x265 HEVC Encoder

[asarian]

I am about to buy an AMD 99503D, which supports AVX-512 very well. I wanted to prepare a command line with --asm avx512, but fail to see avx512 as a valid string in the documentation any more. Are there still versions of x265 that support this?

[LigH]

Support for AVX512 certainly still exists, I just read it in the sources. But it will not be auto-detected. So add the parameter explicitly; if it is not supported either by hardware or by your OS, or being removed from the CLI without my knowledge, x265 will tell you when you execute it.

[asarian]

Thank you.

[TR-7970X]

Originally Posted by asarian

Thank you.

I also have a 9950X3D, and I have this:- --asm avx512 in my x265 command line.

The documentation needs to be updated.

[asarian]

Originally Posted by TR-7970X

I also have a 9950X3D, and I have this:- --asm avx512 in my x265 command line.

Good to hear.

Can you tell me something about the performance boost using AVX512 gives? There are many charts on the 99050X3D, and even on them using x265, but never with AVX512 being used. I feel it will make quite a difference.

[LigH]

In general: In the past, the difference between AVX2 and AVX512 used to be marginal, and AVX512 has a higher risk of thermal throttling.

I can't tell you current results though, sorry. Try to search this thread for AVX512, I believe there have been reports a few months ago.

Like here

[TR-7970X]

Originally Posted by asarian

Good to hear.

Can you tell me something about the performance boost using AVX512 gives? There are many charts on the 99050X3D, and even on them using x265, but never with AVX512 being used. I feel it will make quite a difference.

I can't give you any info on performance, as I haven't been doing much recently, and I guess I just use it because I can.

I never noticed any thermal throttling with the 7950X, and I would suggest that the 9950X3D & the 7970X won't either, but I do have pretty good cooling.

But one thing I have noticed is 2 major typos in your post's, when referencing the 9950X3D !!!!

[Emulgator]

On a i9-11900K I found a few percent speed gain using -- asm avx512, limited by some of Intels safety features (limiting to base clock)
and the available cooling of a 125W CPU+115W GPU notebook system, although with plenty air intake, exhaust and 2x 5000rpm fan).

On a real desktop AMD 7950X/9950X I would expect considerable speed gains in the range of >25%, so well worth it.

[Z2697]

Unfortunatelly they are nowhere near 25% (compared to avx2)

[asarian]

Originally Posted by LigH

In general: In the past, the difference between AVX2 and AVX512 used to be marginal, and AVX512 has a higher risk of thermal throttling.

That's the beauty if the 9950X3D! Allegedly it includes an extremely efficient implementation of AVX512, where the CPU doesn't reduce clockspeed for them, and only consumes a few Watt extra on AVX512. So, this should go endlessly better on the 9950X3D.

[rwill]

Originally Posted by Z2697

Unfortunatelly they are nowhere near 25% (compared to avx2)

Yeah, there are not that many situations in a HEVC encoder where one needs 512bit wide registers.

[Emulgator]

Yeah, there are not that many situations in a HEVC encoder where one needs 512bit wide registers.

True. I should have gone down to 15% in my estimate.
How many % speed advantage AVX2 vs AVX512 in the described scenario do 7950X/9950X owners note ?

[Z2697]

1%-1.5% and 5%-8% respectively, for at least the set of settings and hardwares that I tested with.
(why there's a range, because lower quality / bitrate have more speedup somehow)

[excellentswordfight]

Originally Posted by Emulgator

True. I should have gone down to 15% in my estimate.
How many % speed advantage AVX2 vs AVX512 in the described scenario do 7950X/9950X owners note ?

Originally Posted by Z2697

1%-1.5% and 5%-8% respectively, for at least the set of settings and hardwares that I tested with.
(why there's a range, because lower quality / bitrate have more speedup somehow)

Sounds about right, I've seen about 5-8% with Xeon (Sapphire Rapids), and only 1-3% with 7000-series (Threadripper Pro). With the avx512 change in Zen5 i assume it should be about the same performance gain as the xeon.

[_DLS_]

Originally Posted by asarian

I am about to buy an AMD 99503D, which supports AVX-512 very well. I wanted to prepare a command line with --asm avx512, but fail to see avx512 as a valid string in the documentation any more. Are there still versions of x265 that support this?

In my tests with the 9950X3D, the speed bump with --asm avx512 can be between 16-36% depending on the other settings too.

on 2160p:

CTU 32, ref 4, subme 4, rd 4, rect, no-amp, aq-mode 2, tu-intra-depth 3, tu-inter-depth 3, max-merge 5, crf 20 => +15.5%
CTU 64, ref 5, subme 5, rd 4, rect, no-amp, aq-mode 2, tu-intra-depth 4, tu-inter-depth 4, max-merge 5, crf 18 => +36%

Temps are manageable on air cooling.

[Z2697]

Originally Posted by _DLS_

In my tests with the 9950X3D, the speed bump with --asm avx512 can be between 16-36% depending on the other settings too.

on 2160p:

CTU 32, ref 4, subme 4, rd 4, rect, no-amp, aq-mode 2, tu-intra-depth 3, tu-inter-depth 3, max-merge 5, crf 20 => +15.5%
CTU 64, ref 5, subme 5, rd 4, rect, no-amp, aq-mode 2, tu-intra-depth 4, tu-inter-depth 4, max-merge 5, crf 18 => +36%

Temps are manageable on air cooling.

Kinda too good to be true...

[Z2697]

x265 will produce nonconformity bitstream when a very sudden change in chroma channel happens and triggered weighted prediction.

https://bitbucket.org/multicoreware/...roma_offset_lx
(Finally remembered my bitbucket account)

Related issue (4 years ago, when I didn't know the root cause): https://bitbucket.org/multicoreware/x265_git/issues/582

Many hardware decoders will fail to decode such frame, and result in corrupted output.

HM will abort due to assertion.

int pred = (128 - ((128 * wp[plane].inputWeight) >> (wp[plane].log2WeightDenom)));
int deltaChroma = (wp[plane].inputOffset - pred);
WRITE_SVLC(deltaChroma, "delta_chroma_offset_lX");

TL;DR: the deltaChroma in the code I referenced above should be in the range of [-512, 511], but there’s no check on the value of pred. For example if the value of pred is a fairly “large” negative number, the deltaChroma can exceed the range.

[benwaggoner]

Originally Posted by LigH

In general: In the past, the difference between AVX2 and AVX512 used to be marginal, and AVX512 has a higher risk of thermal throttling.

More to the point "the difference between AVX2 and AVX512 used to be marginal, BECAUSE of AVX512 thermal throttling."

The per clock throughput improvements were solid, but the throttling reductions in instructions per second nearly cancelled those out. An implementation that maintains IPS with the IPC gains would be quite impressive.

[benwaggoner]

Originally Posted by Z2697

TL;DR: the deltaChroma in the code I referenced above should be in the range of [-512, 511], but there’s no check on the value of pred. For example if the value of pred is a fairly “large” negative number, the deltaChroma can exceed the range.

Good catch!

[asarian]

Originally Posted by _DLS_

In my tests with the 9950X3D, the speed bump with --asm avx512 can be between 16-36% depending on the other settings too.

on 2160p:

CTU 32, ref 4, subme 4, rd 4, rect, no-amp, aq-mode 2, tu-intra-depth 3, tu-inter-depth 3, max-merge 5, crf 20 => +15.5%
CTU 64, ref 5, subme 5, rd 4, rect, no-amp, aq-mode 2, tu-intra-depth 4, tu-inter-depth 4, max-merge 5, crf 18 => +36%

Temps are manageable on air cooling.

Now we're talking!

I heard +37% elsewhere (but didn't mention that yet, as I was looking for objective feedback). So, it now seems that number was quite accurate!