NSGminer v0.9.4: The Fastest NeoScrypt GPU Miner

[Conqueror]

Fundraising for Ghostlander's new GPU is gaining some traction.

Consider sending few FTC to his add: 733GhostxAXH9DEoTh2SzpcP9xh7CAeURF

Thank you.

[1714060294]

1714060294

[1714060294]

1714060294

[chrysophylax]

In future you should advice these coin devs on their algos cuz quite often they fail on their own.

A memory intensive algo being compute strained lol  

Quoted for emphasis. With all those rounds of salsa/chacha no idea how they managed to get different expectations... scrypt was already compute bound with GAP 2 and NeoScrypt is ~4 times more intensive!

However Scrypt wasn't compute bound without gapping, was it?

are there any plans on an nvidia ( cuda ) based miner? ...

just asking out of curiosity ...

#crysx

[ghostlander]

In future you should advice these coin devs on their algos cuz quite often they fail on their own.

A memory intensive algo being compute strained lol  

Quoted for emphasis. With all those rounds of salsa/chacha no idea how they managed to get different expectations... scrypt was already compute bound with GAP 2 and NeoScrypt is ~4 times more intensive!

However Scrypt wasn't compute bound without gapping, was it?

are there any plans on an nvidia ( cuda ) based miner? ...

just asking out of curiosity ...

#crysx

I'd like to return the question. If NSGminer with OpenCL starts to offer comparable performance to ccminer some day, will be a direct CUDA support necessary?

[chrysophylax]

In future you should advice these coin devs on their algos cuz quite often they fail on their own.

A memory intensive algo being compute strained lol  

Quoted for emphasis. With all those rounds of salsa/chacha no idea how they managed to get different expectations... scrypt was already compute bound with GAP 2 and NeoScrypt is ~4 times more intensive!

However Scrypt wasn't compute bound without gapping, was it?

are there any plans on an nvidia ( cuda ) based miner? ...

just asking out of curiosity ...

#crysx

I'd like to return the question. If NSGminer with OpenCL starts to offer comparable performance to ccminer some day, will be a direct CUDA support necessary?

of course not ... but that day will probably never come ...

due to nvidia being - well nvidia - the opencl 'market' will be pretty much behind the cuda 'market as far as cuda is concerned ... unless someone can come up with a direct way of implementing opencl 'properly' ( and i dont mean nvidia way of doing it ) - then we will always be stuck with a mediocre solution for opencl on nvidia cards ...

dont get me wrong - i think nvidia make massive improvements on their technologies on the cards ... but releasing opensource software for the benefit of the community - and not making money out of it? ... hehehe ... yup - thats going to happen from nvidia ...

so the cuda toolkit ( currently 7.5 ) is the closest anyone has of getting optimum performance from the nvidia based cards ...

so the question remains ...

nsgminer is opencl based and for the time being is NOT comparable to nvidia based cuda performance ... will there be such a performance gain ( whether opencl or cuda ) ? ...

#crysx

[ghostlander]

so the cuda toolkit ( currently 7.5 ) is the closest anyone has of getting optimum performance from the nvidia based cards ...

so the question remains ...

nsgminer is opencl based and for the time being is NOT comparable to nvidia based cuda performance ... will there be such a performance gain ( whether opencl or cuda ) ? ...

#crysx

CUDA allows inline PTX asm code, that's the most important advantage. Otherwise not much difference. Although their CUDA toolkits focus on the newest hardware only. 7.5 doesn't support anything older than Fermi, while OpenCL is supported even for the ancient 8800 series and can run NSGminer. CUDA support in NSGminer is possible. Depends on demand and community support.

[ghostlander]

so the cuda toolkit ( currently 7.5 ) is the closest anyone has of getting optimum performance from the nvidia based cards ...

so the question remains ...

nsgminer is opencl based and for the time being is NOT comparable to nvidia based cuda performance ... will there be such a performance gain ( whether opencl or cuda ) ? ...

#crysx

CUDA allows inline PTX asm code, that's the most important advantage. Otherwise not much difference. Although their CUDA toolkits focus on the newest hardware only. 7.5 doesn't support anything older than Fermi, while OpenCL is supported even for the ancient 8800 series and can run NSGminer. CUDA support in NSGminer is possible. Depends on demand and community support.

PTX isn't all that awesome. There's no Nvidia SASS assembler; PTX is too high level.

Now, with AMD...

Yeah, I know, it's still better than nothing.

-----

About 40K FTC received in donations, that's about 0.3 BTC. Keep'em coming

I’ve ordered a Gigabyte GTX 750 Ti (GV-N75TOC-2GI). Have got it today and testing now. I'm able to get 115KH/s at 1200MHz shaders or 133KH/s at 1400MHz shaders (~70% TDP). There are minor issues, but shares get solved just fine, no HW errors.

[Conqueror]

About 40K FTC received in donations, that's about 0.3 BTC. Keep'em coming

I’ve ordered a Gigabyte GTX 750 Ti (GV-N75TOC-2GI). Have got it today and testing now. I'm able to get 115KH/s at 1200MHz shaders or 133KH/s at 1400MHz shaders (~70% TDP). There are minor issues, but shares get solved just fine, no HW errors.

I suggest to point your miners to ghost's add for few days.
He doubled your hashrate 
733GhostxAXH9DEoTh2SzpcP9xh7CAeURF

[sp_]

About 40K FTC received in donations, that's about 0.3 BTC. Keep'em coming
I’ve ordered a Gigabyte GTX 750 Ti (GV-N75TOC-2GI). Have got it today and testing now. I'm able to get 115KH/s at 1200MHz shaders or 133KH/s at 1400MHz shaders (~70% TDP). There are minor issues, but shares get solved just fine, no HW errors.

I suggest to point your miners to ghost's add for few days.
He doubled your hashrate  
733GhostxAXH9DEoTh2SzpcP9xh7CAeURF

Good job. The 750ti does around 200khash @ 1200 and 240@1400 in djm34's opensource ccminer. the 980ti does 1MHASH.

[ghostlander]

Making progress.

GTX 750 Ti @ 1200 = 135KH/s, @ 1400 = 155KH/s

R9 280X @ 1000 = 480KH/s

A nice improvement for older Radeons: HD6970 @ 925 = 188KH/s

[chrysophylax]

Making progress.

GTX 750 Ti @ 1200 = 135KH/s, @ 1400 = 155KH/s

R9 280X @ 1000 = 480KH/s

A nice improvement for older Radeons: HD6970 @ 925 = 188KH/s

thats great to hear ...

looking forward to seeing it in action ...

#crysx

[KloNEM]

Hi gents,

I tried mining with CPU version (of NeoScrypt algo), it works like a charm. I'd like to mine also on GPU - but with OSS driver (radeon.ko), not these closed one (Catalyst). Of course I understand, that proprietary driver is (and probably will be forever) more efficient for these purposes, but a/ I'd like to see a progress within OSS drivers, b/ I don't want / cannot built into server closed drivers, etc. In general - would it be possible ? My system is Fedora 23, with latest packages (Mesa 11.1.0-2, kernel 3.4.3-300, etc.).

I grab latest NeoScrypt GPU Miner from git repo

Code:

git clone https://github.com/ghostlander/nsgminer

,  everything went fine.

Code:

./autogen.sh

was also without troubles :

Code:

------------------------------------------------------------------------
nsgminer 0.9.1
------------------------------------------------------------------------


Configuration Options Summary:

  curses TUI...........: FOUND: ncursesw5

  NeoScrypt............: Enabled
  Scrypt...............: Enabled

  OpenCL...............: Enabled
    ADL monitoring.....: Enabled

  BitForce FPGAs.......: Disabled
  Icarus FPGAs.........: Disabled
  ModMiner FPGAs.......: Disabled
  X6500 FPGAs..........: Disabled
  ZTEX FPGAs...........: Disabled
  libudev detection....: yes

...and the same for

Code:

make

. But when I run nsgminer :

Code:

./nsgminer --neoscrypt -g 2 -I 8 -o stratum+tcp://...

I got this error :

Code:

[05:52:21] Probing for an alive pool
[05:52:23] Error -11: Building Program (clBuildProgram)
[05:52:23] input.cl:21:2: error: "Device does not support unaligned stores"
input.cl:68:1: error: OpenCL does not support the 'static' storage class specifier
input.cl:74:1: error: OpenCL does not support the 'static' storage class specifier
input.cl:495:86: error: use of undeclared identifier 'MAX_GLOBAL_THREADS'
input.cl:513:11: warning: incompatible pointer types passing '__global ulong16 *' to parameter of type '__global uint16 *
'
input.cl:469:39: note: passing argument to parameter 'V' here

[05:52:23] Failed to init GPU thread 0, disabling device 0
[05:52:23] Restarting the GPU from the menu will not fix this.
[05:52:23] Try to restart the miner.

Is there any way how to fix this? Thanks for any suggestions!

[ghostlander]

Code:

./nsgminer --neoscrypt -g 2 -I 8 -o stratum+tcp://...

I got this error :

Code:

[05:52:21] Probing for an alive pool
[05:52:23] Error -11: Building Program (clBuildProgram)
[05:52:23] input.cl:21:2: error: "Device does not support unaligned stores"
input.cl:68:1: error: OpenCL does not support the 'static' storage class specifier
input.cl:74:1: error: OpenCL does not support the 'static' storage class specifier
input.cl:495:86: error: use of undeclared identifier 'MAX_GLOBAL_THREADS'
input.cl:513:11: warning: incompatible pointer types passing '__global ulong16 *' to parameter of type '__global uint16 *
'
input.cl:469:39: note: passing argument to parameter 'V' here

[05:52:23] Failed to init GPU thread 0, disabling device 0
[05:52:23] Restarting the GPU from the menu will not fix this.
[05:52:23] Try to restart the miner.

Is there any way how to fix this? Thanks for any suggestions!

That's interesting. I haven't tried it myself with an open source Radeon driver even though I support open source development in many ways.

First of all, -g 2 -I 8 is no good. Start with -g 1 -I 10. As far as I can tell, it allows to maintain desktop interactivity while doing most office tasks, watching online videos, etc. without a significant discomfort.

Second, would be kind enough to tell us that you attempt to use Wolf0's old kernel rather than my one bundled with NSGminer.

"Device does not support unaligned stores" refers to the cl_khr_byte_addressable_store extension disabled or missing. It's required for all released NeoScrypt kernels, though will be unnecessary for my upcoming v7 kernel. Try to enable this extension:

Code:

#pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable

Next, "error: OpenCL does not support the 'static' storage class specifier". It refers to the following code:

Code:

/* Initialisation vector */
static const __constant uint8 blake2s_IV4[1] = {
    (uint8)(0x6A09E667, 0xBB67AE85, 0x3C6EF372, 0xA54FF53A,
            0x510E527F, 0x9B05688C, 0x1F83D9AB, 0x5BE0CD19)
};

static const __constant uchar blake2s_sigma[10][16] = {
    {  0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15 } ,
    { 14, 10,  4,  8,  9, 15, 13,  6,  1, 12,  0,  2, 11,  7,  5,  3 } ,
    { 11,  8, 12,  0,  5,  2, 15, 13, 10, 14,  3,  6,  7,  1,  9,  4 } ,
    {  7,  9,  3,  1, 13, 12, 11, 14,  2,  6,  5, 10,  4,  0, 15,  8 } ,
    {  9,  0,  5,  7,  2,  4, 10, 15, 14,  1, 11, 12,  6,  8,  3, 13 } ,
    {  2, 12,  6, 10,  0, 11,  8,  3,  4, 13,  7,  5, 15, 14,  1,  9 } ,
    { 12,  5,  1, 15, 14, 13,  4, 10,  0,  7,  6,  3,  9,  2,  8, 11 } ,
    { 13, 11,  7, 14, 12,  1,  3,  9,  5,  0, 15,  4,  8,  6,  2, 10 } ,
    {  6, 15, 14,  9, 11,  3,  0,  8, 12,  2, 13,  7,  1,  4, 10,  5 } ,
    { 10,  2,  8,  4,  7,  6,  1,  5, 15, 11,  9, 14,  3, 12, 13 , 0 } ,
};

Well, this is a bug of your compiler actually. According to Khronos, Storage-class Qualifiers, static is allowed for global variables and constants. However you may remove it safely from the source code as const __constant is good enough to describe this data. The LLVM based AMD compiler doesn't care.

"error: use of undeclared identifier 'MAX_GLOBAL_THREADS' -- neither NSGminer nor my kernel use it.

The last warning may be disregarded.

[ghostlander]

More improvements.

GTX 750 Ti @ 1400 = 180KH/s

R9 280X @ 1000 = 500KH/s

Older Radeons boosted up again: HD6970 @ 925 = 255KH/s

[chrysophylax]

More improvements.

GTX 750 Ti @ 1400 = 180KH/s

R9 280X @ 1000 = 500KH/s

Older Radeons boosted up again: HD6970 @ 925 = 255KH/s

thats some nice improvements ...

checkout sp thread for ccminer ( which has djm34 neoscrypt included ) and see what you get from that? ...

djm34 released his neoscrypt kernel only recently mate ...

so a lot can be had ( and maybe even improved - though djm34 does a very thorough job ) from the kernel ...

it is cuda based - but im sure the opensource kernel could help in some ways ...

really impressed with your improvements ... when you decide you would like to get some work outside of here done - let me know ... i could use your help and optimizations with granite ...

tanx ...

#crysx

[ghostlander]

More improvements.

GTX 750 Ti @ 1400 = 180KH/s

R9 280X @ 1000 = 500KH/s

Older Radeons boosted up again: HD6970 @ 925 = 255KH/s

thats some nice improvements ...

checkout sp thread for ccminer ( which has djm34 neoscrypt included ) and see what you get from that? ...

djm34 released his neoscrypt kernel only recently mate ...

so a lot can be had ( and maybe even improved - though djm34 does a very thorough job ) from the kernel ...

it is cuda based - but im sure the opensource kernel could help in some ways ...

really impressed with your improvements ... when you decide you would like to get some work outside of here done - let me know ... i could use your help and optimizations with granite ...

tanx ...

#crysx

Couldn't even compile his ccminer for Windows XP where I have my 750 Ti running now. NVCC rejects MinGW and insists on M$ Visual Studio. VS2013 doesn't produce valid code even with vs120_xp target. Missing entry points in kernel32.dll like GetTickCount64 or InitializeCriticalSectionEx. So I had to strip ccminer down to NeoScrypt only and patch for VS2010 compatibility.

[chrysophylax]

More improvements.

GTX 750 Ti @ 1400 = 180KH/s

R9 280X @ 1000 = 500KH/s

Older Radeons boosted up again: HD6970 @ 925 = 255KH/s

thats some nice improvements ...

checkout sp thread for ccminer ( which has djm34 neoscrypt included ) and see what you get from that? ...

djm34 released his neoscrypt kernel only recently mate ...

so a lot can be had ( and maybe even improved - though djm34 does a very thorough job ) from the kernel ...

it is cuda based - but im sure the opensource kernel could help in some ways ...

really impressed with your improvements ... when you decide you would like to get some work outside of here done - let me know ... i could use your help and optimizations with granite ...

tanx ...

#crysx

Couldn't even compile his ccminer for Windows XP where I have my 750 Ti running now. NVCC rejects MinGW and insists on M$ Visual Studio. VS2013 doesn't produce valid code even with vs120_xp target. Missing entry points in kernel32.dll like GetTickCount64 or InitializeCriticalSectionEx. So I had to strip ccminer down to NeoScrypt only and patch for VS2010 compatibility.

damn i like your style ...

and the resultant strip down and test? ...

btw - its one of the many reasons why i dont use windows ... too messy with compilations ...

#crysx

[ghostlander]

Couldn't even compile his ccminer for Windows XP where I have my 750 Ti running now. NVCC rejects MinGW and insists on M$ Visual Studio. VS2013 doesn't produce valid code even with vs120_xp target. Missing entry points in kernel32.dll like GetTickCount64 or InitializeCriticalSectionEx. So I had to strip ccminer down to NeoScrypt only and patch for VS2010 compatibility.

damn i like your style ...

and the resultant strip down and test? ...

btw - its one of the many reasons why i dont use windows ... too messy with compilations ...

#crysx

GTX 750 Ti @ 1400 = 285KH/s with CUDA 6.5 or 280KH/s with CUDA 7.5

Although I've somehow broken Stratum in the process, so it's solo mining through Getwork now.

[chrysophylax]

Couldn't even compile his ccminer for Windows XP where I have my 750 Ti running now. NVCC rejects MinGW and insists on M$ Visual Studio. VS2013 doesn't produce valid code even with vs120_xp target. Missing entry points in kernel32.dll like GetTickCount64 or InitializeCriticalSectionEx. So I had to strip ccminer down to NeoScrypt only and patch for VS2010 compatibility.

damn i like your style ...

and the resultant strip down and test? ...

btw - its one of the many reasons why i dont use windows ... too messy with compilations ...

#crysx

GTX 750 Ti @ 1400 = 285KH/s with CUDA 6.5 or 280KH/s with CUDA 7.5

Although I've somehow broken Stratum in the process, so it's solo mining through Getwork now.

i actually get higher with c75 than with c65 - when compile under fedora 23 x64 ...

i dunno what it is at the moment- but if you want to know - i can pull a miner off decred for testing neoscrypt if you like ...

but without stratum - its a pretty mess on getwork only ...

but you seem to be making leaps with this ...

btw - i dont oc anything ... just factory clocks ... my cards are gigabyte 750ti oc lp ( non powered ) ...

#crysx

[KloNEM]

Hi ghostlander,

many thanks for your help, I appreciate it! Unfortunately I'm still not running (see comments below), but I believe there will be the "right way", how to do it

That's interesting. I haven't tried it myself with an open source Radeon driver even though I support open source development in many ways.

First of all, -g 2 -I 8 is no good. Start with -g 1 -I 10. As far as I can tell, it allows to maintain desktop interactivity while doing most office tasks, watching online videos, etc. without a significant discomfort.

Yes, of course, it was just an example, in "production environment" I'll change these values to something "work & real".

Second, would be kind enough to tell us that you attempt to use Wolf0's old kernel rather than my one bundled with NSGminer.

"Device does not support unaligned stores" refers to the cl_khr_byte_addressable_store extension disabled or missing. It's required for all released NeoScrypt kernels, though will be unnecessary for my upcoming v7 kernel. Try to enable this extension:

Code:

#pragma OPENCL EXTENSION cl_khr_byte_addressable_store : enable

I added this line everywhere where I found that line :

Code:

#pragma OPENCL EXTENSION cl_amd_media_ops : enable

In fact I found it in four source files : diablo.cl, diakgcn.cl, phatk.cl, poclbm.cl.

But after a compilation, this error is still present...

Next, "error: OpenCL does not support the 'static' storage class specifier". It refers to the following code:

Code:

/* Initialisation vector */
static const __constant uint8 blake2s_IV4[1] = {
    (uint8)(0x6A09E667, 0xBB67AE85, 0x3C6EF372, 0xA54FF53A,
            0x510E527F, 0x9B05688C, 0x1F83D9AB, 0x5BE0CD19)
};

static const __constant uchar blake2s_sigma[10][16] = {
    {  0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15 } ,
    { 14, 10,  4,  8,  9, 15, 13,  6,  1, 12,  0,  2, 11,  7,  5,  3 } ,
    { 11,  8, 12,  0,  5,  2, 15, 13, 10, 14,  3,  6,  7,  1,  9,  4 } ,
    {  7,  9,  3,  1, 13, 12, 11, 14,  2,  6,  5, 10,  4,  0, 15,  8 } ,
    {  9,  0,  5,  7,  2,  4, 10, 15, 14,  1, 11, 12,  6,  8,  3, 13 } ,
    {  2, 12,  6, 10,  0, 11,  8,  3,  4, 13,  7,  5, 15, 14,  1,  9 } ,
    { 12,  5,  1, 15, 14, 13,  4, 10,  0,  7,  6,  3,  9,  2,  8, 11 } ,
    { 13, 11,  7, 14, 12,  1,  3,  9,  5,  0, 15,  4,  8,  6,  2, 10 } ,
    {  6, 15, 14,  9, 11,  3,  0,  8, 12,  2, 13,  7,  1,  4, 10,  5 } ,
    { 10,  2,  8,  4,  7,  6,  1,  5, 15, 11,  9, 14,  3, 12, 13 , 0 } ,
};

Well, this is a bug of your compiler actually. According to Khronos, Storage-class Qualifiers, static is allowed for global variables and constants. However you may remove it safely from the source code as const __constant is good enough to describe this data. The LLVM based AMD compiler doesn't care.

I tried to remove this code from neoscrypt.cl with two variants - remove both paragraphs, and remove just the second one (beginning with ...

Code:

static const __constant uchar blake2s_sigma[10][16] = {

), but after make procedures and run ./nsgminer, that error is still persists too...

"error: use of undeclared identifier 'MAX_GLOBAL_THREADS' -- neither NSGminer nor my kernel use it.

The last warning may be disregarded.

Yep, I believe this error is just a minor issue, when I'll be able to solve the previous two, this would be "a quick action" - I hope!


Many thanks again for your help and co-operation!

[ghostlander]

cl_amd_media_ops is for bitalign/bytealign mostly which are not used in v6 directly. The compiler is supposed to take care of this, but it doesn't do well in the drivers newer than 14.7. It won't be an issue in the next release.