Evaluation of HEVC decoders (SW, Hybrid and HW)

[ibius]

Originally Posted by huhn

they all use different APIs. looks like amd and intel have the similar APIs or the same that are needed for this program while nvidia has it own to get the same results.

It was a rhetorical question. Two independent decoders that utilize OpenCL work on AMD and Intel, but not on NVIDIA.

Like I said before at least for hybrid decoding purposes NVIDIA's OpenCL is a joke, although I'm pretty happy with my GTX 550Ti madVR-wise, I'm gonna have to upgrade to GTX 960 or similar when HEVC encodes hit main street.

[huhn]

Originally Posted by ibius

It was a rhetorical question. Two independent decoders that utilize OpenCL work on AMD and Intel, but not on NVIDIA.

Like I said before at least for hybrid decoding purposes NVIDIA's OpenCL is a joke, although I'm pretty happy with my GTX 550Ti madVR-wise, I'm gonna have to upgrade to GTX 960 or similar when HEVC encodes hit main street.

these decoder are decoder for AMD and intel and simply not nvidia so they are not independent at all.

nvidia DXVA for HEVC using DXVA (except 960) is hybrid like these other openCL decoder. compared to lavfilter x64 or the windows 10 decoder all hybrid decoder are looking like a joke. maybe faster but at which cost?

and nvidia is usually no challenge for AMD in term of GPGPU processing. no secret and cuda doesn't change that.

[ibius]

Originally Posted by huhn

these decoder are decoder for AMD and intel and simply not nvidia so they are not independent at all.

That's a nice phrase "for AMD and Intel and simply not Nvidia", so Strongene and CyberLink developed their OpenCL HEVC decoders specifically for AMD and Intel, but not Nvidia?
And how come it follows that they are not idependent?

In Strongene's OpenCL HEVC decoder readme file there's a note that says

The OpenCL Version Lentoid HEVC decoder now only guarantees the support for AMD dGPUs and APUs (Radeon HD 5000 series or above).

Which implies that an earlier version wasn't AMD-specific.

And I didn't see any AMD-only developer notes from CyberLink, so how do you know that both Strongene and CyberLink only implemented AMD/Intel-specific APIs?
If two different decoders with OpenCL support only work with AMD and Intel cards doesn't mean both decoders were developed only for AMD and Intel cards.
Maybe if Nvidia gets to OpenCL 2.0 in a few years, these then old decoders will suddenly start working.

Originally Posted by huhn

nvidia DXVA for HEVC using DXVA (except 960) is hybrid like these other openCL decoder. compared to lavfilter x64 or the windows 10 decoder all hybrid decoder are looking like a joke.

That's not what I meant or said at all, please pay more attention next time.

[huhn]

if they wanted to write a openCL decoder for nvidia than they could have done that. but they didn't. why? feel free to ask them.
but it is possible! and openCL 1.2 or 2.0 is for sure not the reason.

[NikosD]

You have to be a little or more masochist to write a program in OpenCL for Nvidia when there is CUDA, that's the truth.

So, you write an OpenCL program for the HW that truly supports it (AMD, Intel) and if you have even more resources available, you write another version for CUDA.

End of story.

[NikosD]

A very interesting update:

PowerDVD 15 seems to support 3 HEVC decoding modes, at least for Intel iGPU (Haswell).

CPU, DXVA and OpenCL.

Expect an updated post testing all these modes.

[huhn]

Originally Posted by NikosD

You have to be a little or more masochist to write a program in OpenCL for Nvidia when there is CUDA, that's the truth.

So, you write an OpenCL program for the HW that truly supports it (AMD, Intel) and if you have even more resources available, you write another version for CUDA.

End of story.

i hope you are aware that there are a lot of programs that work totally fine on all 3 GPU vendor. and i hope these people ignore such commentss and still give nvidia users the chance of using there openCL programs.

[ibius]

And I hope you are aware that this is a thread for HEVC decoder evaluation, and the only two OpenCL decoders available don't work on Nvidia hardware, period.

The only HW-assisted HEVC decoding for Nvidia users is through DXVA2 (w/ LAV and maybe PowerDVD 15 now) on high-end Kepler and Maxwell cards or GTX 960 with dedicated ASIC.

[huhn]

Originally Posted by ibius

And I hope you are aware that this is a thread for HEVC decoder evaluation, and the only two OpenCL decoders available don't work on Nvidia hardware, period.

The only HW-assisted HEVC decoding for Nvidia users is through DXVA2 (w/ LAV and maybe PowerDVD 15 now) on high-end Kepler and Maxwell cards or GTX 960 with dedicated ASIC.

and i wonder why?

The OpenCL version is developed jointly by Strongene and AMD

[huhn]

mircosoft is faster on this sample than lavfilter on my system.

i find this pretty strange and i don't see a reason for this.
and i got UHDcode working and it is indeed hybrid decoding. but with no real impact

i3 4130 r9 270
beauty
playback 1280x720

microsoft
59 FPS 89% CPU 26% GPU
UHDcode
34.5 FPS 65% CPU 51% GPU
LAV
46.5 FPS CPU 73% ~8GPU

DECODE
microsoft
59.5 FPS 89% CPU - GPU
UHDcode
36.5 FPS 68% CPU 30 % GPU
LAV
49 FPS 73% CPU - GPU

is it new that microsoft MF can read HEVC in MKV?

[NikosD]

I've finished my little survey today about HEVC CPU decoders, updated with new decoders and new tools here:
http://forum.doom9.org/showthread.ph...38#post1717938

I have to admit that I was strongly impressed by Lentoid x86 decoder, more than anything else.

So, I searched a little bit in the past in order to see some signs of this incredible performance.

I had already tested the previous version 2.0.2.6 twice, first time in the initial post and second time here:
http://forum.doom9.org/showthread.ph...52#post1705352

The results weren't that impressive and now I can tell that there are two reasons behind that.

In order to see for yourselves the two reasons I tested again head to head the two Lentoid versions 2.0.2.6 vs 2.0.3.x in all three systems on two files - 4 and 5 - for both modes playback (using scaling to 1280x720) and decode:

* for Lentoid 2.0.3.3 I used latest DXVA Checker v3.50 (not the older 3.4.1)

System 1

Core i7-4790(AVX2) - Win 8.1 Pro x64 - iGPU@1.5GHz - v4156


4. Girls -1080p60 fps -11Mbps

Lentoid x86 v2.0.3.3(decode) 208/276/332 CPU 84%
Lentoid x86 v2.0.3.2(decode) 214/289/361 CPU 85%
Lentoid x86 v2.0.3.1(decode) 213/288/358 CPU 86%
Lentoid x86 v2.0.2.6(decode) 94/154/210 CPU 35%


Lentoid x86 v2.0.3.3(playback) 89/107/109 CPU 21%
Lentoid x86 v2.0.3.2(playback) 89/107/108 CPU 21%
Lentoid x86 v2.0.3.1(playback) 88/108/109 CPU 20%
Lentoid x86 v2.0.2.6(playback) 74/100/109 CPU 22%


5.Beauty-2160p@30fps-12.3Mbps


Lentoid x86 v2.0.3.3(decode) 65/81/86 CPU 79%
Lentoid x86 v2.0.3.2(decode) 68/83/88 CPU 79%
Lentoid x86 v2.0.3.1(decode) 68/83/89 CPU 80%
Lentoid x86 v2.0.2.6(decode) 35/45/48 CPU 36%


Lentoid x86 v2.0.3.3(playback) 28/48/49 CPU 37%
Lentoid x86 v2.0.3.2(playback) 27/47/47 CPU 35%
Lentoid x86 v2.0.3.1(playback) 25/47/50 CPU 36%
Lentoid x86 v2.0.2.6(playback) 15/33/35 CPU 24%



System 2

Core 2 Quad@2.26GHz(SSE4.1) - Win 10 x64 build 10049 - GT610 - 347.25


4. Girls -1080p60 fps -11Mbps

Lentoid x86 v2.0.3.2(decode) 55/79/93 CPU 91%
Lentoid x86 v2.0.3.1(decode) 56/75/92 CPU 90%
Lentoid x86 v2.0.2.6(decode) 38/68/94 CPU 72%


Lentoid x86 v2.0.3.2(playback) 29/44/45 CPU 43%
Lentoid x86 v2.0.3.1(playback) 32/44/44 CPU 45%
Lentoid x86 v2.0.2.6(playback) 32/43/45 CPU 44%


5.Beauty-2160p@30fps-12.3Mbps

Lentoid x86 v2.0.3.2(decode) 4/19/23 CPU 83%
Lentoid x86 v2.0.3.1(decode) 4/20/23 CPU 82%
Lentoid x86 v2.0.2.6(decode) 8/18/21 CPU 71%


Lentoid x86 v2.0.3.2(playback) 1/13/13 CPU 40%
Lentoid x86 v2.0.3.1(playback) 1/13/15 CPU 41%
Lentoid x86 v2.0.2.6(playback) 3/12/13 CPU 39%



System 3

Core 2 Duo@2.33GHz (SSSE3) - Win 8 Pro x64 - iGPU 945GM - 8.15.10.1930



4. Girls -1080p60 fps -11Mbps


Lentoid x86 v2.0.3.3(decode) 26/46/90 CPU 89%
Lentoid x86 v2.0.3.2(decode) 28/47/76 CPU 91%
Lentoid x86 v2.0.3.1(decode) 28/46/75 CPU 90%
Lentoid x86 v2.0.2.6(decode) 24/41/81 CPU 88%


Lentoid x86 v2.0.3.3(playback) 27/38/45 CPU 72%
Lentoid x86 v2.0.3.2(playback) 26/37/45 CPU 72%
Lentoid x86 v2.0.3.1(playback) 22/37/45 CPU 73%
Lentoid x86 v2.0.2.6(playback) 20/30/44 CPU 63%


5.Beauty-2160p@30fps-12.3Mbps


Lentoid x86 v2.0.3.3(decode) 5/15/17 CPU 92%
Lentoid x86 v2.0.3.2(decode) 3/14/16 CPU 91%
Lentoid x86 v2.0.3.1(decode) 3/14/16 CPU 93%
Lentoid x86 v2.0.2.6(decode) 4/12/14 CPU 84%


Playback mode using EVR not supported by iGPU for UHD.


A careful reader will see the two points that made new Lentoid x86 2.0.3.x extremely fast:

1) In the first two benchmarks tests, I used DXVA processing or Playback benchmark mode as it is called now.

Well, I don't know if it's a problem of DXVA Checker or of Lentoid x64/x86 decoder, but in this mode the CPU utilization of Lentoid is low and the performance is low too.
It's like a bottleneck across all systems and both versions old (2.0.2.6) and new (2.0.3.x)

I've sent an email to the DXVA Checker's developer about this issue, but he hasn't confirmed anything yet.

In the last test, I used Decode Benchmark mode, where the Lentoid x64/x86 decoders really shine, with no bottlenecks and high CPU utilization.

2) In my comments of second test on January 2015 I wrote:

Lentoid (Strongene) seems to be the fastest x86 decoder, but not with HT CPUs. It looks like a problem with multi-threading.

Well, it seems that Lentoid has already seen that weakness of their decoder (or they read my benchmarks ) and they fixed exactly that.

See for yourself the double speed of Lentoid x86 in Decode mode for System 1 (Core i7 4790 - 8 threads) because of the double CPU utilization of the new version compared to the old one on the same system.

Also for the first system, there is an improvement in Playback Benchmark mode, compared to the older platforms, but only for the 2160p file and it's still too low performance for such a fast decoder.

I hope a few things are better clarified now.

[NikosD]

Originally Posted by huhn

is it new that microsoft MF can read HEVC in MKV?

From my 8 HEVC samples, the 5 of them are .mkv and they were decoded by Microsoft x64/x86 HEVC decoder using LAV splitter and Win 10 Build 10049 - System 2

[ksio89]

I'm trying to play some video samples encoded in HEVC but DXVA2 is not being activated, does anyone know why? Both HEVC and UHD are available and activated in LAV (both x86 and x64) when DXVA2 (native) is chosen as the HW decoder. BTW, I'm using MPC-HC x64.

Some time ago I posted a similar question, but it was an UHD video in H.264 codec instead of HEVC.

The video samples are available at http://www.elecard.com/en/download/videos.html

Tried HEVC in UHD, FHD, and HD, DXVA2 isn't activated with none of them.


These are the specs of my laptop:

Dell Inspiron 15R-5537

CPU: core i7-4500U @ 1,8-2,7 GHz
GPU: AMD Radeon HD 8870M
RAM: 8 GB DDR3 @ 1600 MHz
OS: Win 8.1 Pro

software info:

codec: K-Lite codec pack 11.1.0 Full
MPC-HC build: 1.7.8.152 (3f78c2a) (develop) x64
LAV filters: 0.64.0.51

This laptop is an Enduro (MUXless dynamic) platform, similar to nVidia Optimus, in which the iGPU is activated all the time, but rendering (gaming, video playback) can be done by GPU via PCI-E bus if set to high perfomance. However, as someone pointed before, the Radeon GPU can only decode HEVC via OpenCL, so DXVA HW decoding of HEVC is only available to HD Graphics 4400 iGPU.

I don't know if HD Graphics 4400 HEVC decoder or the CPU is fast enough to play 4K, but DXVA should be at least activated, because HEVC is shown as available when DXVA2 native/copy-back HW decoder is selected. Thanks in advance.

[NikosD]

Can you post a DXVA checker first screen screenshot ?

I think your Radeon card "overlaps" your iGPU so only DXVA capabilities of Radeon would be available.

If that's true, you have to wait for egur, Intel's employee, to add to his QS decoder HEVC decoding in order to use it with your laptop.

[NikosD]

Originally Posted by ksio89

BTW, I'm using MPC-HC x64.

Some time ago I posted a similar question, but it was an UHD video in H.264 codec instead of HEVC.

If you go to LAV Video properties inside MPC-HC and choose QS decoder (Intel QuickSync) and you tick UHD (4K), you will have HW acceleration for H.264 UHD (4K) files.

[littleD]

Did anybody notice that Strongene Lentoid decoder has also 64 bit version now? And not suprisingly, it is a bit faster than x86 version. Just tiny behind Lav 64 bit.

[wanezhiling]

Originally Posted by littleD

Did anybody notice that Strongene Lentoid decoder has also 64 bit version now? And not suprisingly, it is a bit faster than x86 version. Just tiny behind Lav 64 bit.

http://forum.doom9.org/showthread.ph...47#post1717847

[NikosD]

This is a post for hybrid decoding (DXVA, OpenCL) on Intel's iGPU HD4600.


Decoders:
LAV x64/x86 DXVA 0.65
MPC x64/x86 DXVA 1.4.4.286
PowerDVD x86 DXVA & OpenCL 15.0.1510.58
Lentoid x86 OpenCL decoder v2.0.0.2



Tool:
DXVA Checker 3.4.1 (Benchmark Decode mode)


Splitter:
LAV Splitter 0.65


All DXVA tests were done in Native mode.
GPU utilization is ~100% in almost all tests during all the time, with the exceptions of LAV/MPC written in comments.


System 1

Core i7-4790 - Win 8.1 Pro x64 - iGPU HD4600@1.5GHz - v4156


1.ProRes-1080p@30fps-2Mbps




MPC DXVA x64 135/203/229 CPU 21%

LAV DXVA x64 131/202/224 CPU 22%

LAV DXVA x86 135/202/227 CPU 24%

PDVD DXVA x86 137/200/237 CPU 25%

PDVD OCL x86 142/198/222 CPU 34%

MPC DXVA x86 145/197/225 CPU 23%

Lentoid OCL x86 (SW fall-back mode)



2.Elephants Dream-1080p@24fps-1.7Mbps



PDVD DXVA x86 132/219/308 CPU 28%

LAV DXVA x64 87/211/304 CPU 23%

MPC DXVA x64 94/208/295 CPU 22%

PDVD OCL x86 120/203/247 CPU 35%

LAV DXVA x86 74/203/290 CPU 25%

MPC DXVA x86 75/202/293 CPU 25%

Lentoid OCL x86 Black screen



3.Big Buck Bunny-1080p@60fps-2Mbps



LAV DXVA x64 180/255/312 CPU 23%

MPC DXVA x64 165/253/308 CPU 27%

PDVD DXVA x86 196/249/300 CPU 28%

MPC DXVA x86 152/242/298 CPU 25%

LAV DXVA x86 135/241/303 CPU 25%

PDVD OCL x86 172/218/241 CPU 33%

Lentoid OCL x86 Not working



4. Girls -1080p60 fps -11Mbps



PDVD DXVA x86 160/199/242 CPU 41%

PDVD OCL x86 140/181/207 CPU 46%

MPC DXVA x64 109/170/218 CPU 27%

LAV DXVA x64 144/166/215 CPU 28%

MPC DXVA x86 91/157/213 CPU 27%

LAV DXVA x86 106/153/196 CPU 27%

Lentoid OCL x86 (SW fall-back mode)



5.Beauty-2160p@30fps-12.3Mbps



LAV DXVA x64 44/54/56 CPU 22%

PDVD DXVA x86 48/53/58 CPU 25%

MPC DXVA x64 18/51/56 CPU 21%

LAV DXVA x86 46/49/52 CPU 25%

MPC DXVA x86 Crashed in 3rd time in a row test

Lentoid OCL x86 (SW fall-back mode)

PDVD OCL x86 (SW fall-back mode)



6.Fitness-2160p@30fps-8Mbps



PDVD DXVA x86 66/75/92 CPU 57%

MPC DXVA x64 41/60/84 CPU 33%

LAV DXVA x64 49/58/82 CPU 34%

LAV DXVA x86 47/57/82 CPU 33%

MPC DXVA x86 37/56/82 CPU 33%

PDVD OCL x86 (SW fall-back mode)

Lentoid OCL x86 Crashed



7.Ducks-2160p@50fps-4Mbps



PDVD DXVA x86 62/67/70 CPU 26%

LAV DXVA x64 63/66/68 CPU 26%

LAV DXVA x86 63/66/68 CPU 27%

MPC DXVA x64 35/63/69 CPU 24%

MPC DXVA x86 Crashed in 3rd time in a row test

Lentoid OCL x86 (SW fall-back mode)

PDVD OCL x86 (SW fall-back mode)




8.Gravity - 2048x858@24fps-5Mbps-10bit

Intel's hybrid decoder doesn't support 10 bit HEVC decoding, so all decoders work but in SW fall-back mode with the exception of Lentoid OCL which crashed.



Comments:



0) 4K60fps decoding is feasible using DXVA hybrid decoders, only with low bitrates on HD4600@1.5GHz.

1) PowerDVD 15 DXVA x86 decoder is not only the fastest x86 DXVA decoder for HEVC, but also the fastest DXVA decoder in general, as it is faster, most of the times, even than LAV/MPC x64 HEVC decoders.

2) On the other hand, PowerDVD 15 OCL x86 decoder although is fast, it uses always more CPU power than DXVA hybrid decoders and mainly it doesn't support 2160p decoding and falls back to SW.

3) LAV/ MPC x64 versions are a little faster than the x86 versions, but not with a big difference.

LAV/MPC x64/x86 had some problems utilizing the GPU completely, sometimes they dropped the load a lot and iGPU frequency dropped too, resulting in lower performance e.g clips 2,4,6.

Optimizations needed for LAV DXVA in order to be competitive with PowerDVD 15 DXVA utilizing GPU at 100% all the time.

4) Lentoid OpenCL is simply not working. Major compatibilities with a lot of HEVC files and when it's working it falls back to SW mode.

5) All DXVA decoders output is NV12, PowerDVD 15 OCL uses YUY2 and Lentoid OCL uses I420

[P.J]

How did you add PDVD to DXVAChecker?

[NikosD]

What do you mean ?

The registration of the DS filter and the use of it by DXVA Checker is exactly the same as PowerDVD 14.