Radeon RX 6800 Series Performance Comes Out Even Faster With Newest Linux Code

[lyamc]

Duby: "This is bad"
Me: "Why?"
Duby: "Look it up"
Me: "What does it say?"
Duby: "Look it up"

If you do not know how to talk about something in simple terms then it means you don't know what you're talking about. Both cb88 and bridgman were able to explain things to you. Why can't you do the same back?

[duby229]

Duby: "This is bad"
Me: "Why?"
Duby: "Look it up"
Me: "What does it say?"
Duby: "Look it up"

If you do not know how to talk about something in simple terms then it means you don't know what you're talking about. Both cb88 and bridgman were able to explain things to you. Why can't you do the same back?

Is it my fault that you refuse to hear me out? I don't think so.

https://www.psychologytoday.com/us/blog/ambigamy/201403/why-won-t-they-listen-reason

[AmericanLocomotive]

Hugely impressive performance from the open source driver. It's really been super impressive watching the AMD OSS driver dramatically improve in performance since AMDGPU was introduced.

EDIT: Have you ever seen generated code? If you have then you'd know exactly what I'm saying. There wouldn't be any need for further explanation.

"Generated Code" is hugely ambiguous, especially when talking about hardware drivers. Tons and tons of the code in the AMD drivers are massive tables setting up hardware registers. So of course there is a massive amount of generated code.

You haven't provided one piece of evidence supporting your argument that ROCm's code is "terrible". Telling people to "look it up" isn't a response to that, neither is looking up "Psychology Today" articles. You claim you don't have the time to look it up, but in the time you spend typing all of your replies, you could have easily and quickly scrolled through a chunk of the source, copied an excerpt and pointed out some things that were "bad".

[duby229]

Hugely impressive performance from the open source driver. It's really been super impressive watching the AMD OSS driver dramatically improve in performance since AMDGPU was introduced.

"Generated Code" is hugely ambiguous, especially when talking about hardware drivers. Tons and tons of the code in the AMD drivers are massive tables setting up hardware registers. So of course there is a massive amount of generated code.

You haven't provided one piece of evidence supporting your argument that ROCm's code is "terrible". Telling people to "look it up" isn't a response to that, neither is looking up "Psychology Today" articles. You claim you don't have the time to look it up, but in the time you spend typing all of your replies, you could have easily and quickly scrolled through a chunk of the source, copied an excerpt and pointed out some things that were "bad".

You got two things very wrong; First, this whole argument started out with me stating my opinions and nothing else. It is in fact my opinion that ROCm code sucks ass, I'm not gonna change that opinion until the day comes that it doesn't suck ass, but that day isn't gonna come, it has no community to fix it. And second, I have no intention of researching a synopsis that will piss me off and won't change a single damn thing. And besides, Clover -is- getting worked on, the day -will- come that Clover can be a functional alternative to ROCm at least for OpenCL. I don't have to even worry about it.

And I guess a third thing might be, telling people to look it up for themselves is almost -always- the right thing to do. Although if I had researched a synopsis it would help make it easier for those people to look it up I guess, still I have no intention of frustrating myself for no good reason.

I guess a fourth thing would be, If I did decide to research a synopsis it would take days to a week to develop a coherent logic, I'm not down for that. If you really want a synopsis that badly, Pay me a fair wage to do it...

(Also that link to Psychology Today about listening to reason, that was specifically meant for the person that I responded to)

[lyamc]

.

It is in fact my opinion

lol

[duby229]

.


lol

I'm entitled to my opinion. I've actually looked at ROCm and Clover. Have you? Are you entitled to your opinion?

[lyamc]

The difference is you treat your opinion like it’s a fact, hence why I laughed

[duby229]

The difference is you treat your opinion like it’s a fact, hence why I laughed

I think it is. But that's just my opinion.

[oleid]

I mean have you even looked at ROCm, like at all? Seriously, look at it.

EDIT: I urge everyone else reading this to do that same. If -ANYONE- thinks it's a project that they can contribute to in a meaningful way, please explain how.

I did debug tensorflow support for my GPU. Was an okay experience. The code, specifically MIOpen and rocBlas, is no piece of art, however, definitely not worse than in other open source projects. Same is true for mesa, by the way.

So what were you saying again?

[bridgman]

You say that Linux graphics components work pretty well without an open development model, except the components that do -ONLY- do because they are submitted to an upstream community project. The ones that aren't, like ROCm, don't...

Not quite - what I said was that the Linux graphics components work pretty well without integration between internal and public bug trackers. They do have an open development model already, and that is what we are trying to extend to the other ROCm components.

EDIT: Please don't misunderstand, I love AMD for their OSS efforts. It's just that ROCm is exactly the wrong way to do open source. You can throw code over the wall as long as there's a community on the other side, ROCm doesn't have a community and it never will.

If "throwing code over the wall" is not the problem then what is the problem ? My impression is that throwing code over the wall ends up being a community-killer even if one tries to form for a while.

Unless ROCm gets drivers for Intel and nVidia and becomes a vendor neutral upstream where OSS devs can collaborate then it will never have a community. Fortunately Mesa is real, it does have a community and that community is working on OpenCL.

ROCm (specifically the core HIP language) does have drivers and libraries for NVidia today - HIP programs can run over NVCC + CUDA libraries/drivers or over HIP/Clang + ROCm libraries/drivers.

HIP was designed to be vendor neutral and to run across AMD, NVidia and potentially Intel hardware.