That's some truly slick updates!
I was indeed planning to do full AES round without t-tables as the amount of masks are nonsensical.
I had the impression the SALU was immensely updated for Tonga given it takes much more VGPRs on the analyzer.
I wonder how to trick the CL compiler in emitting this code.
But most importantly, what are they waiting to just make an AMD_GCN_swizzle extension!?
Topic: [ANN][GRS][DMD][DGB] Pallas optimized groestl opencl kernels - page 5. (Read 61242 times)
Hi,
Have you checked the new GCN3 ISA manual? http://amd-dev.wpengine.netdna-cdn.com/wordpress/media/2013/07/AMD_GCN3_Instruction_Set_Architecture.pdf
It has some really useful things like:
- Bytepermute (no more shifts and masks)
- VOP_DPP: It actually does 2 ds_swizzle in the instruction in no time, so optimizing a single thread for 4 lanes costs no more cycles.
- VOP_SDWA: access a word or a byte in the 32bit inputs and in the output too. (again: no more shifts and masks)
- S alu can write memory
No 3 op add, and 3 op bitwise, though.
And they altered some instruction encodings, so I guess my asm will crash on GCN3 immediately.
Have you checked the new GCN3 ISA manual? http://amd-dev.wpengine.netdna-cdn.com/wordpress/media/2013/07/AMD_GCN3_Instruction_Set_Architecture.pdf
It has some really useful things like:
- Bytepermute (no more shifts and masks)
- VOP_DPP: It actually does 2 ds_swizzle in the instruction in no time, so optimizing a single thread for 4 lanes costs no more cycles.
- VOP_SDWA: access a word or a byte in the 32bit inputs and in the output too. (again: no more shifts and masks)
- S alu can write memory
No 3 op add, and 3 op bitwise, though.
And they altered some instruction encodings, so I guess my asm will crash on GCN3 immediately.
Good work on the groest. Smolens quark miner does around 2 mhash on the 280x.
My gtx 980 does 20mhash. The competition is sleeping...
My gtx 980 does 20mhash. The competition is sleeping...
I think that just applying some well known tricks, already available on public kernels, will bring quark hashrate to around 10.
Thing is, it's not funny. Optimizing single kernel algos is much more interesting, imho.
I agree about the easy Quark speedups. Doing this was fun, doing the foundation for Quark would be simply boring. Sure, it would get fun later, when working on one algo at a time, but to get there...
Just wanted to say I've tried applying some of the tricks I learnt working on whirlpoolx to the groestl kernel, but it's not so simple.
This kernel is much bigger in size so you can't just copy some good lines of code and it runs faster. Furthermore some of the optimizations I made in the past, make it more time consuming to apply some apparently simple hacks. Wolf0 I'm sure you know what I mean ;-)
Still there is room for improvement, I have some ideas, but the question is: when the profit is gone, and the fun is gone, is it still worth?
Another trick, not for speed, but for cleaning the code - when you want to postpone sboxing of byte, put preimage of zero (0x81 in Whirlpool) there. This kernel is much bigger in size so you can't just copy some good lines of code and it runs faster. Furthermore some of the optimizations I made in the past, make it more time consuming to apply some apparently simple hacks. Wolf0 I'm sure you know what I mean ;-)
Still there is room for improvement, I have some ideas, but the question is: when the profit is gone, and the fun is gone, is it still worth?
I get 26MHs on 280x mining groestl however I have quit groestl mining of DMD for the moment till diff drops back into the teens. For some reason ASM kernel crashes 7950 within a few minutes ... I am mining nneoscrypt on yaamp at present and also selling neo on westhash 
Buying more DMD than I used to mine direct
Will see what happens in next week or so as miners drop like flies on DMD ...
Buying more DMD than I used to mine direct
Will see what happens in next week or so as miners drop like flies on DMD ... @wolf0 do you have anything better than the neocrypt kernel u leaked on feathercoin thread? I am getting 278KHs on 7950 and 295 on 280x
I didn't leak that, I released it. Checking my records...
EDIT: Okay, most recent record of Neoscrypt I have is 12/23/2014 (NSFW): https://ottrbutt.com/miner/neoscryptwolf-12232014.png
Same goes for groestl Pallas U are my heroes And realhet who understands AMD GPU coding better than all of us
realhet hetpas assembly kernel still best for 280x and other Tahiti cards AFAIK Nope, I have 21MH/s out of a 7950 at 1125/1250, IIRC, using OpenCL.
Looking ... 1160/1500
I have modded card a bit for better cooling Any chance of getting your latest OCL source to try on 280x (Hawaii)
I assume you meant Tahiti.
I've acquired a 280x myself: it's not worth using v2 on it, hashrate is lower than with v1.
Doh, yeah Tahiti 2 wavefronts not possible?
Both me and Wolf0 tried that and (at least for me) stopped trying after a while. Funny no longer ;-)
Just did it.
2 waves in flight on Tahiti, for 21MH/s on a 7950 @ 1125/1250. Screenshot (NSFW): https://ottrbutt.com/miner/groestlcoinwolf-04082015.png
Just wanted to say I've tried applying some of the tricks I learnt working on whirlpoolx to the groestl kernel, but it's not so simple.
This kernel is much bigger in size so you can't just copy some good lines of code and it runs faster. Furthermore some of the optimizations I made in the past, make it more time consuming to apply some apparently simple hacks. Wolf0 I'm sure you know what I mean ;-)
Still there is room for improvement, I have some ideas, but the question is: when the profit is gone, and the fun is gone, is it still worth?
This kernel is much bigger in size so you can't just copy some good lines of code and it runs faster. Furthermore some of the optimizations I made in the past, make it more time consuming to apply some apparently simple hacks. Wolf0 I'm sure you know what I mean ;-)
Still there is room for improvement, I have some ideas, but the question is: when the profit is gone, and the fun is gone, is it still worth?
I expect DMD to drop into low teens difficulty after a week or so
If it does not mining is dead LOL. I have a direct interest in this as a partner on donkypool ... 12 miners up from 6 a few weeks ago ... I am currently mining neoscrypt for sale on westhash lol and p=4.8 selling anything less goes to yaamp ... Any chance of getting your latest OCL source to try on 280x (Hawaii)
I assume you meant Tahiti.
I've acquired a 280x myself: it's not worth using v2 on it, hashrate is lower than with v1.
Doh, yeah Tahiti 2 wavefronts not possible?
Both me and Wolf0 tried that and (at least for me) stopped trying after a while. Funny no longer ;-)
I think I've beaten your ASM with pure OpenCL on 290X.
Some of your last tips (and smolen's) can be applied to this kernel as well, I think it can reach 38/40 Mh/s ;-)
Possibly - but turns out, it didn't reach the speed I thought it did; it's still slightly under. Damn.
Just wanted to say I've tried applying some of the tricks I learnt working on whirlpoolx to the groestl kernel, but it's not so simple.
This kernel is much bigger in size so you can't just copy some good lines of code and it runs faster. Furthermore some of the optimizations I made in the past, make it more time consuming to apply some apparently simple hacks. Wolf0 I'm sure you know what I mean ;-)
Still there is room for improvement, I have some ideas, but the question is: when the profit is gone, and the fun is gone, is it still worth?
This kernel is much bigger in size so you can't just copy some good lines of code and it runs faster. Furthermore some of the optimizations I made in the past, make it more time consuming to apply some apparently simple hacks. Wolf0 I'm sure you know what I mean ;-)
Still there is room for improvement, I have some ideas, but the question is: when the profit is gone, and the fun is gone, is it still worth?
@wolf0 do you have anything better than the neocrypt kernel u leaked on feathercoin thread? I am getting 278KHs on 7950 and 295 on 280x
I didn't leak that, I released it. Checking my records...
EDIT: Okay, most recent record of Neoscrypt I have is 12/23/2014 (NSFW): https://ottrbutt.com/miner/neoscryptwolf-12232014.png
Same goes for groestl Pallas U are my heroes And realhet who understands AMD GPU coding better than all of us
realhet hetpas assembly kernel still best for 280x and other Tahiti cards AFAIK Any chance of getting your latest OCL source to try on 280x (Hawaii)
I assume you meant Tahiti.
I've acquired a 280x myself: it's not worth using v2 on it, hashrate is lower than with v1.
Doh, yeah Tahiti 2 wavefronts not possible?
Both me and Wolf0 tried that and (at least for me) stopped trying after a while. Funny no longer ;-)
I think I've beaten your ASM with pure OpenCL on 290X.
@wolf0 do you have anything better than the neocrypt kernel u leaked on feathercoin thread? I am getting 278KHs on 7950 and 295 on 280x
Pallas, can you rewrite this groesl-256 implementation to a groestl-512 and add it to sgminer (x11,x13,x15).?
Sorry for the delay.
That would be nice, but everybody's using wolf0's binaries, so why? It would make sense if there is a plan to opensource optimized versions of most of the algos.
Wolf0 claims to know aes from the inside backwords and forwards. Me too.
The answer is SEA :-)
Yes, that makes the game damn addictive.The answer is SEA :-)
Look, you told us about wide tables, great idea, but to skip sboxing with it couple more inches deeper inside AES is needed
Some of competitors are awake, taking exercises with pen and paper to get all AES-wannabees at once
Wolf0 claims to know aes from the inside backwords and forwards. Me too.
The answer is SEA :-)
Pallas, can you rewrite this groesl-256 implementation to a groestl-512 and add it to sgminer (x11,x13,x15).?
The need for an improved goestl kernel is now immediate ... please do what u can ... I am just C, C++ coder and am not fully into multi thread GPU coding ...
Hmm... it's one hell of a lot harder than I anticipated to lose two goddamned VGPRs than I thought it'd be.
Have you rotated table values left by 3 bits? 
Not sure it will help with register usage through...Rotations seem to hurt reg usage a bit. The source REALLY needs cleaning, but IMO, it's rather well done code by Pallas. I'm not really used to seeing anyone with a semblance of clue doing AMD miners.
Now I've put some parts of the code (ex. the list of rbtts) in pragma unrolled for loops and it looks much better ;-)
Nice - now, I haven't tried this, so the OpenCL compiler may mangle the shit out of it (unpredictable little fucker) - but you have vector types. They look a lot nicer than for loops.
That's about as far as my parse got before I went, "Is that a fucking NULL pointer dereference?"
Yes Indexed address is calculated in bitselect. LUT0 and LUT4 indexing is just single AND operation.EDIT: Oh, wait UINT8 is byte, not int vector. I probably went too far redefining every type
Okay... I'm guessing that you've removed bits from the tables and are regenerating them on the fly, but I can't quite figure out how. Then again, bitwise ops aren't really my best subject...
Maybe not: people are using wolf0's precompiled x11 binaries, just adding your trick to stock kernels will not come close to them speed-wise.
Jump to: