x265 HEVC Encoder

[LeXXuz]

Originally Posted by jpsdr

@LeXXuz
For the same command line effect, you have to replace --sbrc --sbrc-aq5 --sbrc-hyst with --aq-auto 10 assuming you don't want the new sbrc (wich has totaly nothing to do with the first patch) and just keep the auto AQ feature.
(And still check with the log file )

Thank you jpsdr. Actually I'm not sure if I want or need the new sbrc as I don't fully understand what it does and if it is beneficial for my usage.

So with the latest version I just add "--aq-mode 5 --aq-auto 10", is that right?

EDIT: Btw the clang version shows "x265 [INFO]: build info [Windows][MSVC 1928][64 bit] 10bit", is that correct

[Boulder]

Originally Posted by LeXXuz

So with the latest version I just add "--aq-mode 5 --aq-auto 10", is that right?

If you want to use the automatic AQ mode with AQ5 + hysteresis enabled, just use --aq-auto 10. --aq-auto 6 for HDR (enables auto mode + hysteresis, no AQ3 or AQ5 used since it's counterproductive).

[jpsdr]

Originally Posted by LeXXuz

EDIT: Btw the clang version shows "x265 [INFO]: build info [Windows][MSVC 1928][64 bit] 10bit", is that correct

Yes, because it's build with llvm but still with Visual Studio (so using clang compiler instead).

[Boulder]

@jpsdr: I've seen mentioned in some places that AQ mode 4 needs a bit lower aq-strength than other modes. Would it be possible to have a separate aq-strength parameter for it in case aq-auto is enabled?

[jpsdr]

@Boulder
Euh... I don't know. I'll see but absolutely not promises... Don't have time for now.

[Boulder]

Originally Posted by jpsdr

@Boulder
Euh... I don't know. I'll see but absolutely not promises... Don't have time for now.

Thank you

I was thinking that it might be as simple as adding the parameter and then conditionally using it as the base strength when the mode for the frame is either X265_AQ_EDGE or X265_AQ_EDGE_BIAS.

[jpsdr]

@Boulder
For X265_AQ_EDGE the calcul is:

                            inclinedEdge = curFrame->m_lowres.edgeInclined[blockXY];
                            qp_adj = curFrame->m_lowres.qpCuTreeOffset[blockXY];
                            if(inclinedEdge && (qp_adj - avg_adj > 0))
                                qp_adj = ((strength + AQ_EDGE_BIAS) * (qp_adj - avg_adj));
                            else
                                qp_adj = strength * (qp_adj - avg_adj);

So, you want something that can change the value of strength when aq-auto is enabled ?
If it's that, 2 options. Somethings like strength2 = strength*coeff or strength2 = strength + bias with coeff=1 or bias=0 when aq-auto is disabled and use strength2 instead of strength in calcul.
And if i add some kind of AQ4 specific parameter, it can be applied to AQ 4 mode, with or without aq-auto enable, if default value keeps original behavior.

[Boulder]

What I was thinking was basically that - having AQ 4 its own strength parameter. Maybe making it equal to aq-strength in case aq-auto is not enabled so any command line would work normally.

I was wondering if it would make sense to just change the calculation of 'strength' here, I haven't had the time to get the whole idea to work because the compiled exe won't accept my new parameters for some reason. What I did was just carbon copy the existing aq-strength and aq-bias-strength parameters in various files and adjust the variable names accordingly, but probably I've missed something. In this example there's also a different value for AQ 5 bias strength.

                    if (param->rc.aqMode == X265_AQ_EDGE || param->rc.aqMode == X265_AQ_EDGE_BIASED)
                        strength = param->rc.aqStrengthEdge * avg_adj;
                    else
                        strength = param->rc.aqStrength * avg_adj;
                    avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - modeTwoConst) / avg_adj;
                    if (param->rc.aqMode == X265_AQ_EDGE_BIASED)
                        bias_strength = param->rc.aqBiasStrengthEdge * param->rc.aqStrengthEdge;
                    else
                        bias_strength = param->rc.aqBiasStrength * param->rc.aqStrength;

[jpsdr]

Is it what you want ?
https://github.com/jpsdr/x265/commit...6dc36cef1d1880

[Boulder]

Originally Posted by jpsdr

Is it what you want ?
https://github.com/jpsdr/x265/commit...6dc36cef1d1880

Yes, that looks exactly what I thought of doing Thank you so much, I really appreciate it

[jpsdr]

I don't know when i'll be able to make a build (as it takes time), maybe this evening, maybe in few days, but you still can make your own build frommy github.

[Boulder]

Originally Posted by jpsdr

I don't know when i'll be able to make a build (as it takes time), maybe this evening, maybe in few days, but you still can make your own build frommy github.

No worries, I created a diff file to use for patching in Media Autobuild Suite so I got a compiled exe to test. First impression looks good, so I believe it's working. I didn't do any frame by frame compares but just watched an encoded clip to see if anything's weird.

[jpsdr]

If you have your own build, in that case, i'll realy take my time to make a new build...

[Boulder]

Originally Posted by jpsdr

If you have your own build, in that case, i'll realy take my time to make a new build...

Yeah, no problem with that

If anyone wants to test the modifications, here's the build I made. 10-bit only, and optimized for Zen 3. I can also provide the patch file I used if needed.

https://drive.google.com/file/d/1WGi...usp=share_link

[vadlerg]

Thank You for sharing the build.

--aq-auto 10 works. It is about 3% faster than --aq-mode 4 and may preserve a bit more detail.

When using --rd 4 --dynamic-rd 3 options I've got the warning

x265 [WARN]: Dynamic-rd disabled, requires RD <= 4, VBV and aq-mode enabled

That seems to be not rational. There must be some typo in the source, the build itself is OK.

[Boulder]

Originally Posted by vadlerg

Thank You for sharing the build.

--aq-auto 10 works. It is about 3% faster than --aq-mode 4 and may preserve a bit more detail.

When using --rd 4 --dynamic-rd 3 options I've got the warning


That seems to be not rational. There must be some typo in the source, the build itself is OK.

The error message is a bit bad English to me. Do you have VBV enabled as well? I believe the AQ-mode checks ok if AQ auto is enabled.

if (m_param->dynamicRd && (!bIsVbv || !p->rc.aqMode || p->rdLevel > 4))
    {
        p->dynamicRd = 0;
        x265_log(p, X265_LOG_WARNING, "Dynamic-rd disabled, requires RD <= 4, VBV and aq-mode enabled\n");
    }

[vadlerg]

VBV is set automatically because of the parameters

--crf 18 --preset slower --profile main10 --level-idc 52

.
aqMode should be 2 by default, but --aq-auto 10 sets it to

AQ: mode / str(edg) / qg-size / cu-tree : auto-hyst-aq5 / 1.0(1.0) / 32 / 1

rdLevel is set to 4 and encoder still disabling dynamic-rd, even if I set --rd to 3.
In the INFO section I see

x265 [INFO]: tools: rect amp limit-modes rd=4 psy-rd=2.00 rdoq=2 psy-rdoq=1.00
x265 [INFO]: tools: rskip mode=2 rskip-edge-threshold=0.03 limit-tu=4 signhide
x265 [INFO]: tools: tmvp b-intra strong-intra-smoothing deblock sao
x265 [INFO]: tools: selective-sao

[Boulder]

Originally Posted by vadlerg

aqMode should be 2 by default

Then it must be because of --aq-auto. I'll take a look at it when I'm back home from work and adjust that check in the code accordingly to make it work again.

[LeXXuz]

Originally Posted by vadlerg

aqMode should be 2 by default, but --aq-auto 10 sets it to

Isn't that what it is supposed to do?
https://forum.doom9.org/showthread.p...00#post1981600

[vadlerg]

Of course it is what it supposed to be but this should not disable dynamicRd.