madVR - high quality video renderer (GPU assisted)

[Duffy Moon]

Originally Posted by Tornado15550

I'm disappointed that there is still NO 64-bit version of madVR available..

Originally Posted by Damien147

there won't be any real advantage.

Additionally, it would take a lot of madshi's time to develop and then support. I think most of us agree that his time is better spent perfecting the 32-bit version first.

[burfadel]

Seek bar doesn't show for me (d3d fullscreen on secondary monitor), and the seek bar in mpc-hc doesn't do anything with madvr...

[mandarinka]

Originally Posted by madshi

madVR v0.84.3 released
(...)
* maybe fixed: black screen after switching display mode

Sadly it doesn't seem to be, at least for me. (A8-3850 CPU/GPU, Windows 8 preview, 64-bit, latest AMD drivers I think).

P.S. I'm fresh to MadVR, so I may be missing something.

[ryrynz]

Originally Posted by wiFFy

v0.84.3 performs worse than v.0.84.2. Started dropping frames when I updated madVR.

Drop your GPU queue size down to 4.

That being said, your really giving the HD4000 a workout with those settings, one small change here and there could cause dropped frames.
Keep in mind that quality is often preferred over performance in MadVR and if things get improved there may be a performance trade off.
Your settings may need to be changed to accommodate new versions as your basically at the limit of what the HD4000 can handle.

[Thunderbolt8]

Originally Posted by Tornado15550

I'm disappointed that there is still NO 64-bit version of madVR available. Madshi, are you planning to take care of this soon? This is because I do not intend on further using 32-bit media players on a 64-bit version of Windows.

bye bye.

[NicolasRobidoux]

I'm going to leave this alone for a while (I'm sure Mathias can do with a break from my interventions), but indeed EWA cubic B-spline windowed Jinc Radius 4 (which is a 4 lobe method) is one of my better low halo but reasonably sharp EWA methods. It's OK with text-like things, if you don't mind some blockiness. The one suggested in "The Recommendations" when halo bothers you, namely EWA quadratic B-spline windowed Jinc 3-lobe (with no, or very very slight, deblur) is less jaggy, but more blurry. So I think that many people may prefer the 4-lobe version.
This being said, madVR has a decent anti-ringing system, that gets improved all the time, so this may not matter here.
-----
P.S.
Since I ended up talking about it, here is the result with EWA quadratic B-spline windowed Jinc 3 lobe with no deblur (and no sigmoidization or colorspace shenanigans):
http://web.cs.laurentian.ca/nrobidou...draticJinc.png
And I may as well add "classic" EWA LanczosSharp:
http://web.cs.laurentian.ca/nrobidou...arpClassic.png

[wiFFy]

Originally Posted by ryrynz

Drop your GPU queue size down to 4.

Thanks for your reply. Whups, for some reason I forgot to mention that "delay playback start until render queue is full" and "delay playback start after seeking, too" are selected, so playback starts with full queues anyway. Changing GPU queue to 4 doesn't really change much as far performance is concerned. I'll edit my original post so there is no more confusion concerning that.

I did a little more testing and noticed that 720p to 1080p upscaling runs without dropped frames using lanczos3 without ar for chroma and jinc3 ar for luma. With lanczos3 ar for chroma and jinc3 without ar for luma it's too slow and it keeps dropping frames. Same goes for lanczos3 ar for chroma and jinc3 ar for luma with all "trade quality for performance" settings enabled.

Originally Posted by ryrynz

That being said, your really giving the HD4000 a workout with those settings, one small change here and there could cause dropped frames.

That's quite correct. According to GPU-Z gpu load is between 89% and 96% using v0.84.2 with lanczos3 ar for chroma and jin3 ar for luma. With v0.84.3 using the same settings gpu load is between 95% and 100%. I think that that hd4000 desktop model would have actually survived these changes from v.0.84.2 to v.084.3. If you got time, take a look at this article. It pretty much explains why I would rather not drop my gpu queue too low when gpu load is reaching maximum.

Originally Posted by ryrynz

Keep in mind that quality is often preferred over performance in MadVR and if things get improved there may be a performance trade off.
Your settings may need to be changed to accommodate new versions as your basically at the limit of what the HD4000 can handle.

I'm hoping that with time we will get configurable antiringing as I think madshi has suggested and I could find some settings that my gpu can handle. I would really hate to completely lose antiringing with these settings.

I think I will keep using the v0.84.2 for now.

[ryrynz]

Originally Posted by wiFFy

Thanks for your reply. Whups, for some reason I forgot to mention that "delay playback start until render queue is full" and "delay playback start after seeking, too" are selected, so playback starts with full queues anyway. Changing GPU queue to 4 doesn't really change much as far performance is concerned. I'll edit my original post so there is no more confusion concerning that.

It did for me on my HD4000, I could run your luma and chroma settings (both with AR) okay from 720 to 1080 upscaling.

Originally Posted by wiFFy

I did a little more testing and noticed that 720p to 1080p upscaling runs without dropped frames using lanczos3 without ar for chroma and jinc3 ar for luma.

I was going to suggest this also, removing AR from chroma gives you a good 10% or so extra GPU on the HD4000, every little bit helps.

Originally Posted by wiFFy

I think I will keep using the v0.84.2 for now.

Or overclock your HD4000, RAM speed may play a part as well.. I'm going to test that.

[wiFFy]

I keep editing my posts way too slow to keep up with you.

Originally Posted by ryrynz

It did for me on my HD4000, I could run your luma and chroma settings (both with AR) okay from 720 to 1080 upscaling.

Hmm, I need to recheck that.

Originally Posted by ryrynz

Or overclock your HD4000, RAM speed may play a part as well.. I'm going to test that.

I don't think that the hd4000 ulv is overclockable. Atleast MSI Afterburner and Intel extreme tuning utility don't allow changing clockrates. Just to be clear, I'm running madvr on this laptop.

Also it's already rather crippled memory bandwidth wise, since it's running on single channel. there is 2gb soldered to motherboard so there went my dual channel when I changed the original 2gb+2gb configuration to 2gb+8gb.

There is also the Geforce GT 620m, but as far as I've tested, it's completely useless when it comes to running madVR.

[ryrynz]

Originally Posted by wiFFy

I don't think that the hd4000 ulv is overclockable. Atleast MSI Afterburner and Intel extreme tuning utility don't allow changing clockrates. Just to be clear, I'm running madvr on this laptop.

Ahh performance may not be identical to my system then (Desktop i5 3570K). I did some benching and RAM speed does help for example setting Jinc 3 AR on luma and Lanczos 8 AR for chroma with my RAM at 1066Mhz 9-9-9-24
for a 30 second clip (720 to 1080 upscaling) I dropped 62 frames and delayed 281.

Setting it to 1600Mhz 9-9-9-24 that went down to 38/239, the command rate had an impact as well further impact as well dropping that to 32 dropped frames over 30 seconds.

But the biggest performance increase I had was overlocking the HD4000 to 1400Mhz which resulted in zero dropped and delayed frames so it's a shame the Intel extreme tuning utility doesn't work for you.

[strumf666]

Is it normal for madvr auto resolution matching to switch to 23,98Hz for 25p sources instead for 50Hz (my projector doesn't support 25p)? If it is, is there a way to bypass/force it?

[madshi]

Originally Posted by Jong

Any chance the frame drop/repeat estimate is wrong in the latest builds? Maybe only when using Reclock?

I've not changed anything in any code related to that. I've only decreased the OS timer interval, but that should help rather than hurt these calculations. These estimations take a looooong time to settle down. Maybe you just need to wait a little bit longer? Or is the different behaviour between the 2 madVR versions 100% reproduceable? Does madVR estimate frame *drops* or *repeats*? FWIW, repeats are estimated, but not actually counted during playback. So if there are repeats, you will not see it in the OSD stats.

Originally Posted by 6233638

Looking at it closer, it seems that the issue is that Jinc ends up with hard-edged ringing, rather than more ringing, which stands out more.

The text in this image shows it quite well. (at this point, it seems like "artificial" tests display these kinds of errors much easier than films)

Ouch. Such computer graphics are terribly hard to handle for the anti-ringing filter. Basically any content which has a lot of "one color pixel next to totally different color pixel next to another totally different color pixel" is hard to handle correctly. I would be dead easy for me to make madVR work great for these type of computer graphics, but then real videos would suffer. Well, maybe I should try to analyze the type of content first and then switch between 2 different anti-ringing filters? I don't know. Will think about that...

Originally Posted by 6233638

I was wondering how you tweak the anti-ringing filter though? Is it that you have the filter constructed, and you are now tweaking values, or is it more complex than that?

It's much more complex than that. I'm playing around with different algorithms, even multi-pass shaders instead of single-pass shaders etc...

Originally Posted by wiFFy

v0.84.3 performs worse than v.0.84.2. Started dropping frames when I updated madVR.

Rendering times in 0.84.3 upscaling 720p to 1080p:
chroma lanczos3 ar, luma jinc3 ar. Rendering time 45ms Constantly dropping frames.
chroma lanczos3 ar, luma lanczos8 ar. Rendering time 37ms ok. Queues are constantly full.

Rendering times in 0.84.2 upscaling 720p to 1080p:
chroma lanczos3 ar, luma jinc3 ar. Rendering time 39ms ok. Queues are constantly full.
chroma lanczos3 ar, luma lanczos8 ar. Rendering time 33ms ok. Queues are constantly full.

There are only 2 things which changed performance in v0.84.3:

(1) Linear light scaling, which should be slightly faster now.
(2) Jinc AR, which might be slightly slower now.

I'm still working on Jinc AR, it might get even slower in a future build, the algorithm is not totally finalized yet. It's unfortunate that it was fast enough for you in v0.84.2 and is just a tiny bit too slow in v0.84.3 now, but I can't stop optimizing it just because the optimization costs 3% more performance or something like that. Quality is more important than a couple percent performance.

All other filters (e.g. Lanczos AR) should perform identical. They haven't changed at all from v0.84.2 to v0.84.3.

Originally Posted by Bandito

when is madVR going to get sharpening filters?

Are you talking about custom pixel shader support, similar to what MPC-HC supports for EVR? Or are you talking about a dedicated sharpening algorithm implemented in madVR, with comparable quality to good AviSynth sharpening scripts? Both might come some day, but custom pixel shader support will likely come first.

Originally Posted by Tornado15550

I'm disappointed that there is still NO 64-bit version of madVR available. Madshi, are you planning to take care of this soon? This is because I do not intend on further using 32-bit media players on a 64-bit version of Windows.

No, I'm not planning to create a 64bit version soon. Maybe some day, but not anytime soon. The reason? Because the possible benefit is not big enough compared to the development time it would cost me.

Originally Posted by burfadel

Seek bar doesn't show for me (d3d fullscreen on secondary monitor), and the seek bar in mpc-hc doesn't do anything with madvr...

You do have the seekbar enabled in the madVR settings, yes? Do you have the "D3D Fullscreen" option activated in the MPC-HC "output" renderer settings? Try deactivating that.

Originally Posted by mandarinka

Sadly it doesn't seem to be, at least for me. (A8-3850 CPU/GPU, Windows 8 preview, 64-bit, latest AMD drivers I think).

Please describe your exact problem in detail.

Originally Posted by NicolasRobidoux

OK: Maybe I'll learn to love 4-lobe windowed Jinc methods, at least when using the "Radius = number of lobes" deblur. Here is the result of enlarging 7x with no colorspace or sigmoidization shenanigans with EWA Spline (meaning cubic B-spline windowing; this is the old ImageMagick "Cubic" filter and window) windowed Jinc 4 lobe with deblur set so that the radius is 4:
http://web.cs.laurentian.ca/nrobidou...incRadius4.png

Looks quite good for this one test image, at least.

Let me double check, just to be safe. This image was created with the following parameters:

filter: B-Spline 4-taps
window: Jinc 4-taps, 0.9431597994328477

Is that correct? Is the filter exactly the same one used for tensor scaling? But you feed the true distance (sqrt(distx^2 + disty^2) into the B-Spline filter, is that correct?

That was one thing I was wondering about, anyway: Can you use any normal tensor filter (like B-Spline, Mitchell, Bicubic, whatever), combine it with Jinc windowing and then feed the true distance into the tensor filter? Or do the normal filters have to be modified to be able to handle "true distance -> weight"? Or do you feed the x and y distance separately in to the filter, just like you would with tensor upscaling, and then just add the Jinc window on top?

Thanks!

[madshi]

Originally Posted by strumf666

Is it normal for madvr auto resolution matching to switch to 23,98Hz for 25p sources instead for 50Hz (my projector doesn't support 25p)? If it is, is there a way to bypass/force it?

3 questions:

(1) What is the list of display modes you entered into the madVR display mode switcher?
(2) What does the madVR debug OSD (Ctrl+J) say about the "movie fps (says source filter)"?
(3) What is the exact file name of the movie you want to play?

[strumf666]

(1) 1080p23, 1080p24, 1080p50, 1080p60
(2) 25fps - from memory
(3) will check when I get from work.

It's the expendables 2 1080p rip, russian source. It's dropping approx. 1 frame per second if refresh is set to 23,98Hz but plays smoothly when I set it to 50Hz.

It's does happen with other 25p sources too, but I rarely stumble upon them.

[burfadel]

Thanks It was because I left D3D fullscreen on! Seek bar works perfectly now. Maybe a ticket should be submitted for MPC-HC, so the D3D option is greyed out when using madVR

With the internal decoder, when going on to the next video, the screen is black until you exit exclusive mode but the audio plays. Video comes back once in windowed mode, and is still there when back in exclusive mode. If the internal decoder is disabled, and ffdshow is used instead (for example), the video plays as it should when the next video plays. This really isn't a big issue since I can use ffdshow, but thought I'd mention it

[nlnl]

Quote:

No, it does not
Please, have a look at the log and freeze report:
https://dl.dropbox.com/u/43376972/madvrnlnl_log_freez%2013-10-12.zip

Thanks!

madshi
madvrnlnl.ax build does not fix the problem, but 84.3 does!
Thanks!
BTW can not activate debug mode using 84.3.

[mandarinka]

Originally Posted by mandarinka

Sadly it doesn't seem to be, at least for me. (A8-3850 CPU/GPU, Windows 8 preview, 64-bit, latest AMD drivers I think).

P.S. I'm fresh to MadVR, so I may be missing something.

Originally Posted by madshi

Please describe your exact problem in detail.

When switching between window and fullscreen, the screen goes black for an instant after the switch to FS. Only for an short instant and not immediately, for a short while the video plays before the black screen. When the video comes back, the OSD message indicating that exclusive mode is used appears, so I guess that's the cause.

As said, this is from test on a Windows 8 release preview install, so it might not be worth to investigate yet, until the RTM version is out and Catalyst drivers are more mature.

I only follow this thread from time to time (since I didn't use MadVR so far). So this might be a known behaviour or simply an unavoidable thing, NVM me in that case. It isn't that annoying if the 3 second delay is disabled, too. When I saw that mention of the black screen in the changelog, I was under the impression that the mention refers to the issue I was seeing.

[noee]

Originally Posted by mandarinka

As said, this is from test on a Windows 8 release preview install, so it might not be worth to investigate yet, until the RTM version is out and Catalyst drivers are more mature.

The latest AMD drivers for Win8 seem quite solid in single monitor setup, but dual-mon is just not right yet if you're running the second monitor at a different refresh rate than the first. It might also be the new feature that puts the taskbar on that second monitor (this can be disabled). Don't know if it's Win8 or just the drivers so far, but I cannot get reliable behavior with Win8 on dual-mon with refresh rate changing and FSE with either of my Llano machines. Look for the coming 12.9 drives to coincide with the Win8 "release", supposedly will address "dual-mon" issues. You could try the beta drivers now to see if there's any improvement.

[strumf666]

Originally Posted by strumf666

Is it normal for madvr auto resolution matching to switch to 23,98Hz for 25p sources instead for 50Hz (my projector doesn't support 25p)? If it is, is there a way to bypass/force it?

Originally Posted by madshi

3 questions:

(1) What is the list of display modes you entered into the madVR display mode switcher?
(2) What does the madVR debug OSD (Ctrl+J) say about the "movie fps (says source filter)"?
(3) What is the exact file name of the movie you want to play?

Originally Posted by strumf666

(1) 1080p23, 1080p24, 1080p50, 1080p60
(2) 25fps - says source filter
(3) avshd-expendables1080p.mkv

It's the expendables 2 1080p rip, russian source. It's dropping approx. 1 frame per second if refresh is set to 23,98Hz but plays smoothly when I set it to 50Hz.

It's does happen with other 25p sources too, but I rarely stumble upon them.

Added missing info.

[NicolasRobidoux]

Originally Posted by madshi

...
Let me double check, just to be safe. This image was created with the following parameters:

filter: B-Spline 4-taps
window: Jinc 4-taps, 0.9431597994328477
...

Mathias: It's going to take me a while to answer your questions making sure that there is absolutely no confusion, in part because we speak slightly different technical languages.
-----
If I use the ImageMagick -define filter:verbose=1 option to "spy" into the filtering, I get this:

convert small.png -define filter:verbose=1 -define filter:filter=Jinc -define filter:lobes=4 -define filter:window=Spline -define filter:blur=.9431597994328477 -distort Resize 700% EWASplineJincRadius4.png
# Resize Filter (for graphing)
#
# filter = Jinc
# window = Cubic
# support = 4.24106
# window-support = 4.24106
# scale-blur = 0.94316
# practical-support = 4
# B,C = 1,0

 0.00	1.0472
 0.01	1.04701
 0.02	1.04646
 0.03	1.04554
 0.04	1.04425
 0.05	1.0426
 0.06	1.04059
 0.07	1.03822
 0.08	1.03549
 0.09	1.0324
 0.10	1.02897
 0.11	1.02518
 0.12	1.02105
 0.13	1.01657
 0.14	1.01176
 0.15	1.00662
 0.16	1.00114
 0.17	0.995345
 0.18	0.989228
 0.19	0.982798
 0.20	0.976058
 0.21	0.969016
 0.22	0.961677
 0.23	0.954048
 0.24	0.946133
 0.25	0.93794
 0.26	0.929475
 0.27	0.920745
 0.28	0.911757
 0.29	0.902518
 0.30	0.893036
 0.31	0.883317
 0.32	0.87337
 0.33	0.863201
 0.34	0.852819
 0.35	0.842232
 0.36	0.831447
 0.37	0.820472
 0.38	0.809317
 0.39	0.797988
 0.40	0.786495
 0.41	0.774845
 0.42	0.763047
 0.43	0.75111
 0.44	0.739042
 0.45	0.726852
 0.46	0.714548
 0.47	0.702138
 0.48	0.689632
 0.49	0.677038
 0.50	0.664364
 0.51	0.65162
 0.52	0.638813
 0.53	0.625952
 0.54	0.613045
 0.55	0.600101
 0.56	0.587129
 0.57	0.574136
 0.58	0.56113
 0.59	0.548121
 0.60	0.535115
 0.61	0.522121
 0.62	0.509147
 0.63	0.496201
 0.64	0.483289
 0.65	0.47042
 0.66	0.457601
 0.67	0.444839
 0.68	0.432141
 0.69	0.419515
 0.70	0.406967
 0.71	0.394505
 0.72	0.382133
 0.73	0.36986
 0.74	0.35769
 0.75	0.345631
 0.76	0.333688
 0.77	0.321867
 0.78	0.310173
 0.79	0.298612
 0.80	0.287189
 0.81	0.275909
 0.82	0.264777
 0.83	0.253797
 0.84	0.242973
 0.85	0.232311
 0.86	0.221814
 0.87	0.211485
 0.88	0.201329
 0.89	0.191349
 0.90	0.181548
 0.91	0.17193
 0.92	0.162496
 0.93	0.15325
 0.94	0.144195
 0.95	0.135331
 0.96	0.126662
 0.97	0.118189
 0.98	0.109914
 0.99	0.101838
 1.00	0.093963
 1.01	0.0862892
 1.02	0.0788179
 1.03	0.0715497
 1.04	0.064485
 1.05	0.0576241
 1.06	0.0509672
 1.07	0.0445141
 1.08	0.0382647
 1.09	0.0322184
 1.10	0.0263747
 1.11	0.0207329
 1.12	0.015292
 1.13	0.0100509
 1.14	0.00500834
 1.15	0.000163021
 1.16	-0.00448675
 1.17	-0.00894254
 1.18	-0.0132062
 1.19	-0.0172799
 1.20	-0.0211654
 1.21	-0.0248651
 1.22	-0.0283815
 1.23	-0.0317169
 1.24	-0.034874
 1.25	-0.0378553
 1.26	-0.0406637
 1.27	-0.0433022
 1.28	-0.0457738
 1.29	-0.0480813
 1.30	-0.0502282
 1.31	-0.0522175
 1.32	-0.0540527
 1.33	-0.0557372
 1.34	-0.0572742
 1.35	-0.0586675
 1.36	-0.0599205
 1.37	-0.0610369
 1.38	-0.0620204
 1.39	-0.0628745
 1.40	-0.0636032
 1.41	-0.06421
 1.42	-0.0646989
 1.43	-0.0650737
 1.44	-0.0653381
 1.45	-0.0654959
 1.46	-0.0655512
 1.47	-0.0655075
 1.48	-0.065369
 1.49	-0.0651393
 1.50	-0.0648222
 1.51	-0.0644217
 1.52	-0.0639414
 1.53	-0.0633851
 1.54	-0.0627566
 1.55	-0.0620595
 1.56	-0.0612976
 1.57	-0.0604744
 1.58	-0.0595936
 1.59	-0.0586587
 1.60	-0.0576731
 1.61	-0.0566404
 1.62	-0.0555638
 1.63	-0.0544468
 1.64	-0.0532926
 1.65	-0.0521043
 1.66	-0.0508854
 1.67	-0.0496384
 1.68	-0.0483668
 1.69	-0.0470733
 1.70	-0.0457607
 1.71	-0.044432
 1.72	-0.0430897
 1.73	-0.0417366
 1.74	-0.040375
 1.75	-0.0390076
 1.76	-0.0376367
 1.77	-0.0362645
 1.78	-0.0348934
 1.79	-0.0335253
 1.80	-0.0321625
 1.81	-0.0308069
 1.82	-0.0294602
 1.83	-0.0281242
 1.84	-0.0268009
 1.85	-0.0254917
 1.86	-0.0241979
 1.87	-0.0229214
 1.88	-0.0216633
 1.89	-0.0204249
 1.90	-0.0192074
 1.91	-0.018012
 1.92	-0.0168396
 1.93	-0.0156912
 1.94	-0.0145677
 1.95	-0.0134699
 1.96	-0.0123984
 1.97	-0.011354
 1.98	-0.0103372
 1.99	-0.00934828
 2.00	-0.00838814
 2.01	-0.00745659
 2.02	-0.00655429
 2.03	-0.00568133
 2.04	-0.00483784
 2.05	-0.00402394
 2.06	-0.00323964
 2.07	-0.00248501
 2.08	-0.00175991
 2.09	-0.00106439
 2.10	-0.000398309
 2.11	0.000238611
 2.12	0.000846515
 2.13	0.00142575
 2.14	0.00197648
 2.15	0.00249902
 2.16	0.00299377
 2.17	0.003461
 2.18	0.00390145
 2.19	0.00431509
 2.20	0.00470278
 2.21	0.00506479
 2.22	0.00540162
 2.23	0.00571386
 2.24	0.00600208
 2.25	0.00626698
 2.26	0.00650877
 2.27	0.00672836
 2.28	0.0069262
 2.29	0.00710303
 2.30	0.00725933
 2.31	0.00739585
 2.32	0.00751322
 2.33	0.00761204
 2.34	0.00769305
 2.35	0.00775688
 2.36	0.00780397
 2.37	0.00783547
 2.38	0.00785152
 2.39	0.00785306
 2.40	0.00784068
 2.41	0.00781506
 2.42	0.00777683
 2.43	0.00772663
 2.44	0.00766511
 2.45	0.00759289
 2.46	0.00751058
 2.47	0.00741883
 2.48	0.00731821
 2.49	0.00720935
 2.50	0.00709278
 2.51	0.00696912
 2.52	0.0068389
 2.53	0.00670271
 2.54	0.00656106
 2.55	0.00641448
 2.56	0.00626344
 2.57	0.0061085
 2.58	0.00595007
 2.59	0.00578869
 2.60	0.00562475
 2.61	0.00545873
 2.62	0.00529103
 2.63	0.00512206
 2.64	0.0049522
 2.65	0.00478184
 2.66	0.00461131
 2.67	0.00444097
 2.68	0.00427118
 2.69	0.0041022
 2.70	0.00393435
 2.71	0.0037679
 2.72	0.00360313
 2.73	0.00344026
 2.74	0.00327957
 2.75	0.00312123
 2.76	0.00296548
 2.77	0.00281249
 2.78	0.00266243
 2.79	0.00251546
 2.80	0.00237176
 2.81	0.00223144
 2.82	0.00209459
 2.83	0.00196137
 2.84	0.00183185
 2.85	0.00170611
 2.86	0.00158422
 2.87	0.00146625
 2.88	0.00135223
 2.89	0.00124221
 2.90	0.00113622
 2.91	0.00103426
 2.92	0.000936347
 2.93	0.000842474
 2.94	0.000752636
 2.95	0.000666805
 2.96	0.000584958
 2.97	0.000507055
 2.98	0.000433064
 2.99	0.000362925
 3.00	0.000296587
 3.01	0.000233983
 3.02	0.000175052
 3.03	0.00011972
 3.04	6.79069e-05
 3.05	1.95334e-05
 3.06	-2.5486e-05
 3.07	-6.7241e-05
 3.08	-0.000105827
 3.09	-0.000141332
 3.10	-0.000173861
 3.11	-0.000203509
 3.12	-0.000230387
 3.13	-0.00025458
 3.14	-0.000276216
 3.15	-0.00029538
 3.16	-0.000312196
 3.17	-0.000326749
 3.18	-0.00033917
 3.19	-0.000349544
 3.20	-0.000357989
 3.21	-0.000364599
 3.22	-0.000369494
 3.23	-0.000372758
 3.24	-0.000374508
 3.25	-0.000374839
 3.26	-0.000373844
 3.27	-0.000371624
 3.28	-0.000368268
 3.29	-0.000363867
 3.30	-0.000358517
 3.31	-0.000352298
 3.32	-0.000345301
 3.33	-0.000337607
 3.34	-0.000329289
 3.35	-0.000320409
 3.36	-0.000311066
 3.37	-0.000301317
 3.38	-0.000291229
 3.39	-0.000280864
 3.40	-0.000270291
 3.41	-0.000259545
 3.42	-0.000248694
 3.43	-0.000237803
 3.44	-0.0002269
 3.45	-0.000216013
 3.46	-0.000205228
 3.47	-0.000194554
 3.48	-0.000184026
 3.49	-0.000173676
 3.50	-0.000163535
 3.51	-0.000153635
 3.52	-0.000144008
 3.53	-0.000134645
 3.54	-0.000125595
 3.55	-0.000116858
 3.56	-0.000108441
 3.57	-0.000100368
 3.58	-9.26422e-05
 3.59	-8.52692e-05
 3.60	-7.82537e-05
 3.61	-7.1605e-05
 3.62	-6.53032e-05
 3.63	-5.9361e-05
 3.64	-5.37823e-05
 3.65	-4.8556e-05
 3.66	-4.36636e-05
 3.67	-3.91173e-05
 3.68	-3.48964e-05
 3.69	-3.09978e-05
 3.70	-2.74054e-05
 3.71	-2.41083e-05
 3.72	-2.1095e-05
 3.73	-1.83746e-05
 3.74	-1.58858e-05
 3.75	-1.36731e-05
 3.76	-1.16617e-05
 3.77	-9.89684e-06
 3.78	-8.31668e-06
 3.79	-6.94366e-06
 3.80	-5.72873e-06
 3.81	-4.68372e-06
 3.82	-3.77891e-06
 3.83	-3.01256e-06
 3.84	-2.37097e-06
 3.85	-1.83769e-06
 3.86	-1.39463e-06
 3.87	-1.04208e-06
 3.88	-7.5514e-07
 3.89	-5.36571e-07
 3.90	-3.62735e-07
 3.91	-2.38204e-07
 3.92	-1.46614e-07
 3.93	-8.73773e-08
 3.94	-4.61199e-08
 3.95	-2.28118e-08
 3.96	-9.10535e-09
 3.97	-2.47706e-09
 3.98	-1.23499e-09
 3.99	-0
 4.00	0

Now, this report truncates floating point values to make them more easy to read, and it actually does not show the LUT itself, just a sampling of the LUT at decimal positions (the LUT, of course, contains a number of values equal to a power of 2).
What it shows is that the "radial weight function" is the product of

a standard B-spline cubic which is stretched by a factor of 2 so it has radius 4 instead of 2 (if I was using this as a tensor windowing function, I'd get a square of "radius" 4 instead of the standard 2, which actually gives a square of width=height=2*4=8, which in my terminology gives a method which is 8 taps in each direction, for a total of at most 64 values used to compute an output pixel value; with EWA, less values are used because they have to be inside the circle inscribed in the square) (this is the window function)

and of

a standard Jinc 4-lobe function shrunk by 0.94316... so that it has radius 4 as well. (this is the windowed truncated low pass filter)

At every point, the LUT is the product of the values of these two functions.

------

For reference:

When using plain (deblur=1) EWA Jinc-windowed Jinc 4-lobe, instead of the stretched B-spline we use the stretched first lobe of Jinc so that it's radius matches exactly the radius of the unstretched and unshrunk Jinc 4-lobe disc (which has radius equal to the 4th root of the Bessel function J_1 divided by the number pi), and you multiply that by the unstretched and unshrunk 4-lobe Jinc.

When using tensor Lanczos 4 (Sinc-windowed Sinc 4-lobe), instead of the stretched B-spline we use the stretched first lobe of the Sinc function, stretched by a factor of 4 because its first root is 1 and the fourth root of Sinc is 4, and we multiply that by the unstretched and unshrunk Sinc 4-lobe function.

-----

If it helps, I can "verbose" the stretched B-spline function and the shrunk Jinc 4-lobe separately so you can see the LUTs. Just ask.

-----

I'll do my best to clarify things further in future posts.

P.S. I had to fix an horrid error in the "number of taps" discussion above, demonstrating that coffee must come before arithmetic.