AMD have released a tool which converts Nvidia's CUDA code to portable C++ code which works on both Nvidia and AMD cards!

[goodtofufriday]

1 minute ago, LAwLz said:

Isn't this really old? Here is a thread about it from over a year ago.

(Also a fun read sicne it involves a certain person refusing to give sources to his claims)

I remember that too. I think then it was being worked on. Here they've actually released it to the public.

[SpaceGhostC2C]

4 minutes ago, AlwaysFSX said:

No, if a company fairly and legally becomes the only one in the market they're free to do as they wish within reason and will not be broken up.

 

Edit: typo, been up a while

The US could also force a company to split. See for example the case of Bell.

[Misanthrope]

Well at the end of the day it won't help that much. Despite their claims I think there's gonna be at least a bit of significant performance impact. AMD is probably banking on the fact that they're quite stronger on compute than similar Nvidia cards and probably just eat this performance and still come out on top or on par but ultimately, developers would actually have a significant advantage with the same card with OpenCL instead.

 

Just speculation though no need to call the engineering LTT team of one to post several PDF documents on how I am wrong okay?

[dalekphalm]

3 hours ago, Helly said:

This is fully legal because AMD obtained a license quite a while ago. I think this is the first thing they've done with it.

 

Click! just over a year ago :). Or was that wrong and not an actual license?

 

 

 

3 hours ago, SpaceGhostC2C said:

That would not make sense, since your code is yours. Remember, this isn't translating anything Nvidia has created into something else. It only translates code you write in CUDA language into something else. It is impossible for it to be illegal. It would mean that if you had an idea for a program ans write it in C, then you can write a CUDA version, but if you write the CUDA version first, then it is illegal to write a C version of your own program. Even if Nvidia would include such thing somewhere in its license agreements, it would be void.

There are no legal issues with translating code form one syntax to another.

 

3 hours ago, ivan134 said:

Uhhhh, yes they do.

 

I would like to point out that there was (in that thread posted last year) NO PROOF that AMD acquired a CUDA license, or even needed one, As @SpaceGhostC2C says, you own your own code that you write - this is pretty much universal IP law (With an exception being some employment contracts will take ownership if created during working hours, etc).

 

NVIDIA does not own the code you write, even if that code is in CUDA. They can put it in a contract or TOS all they want - but that contract would be illegal and would be thrown out if challenged.

 

AMD doesn't need a CUDA license to make a code translator, because it's taking the code YOU OWN and translating it into C (whatever version of C is used - C++?) - Patrick claimed in that other thread that AMD got a license, but could never substantiate that claim or back it up with any facts or sources.

[samcool55]

There is a nice example of it btw:

So to keep it simple, if company X is running a program that requires cuda, and wants to go AMD, they can do that now because they can port over the program without any issues. At least if they have the binary code/source code (don't know if that's the same?)

 

Best thing is that if you have the program in c++ code, you can compile it for AMD or nvidia, doesn't matter, that's up to you.

[Belgarathian]

AMD/RTG, I love you.

 

 

[Zodiark1593]

4 hours ago, AlwaysFSX said:

Isn't that illegal without AMD having a license?

Had they reverse engineered without insider knowledge, I'd assume this to be completely legal. 

 

Were this the case, and performance + stability are good, I'd have to start kicking myself a bit for getting a GTX 960. 

[dalekphalm]

Just now, Zodiark1593 said:

Had they reverse engineered without insider knowledge, I'd assume this to be completely legal. 

 

Were this the case, and performance + stability are good, I'd have to start kicking myself a bit for getting a GTX 960. 

Honestly you don't even need to reverse engineer anything to make this work. Any competent programmer could take their CUDA source code and port that code over to C++ or whatever. It's just time consuming and manpower intensive, which is why a lot don't bother.

[Yamoto42]

@dalekphalm and @SpaceGhostC2C, I;m no legal expert, but it still may not be that cut and dry.

 

Although it has been resolved for now, with CUDA being an API it could potentially fall under the same legal category as Oracle vs. Google in whether or not the API itself could be subject to copyright.  Granted the courts have, for now, ruled (with my personal opinion) that it cannot be copyrighted.

Remember even in that case, JVM and Dalvik compile to different bytecode.  The API consists of "expected behavior" and method signatures.  If the decision were overturned (deeming the CUDA API copyright-able) then any third party compiling CUDA specific source code may be guilty of infringement.  The important part here being (as it is an API) the language the developers used to write it, not the compiler or library implementation.

 

To bring it back to the Oracle v Google example: People writing android apps were not liable.  Google was for interpreting the code...


That's far fetched...but worth a moment to consider.

[dalekphalm]

3 minutes ago, Yamoto42 said:

@dalekphalm and @SpaceGhostC2C, I;m no legal expert, but it still may not be that cut and dry.

 

Although it has been resolved for now, with CUDA being an API it could potentially fall under the same legal category as Oracle vs. Google in whether or not the API itself could be subject to copyright.  Granted the courts have, for now, ruled (with my personal opinion) that it cannot be copyrighted.

Remember even in that case, JVM and Dalvik compile to different bytecode.  The API consists of "expected behavior" and method signatures.  If the decision were overturned (deeming the CUDA API copyright-able) then any third party compiling CUDA specific source code may be guilty of infringement.  The important part here being (as it is an API) the language the developers used to write it, not the compiler or library implementation.

 

To bring it back to the Oracle v Google example: People writing android apps were not liable.  Google was for interpreting the code...


That's far fetched...but worth a moment to consider.

That would be very unlikely - but hey, if you can patent rounded edges (even temporarily), then anything is possible.

[Prysin]

4 hours ago, AlwaysFSX said:

Isn't that illegal without AMD having a license?

no. Because AMD isnt the one running or modifying the code. The tool is used by people who already owns the license. All AMD allowed them to do, is to "quickly" rewrite the code they already wrote themselves.

 

unlike Nvidia who tried to make that X86 translator in hardware. And that is something else

[Yamoto42]

3 minutes ago, dalekphalm said:

That would be very unlikely - but hey, if you can patent rounded edges (even temporarily), then anything is possible.

That's why I tried to use an abundance of probabilistic quantifiers, such as "if", "may", and "potentially".


In this case it would probably against Nvidia's best interest to pursue it, even if it had a relatively sound legal foundation.  

[dalekphalm]

1 minute ago, Nicholatian said:

Can someone explain to me what all of the fuss is about? Why does CUDA matter so much when things like OpenCL exist?

For a few reasons:

1. Whether we want to believe it or not, CUDA seems to be a more efficient GPGPU Compute language, which leads into:

2. CUDA is pretty much the "go-to" Compute language, and finally

3. CUDA is exclusive to NVIDIA - and they have not given AMD licenses required to compute CUDA code on AMD hardware

 

So AMD has found a good work around: A code translator that takes raw CUDA source code, and converts that code into raw C++ source code, which can then be compiled into an OpenCL application/service/whatever.

 

AMD hardware doesn't touch or run the CUDA code at all. The AMD translator takes the raw cuda code and turns it into C++ code, which is then compiled and can run on AMD hardware. Since the author of the code is the legal rightful owner of said code, they do not need permission from NVIDIA (or anyone else) to translate their code into other languages.

[Kobathor]

47 minutes ago, dalekphalm said:

They can put it in a contract or TOS all they want - but that contract would be illegal and would be thrown out if challenged

unless you live in South Korea where Terms of Service agreements for code developers are being backed by law  

[SpaceGhostC2C]

22 minutes ago, Yamoto42 said:

@dalekphalm and @SpaceGhostC2C, I;m no legal expert, but it still may not be that cut and dry.

 

Although it has been resolved for now, with CUDA being an API it could potentially fall under the same legal category as Oracle vs. Google in whether or not the API itself could be subject to copyright.  Granted the courts have, for now, ruled (with my personal opinion) that it cannot be copyrighted.

Remember even in that case, JVM and Dalvik compile to different bytecode.  The API consists of "expected behavior" and method signatures.  If the decision were overturned (deeming the CUDA API copyright-able) then any third party compiling CUDA specific source code may be guilty of infringement.  The important part here being (as it is an API) the language the developers used to write it, not the compiler or library implementation.

 

To bring it back to the Oracle v Google example: People writing android apps were not liable.  Google was for interpreting the code...


That's far fetched...but worth a moment to consider.

I really don't know enough either to guess where a judge could draw the line. But a translator is essentially a text replacement tool: it reads text from a (user-made) file, and creates another file with different text. I could create a text that generates a text file by replacing the word "subroutine" for the word "turnip". It would be useless, but possible. I don't think you can prohibit the tool on the basis of whether the resulting file after all he text replacements is usable or not (i.e., it can be compiled in another language).

The way I see it, if that was possible then you could in principle also prohibit winrar or similar from compressing files containing CUDA source code if you so wanted to, since that's also a tool that reads the original file and generates a different, usable file.

Could Tolkien's heirs sue you for making a tool that reads a file in Sindarin and creates a file in English?  

[Sauron]

4 hours ago, AlwaysFSX said:

Isn't that illegal without AMD having a license?

No, it's just converting text to different text.

[Yamoto42]

@SpaceGhostC2C Again, noting my admission that this is unlikely:

With an API being the signature AND behavior...

 

The infringing situation would be:

Developer writes code that calls function X to do action Y with result Z.

Developer runs code through tool chain to create executable program.

Executable program is run and outputs result Z.

 

Agian, although Andriod apps are written in Java, they compile to Dvalik, not Java bytecode.  It is about allowing a developer accomplish a specific task by using a specific phrase with specific inputs.  Note the three times I used the word "specific".

 

About WinRar, no.  It reads the signature, but does not induce the behavior in it's output.  The concept of parsing is a significant note here...especially since it's likely to be far more than just simple renaming.  winrar just sees a byte sequence, a parser acts.

 

Translating Sindarin? ...actually (assuming languages to be copyright-able) yes...but for different reasons.  In this case there is no analogous concept to the "behavior" of a program.  To quantify the behavior of programs you really have to go back to Church and Turing; think of a program as a computational machine, which is a concept human languages lack.

[SpaceGhostC2C]

2 minutes ago, Yamoto42 said:

 

About WinRar, no.  It reads the signature, but does not induce the behavior in it's output.  The concept of parsing is a significant note here...especially since it's likely to be far more than just simple renaming.  winrar just sees a byte sequence, a parser acts.

 

Translating Sindarin? ...actually (assuming languages to be copyright-able) yes...but for different reasons.  In this case there is no analogous concept to the "behavior" of a program.  To quantify the behavior of programs you really have to go back to Church and Turing; think of a program as a computational machine, which is a concept human languages lack.

But here is where I don't see it. A translator (not a compiler) creates text from text, not function. winrar creates a rar, the translation creates source code, the soruce code does nothing. There is the potential for behavior if you then compile it. That's why I compared it to languages: the behavior would be reading, but the program just creates the English (or C++) text.

All the translator software does is to automate something that you can legally do by hand: as a developer you can open your own (could be someone else's, as long as you have the right to read it) CUDA source code, open another text file, and type code doing the same in a different language. A code that automatically does the syntax substitutions you would manually do still isn't a compiler, and doesn't create executable files from the CUDA source code.

[cj09beira]

5 hours ago, TheNuzziNuzz said:

There's always a way...

they have.

[Yamoto42]

DISCLAIMER FOR PEOPLE JUST READING THIS POST:  PURELY HYPOTHETICAL SITUATION.

13 minutes ago, SpaceGhostC2C said:

A translator (not a compiler)

That distinction doesn't exist.  A compiler IS a translator...or more specifically (in a typical implementation) multiple translators chained together.  Side note:  compiler design is a pretty fun topic...
 

This becomes a problem of token recognition:  Does a party X have permission to mechanically identify and take a defined action on token Y that leads to result Z.

 

WinRAR makes no effort to identify Y or produce result Z.
A programmer manually re-writing code works towards result Z independently of token Y.

 

But an AUTOMATED translation process depends on token Y and is not only aware of, but actively attempts to [upon detecting Y] lead to result Z...a translator isn't any good if it's output doesn't give the same result.

 

An THAT is the relation to Oracle v Google, can someone own "say Y to do Z".

[SpaceGhostC2C]

1 minute ago, Yamoto42 said:

 

That distinction doesn't exist.  A compiler IS a translator...or more specifically (in a typical implementation) multiple translators chained together.  Side note:  compiler design is a pretty fun topic...

I still don't see it. Before we were discussing function, and to me the difference between translating (still source code) and compiling (executable output) is a functional one. A compiler produces output with a different functionality than the source code, while a translation keeps the same functionality.

 

Quote

This becomes a problem of token recognition:  Does a party X have permission to mechanically identify and take a defined action on token Y that leads to result Z.
 
WinRAR makes no effort to identify Y or produce result Z.
A programmer manually re-writing code works towards result Z independently of token Y.
 
But an AUTOMATED translation process depends on token Y and is not only aware of, but actively attempts to [upon detecting Y] lead to result Z...a translator isn't any good if it's output doesn't give the same result.

 

Again, you may be right, but I have problem seeing that: since source code can be saved in a text file, there is a one-to-one mapping between certain bits (as winrar reads) and the lines of code being translated (then you would keep even arbitrary spaces and line breaks). I see the human programmer re-writing the code "independently", but that's just one possible case: the programmer could also mechanically read lines from one file and write corresponding lines in another file.

Imagine the following situation: I start a translation service, telling people they can send me their CUDA code and I'll provide C++ versions. Would it be legal? Would that depend on whether I'm doing it manually or i have automated the process somehow? How much of the process can I automate?

 

And to be clear, I'm not saying you are wrong, because I don't really know what would happen in court, and I take your point that it could end up being grayer than I thought. I just can't think of a law that would lead to it being illegal and that i would agree to at the same time  

[ThinkWithPortals]

6 hours ago, AlwaysFSX said:

Isn't that illegal without AMD having a license?

I doubt AMD would have knowingly done this illegally. That would be both financial and PR suicide for them, they can't be that stupid.

[patrickjp93]

6 hours ago, Joe_MacDougall said:

They would force NVIDIA to split into two companies which would probably hurt them a lot more. At least in the UK that would happen because of Competition laws.

No it wouldn't. Intel would just buy up Radeon in the fallout.

[patrickjp93]

40 minutes ago, ThinkWithPortals said:

I doubt AMD would have knowingly done this illegally. That would be both financial and PR suicide for them, they can't be that stupid.

Barcelona and Mullins say hi.

[Joe_MacDougall]

Just now, patrickjp93 said:

No it wouldn't. Intel would just buy up Radeon in the fallout.

True. I mean they are putting Radeon tech in new Intel CPUs apparently.