D Cores

RetroArch got its first 3D core recently in the form of TyrQuake and it seems most of the existing pixel shaders either look unimpressive or don't work at all. Thankfully, shaders written for Pete's OpenGL plugin for PSX emulators are easy to port to RetroArch's XML format and they seem to generally work well enough.

Click here to download all of the shaders covered in this post.

No Shader (for comparison)

A Lovely Bloom

One of the few conventional shaders that still functions as expected, this one looks really nice, and the screenshot doesn't really do it justice. It gives light sources a nice bloomy glow that softens things up a bit. 

 HDRTV

This shader saturates colors somewhat and adds faint scanlines. It has been around for a long time (it was one of the original shaders available for bsnes when it first got pixel shader support) but the scanlines weren't visible on 2D sources, so I was surprised to see them here.

NV+MSAA

This shader uses the "natural vision" algorithm to do some color adjustment with added post-proc anti-aliasing.

Cartoon

This one darkens the image too much to be very useful in an already dark game like Quake but may be good for future cores that have brighter games.

Edge Detection

While not necessarily practical, this shader draws a line at each color intersection to produce a cool, old-timey effect.

Scale4xHQ

One of the only scale/interpolation shaders that has any noticeable effect, this one is a bit blurry for my tastes but it evens out a lot of the jaggies.

Scanline3x

I'm a sucker for scanlines and these normally crummy-looking scanline filters I made a while back actually look pretty nice here. (Note: I think I got the 3x and 4x mixed up; 3x should have fewer, larger scanlines than 4x)

Scanline4x

Themaister Waterpaint Scanlines

While the effect isn't as pronounced as with 2D cores, this shader does some smoothing with scanlines and ends up with a sort of screen-door effect that nevertheless looks pretty good.

CPU Filters

Some of the bsnes-style CPU filters work, but not that well. The NTSC variants all perform their normal functions, but the the image is off-center and cuts some things off, as pictured. The scanline filters work pretty well, too, though they're sort of hard to see. Interpolation filters like HQ2X seem to have no visible effect, and SuperEagle crashed my game :/

XRandR Overscan Fix for Intel HD4000 Ubuntu

I have my television hooked up to a computer running Ubuntu 12.10 Quantal and have had issues with overscan across a variety of chipsets. For Nvidia graphics, the proprietary driver sometimes has an overscan compensation slider or at least something. Intel drivers are open source and very convenient but don't seem to have much in the way of driver configuration outside of what's exposed in the normal settings applications.

If you have an Intel chipset, such as the built-in HD4000 graphics that come with Ivy Bridge CPUs, there are a few angles you can try that center around using xrandr, which stands for 'X Resize, Rotate and Reflection.' This program has a bunch of different options, and I strongly recommend spending some time with the man page. No matter which options you try first, you'll need to query the program and see which display device is being used. You can do this by typing into a terminal:
xrandr -q
For me, HDMI3 is the one that's hooked up, so it lists all of the available modelines and some other information.

Now that we know which device is being used, we can invoke xrandr with this syntax:
xrandr --output [device] --[option]

The option that worked for me is --transform, which performs some crazy matrix math that I don't understand (I'll update this post if someone tells me wtf is going on) on each pixel, apparently kinda like the pixel shader stuff I've written about elsewhere, I think...? Anyway (replace HDMI3 with whatever your device is called):
xrandr --output HDMI3 --transform 1.05,0,-35,0,1.05,-19,0,0,1
This doesn't make it fit exactly on my screen but it's much closer than it was. Unfortunately, it gets wiped out every time you logout, so we need to do something to make it stick. You can add it to your startup items or tack it onto your lightdm window manager's configuration file by typing into a terminal:
sudo nano /etc/lightdm/lightdm.conf
Navigate to the end of the file and start a new line and paste this in:
display-setup-script=xrandr --output HDMI3 --transform 1.05,0,-35,0,1.05,-19,0,0,1
Save and exit and the transform command should kick in every time you login.

Other options that may help:
The 'scale' option should ideally just scale your screen down, but in my experience, it makes everything (text, window borders , etc) bigger, similar to changing the DPI in Windows, without actually changing the visible area.

The 'border' option also seems like it should do exactly what we want, but I haven't seen it actually work anywhere. On my system, it fails with this error:

X Error of failed request:  BadName (named color or font does not exist)
  Major opcode of failed request:  140 (RANDR)
  Minor opcode of failed request:  11 (RRQueryOutputProperty)
  Serial number of failed request:  29
  Current serial number in output stream:  29

which seems to be a catch-all for unrecognized commands :/

Generally, I would recommend skipping all of this and attempting to fix it through your TV's built-in options (generally called "just scan" mode or "pixel perfect" or something like that). The only time you should attempt to do it in software is when your TV is too crummy (like mine) and doesn't have any of those options.

UPDATE (5/20/2014): A helpful anonymous commentor explains the transform matrix thusly:

The transformation matrix seems pretty straightforward to me:
It's just an affine 2D transform. Think of 1.05,0,-35,0,1.05,-19,0,0,1 as the following matrix:
|1.05 0 -35|
|0 1.05 -19|
|0   0      0 |
The upper two diagonal coefficients are scaling factors, i.e. in this case, the picture would be expanded by 5% horizontally and vertically. The last column is a translation vector, in this case causing the picture to be transformed by 35 pixels to the left and 19 pixels upwards (or downwards, depending on the coordinate system).
The last row needs to be (0, 0, 1), or combining two affine transforms will not work correctly. They are not relevant for the actual transformation.