MPC: Half/Full Float Processing Point?

[flashandpan007]

I use MPC with LAV Filters. What is the benefit of Half Float Processing Point and Full Float Processing Point?

[Hoser Rob]

Google is your friend.

[flashandpan007]

I already tried but i don't understand it. Also read the mpc wiki Render settings. Please help me.

[flashandpan007]

I also don't know what is different when no one is selected.

[hello_hello]

Google is your friend.

Apparently not a good enough friend to allow you to provide a link to the answer.

[deadrats]

traditionally full floating point refers to 32 bit floating point precision when carrying out calculations, half point would be 16 bit floating point precision. in theory, 32 bit floating point precision would lead to the highest quality because of less rounding of calculation. 32 bit precision during rendering is usually only available with pro caliber apps.

with regards to MPC, i assume you mean media player classic, it refers to the color depth that the video is displayed at, much like with your monitor you can set it to 32 bit (full color) or 16 bit (less color fidelity).

use 32 bit unless your monitor is ancient and doesn't support such a color depth.

[flashandpan007]

Most video files i have are 8bit 4:2:0. what is the difference of 32 full / 16 half / no selection of these two? Is it visible? There is also 10bit Output Setting in this menu. What Color depth is used when no one is selected?

[flashandpan007]

Can somebody confirm the following?
- nothing selected -> 8bit rendering
- Half Floating Point Processing -> 16bit rendering
- Full Floating Point Processing -> 32bit rendering

[pandy]

Can somebody confirm the following?
- nothing selected -> 8bit rendering
- Half Floating Point Processing -> 16bit rendering
- Full Floating Point Processing -> 32bit rendering

No,

floating is floating - half - use 16 bit where full use 32 bits - output can be rendered as 8 or 10 bits (higher?) with dither (noiseshaping ? - error distribution)

http://en.wikipedia.org/wiki/Half-precision_floating-point_format
http://www.nvnews.net/vbulletin/showthread.php?t=18997

[flashandpan007]

What should I use?
Reading your links is now in progress

[pandy]

https://trac.mpc-hc.org/wiki/New_Renderer_Settings

"Full Floating Point Processing

This option enables 32-bit floating point RGB surfaces. Mathematical outputs from the internal filters do not receive any rounding errors in between with this surface format. Dithering can be enabled with this mode.

Half Floating Point Processing

This option enables 16-bit floating point RGB surfaces. Mathematical outputs from the internal filters do not receive much rounding errors in between with this surface format. It's a bit lighter on memory throughput than the Full Floating Point Processing option. Dithering can be enabled with this mode.
"

[enim]

Does anyone know -> How much maximum Floating Point Processing Accuracy can be expected from a generic C or CPP program?

[flashandpan007]

https://trac.mpc-hc.org/wiki/New_Renderer_Settings
@ pandy: I read this already 20 times since I found it 2 days ago but I don't get it. I tried several translators but with no luck. It would help me a lot if you and/or some other in this forum could explain me that.

[pandy]

Does anyone know -> How much maximum Floating Point Processing Accuracy can be expected from a generic C or CPP program?

http://en.wikipedia.org/wiki/IEEE_floating_point#Basic_formats

single 32 bit and double 64 bit

http://en.wikipedia.org/wiki/Single_precision
http://en.wikipedia.org/wiki/Double-precision_floating-point_format

[pandy]

https://trac.mpc-hc.org/wiki/New_Renderer_Settings
@ pandy: I read this already 20 times since I found it 2 days ago but I don't get it. I tried several translators but with no luck. It would help me a lot if you and/or some other in this forum could explain me that.

i would say - for me going for FP is overkill when compressed video is considered as a source to display on ordinary LCD.
FP16 reduce requirement on bandwidth almost twice when compared to FP32 thus i would say: i don't see any added value to go for FP32 but if your hardware is powerful enough then go for FP32. For me correctly used 16 bit integer is more than required to have high quality video - it is hard for me to imagine algorithm that require FP32 (i mean this is even overkill for average consumer audio).
Perhaps im wrong, perhaps there are some benefits from FP still there is only one limitation - your hardware.

btw did you noticed this?:

"Mathematical outputs from the internal filters do not receive any rounding errors in between with this surface format."
and
"Mathematical outputs from the internal filters do not receive much rounding errors in between with this surface format."


I would say that even theoretically possible rounding errors will be not perceived in average customer scenario.
Perhaps this can be important for high complexity filtering and multiple iteration or some folded calculation but most of cases for simple, linear processing i would go to the fastest one (probably also less energy consumed)...

[flashandpan007]

@ pandy
Ok, that helped me.
Thank you

What do you use (if you're a MPC-HC/MPC-BE-user)?

Final question: What does the renderer when no one of this two is selected in MPC?

[enim]

Why theories goes on left while keeping practice on the right?
I mean did anyone really observed these theoretical floating point accuracy in a generic C or CPP program output?
I never did beyond XX.123456789012xxxxxxxxxxxxxxx (where xxxx = errands - error garbage bits).

WTF , We are taking about here.
If you claim more than this, prove it by a simple program adding two real numbers.


---------------------------------------------------------------------------------------------------------------------------------------------------------
This world is full of BuNcH oF IDiOTs (me included too, as usual), I do not blame who develop the standards coz we are far above.

[pandy]

@ pandy
Ok, that helped me.
Thank you

What do you use (if you're a MPC-HC/MPC-BE-user)?

Final question: What does the renderer when no one of this two is selected in MPC?

I would check from low to high and give up when i can't see difference - remain time i will use to enjoy life - take seriously my advise - don't waste time to search for something hidden somewhere at 0.000000x .

[pandy]

Why theories goes on left while keeping practice on the right?
I mean did anyone really observed these theoretical floating point accuracy in a generic C or CPP program output?
I never did beyond XX.123456789012xxxxxxxxxxxxxxx (where xxxx = errands - error garbage bits).

WTF , We are taking about here.
If you claim more than this, prove it by a simple program adding two real numbers.

Sometimes high accuracy is required - typical case is high order IIR/FIR filters - float numbers have own issues "Floating-point arithmetic,on the other hand, requires a quantization after every multiply and after every add in the convolution summation." (source Bomar, B.W. “Finite Wordlength Effects” Digital Signal Processing Handbook Ed. Vijay K. Madisetti and Douglas B. Williams).

But once again usually this is not case for video signal on consumer applications.

Float sometimes are bad also because less experienced developers overestimate they precision and confusing dynamic range with accuracy.

btw
For audio applications sometimes even up to 56 - 76 bits integer precision is used (devices such as TAS3204, CS47048, ADAU1701)

[enim]

Sometimes high accuracy is required - typical case is high order IIR/FIR filters - float numbers have own issues "Floating-point arithmetic,on the other hand, requires a quantization after every multiply and after every add in the convolution summation." (source Bomar, B.W. “Finite Wordlength Effects” Digital Signal Processing Handbook Ed. Vijay K. Madisetti and Douglas B. Williams).

But once again usually this is not case for video signal on consumer applications.

Float sometimes are bad also because less experienced developers overestimate they precision and confusing dynamic range with accuracy.

I am just replying you, nothing personal.

In my inner-close group all are GuRuS and Pros in their respective fields we can either call them Dr or Scientists.
We know all in-and-out of the subject we touch to the root levels.
We know all Blah... Blah...Blah... and Boo...Boo...Boos L@@Ks good on paper.
It makes only difference when it comes to real world practice.

But once again usually this is not case for video signal on consumer applications.

Even Rocket & Missile Technology does not need precision more than 10 digits, in most of the case.
Applications that runs universally without hunger of additional resources are always welcomed widely.
Then why waste of too much energy, power, time and money?

[flashandpan007]

I don't see any difference when I use 16 or 32 bit processing, so I think I leave it like I used it for the last years.

I use a Samsung T27A950 and the display stats of MPC-BE (had to move on from MPC-HC because it is very unstable now, perhaps it will be more stable, but for now I'm happy with MPC-BE) that the display is "X8R8G8B8" so why should I use 16 or 32 bit processing when the display and backprocessing is "X8R8G8B8"? Only the surface is changed