RiVa AMMX Benchmarks

[meynaf]

Originally Posted by grond

You probably also use "cat" as a video player on linux, so many fps!

No, i've simply used unmodified Riva under winuae, with Tuko's settings.
Just about everyone can download the program and try the same.

[wawa]

I didnt know we are benchmarking pc hardware with uae here, are we?

[ShK]

080 has already compatibility for the old RiVA. Let's test how RiVA AMMX runs on an Amiga 2000:

[meynaf]

Originally Posted by grond

BTW, the fact that you had to use the dither=gray option to make your UAE beat the 080 proves two points: AMMX is a very powerful extension to the 68k ISA and you need to buy a newer PC.

It proves absolutely nothing. I just copy-pasted the settings

4.Upload:> riva-0.50 verbose fps=1000 noskip noaudio shk_topgun_320.mpg            
 Video: 320x176, 24.000 fps
 Audio: <NONE>

 Number of frames played:  2143
 Number of frames skipped: 0
 Total number of frames:   2143

 Total playback time: 13.8367 seconds.
 Average framerate:   154.8777 fps
 Displayed framerate: 154.8777 fps

4.Upload:>

Satisfied ?

Originally Posted by wawa

I didnt know we are benchmarking pc hardware with uae here, are we?

Now you know this exists too.

[flype]

Just some facts :

RiVA 0.50 in Aminet is already compatible with any Amiga. The work done by Stephen Fellner yet push your Amiga to the max. This is the faster MPG-1 player available on the classics Amiga - by far. This is very good work he did yet. Our RiVA do not hurts the existing situation, it just brings improvements to a hardware that can't runs @ Ghz.

RiVA offers some options - Accupak or PIP option - which are compatibles only Picasso IV Gfx card. These are options for users who have the hardware for. We propose in a elegant manner a fast solution for a dedicated hardware. This is also quite equivalent when using dedicated hardware decoder to listen MP3 (MHI).

[matthey]

Some of meynaf's questions and points are valid. I would also be interested to have some answers.

1) What is the status of the 68k RiVa? Is the source available anywhere?

2) There is mention of improvements to superscalar execution which benefit the 68060 as well. Will this update be released? What are the SIMD isolated results when the 68060 executable with superscalar improvements is compared to the SIMD AMMX enhanced version?

3) What is the plan for future SIMD extensions to the ISA? How would floating point support in the SIMD be accomplished without putting floating point values in integer registers with a shared integer/SIMD ISA? How would wider SIMD registers be supported with a shared integer/SIMD ISA? Is the Apollo-core documentation of the ISA/ABI ever really going to be made public? Why are the comments for the SIMD instructions more readable than the assembler instructions and appear alien to the 68k?

@ meynaf :
obviously, you don't like what the vampire team do and what the Apollo 080 is, right ? It's your choice, and you are free to think so. But the fact is, that this team has brought the first innovative and by the way the far fastest accelerator to the classic Amiga since decades. The vampire is a nice and fast piece of hardware, and the apollo core is the most advanced 68k itteration now. And the only thing you do is to compare it with an UAE running on a super fast PC and complaining that those AMMX instructions are not a good idea ? Comme on, if you're not happy with the vampire, go play with WinUAE or OS4 or MOS or Aros, but don't say that what the team do is pointless, because this is clearly not the case. I have a vampire 600 and a blizzard ppc 060, and I can tell you that the speed feeling is clearly in favor of the vampire, no discution ! Je ne comprends pas comment on peut être aussi mauvais envers le travail des autres !!!!

[meynaf]

You're making abusive generalization here. I'm not saying the 080 is a bad thing as a whole - i'm just saying these ammx instructions are not a good idea.
Sure it's fast. But the speed feeling you have does not come from ammx.

Remember that winuae is something everyone can verify.
Can everyone verify the ammx ?

I've been in the Apollo team for some time - probably longer than most team members posting here, in fact -, so it's not exactly "le travail des autres".

[grond]

Originally Posted by meynaf

You're making abusive generalization here. I'm not saying the 080 is a bad thing as a whole - i'm just saying these ammx instructions are not a good idea.
Sure it's fast. But the speed feeling you have does not come from ammx.

The double frame rate in movie replay does come from ammx. And your point about arbitrary instructions for single purpose use is moot. MMx instructions have been used in millions of programs and all sorts of codices.

Can everyone verify the ammx ?

Several hundred vampire owners will be able to do this shortly when GOLD2 is released.

I've been in the Apollo team for some time - probably longer than most team members posting here, in fact -, so it's not exactly "le travail des autres".

So now you are claiming part of the apollo core as your work?

Originally Posted by meynaf

You're making abusive generalization here. I'm not saying the 080 is a bad thing as a whole - i'm just saying these ammx instructions are not a good idea.
Sure it's fast. But the speed feeling you have does not come from ammx.

Yes, it does : I'm beta testing Riva too, and I can tell you that the last beta version has a big speed improvement over the latest that didn't use the MMX instructions, this on the same machine. I have demonstrated it the last week at the Micro Alchimie in France, many people was impressed with this speed bump.
You can dislike the MMX addition, but the fact is, that RIVA is faster thanks those instructions. Period. It's the only thing that TuKo said on the first post, we didn't claim anything else.

[prowler]

WARNING

From this point on, all trolling will be removed to promote legitimate discussion only.

[flype]

Originally Posted by matthey

Some of meynaf's questions and points are valid. I would also be interested to have some answers.

1) What is the status of the 68k RiVa? Is the source available anywhere?

2) There is mention of improvements to superscalar execution which benefit the 68060 as well. Will this update be released? What are the SIMD isolated results when the 68060 executable with superscalar improvements is compared to the SIMD AMMX enhanced version?

3) What is the plan for future SIMD extensions to the ISA? How would floating point support in the SIMD be accomplished without putting floating point values in integer registers with a shared integer/SIMD ISA? How would wider SIMD registers be supported with a shared integer/SIMD ISA? Is the Apollo-core documentation of the ISA/ABI ever really going to be made public? Why are the comments for the SIMD instructions more readable than the assembler instructions and appear alien to the 68k?

Hi,

About the original 0.50 sources, RiVA was not free until 2014, but there was a demo with some limitations. Then it became free. But still closed-sources. Current status is i asked yesterday the opinion of the author about a public release. I know he is thinking of it and is open to. In itself this is good news.

Sure, there will be a release. We aim this with GOLD 2 release.

About the comments on SIMD instructions more readable ; it is because code use MACROs.

Can't help more on the others questions, quite techy for my knowledge.

[meynaf]

Originally Posted by grond

The double frame rate in movie replay does come from ammx.

It comes mostly from other optimizations, like better superscalar execution. See the comparison for 68060.
And even if it did. Creating instructions for a mere x2 isn't worth it. If it were x50 or even x10, i'd have said : wow. But x2 ?

These instructions will become obsolete as soon as another, larger simd comes out. But the cost of supporting these will still be here.
x86 is crippled with more than 1,000 instructions. I don't want the 68k to become like that.

Originally Posted by grond

And your point about arbitrary instructions for single purpose use is moot. MMx instructions have been used in millions of programs and all sorts of codices.

Really, no. SSE2 maybe, and not in millions of programs (and mostly to replace old x87). MMX certainly not.

Originally Posted by grond

Several hundred vampire owners will be able to do this shortly when GOLD2 is released.

Only if the program is available for them to check.

Originally Posted by grond

So now you are claiming part of the apollo core as your work?

My influence was small, but not zero.
Do you know who fought the most with Gunnar so that all instructions get implemented with no emulation (which they hopefully are today) ?
Do you know who gave him instruction usage statistics ?
Look in the docs for new stuff. Perhaps some was my idea, or was inspired by me.
Me and Matt Hey were in the "team" part at apollo-core.com before Gunnar erased it. I can't believe nothing came out of this.

Originally Posted by guibrush

Yes, it does : I'm beta testing Riva too, and I can tell you that the last beta version has a big speed improvement over the latest that didn't use the MMX instructions, this on the same machine. I have demonstrated it the last week at the Micro Alchimie in France, many people was impressed with this speed bump.
You can dislike the MMX addition, but the fact is, that RIVA is faster thanks those instructions. Period. It's the only thing that TuKo said on the first post, we didn't claim anything else.

Then why being so reluctant to publish that ammx code, even by PM ?

Originally Posted by flype

About the original 0.50 sources, RiVA was not free until 2014, but there was a demo with some limitations. Then it became free. But still closed-sources. Current status is i asked yesterday the opinion of the author about a public release. I know he is thinking of it and is open to. In itself this is good news.

I'm not asking you to share the whole code if you can't, but, at least, the parts that got rewritten - especially the ammx ones - would help.

The 68080 also has new stuff that's not ammx. What proves me it's ammx more than other stuff that could make Riva faster ? Only the code can.

And if you don't share the source, at least you can distribute the binary.

[wawa]

Originally Posted by meynaf

Creating instructions for a mere x2 isn't worth it. If it were x50 or even x10, i'd have said : wow. But x2 ?

These instructions will become obsolete as soon as another, larger simd comes out. But the cost of supporting these will still be here.
x86 is crippled with more than 1,000 instructions. I don't want the 68k to become like that.

Meynaf has a point here in my eyes. Instruction set extensions should be treated very carefully and with perspectivic foresight, rather than for lesser immediate advantage. 080 core doesnt need minor publicity stunts if they might block the future development or unneccesarily waste instruction space, it provesto be good enough to gather enough interest anyway. So please consider this for own good of the project.

Originally Posted by flype

Hi,

About the original 0.50 sources, RiVA was not free until 2014, but there was a demo with some limitations. Then it became free. But still closed-sources. Current status is i asked yesterday the opinion of the author about a public release. I know he is thinking of it and is open to. In itself this is good news.

Good to hear. Opening sources would be a good thing even if in asm. One way or the other, though, please avoid to release hardware dedicated binaries. Release one consistent update for all systems, handling extensions on itself. Even if it costs some time, effort and a bigger binary, avoid the versions hell. And please cooperate where possible.

[Zetr0]

@Tuko / Apollo Team

Impressive stuff as always! the last video running on an A2000 just blew me a way a bit! can I ask is the Audio going through the native 22KHz or through a dedicated sound card?


@Maynaf

Breathe my friend, breath.....

Implementation of a register set or SIMD in this case isn't anything to worry about - even if it only offloads enough giving a x2 performance - surely its better than nothing. Although I will say SIMD before FPU was a shock - but I am pretty sure it was because it was "easier" to implement, but I have worked with SIMD before on the PC's so there is a potential even it was stuck in this early stage / state.

Sometimes projects evolve outside of our influence, even if we have put so much in to begin. I am firm believer in that if a project is too far from where you want it to be - its time to start again. Fortunately there is a lot of ground work for a "new / improved" 68k ISA - (please include a FPU component or VPU - that would be nice)
I have been reading yours / Mattey's 68k ISA with much interest - I do hope you move to some testing of concepts soon =)

[flype]

Originally Posted by meynaf

I'm not asking you to share the whole code if you can't, but, at least, the parts that got rewritten - especially the ammx ones - would help.

The 68080 also has new stuff that's not ammx. What proves me it's ammx more than other stuff that could make Riva faster ? Only the code can.

And if you don't share the source, at least you can distribute the binary.

The decision to share the RiVA sources do not belong to your willing. This belongs to the author, and is for the whole community. We were thinking of this much before this discussion.

[wawa]

I dont think the discussion is about what anyone wants personally, but about how to progress best in general. I think the ammx set is now part of 080 set in stone. But i think informed discussion about such matters can still be benefitial, as it was with implementing the complete 68kinstruction set in hardware.

[flype]

Originally Posted by meynaf

Then why being so reluctant to publish that ammx code, even by PM

Reluctant to publish code ? An example was published. This thread was created 2 days ago and you speak like if you were waiting them for 6 months. TuKo presented results of our work, that's the point of this thread : being able to plays videos smoothly on one dedicated hardware.

The work is experimental and still WIP, just wait quietly until we release it. About SS optims, they are clearly not enough to obtain smooth playback with a Cyclone 3. Please, keep in mind this results are running on a very cheap FPGA, with limited ressources, not a GHz one. This is out of the point to tell legacy instructions are enough. You could be 200% right with unlimited MHz, this is not correct with limited MHz. Goal is to be able to let Vampire users have fun with their real HW, not with UAE.

[meynaf]

Originally Posted by flype

The decision to share the RiVA sources do not belong to your willing. This belongs to the author, and is for the whole community. We were thinking of this much before this discussion.

I don't think the ammx changes belong to the original author...

Originally Posted by flype

Reluctant to publish code ? An example was published. This thread was created 2 days ago and you speak like if you were waiting them for 6 months. TuKo presented results of our work, that's the point of this thread : being able to plays videos smoothly on one dedicated hardware.

Why advertising for a work in progress ? What does that bring ?
If it's too early for code/doc release, then there is no point coming here with benchmarks. The work must be finished first.

Originally Posted by flype

The work is experimental and still WIP, just wait quietly until we release it.

Again, if you want me to wait quietly you should have done the same and not start your ad campaign here.

Originally Posted by flype

About SS optims, they are clearly not enough to obtain smooth playback with a Cyclone 3. Please, keep in mind this results are running on a very cheap FPGA, with limited ressources, not a GHz one. This is out of the point to tell legacy instructions are enough. You could be 200% right with unlimited MHz, this is not correct with limited MHz. Goal is to be able to let Vampire users have fun with their real HW, not with UAE.

Hardware acceleration for video playback can be done in a better way than adding instructions.

[flype]

Originally Posted by meynaf

I don't think the ammx changes belong to the original author...

It belongs to the author(s) depending on the license form choice. So i respect first opinions of author(s) before telling irrelevant informations.

Originally Posted by meynaf

Again, if you want me to wait quietly you should have done the same and not start your ad campaign here.

Ah, now it is forbidden. Didn't knew that.

Originally Posted by meynaf

Hardware acceleration for video playback can be done in a better way than adding instructions.

Again, i told that a Cyclone have limited ressources, you continue to ignore obvious things.