[ANN][BURST] Burst | Efficient HDD Mining | New 1.2.3 Fork block 92000

Hi everyone,

After many hours of setup I finally made it. I have a 1Tb generation in progress and 3x100Gb already finished.
I would like to test the V2 pool but I haven't any BURST for now. Could someone send me 1 BURST to test it please ? Here is my address : BURST-YA29-QCEW-QXC3-BKXDL.

Regarding the plot generation, I found an OpenCL implementation of Shabal (https://github.com/aznboy84/X15GPU/blob/master/kernel/shabal.cl) that could be used to make a GPU version of the generator. I will try to work on it when I have some free time.

Regards

Hi everyone,

As promised I have been working on a GPU plot generator on the last few days. I made a little program built on top of OpenCL, and it seems to work pretty well in CPU mode. Unfortunately, I can't test the GPU mode as it requires a very powerfull graphic card (with at least 46kB private memory per compute unit, because the algorithm needs at least 4096*64 static bytes to store an entire plot).

Here is a preview you can test for now :
gpuPlotGenerator-src-1.0.0.7z : https://mega.co.nz/#!bcF2yKKL!3Ud86GaibgvwBehoxkbO4UNdiBgsaixRx7ksHrgNbDI
gpuPlotGenerator-bin-win-x86-1.0.0.7z : https://mega.co.nz/#!HJsziTCK!UmAMoEHQ3z34R4RsXoIkYo9rYd4LnFtO_pw-R4KObJs

I will build another release in the end of the day with some minor improvements (threads per compute unit selection, output of OpenCL error codes, improvement of the Makefile to generate the distribution directly).
I will also try to figure out another mean to dispatch the work between the GPU threads to reduce the amount of private memory needed by the program.

For the windows people, you can use the binary version directly.
For the linux people, just download the source archive, make sure to modify the OpenCL library and lib path in the makefile (and maybe the executable name), and build the project via "make". To run the program, you need the "kernel" and the "plots" directories beside the executable.

The executable usage is : ./gpuPlotGenerator <address> <start nonce> <nonces> <stagger size>
The parameters are the same as the original plot generator, without the threads number.

If you find bugs or if you want some new features, let me now.

If you want to support me, here are my Bitcoin and Burst addresses :
Bitcoin: 138gMBhCrNkbaiTCmUhP9HLU9xwn5QKZgD
Burst: BURST-YA29-QCEW-QXC3-BKXDL

Regards

Quote from: bipben on September 08, 2014, 09:02:10 AM

Quote from: burstcoin on September 08, 2014, 08:43:32 AM

Quote from: bipben on September 08, 2014, 08:20:35 AM

Unfortunately, I can't test the GPU mode as it requires a very powerfull graphic card (with at least 46kB private memory per compute unit, because the algorithm needs at least 4096*64 static bytes to store an entire plot).

It's nice to see someone else working on this, since I seem to have failed in it.

Private memory is actually part of global on AMD cards, so storing it in private isn't any better than just using global for everything; it's local that needs to aimed for for the massive speedup. No AMD cards have more than 64KB local per workgroup, which makes storing it all in local impossible however.

I haven't tried your implementation yet, but on my own first attempt, I also used global on everything also, and the result was faster than the java plotter, but slower than dcct's c plotter. My 2nd attempt used a 32KB local buffer I rotated through for storing the currently being hashed stuff, however I couldn't figure out how to get it copied also to global fast enough, and the local -> global copy killed the performance.

You might be interested in those kernels here: https://bitcointalk.org/index.php?topic=731923.msg8695829#msg8695829

Thanks, I will look at your kernels to see if I can find a better solution.

Here is the new version. I reduced the amount of memory used from 40KB to about 1KB per unit. The only drawback is that it requires twice the global memory as before. I will search a mean to reduce this overhead later.
In CPU mode, it all goes pretty well (when no graphic card is detected).
The GPU mode is still kind of buggy on my graphic card (an old GeForce 9300M GS), don't know the exact reason yet. Sometimes it works, sometimes not. I will try to fix this issue tomorrow.

Here are the files :
gpuPlotGenerator-src-1.1.0.7z : https://mega.co.nz/#!iYFWAL5B!BvtmRQ5qGq4gGwjDglFNtDtNIX4LDaUvATBtClBdTlQ
gpuPlotGenerator-bin-win-x86-1.1.0.7z : https://mega.co.nz/#!aBVGBBQD!tBsRtb8VrHR12_anrFTrl41U0fPQu_OqFnxyi5nCyBY

For the linux users, the Makefile has a new target named "dist" that builds and copy/paste all the necessary files to the "bin" directory.

The executable usage is : ./gpuPlotGenerator <path> <address> <start nonce> <nonces> <stagger size> <threads>
<path> : the path to the plots directory
<threads> : number of parrallel threads for each work group

Found the "randomness" cause. NVIDIA is caching the kernel after the first build and rebuild it from time to time. By cleaning the cache, I can force the kernel build and speed up the debugging process.
I will notify you as soon as the crash cause is found and corrected.

Bad news guys. There is no actual "bug" in the implementation. Seems like the graphic card is beeing streesed too much by the shabal core, thus the driver is shutting down the kernel (there is a watch-dog timer for this purpose hard coded in the display driver to ensure that the display don't freeze too much). I will try to improve the whole algorithm and memory consumption to the needed graphic card power.

In the meantime, I found this thread (http://stackoverflow.com/questions/12259044/limitations-of-work-item-load-in-gpu-cuda-opencl) that speak about this particular issue. The available options are :
- If you have more than one graphic card, you can launch the plotter on the one that does not hold the display. There is still no option to select the graphic card in the plotter, but I will code it soon so that you can test it in a multi-GPU environment.
- You can try to turn-off the watchdog timer by following the provided link, but be CAREFUL, you may experience terrible display lags, or even full black screens until the plotter process finishes its work.

You don't need to improve it to avoid this issue, just split it. One kernel for first half, one kernel for second half.

The new major update is in progress, thanks to burstcoin advice. I think that a lot more of graphic cards will be compatible with this version (at least I hope so).

I'm a Linux n00b, but I managed to get Ubuntu up and running so that I could use dcct's plot generator.
...I'm currently trying to use your v1.1.0 of the GPU Plot Generator with my AMD (ATI) Radeon 7800.
I've edited/made the following changes to the "Makefile" file (I chose the default install location for AMDAPPSDK):

Quote

CC = g++
CC_FLAGS = -ansi -pedantic -W -Wall -std=c++0x -O3 -I../opt/AMDAPPSDK-2.9-1/include
LD = g++
LD_FLAGS = -fPIC -L../opt/AMDAPPSDK-2.9-1/lib/x86_64 -static-libgcc -static-libstdc++ -lopencl

I keep getting this error now when I issue the "make" command:

Quote

Linking [bin/gpuPlotGenerator.exe]
/usr/bin/ld: cannot find -lopencl
collect2: error: ld returned 1 exit status
make: *** [bin/gpuPlotGenerator.exe] Error 1

I've tried Googling the error and see that a bunch of posts on the issue, but can't seem to correct this "/usr/bin/ld: cannot find -lopencl" error. Any suggestions?