mfakto: an OpenCL program for Mersenne prefactoring

[kracker]

Originally Posted by Bdot

Hmm. Weird. Now it is at 424 GHz ... I'll monitor it. Looks like something ran in the background ...

Wow... this is weird. I had reverted to a older driver(14.3) and it still gave me errors when I ran -st2 just now. Going to do a little investigating...

[kracker]

Hmm, I don't know what's up. -st2 with GCN2 works, and mfakto 0.14 works too...

[Bdot]

Originally Posted by kracker

Hmm, I don't know what's up. -st2 with GCN2 works, and mfakto 0.14 works too...

Now my GPU slowed down again. Seems to be the GPU memory access that almost blocks it. I still blame the driver.

Anyway, I now reproduced the error (I just needed to set GPUType=GCN ). I did not yet check why 0.14 does not show this, but the reason is that I missed a part in my error estimation. The 15-82, 15-83 and 15-88 bit kernels can have a higher error than I expected - so they cannot reliably do the last bitlevel they're enabled for. In other words, if you've done TF tests beyond 81 bits using a GCN GPU, there's a small chance of missed factors. Not sure how small, but most of the selftests do succeed. I'll try to find out about 0.14 tonight.

To fix it, I can either reduce the kernels' bitlevel, or add code to reduce the error, which will slow it down a bit. I have not yet decided.

[Bdot]

mfakto version 0.14 did not have these errors because the failing tests were part of another 2300 tests that I only added for 0.15. I built a version of 0.14 with just these tests added, and they are failing too.

This means that these tests were unreliable:

• run 81-82, 82-83 or 87-88 bits on any GCN GPU (other GPUs would select a different kernel)
• out of 262 test cases in these 3 ranges, 9 are failing (~3.5%). Due to the small number of tests, the error percentage is rather inexact.

Please do not use mfakto 0.14 for tests beyond 80 bits on GCN. I will reduce the bitlevels of these kernels to a safe level for 0.15 and add another kernel that has the additional code to safely run to 88 bits.

I'm sorry.

[LaurV]

Very good job Bdot!
(it does not affect me, due to the speed penalty at higher bitlevels, I used my 7970 higher than 72 bits only very seldom, but the effort to fix the tool and the recognizing of own slips, this scales very high in my book!)

[tului]

My Beema(A8-6410) takes a crap trying to run mfakto built from the github source. Is there a linux binary available or any tests and log files you'd like me to run on my builds?

[Bdot]

Originally Posted by tului

My Beema(A8-6410) takes a crap trying to run mfakto built from the github source. Is there a linux binary available or any tests and log files you'd like me to run on my builds?

Sorry to say, but I currently have no time to really help here, especially when the error description is as vage as "takes a crap" (I could not find this string in the source code).

[tului]

I was asking if there is any -st -st2 type stuff you'd like me to log and report back.Of course I can give you the exact error message as well. Wasn't thinking about that when I posted.

[TheMawn]

Originally Posted by Bdot

"takes a crap" (I could not find this string in the source code).

[tului]

Originally Posted by TheMawn

We could always do a git merge :-P

[frmky]

I tried playing with this a bit today, but I'm getting an error:

mfakto 0.15pre5 (64bit build)


Runtime options
  Inifile                   mfakto.ini
  Verbosity                 1
  SieveOnGPU                yes
  MoreClasses               yes
  GPUSievePrimes            81157
  GPUSieveProcessSize       24Ki bits
  GPUSieveSize              96Mi bits
  FlushInterval             0
  WorkFile                  worktodo.txt
  ResultsFile               results.txt
  Checkpoints               enabled
  CheckpointDelay           300s
  Stages                    enabled
  StopAfterFactor           class
  PrintMode                 compact
  V5UserID                  none
  ComputerID                none
  TimeStampInResults        yes
  VectorSize                2
  GPUType                   AUTO
  SmallExp                  no
  UseBinfile                mfakto_Kernels.elf
Compiletime options
  DEBUG_FACTOR_FIRST        enabled (DEBUG option)

Select device - Get device info:

OpenCL device info
  name                      Hawaii (Advanced Micro Devices, Inc.)
  device (driver) version   OpenCL 2.0 AMD-APP (1642.5) (1642.5 (VM))
  maximum threads per block 256
  maximum threads per grid  16777216
  number of multiprocessors 44 (2816 compute elements)
  clock rate                1000MHz

Automatic parameters
  threads per grid          0
  optimizing kernels for    GCN3

Loading binary kernel file mfakto_Kernels.elf
Compiling kernels.
Started a simple selftest ...
ERROR: selftest failed for M1031831 (cl_barrett15_69_gs)
  no factor found
Selftest statistics                                    
  number of tests           30
  successful tests          29
  no factor found           1

selftest FAILED!

Version 0.14 passes both -st and -st2. Is this the error mentioned earlier?