Scaler

[Grabulosaure]

You like scalers, right ?

I have made one for my cores.

It is not finished, it only does nearest and bilinear, and it is not well tested.
Anyway, here is a preliminary version (ascal.vhd), with the Megadrive/Genesis as an example.

http://temlib.org/pub/mister/ascal/

[Sorgelig]

Looks very interesting! I will try it.
Thanks!

[Sorgelig]

I've tested your precompiled lite version - impressive!
two notes:
- HQ2x produces messed graphics - may be input resolution is beyond expected max..
- I see tearing - tripple frame buffer should solve this problem.

May be you have de-interlacer as well? Some cores like Amiga, TGFX, Genesis require deinterlacer. Especially Amiga.

[olin]

Would that mean a time unlimited scaler (this new one) could possibly be compiled on a free version of Quartus? Any other benefits like shorter build time of non-lite cores?

[vanfanel]

- I see tearing - tripple frame buffer should solve this problem.

Triple buffer will add 1 frame of lag. V-synced double buffer is a better solution (unless there's something I'm missing here).

[ijor]

Triple buffer will add 1 frame of lag. V-synced double buffer is a better solution (unless there's something I'm missing here).

I agree. If you are "vsynced" you shouldn't need triple frame buffer. You actually don't need double buffer either if you can "sync" the frame buffer correctly.

[Sorgelig]

In general case you cannot vsync HDMI to retro video as it usually doesn't follow HDMI requirements.
vsync_adjust is an option for compatible monitors to lock vsync for smooth scrolls but it cannot be a requirement.

[vanfanel]

In general case you cannot vsync HDMI to retro video as it usually doesn't follow HDMI requirements.
vsync_adjust is an option for compatible monitors to lock vsync for smooth scrolls but it cannot be a requirement.

Of course it can not be a requeriment, but it's one of the coolest things FPGA implementations can do: lagless video. So this option is "the magic"

I agree. If you are "vsynced" you shouldn't need triple frame buffer. You actually don't need double buffer either if you can "sync" the frame buffer correctly.

Exactly!

[Sorgelig]

it only does nearest and bilinear, and it is not well tested.

is it hard to add polyphase filtering? That would make it as replacement of VIP scaler. NN filter is good. Bilinear probably hard to use as it gives too blurry image. Need an option for those who don't like sharp NN look. Also computer cores use text a lot, so NN woudn't be a best choice there.
It would be really wonderful to replace the whole VIP and save a lot of resources with reduced compilation time. ao486 will especially benefit from it.

[ijor]

is it hard to add polyphase filtering? That would make it as replacement of VIP scaler... It would be really wonderful to replace the whole VIP

A high quality, free, open source, polyphase scaler would be fabulous. Doesn't sound easy. As you said, it requires math skills, probably more than anything else.

May be we should try a joint collaborative effort with some of the mathematician that have been posting in the scaler threads. I googled the issue some time ago. Of course that I couldn't find any open source implementation. But also I found some posts that seem to suggest that the VIP scaler, or at least parts of it, it's written in some math oriented high level language (not on any standard HDL).

[Grabulosaure]

I will first try BiCubic, I think I know how to do it. AFAIU polyphase is not very different.

For interlaced video, is the "field" input sufficient? Decoding composite PAL or NTSC synchro, with equalization pulses, is not fun.

I have updated the Megadrive/Genesis .RBF with support for triple buffering.

[Sorgelig]

I will first try BiCubic, I think I know how to do it. AFAIU polyphase is not very different.

according to Altera docs polyphase uses previous data(pixels?) to calculate current pixel(s) while bicubic use only current data.
Current Genesis version includes coefficients in BRAM and receives it from MiSTer (if supplied), 16 phase with 4 taps. May be you can re-use it in your bicubic algo.

For interlaced video, is the "field" input sufficient? Decoding composite PAL or NTSC synchro, with equalization pulses, is not fun.

Yes. Field is enough for determining the second field of interlace video. Progressive video always set it to 0. VIP scaler also use this signal to detect the interlace. Synchro signals aren't used for interlace detection.

I have updated the Megadrive/Genesis .RBF with support for triple buffering.

Can you fork my Genesis repository on github and apply your changes with scaler there? It would be easier to track the changes.
If it's not convenient for you then fine. Can use your site.

[ijor]

I will first try BiCubic, I think I know how to do it. AFAIU polyphase is not very different.

That would be fabulous!

This is an Altera/Intel article about Bicubic vs Polyphase:
https://www.intel.com/content/www/us/en ... 51468.html

[ghogan42]

I will first try BiCubic, I think I know how to do it. AFAIU polyphase is not very different.

For interlaced video, is the "field" input sufficient? Decoding composite PAL or NTSC synchro, with equalization pulses, is not fun.

I have updated the Megadrive/Genesis .RBF with support for triple buffering.

Great Work!

I don't have any experience with fpgas, so I don't know how easy/hard anything is to implement. But if you have a working bilinear filter then you might be interested in what I do in my retropie pixel shaders to get improved scaling just using the built-in bilinear filtering hardware.

Here's a screenshot illustrating what I mean (sorry if it doesn't appear in-line):

pixel_scaling.jpg

The picture shows (in order):

• Regular Bilinear
• Inigo Quilez's "Improved Texture Interpolation" from here: http://www.iquilezles.org/www/articles/ ... exture.htm
• The modified sharper version of the above. This is the one I actually use in my pixel shaders
• An even sharper version that can be used. But this is too close to nearest neighbor and shows shimmering in motion.

I put it in a shadertoy here if you want code: https://www.shadertoy.com/view/Xl2czV

I'll post the algorithm here too. I'm not trying to force you into implementing it. I just figure it's faster to post now then to ask you if you were interested and then waiting for your reply.

Here's how it works. I'm assuming you run the interpolation twice (once in x and once in y):

Code: Select all

Suppose you have a game resolution of 320x240.

Suppose the pixel in the display maps to an x coordinate of 112.13 (just an example).  Also, in pixel shaders we say 112.5 is the center of pixel 112, so x=112.13 would be a distance of 0.37 from pixel center 112 and a distance of 0.63 from pixel center 111.

Instead of linearly interpolating between pixels 111 and 112 with (0.37)(Pixel 111) + (0.63)(Pixel 112) we do this instead:

Adjust the coordinates as follows:
	X = 112.13
	C = floor(X)+0.5  (Center of the current Source Pixel)
	Dist = X - C   (Distance from the Pixel Center)
	Dist = 4.0*Dist^3  (Adjusts the distance from center.  See below for explanation)
	X = C + Dist   (Put the x coordinate back together)

The result from the above would map 112.13 to 112.297

Then you linearly interpolate with standard linear interpolation: (0.203)(Pixel 111) + (0.797)(Pixel 112)

Note that what this is doing is mapping "Dist" from [-0.5,0.5] back onto [-0.5,0.5]. But numbers close to zero get mapped even closer to zero.
And -0.5 maps to -0.5 and 0.5 maps to 0.5 because interpolation would break without that.

You can look here: https://www.desmos.com/calculator/alxmsdemu3 to see how I modify the "Distance from Pixel Center" before linear interpolation. The flatter it is through (0,0) the sharper the scaling

Let me know what you think. Feel free to ignore this and implement whatever you want though.

[Sorgelig]

I will first try BiCubic, I think I know how to do it. AFAIU polyphase is not very different.

is there any progress?

[cacophony]

Part of the motivation for using a custom scaler:
https://twitter.com/RetroRGB/status/1058093654554894336

According to RetroRGB:
"I did some MiSTer lag tests. In it's current state, the MiSTer's HDMI output ads about 2 frames of lag. Future community-built scalers can significantly lower this though."

[Grabulosaure]

is there any progress?

I was away from Internet for a few days... But I had my computer!

Here is the current state of the scaler :

http://temlib.org/pub/mister/ascal/

Big changes. Bugs.

[BlockABoots]

So this scaler can be implemented into and core, or is it just Genesis?

[coughlanio]

I've noticed some severe image retention with the Genesis core using the new scaler on my monitor. It's an LG 27UK850W. I tried to see if I could reproduce the burn-in with the standard Genesis core, but it seems to be exclusive to the new scaler. Speaking with some people, it could be due to weird video timings?

[Grabulosaure]

I have just fixed a bug and updated the files.

I've noticed some severe image retention with the Genesis core using the new scaler on my monitor. It's an LG 27UK850W. I tried to see if I could reproduce the burn-in with the standard Genesis core, but it seems to be exclusive to the new scaler. Speaking with some people, it could be due to weird video timings?

What do you mean by image retention? Do the "triple buffer" option change anything ? The scaler should use the same video modes as Sorgelig's Genesis core.

So this scaler can be implemented into and core, or is it just Genesis?

It can be used with any core, replacing the Intel/Altera scaler.

According to RetroRGB:
"I did some MiSTer lag tests. In it's current state, the MiSTer's HDMI output ads about 2 frames of lag. Future community-built scalers can significantly lower this though."

This scaler doesn't try to minimise lag yet. It shouldn't be too difficult, assuming that the output vertical frequency is locked to the input frequency. Currently, with the "triple buffer" mechanism, there is a 1 to 2 images delay.

[Sorgelig]

I have just fixed a bug and updated the files.

huge progress!
There are image corruptions on edges - may be it's possible to extend internal buffer to include the blanking area and then cut it on output, so these artifacts will be outside of active area. Is possible to add Width/Height of active image within a current resolution with adding black borders? It's needed for aspect ratio compensation and integer vertical scaling.

[ghogan42]

I have just fixed a bug and updated the files.

I've noticed some severe image retention with the Genesis core using the new scaler on my monitor. It's an LG 27UK850W. I tried to see if I could reproduce the burn-in with the standard Genesis core, but it seems to be exclusive to the new scaler. Speaking with some people, it could be due to weird video timings?

What do you mean by image retention? Do the "triple buffer" option change anything ? The scaler should use the same video modes as Sorgelig's Genesis core.

So this scaler can be implemented into and core, or is it just Genesis?

It can be used with any core, replacing the Intel/Altera scaler.

According to RetroRGB:
"I did some MiSTer lag tests. In it's current state, the MiSTer's HDMI output ads about 2 frames of lag. Future community-built scalers can significantly lower this though."

This scaler doesn't try to minimise lag yet. It shouldn't be too difficult, assuming that the output vertical frequency is locked to the input frequency. Currently, with the "triple buffer" mechanism, there is a 1 to 2 images delay.

I know you have more interesting things to work on. Good job on all you've done! But I notice that the method I described isn't working. I don't know VHDL but based on how the vip scaler seems to work, a 4bit LUT should implement my function well enough. I'm just going to copy/paste the function here for you. I tried it and it works correctly.

Code: Select all

  -----------------------------------------------------------------------------
  -- Ghogan
  -- Shift Bilinear weights away from the center (towards 0 or 1) to
  -- make scaling sharper
  FUNCTION gho_calc(frac : unsigned(11 DOWNTO 0)) RETURN unsigned IS
  BEGIN
  	case frac(11 DOWNTO 8) is  --We don't need so many bits
		when "0000" => Return "00000000";
		when "0001" => Return "00000001";
		when "0010" => Return "00000010";
		when "0011" => Return "00000110";
		when "0100" => Return "00010000";
		when "0101" => Return "00100000";
		when "0110" => Return "00110110";
		when "0111" => Return "01010101";
		when "1000" => Return "10000000";
		when "1001" => Return "10101010";
		when "1010" => Return "11001010";
		when "1011" => Return "11100000";
		when "1100" => Return "11110000";
		when "1101" => Return "11111010" ;
		when "1110" => Return "11111110";
		when "1111" => Return "11111111";
    end case;
  END FUNCTION;
  
  SIGNAL o_h_gho_pix,o_v_gho_pix : type_pix;
  SIGNAL o_h_gho_frac,o_v_gho_frac : unsigned(7 DOWNTO 0);

Also, it's worth noting that the way the vip polyphase filter works similarly with a 4bit lut. I think it's pretty much like this for horizontal scaling. Does this sound right to you? I'm still wrapping my head around it.

-----------------------------------------------------------------
VIP Polyphase Filter with 4 taps and 16 Phases?
-----------------------------------------------------------------
(1) (Currrent Output X)/(Output X Res) * (Input Res X) - 0.5 = Position
(2) Split Position into Integer Part and Fractional Part -- I'll call them Int_Pos and Frac_Pos
(3) Quantize Frac_Pos down to 4 bits if it isn't already
(4) Use these four bits to look up a row in the coefficient table. Call them C0, C1, C2, C3. Use however many bits you want. These need to be signed. Example Table For Bicubic Scaling here:

"0000": 0, 128, 0, 0
"0001": -4, 127, 5, 0
"0010": -7, 124, 12, -1
"0011": -10, 119, 21, -2
"0100": -11, 112, 31, -4
"0101": -11, 103, 41, -5
"0110": -11, 94, 52, -7
"0111": -11, 84, 63, -8
"1000": -10, 74, 74, -10
"1001": -8, 63, 84, -11
"1010": -7, 52, 94, -11
"1011": -5, 41, 103, -11
"1100": -4, 31, 112, -11
"1101": -2, 21, 119, -10
"1110": -2, 13, 124, -7
"1111": 0, 5, 127, -4

(5) Fetch Pixels T0, T1, T2,, T3 from the source buffer. Pixel T1 is at Int_Pos. So T0 is to the left and T2 and T3 are to the right. Duplicate pixels if you run off the edge of the buffer?
(6) Your Output Pixel Color is C0*T0 + C1*T1 + C2*T2 + C3*T3
---------------------------------------------------------------------
Can anybody see any mistakes that I made? Since all of the useful filters can be implemented with a 4 tap filter like this, I might try adding it as an extra option to the ascal scaler. I might not succeed because simple things in VHDL are hard for me since I have never used it before today. But if it works then we have all of the filters that I made over in the other thread: http://www.atari-forum.com/viewtopic.ph ... 00#p356779 in addition to the ones implemented directly in ascal.

[Sorgelig]

I believe all these filters including bilinear can be implemented by just different coefficients.
Recent MiSTer binary already use coeff.txt file from Genesis folder and push them to the core.
So you don't need to hardcode coefficients in HDL. Just receive them like it's done in my recent Genesis core.

[kitrinx]

The menu option for changing the coefficients is actually a pretty nice user experience feature. The coeff.txt method provides the maximum flexibility and the ability to experiment and grow, but having that menu option with those presets makes life a lot easier, especially for less experienced users. Maybe some middle ground could be found?

[coughlanio]

I've noticed some severe image retention with the Genesis core using the new scaler on my monitor. It's an LG 27UK850W. I tried to see if I could reproduce the burn-in with the standard Genesis core, but it seems to be exclusive to the new scaler. Speaking with some people, it could be due to weird video timings?

What do you mean by image retention? Do the "triple buffer" option change anything ? The scaler should use the same video modes as Sorgelig's Genesis core.

Image retention still occurs when using triple buffered mode.

Here's an example of what I can see when I switch back to my desktop:


It seems to resolve itself after a minute or two.