Gateless Gate Sharp 1.3.8: 30Mh/s (Ethash) on RX 480!

[zawawa]

Now my custom LLVM/Clang is capable of generating assembly source codes compatible with AMD's proprietary OpenCL drivers. The only remaining work is to add CLRX-specific stuff to LLVM. I really want to get this done by tomorrow so that I can go back to miner programming.

[h311m4n]

Hi Zawawa,

First off, thanks for this program, works well.

I'm using version 0.1.3-pre1 and on a 6 GPU rig with Sapphire RX470 8GB, the fans seem to go crazy after some time. Like RPM goes to 0 then goes full speed, then back to 0 and so on. I don't think all the GPUs are affected. Also it seems GG seems to hang at some point. The computer is still available and closing re-opening GG get's it going again.

I do not have this issue with a 4 GPU RX470 rig though.

Anyone else having a similar issue?

Thanks

[VaynardW]

I had the same exact issue at first, it also happens to one of my rig,
It turns out that rig still using Catalyst 16.9, after a DDU uninstall and upgrading to 16.11, the problem seems to disappear

[fr4nkthetank]

I had the same exact issue at first, it also happens to one of my rig,
It turns out that rig still using Catalyst 16.9, after a DDU uninstall and upgrading to 16.11, the problem seems to disappear

Yeah get the 16.11 drivers and I would recommend a custom fan profile with MSI afterburner.  Unless you are on linux I guess.  It gets the job done well.

[zawawa]

Building LLVM is taking forever, so I have to entertain myself with something else...
So I have been thinking for quite some time about adding a nice graphical user interface/remote monitor to Gateless Gate, and I think I want to go with MultiMiner as somebody suggested here a while ago:

https://github.com/nwoolls/MultiMiner


It's open-source, compatible with sgminer, and seems to be feature-complete with automatic currency-switching.
What do you guys think?

[zawawa]

Oh, I forgot to add that MultiMiner is cross-platform as well.
Actually, it seems to be working quite well with GG...

[nerdralph]

I finally tracked down the last bug and was able to confirm that LLVM/Clang can be used to generate binaries for AMD's proprietary OpenCL drivers. Since LLVM is only capable of generating ROCm binaries, I had to insert a thin ABI translation layer at the beginning of the kernel.

According to a source at AMD, the ROCm ABI is the same as the HSAIL/brig ABI.  When I used CodeXL to generate a kernel binary on a system with Crimson 16.10 drivers, the binary contained HSAIL as well as brig symbols.  Since AMDs drivers can handle HSA/brig kernels, I would think they should also handle ROCm binaries.

What I found out was that AMD's drivers add extra hidden arguments to kernels, which completely messes up the order of arguments. Therefore, modifications to LLVM are necessary in order to use it for AMD's proprietary drivers.

Interesting.  I started to reverse-engineer the fglrx drivers but only got part-way through it.  It appears most of the clang/llvm mods by AMD are open source, but it's rather difficult to figure out which branch with which patches are used in which driver release.  Here's an example of how AMD has been reasonably transparent about what they've been doing with llvm:
http://llvm.org/devmtg/2013-11/slides/Stellard-R600.pdf

[YIz]

Thanks for the good work zawawa and nerdalph, seems to be very close to Claymore's on my hardware and without the fee, but I'm unsuccessful in trying to make it start with Windows.

[zawawa]

Thanks for the good work zawawa and nerdalph, seems to be very close to Claymore's on my hardware and without the fee, but I'm unsuccessful in trying to make it start with Windows.

My crystal ball has been abused with assembly programming, so I need to know the following:

(1) your OS,
(2) the coin you tried to mine, and
(3) the display driver you are using.

[zawawa]

I finally tracked down the last bug and was able to confirm that LLVM/Clang can be used to generate binaries for AMD's proprietary OpenCL drivers. Since LLVM is only capable of generating ROCm binaries, I had to insert a thin ABI translation layer at the beginning of the kernel.

According to a source at AMD, the ROCm ABI is the same as the HSAIL/brig ABI.  When I used CodeXL to generate a kernel binary on a system with Crimson 16.10 drivers, the binary contained HSAIL as well as brig symbols.  Since AMDs drivers can handle HSA/brig kernels, I would think they should also handle ROCm binaries.

What I found out was that AMD's drivers add extra hidden arguments to kernels, which completely messes up the order of arguments. Therefore, modifications to LLVM are necessary in order to use it for AMD's proprietary drivers.

Interesting.  I started to reverse-engineer the fglrx drivers but only got part-way through it.  It appears most of the clang/llvm mods by AMD are open source, but it's rather difficult to figure out which branch with which patches are used in which driver release.  Here's an example of how AMD has been reasonably transparent about what they've been doing with llvm:
http://llvm.org/devmtg/2013-11/slides/Stellard-R600.pdf

I don't know... They just moved implicit arguments at the end of the argument list to to the beginning. This does not make any sense to me, and I did not find any traces or mentions of this seemingly arbitrary change. Things like this make me believe that open-source tools are crucial for GPGPU and AMD is not doing enough.

In any case, it doesn't matter to me because I'm almost done with my custom LLVM. Can't wait to actually use it...

[hannusolo]

Building LLVM is taking forever, so I have to entertain myself with something else...
So I have been thinking for quite some time about adding a nice graphical user interface/remote monitor to Gateless Gate, and I think I want to go with MultiMiner as somebody suggested here a while ago:

https://github.com/nwoolls/MultiMiner


It's open-source, compatible with sgminer, and seems to be feature-complete with automatic currency-switching.
What do you guys think?

This is a good idea, it would make it more user friendly, which would help the average people who might have a good gpu but is using it only for gaming and doesn't want to mess with command windows.

But I think you shouldn't bother with comfort functions like this only if you really have notjhng to do

[h311m4n]

I had the same exact issue at first, it also happens to one of my rig,
It turns out that rig still using Catalyst 16.9, after a DDU uninstall and upgrading to 16.11, the problem seems to disappear

Alright thanks for the tip!

[zawawa]

But I think you shouldn't bother with comfort functions like this only if you really have notjhng to do

I know, I know... I'm still debugging my custom LLVM, but it is coming along.
I stripped away most of the unnecessary targets and tools to make it relatively lightweight.
I am taking my time here because it is crucial for miner development to have inline GCN assembly and a set target as opposed to ever-changing AMD drivers.
This may seem like a detour, but it is actually a necessary step for me to continue miner development without losing sanity.

[jstefanop]

Seems like claymore is in the same spot...wouldnt be surprised if his recent "eureka" moment is from the discussion happening in this thread.

[zawawa]

Seems like claymore is in the same spot...wouldnt be surprised if his recent "eureka" moment is from the discussion happening in this thread.

Oh, I'm now pretty sure he reads this thread. It's really funny in an ironic sense, but it does make things a bit more challenging for me. It's like playing poker with you being the only player with open hands. Oh well.

I am not entirely sure if I would take his current approach, though. I'm not going to feed my competition this time around, so he has to figure things out himself...

[zawawa]

Hallelujah! My GCN compiler is finally working!
Well, it is a Frankenstein of LLVM/Clang and CLRadeonExtender, but as long as it works...

[m1n1ngP4d4w4n]

Hallelujah! My GCN compiler is finally working!
Well, it is a Frankenstein of LLVM/Clang and CLRadeonExtender, but as long as it works...

Good work man, now dazzle us   

[fr4nkthetank]

Hallelujah! My GCN compiler is finally working!
Well, it is a Frankenstein of LLVM/Clang and CLRadeonExtender, but as long as it works...

Cant wait to test out your next update !!!  I just can't get enough of em flashy colors

[xeridea]

Seems like claymore is in the same spot...wouldnt be surprised if his recent "eureka" moment is from the discussion happening in this thread.

Oh, I'm now pretty sure he reads this thread. It's really funny in an ironic sense, but it does make things a bit more challenging for me. It's like playing poker with you being the only player with open hands. Oh well.

I am not entirely sure if I would take his current approach, though. I'm not going to feed my competition this time around, so he has to figure things out himself...

I find these assumptions a bit odd, he has been fastest 95% of the time since launch, and people assume he uses methods from slower miners.  I think recent improvements are just to keep up with Optiminer, and timing is a coincidence.  He already had ASM working well many weeks ago for all cards.  I can't know for sure, I just find it odd that every time there is a CM speedup, some assume automatically it is from others.

Keep up the good work, just posting my thoughts.

[FFI2013]

Seems like claymore is in the same spot...wouldnt be surprised if his recent "eureka" moment is from the discussion happening in this thread.

Oh, I'm now pretty sure he reads this thread. It's really funny in an ironic sense, but it does make things a bit more challenging for me. It's like playing poker with you being the only player with open hands. Oh well.

I am not entirely sure if I would take his current approach, though. I'm not going to feed my competition this time around, so he has to figure things out himself...

I find these assumptions a bit odd, he has been fastest 95% of the time since launch, and people assume he uses methods from slower miners.  I think recent improvements are just to keep up with Optiminer, and timing is a coincidence.  He already had ASM working well many weeks ago for all cards.  I can't know for sure, I just find it odd that every time there is a CM speedup, some assume automatically it is from others.

Keep up the good work, just posting my thoughts.

yes but zawawa was talking asm than suddenly claymore implemented it that's just weird timing but could be a coincidence