mfakto: an OpenCL program for Mersenne prefactoring - Page 120

glebm · 2015-07-25, 22:26

Fixing int sign change warnings had no effect. Since GPU sieve works, I set GWDEBUG in gpusieve.cl and got lots of output like this:

Code:

sloppy_mod_p: p doesn't match pinv!! p = 713509, pinv = 6018 (should be 6019)

When running without any arguments (a simple selftest) I get:

Code:

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!

maxa · 2015-08-07, 04:09

When I try to run mfakto on a R9 Fury X it tells me:

Code:

 
Select device - Get device info - Loading binary kernel file mfakto_Kernels.elf
Compiling kernels.
WARNING: Unknown GPU name, assuming GCN. Please post the device name "Fiji (Advanced Micro Devices, Inc.)" to http://www.mersenneforum.org/showthread.php?t=15646 to have it added to mfakto. Set GPUType in mfakto1.ini to select a GPU type yourself to avoid this
 warning.
OpenCL device info
  name                      Fiji (Advanced Micro Devices, Inc.)
  device (driver) version   OpenCL 2.0 AMD-APP (1800.8) (1800.8 (VM))
  maximum threads per block 256
  maximum threads per grid  16777216
  number of multiprocessors 64 (4096 compute elements)
  clock rate                1050MHz
Automatic parameters
  threads per grid          256
  optimizing kernels for    GCN
Started a simple selftest ...
######### testcase 4/17 (M50863909[69-70]) #########
mfakto will exit once the current test is finished.

and then hangs in testcase 4/17.
Is there any setting I can change to make it work or does it require an updated version of mfakto?

Bdot · 2015-08-09, 19:23

Quote:

Originally Posted by glebm

The latest source from GitHub, compiled Visual Studio 2013 and App SDK 3.0 Beta running on Tonga (R9 380) fails both selftests (0.14) works.

Version 0.14 is currently the latest stable one. Many of the features for 0.15 are unfinished yet; the version from github cannot be used right now (and it's not because of the unsigned warnings).

I need to find some time to fix the new features ...

Bdot · 2015-08-09, 19:30

Quote:

Originally Posted by maxa

When I try to run mfakto on a R9 Fury X it tells me:

Code:

 
Select device - Get device info - Loading binary kernel file mfakto_Kernels.elf
Compiling kernels.
WARNING: Unknown GPU name, assuming GCN. Please post the device name "Fiji (Advanced Micro Devices, Inc.)" to http://www.mersenneforum.org/showthread.php?t=15646 to have it added to mfakto. Set GPUType in mfakto1.ini to select a GPU type yourself to avoid this
 warning.
OpenCL device info
  name                      Fiji (Advanced Micro Devices, Inc.)
  device (driver) version   OpenCL 2.0 AMD-APP (1800.8) (1800.8 (VM))
  maximum threads per block 256
  maximum threads per grid  16777216
  number of multiprocessors 64 (4096 compute elements)
  clock rate                1050MHz
Automatic parameters
  threads per grid          256
  optimizing kernels for    GCN
Started a simple selftest ...
######### testcase 4/17 (M50863909[69-70]) #########
mfakto will exit once the current test is finished.

and then hangs in testcase 4/17.
Is there any setting I can change to make it work or does it require an updated version of mfakto?

I'll add Fiji to the list of known GCN chips. It would be good to have some performance test so I know how to best use the chip, but I guess for that the other issue needs to be fixed first:

Could you please run

Code:

 mfakto -i mfakto1.ini -st2

this should show exactly at which kernel it hangs,

And maybe give the perftestmfakto.cmd from http://mersenneforum.org/mfakto/mfak...o-0.15pre5.zip a chance - if it does not stop right away, the output might be helpful.

Bdot · 2015-08-09, 19:54

Quote:

Originally Posted by Ethan (EO)

I've pulled the card now, but did manage to run the 0.15pre5 benchmarking script first; results attached.

Thank you for the HD6950 Cayman results.

Code:

Resulting speed for M66362159:
bit_min - bit_max  GHz-days/day  kernelname
     60 -      69       209.504  cl_barrett15_69_gs  
     69 -      70       196.994  cl_barrett15_71_gs  
     70 -      74       185.303  cl_barrett15_74_gs  
     74 -      77       162.719  cl_barrett32_77_gs  
     77 -      88       147.887  cl_barrett32_88_gs  
     88 -      92       117.277  cl_barrett32_92_gs

They confirm that the burnt electricity is probably not worth it - there are way faster and more efficient cards these days. The 205 GHz-days/day that are listed on http://www.mersenne.ca/mfaktc.php for the 6950 are probably a bit optimistic and can only be achieved when factoring up to 69 bits.

airsquirrels · 2015-08-22, 14:03

I am seeing something interested and unexpected with mfakto and multiple GPUs. On one much older host system, the PCIe links are reducing to 5GT/s x16 for the primary card and 5GT/s x4 for the secondary. What is surprising is this has a pretty massive effect on mfakto's performance. This is strange because mfakto should not need much bandwidth to the cards at all, and to my understanding of the code should only be making a (trivial) call for each class.

These are Fury X cards which normally get around 1000GhzDay/Day at 8GT/s x16 PCIe speeds.

I will see a few different tiers of speeds that directly correlate to negotiated PCIe speeds, all factoring using cl_barret15_73_gs_2.

PCIe Speed - GhzDay/day
8GT/s x16 - 1000
5GT/s x16 - 890
5GT/s x8 - 690
5GT/s x4 - 480
2.5GT/s x4 - 190

Increasing the number of streams did not seem to have any appreciable effect.

LaurV · 2015-08-22, 17:35

That is strange! I always have a mixture of x8 and x16 cards and don't see any difference in output for my gtx580s...
Are you sure you call mfaktX with appropriate "-d gpu_number"?

airsquirrels · 2015-08-22, 19:01

Yes, I can verify the correct card is initialized and the performance matches the PCIe lane speed even if the other card is idle. With my slower cards this is not noticeable until 5GT/4x or even 2.5GT/4x but the Fury X sees the penalty much earlier. I might try Windows instead of Debian to see about the Catalyst driver version.

I have switched the two cards and the performance definitely follows the slot, this is true in two different systems as well so that rules out the motherboard/chipset. I will say both of these systems are AMD CPUs, but my Intel systems are all much faster systems overall with plenty of PCIe lanes.

I did notice that we do actually do multiple kernel schedulings per class as it works through the K range, but adjusting the sieve size parameters changed performance an order of magnitude less than the PCIe slot.

chalsall · 2015-08-22, 19:21

Quote:

Originally Posted by airsquirrels

I have switched the two cards and the performance definitely follows the slot, this is true in two different systems as well so that rules out the motherboard/chipset.

"The most exciting phrase to hear in science, the one that heralds new discoveries, is not 'Eureka!' but 'That's funny...'" -- Isaac Asimov

axn · 2015-08-23, 05:53

Quote:

Originally Posted by airsquirrels

Yes, I can verify the correct card is initialized and the performance matches the PCIe lane speed even if the other card is idle. With my slower cards this is not noticeable until 5GT/4x or even 2.5GT/4x but the Fury X sees the penalty much earlier. I might try Windows instead of Debian to see about the Catalyst driver version.

I have switched the two cards and the performance definitely follows the slot, this is true in two different systems as well so that rules out the motherboard/chipset. I will say both of these systems are AMD CPUs, but my Intel systems are all much faster systems overall with plenty of PCIe lanes.

I did notice that we do actually do multiple kernel schedulings per class as it works through the K range, but adjusting the sieve size parameters changed performance an order of magnitude less than the PCIe slot.

I guess it doesn't hurt to ask. Are you using GPU Sieve or CPU Sieve?

LaurV · 2015-08-23, 14:37

Quote:

Originally Posted by axn

I guess it doesn't hurt to ask. Are you using GPU Sieve or CPU Sieve?

Bingo, good question! (as usual axn to the point!).

2015-07-25, 22:26	#1310
glebm "Gleb Mazovetskiy" Jul 2015 London, UK 2² Posts	Fixing int sign change warnings had no effect. Since GPU sieve works, I set GWDEBUG in gpusieve.cl and got lots of output like this: Code: sloppy_mod_p: p doesn't match pinv!! p = 713509, pinv = 6018 (should be 6019) When running without any arguments (a simple selftest) I get: Code: 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!

2015-08-22, 14:03	#1315
airsquirrels "David" Jul 2015 Ohio 11·47 Posts	I am seeing something interested and unexpected with mfakto and multiple GPUs. On one much older host system, the PCIe links are reducing to 5GT/s x16 for the primary card and 5GT/s x4 for the secondary. What is surprising is this has a pretty massive effect on mfakto's performance. This is strange because mfakto should not need much bandwidth to the cards at all, and to my understanding of the code should only be making a (trivial) call for each class. These are Fury X cards which normally get around 1000GhzDay/Day at 8GT/s x16 PCIe speeds. I will see a few different tiers of speeds that directly correlate to negotiated PCIe speeds, all factoring using cl_barret15_73_gs_2. PCIe Speed - GhzDay/day 8GT/s x16 - 1000 5GT/s x16 - 890 5GT/s x8 - 690 5GT/s x4 - 480 2.5GT/s x4 - 190 Increasing the number of streams did not seem to have any appreciable effect. Last fiddled with by airsquirrels on 2015-08-22 at 14:03

2015-08-22, 17:35	#1316
LaurV Romulan Interpreter Jun 2011 Thailand 3²·29·37 Posts	That is strange! I always have a mixture of x8 and x16 cards and don't see any difference in output for my gtx580s... Are you sure you call mfaktX with appropriate "-d gpu_number"? Last fiddled with by LaurV on 2015-08-22 at 17:36

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
gpuOwL: an OpenCL program for Mersenne primality testing	preda	GpuOwl	2718	2021-07-06 18:30
mfaktc: a CUDA program for Mersenne prefactoring	TheJudger	GPU Computing	3497	2021-06-05 12:27
LL with OpenCL	msft	GPU Computing	433	2019-06-23 21:11
OpenCL for FPGAs	TObject	GPU Computing	2	2013-10-12 21:09
Program to TF Mersenne numbers with more than 1 sextillion digits?	Stargate38	Factoring	24	2011-11-03 00:34

2015-08-22, 19:01	#1317
airsquirrels "David" Jul 2015 Ohio 1000000101₂ Posts	Yes, I can verify the correct card is initialized and the performance matches the PCIe lane speed even if the other card is idle. With my slower cards this is not noticeable until 5GT/4x or even 2.5GT/4x but the Fury X sees the penalty much earlier. I might try Windows instead of Debian to see about the Catalyst driver version. I have switched the two cards and the performance definitely follows the slot, this is true in two different systems as well so that rules out the motherboard/chipset. I will say both of these systems are AMD CPUs, but my Intel systems are all much faster systems overall with plenty of PCIe lanes. I did notice that we do actually do multiple kernel schedulings per class as it works through the K range, but adjusting the sieve size parameters changed performance an order of magnitude less than the PCIe slot.