CCminer(SP-MOD) Modded GPU kernels.

[joblo]

Compiling on Windows is a pain. I have to rebuild my compile environment every month because VS shuts down
unless I register. I had to create a virtual machine snapshot before installing VS the first time, otherwise the
tombstone from the previous install would trigger the forced registration imediately.

Compile with visual studio express 2013. It is free

VS community is advertised as free but it only works for a month without registering. The registration might be
free but I haven't tried it. I don't recall beig able to download VS Express, I'll take another look, thanks.

Got the test results on a EVGA 980 reference standard clocks.

                 76-7.5       76-6.5        74-6.5
quark        19.9          19.3           19.3
x11           9850         9920          10000
lyra2v2     10.7          11.4           11.6
neo           220           635            640

These results confirm the increase in quark is purely due to cuda 7.5. It also shows no degradation
in cuda 6.5 performance, a win-win for quark.

The neo degradation is also purely due to cuda 7.5 with no significant difference between r74 & r76
when compiled with cuda 6.5.

X11 is interesting, the 76-6.5 performance is lower than 74-6.5 and the 76-7.5 performance is lower still, a lose-lose.

Since quark was the focus of the most recent changes it proves that cuda 7.5 can perform better than 6.5. I hope
these results translate to the other algos.

[pallas]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.
Tested working on Linux. Here's the procedure:

- build ccminer with cuda 7.5 as usual
- remove all the object files in the neoscrypt folder: rm neoscrypt/*.o
- edit Makefile
- replace all the instances of "7.5" with "6.5"
- run make again
- you just made a ccminer executable with all the algos on 7.5 except neoscrypt on 6.5 :-)
- revert the Makefile changes to build it again in the future

If you find this useful, please donate to the BTC address in my signature.

[sp_]

Got the test results on a EVGA 980 reference standard clocks.

                 76-7.5       76-6.5        74-6.5
quark        19.9          19.3           19.3
x11           9850         9920          10000
lyra2v2     10.7          11.4           11.6
neo           220           635            640

Thanks for testing. Can you please try to compile release 74 with x86 build and cuda 7.5?

[sp_]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.

This is not a solution, this is a workaround.

x11 was hashing @ 1900KHASH in release 74 on the 750ti when building with cuda 7.5(x86) I modded it to go to 2940 in release 76, wich is a bit below release 74 buildt with cuda 6.5.

After changing the compiler version in the project, all the kernals need to be retuned.
I haven't finished yet.

My 5% increase in the quark hashrate has only given me a little more than 0.1 btc in donations.
Please donate some more guys. I need your support.

Here you can see my latest opensource work:

https://github.com/sp-hash/ccminer/commits/windows

...

Since cude 7.5 is increasing the register and stack usage in the table based AES algos, I reduced the number of threads per block to compensate. I also had to move some precalc tables into the instruction cache. (inlined in the instructions.) I also have reduced the codesize abit since cuda 7.5 seems to loop bether. etc..etc..

[joblo]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.
Tested working on Linux. Here's the procedure:

- build ccminer with cuda 7.5 as usual
- remove all the object files in the neoscrypt folder: rm neoscrypt/*.o
- edit Makefile
- replace all the instances of "7.5" with "6.5"
- run make again
- you just made a ccminer executable with all the algos on 7.5 except neoscrypt on 6.5 :-)
- revert the Makefile changes to build it again in the future

If you find this useful, please donate to the BTC address in my signature.

If I undertstand the result neo is compiled with 6.5 and eveyrthing else with 7.5. Then it is all linked
with 6.5. I'm not sure linking object files from different compilers is safe.

I prefer to use a script to select the prefered executable based on the algo. Less work, less risk, more
flexible.

Edit: but it's still a workaround.

[pallas]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.
Tested working on Linux. Here's the procedure:

- build ccminer with cuda 7.5 as usual
- remove all the object files in the neoscrypt folder: rm neoscrypt/*.o
- edit Makefile
- replace all the instances of "7.5" with "6.5"
- run make again
- you just made a ccminer executable with all the algos on 7.5 except neoscrypt on 6.5 :-)
- revert the Makefile changes to build it again in the future

If you find this useful, please donate to the BTC address in my signature.

If I undertstand the result neo is compiled with 6.5 and eveyrthing else with 7.5. Then it is all linked
with 6.5. I'm not sure linking object files from different compilers is safe.

I prefer to use a script to select the prefered executable based on the algo. Less work, less risk, more
flexible.

Edit: but it's still a workaround.

Linking object files from different compilers: I've often linked object files create with a C compiler and others created with an assembler.
In ccminer, some objects are compiled with gcc, some others with nvcc... you get the picture.

[sp_]

Since quark was the focus of the most recent changes it proves that cuda 7.5 can perform better than 6.5. I hope these results translate to the other algos.

I have showed that it can be done with quark.
I believe the other algos can be tuned faster as well with more work..

0.01 BTC guys. This is all I am asking

[pallas]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.

This is not a solution, this is a workaround.

I don't see a problem.
Most people already have both cuda 6.5 and 7.5.
With some little changes to the Makefile, you could compile each kernel with its best cuda version, in a single executable.

[sp_]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.

This is not a solution, this is a workaround.

I don't see a problem.
Most people already have both cuda 6.5 and 7.5.
With some little changes to the Makefile, you could compile each kernel with its best cuda version, in a single executable.

Don't you want to know if cuda7.5 can make your code faster?. A recoding is probobly required.

[pallas]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.

This is not a solution, this is a workaround.

I don't see a problem.
Most people already have both cuda 6.5 and 7.5.
With some little changes to the Makefile, you could compile each kernel with its best cuda version, in a single executable.

Don't you want to know if cuda7.5 can make your code faster?. A recoding is probobly required.

You can do both things.
On amd we've been doing this forever.

[sp_]

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-19805474
// Cuda compilation tools, release 7.5, V7.5.16
// Based on LLVM 3.4svn
//
cuda 6.5 is 1.5 years old.

1.5 year development of compiler technology.

I think we need this to reach the optimal kernal speed.. It worked in the quark kernal, and it should work in the others as well.

[joblo]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.
Tested working on Linux. Here's the procedure:

- build ccminer with cuda 7.5 as usual
- remove all the object files in the neoscrypt folder: rm neoscrypt/*.o
- edit Makefile
- replace all the instances of "7.5" with "6.5"
- run make again
- you just made a ccminer executable with all the algos on 7.5 except neoscrypt on 6.5 :-)
- revert the Makefile changes to build it again in the future

If you find this useful, please donate to the BTC address in my signature.

If I undertstand the result neo is compiled with 6.5 and eveyrthing else with 7.5. Then it is all linked
with 6.5. I'm not sure linking object files from different compilers is safe.

I prefer to use a script to select the prefered executable based on the algo. Less work, less risk, more
flexible.

Edit: but it's still a workaround.

Linking object files from different compilers: I've often linked object files create with a C compiler and others created with an assembler.
In ccminer, some objects are compiled with gcc, some others with nvcc... you get the picture.

I get the picture. I got a pascal program to call a fortran subroutine several decades ago. No available linker could handle it I had to
link it manually.

The examples you gave are where it is explicitly supported by the respective linkers. I wouldn't expect that support in different
major versions of the same compiler. ABI changes can be introduced.

I'm not saying it can't be done, just that it's probably not supported and might not always work.

[djm34]

Since quark was the focus of the most recent changes it proves that cuda 7.5 can perform better than 6.5. I hope these results translate to the other algos.

I have showed that it can be done with quark.
I believe the other algos can be tuned faster as well with more work..

0.01 BTC guys. This is all I am asking

so you need donation because you believe it can be done ?

[djm34]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.
Tested working on Linux. Here's the procedure:

- build ccminer with cuda 7.5 as usual
- remove all the object files in the neoscrypt folder: rm neoscrypt/*.o
- edit Makefile
- replace all the instances of "7.5" with "6.5"
- run make again
- you just made a ccminer executable with all the algos on 7.5 except neoscrypt on 6.5 :-)
- revert the Makefile changes to build it again in the future

If you find this useful, please donate to the BTC address in my signature.

If I undertstand the result neo is compiled with 6.5 and eveyrthing else with 7.5. Then it is all linked
with 6.5. I'm not sure linking object files from different compilers is safe.

I prefer to use a script to select the prefered executable based on the algo. Less work, less risk, more
flexible.

Edit: but it's still a workaround.

Linking object files from different compilers: I've often linked object files create with a C compiler and others created with an assembler.
In ccminer, some objects are compiled with gcc, some others with nvcc... you get the picture.

I get the picture. I got a pascal program to call a fortran subroutine several decades ago. No available linker could handle it I had to
link it manually.

The examples you gave are where it is explicitly supported by the respective linkers. I wouldn't expect that support in different
major versions of the same compiler. ABI changes can be introduced.

I'm not saying it can't be done, just that it's probably not supported and might not always work.

used for quite some time c++ programme linked to old fortran libraries, this was done automatically (well in a script).

Also regarding nvcc, there are different way to link and compile. You can very well compile one part or the other with one cuda version or another and do the linking with gcc (and I am not entirely sure, but I think for linking cuda is just calling the default compiler (gcc or visual stuff... so actually it doesn't really matter that you compile with various cuda version...

[joblo]

Got the test results on a EVGA 980 reference standard clocks.

                 76-7.5       76-6.5        74-6.5    74-7.5
quark        19.9          19.3           19.3        19.7
x11           9850         9920          10000     7680
lyra2v2     10.7          11.4           11.6        10.9
neo           220           635            640         220

Thanks for testing. Can you please try to compile release 74 with x86 build and cuda 7.5?

Done, results included above. The quark rate with 74-7.5 was unexpected, some of the previous changes
must have provided a bigger improvement on cuda 7.5 than 6.5

[scryptr]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.

This is not a solution, this is a workaround.

I don't see a problem.
Most people already have both cuda 6.5 and 7.5.
With some little changes to the Makefile, you could compile each kernel with its best cuda version, in a single executable.

PLEASE POST THE FRANKEN-WEENIE WORKAROUND--

If you could post a Windows executable with the promised features, I'll donate.  Quark is faster with release dot 76.  I want to use the executable with NiceHash Miner and auto-switch.  On Linux, I could compile it myself... but I don't have a working autoswitch script.       --scryptr

[pallas]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.

This is not a solution, this is a workaround.

I don't see a problem.
Most people already have both cuda 6.5 and 7.5.
With some little changes to the Makefile, you could compile each kernel with its best cuda version, in a single executable.

PLEASE POST THE FRANKEN-WEENIE WORKAROUND--

If you could post a Windows executable with the promised features, I'll donate.  Quark is faster with release dot 76.  I want to use the executable with NiceHash Miner and auto-switch.  On Linux, I could compile it myself... but I don't have a working autoswitch script.       --scryptr

I never built ccminer on windows. I don't even have a working windows environment with cuda and nvidia drivers. Sorry.
Maybe I could do a Makefile (or whatever) patch to make neoscrypt compile on 6.5 and post a pull request? That way sp_ can build the win binary for everybody. But I don't think he agrees on "the method".

[sp_]

Got the test results on a EVGA 980 reference standard clocks.
                 76-7.5       76-6.5        74-6.5    74-7.5
quark        19.9          19.3           19.3        19.7
x11           9850         9920          10000     7680
lyra2v2     10.7          11.4           11.6        10.9
neo           220           635            640         220

Thanks for testing. Can you please try to compile release 74 with x86 build and cuda 7.5?

Done, results included above. The quark rate with 74-7.5 was unexpected, some of the previous changes
must have provided a bigger improvement on cuda 7.5 than 6.5

My tuning rig doesn't contain a gtx 980. I only have 750ti, 960, 970x2. your results are different than mine in the quark algo.

As you can see I have optimized x11 30% faster(than release 74 compiled with cuda 7.5), but still slower than cuda 6.5.


(but only 4 of the 11 x11 kernals have been changed)

[scryptr]

MINERS WILL DONATE FOR FRANKEN-WEENIE--

No waiting. Work on the purist executable at your own pace.  Maybe DJM will flip out some useful code for neoscrypt.  Maybe we will all just get along.       --scryptr

[joblo]

I've found a solution to the neoscrypt problem: building a cuda 6.5/7.5 hybrid.
Tested working on Linux. Here's the procedure:

- build ccminer with cuda 7.5 as usual
- remove all the object files in the neoscrypt folder: rm neoscrypt/*.o
- edit Makefile
- replace all the instances of "7.5" with "6.5"
- run make again
- you just made a ccminer executable with all the algos on 7.5 except neoscrypt on 6.5 :-)
- revert the Makefile changes to build it again in the future

If you find this useful, please donate to the BTC address in my signature.

If I undertstand the result neo is compiled with 6.5 and eveyrthing else with 7.5. Then it is all linked
with 6.5. I'm not sure linking object files from different compilers is safe.

I prefer to use a script to select the prefered executable based on the algo. Less work, less risk, more
flexible.

Edit: but it's still a workaround.

Linking object files from different compilers: I've often linked object files create with a C compiler and others created with an assembler.
In ccminer, some objects are compiled with gcc, some others with nvcc... you get the picture.

I get the picture. I got a pascal program to call a fortran subroutine several decades ago. No available linker could handle it I had to
link it manually.

The examples you gave are where it is explicitly supported by the respective linkers. I wouldn't expect that support in different
major versions of the same compiler. ABI changes can be introduced.

I'm not saying it can't be done, just that it's probably not supported and might not always work.

used for quite some time c++ programme linked to old fortran libraries, this was done automatically (well in a script).

Also regarding nvcc, there are different way to link and compile. You can very well compile one part or the other with one cuda version or another and do the linking with gcc (and I am not entirely sure, but I think for linking cuda is just calling the default compiler (gcc or visual stuff... so actually it doesn't really matter that you compile with various cuda version...

Just for kicks I googled it. nvcc does do a compatibility check on g++ (in the example below) versions so they do explicitly support
it and there are versions that don't work.

http://stackoverflow.com/questions/9421108/how-can-i-compile-cuda-code-then-link-it-to-a-c-project

I don't know for sure whether nvcc supports different versions of itself, it may have checked and passed in your case.
If I was a compiler developper it would be a very low priority as it would complicate changing the ABI and is generally only useful
when the original source code is not available.