mfakto: an OpenCL program for Mersenne prefactoring

[aheeffer]

Originally Posted by aheeffer

Thanks! Your post made me realize I could make some use of the idle APU of my Ryzen5 3400g box which is running with two Radeon VII cards. To my great surprise Mfacto 015pre7 runs between 420 and 480 Ghz-d/day (depending on the factor size) on the Vega 11 APU of the 3400G! Not bad for a 140 € cpu.

I was wrong. Mfacto was stealing some cycles from the Radeon VII card. The gpu id's seem to be different from what gpuowl is using. And I can't get it to work with '-d 2'.

[kriesel]

Originally Posted by aheeffer

I can't get it to work with '-d 2'.

Mfakto device callout is a little different. Read the readme or https://www.mersenneforum.org/showpo...58&postcount=1

Originally Posted by DrobinsonPE

I assume that the >>selftest.txt sends the output to a text file so I will try that

>filename is redirect (overwriting whatever might have been in the destination before)
>>filename is redirect via append, preserving any previous selftest results, for example.
Those both affect stdout, while stderr still goes to the console.
(There are a few cases in GIMPS software where part of an error message is to stdout, and part to stderr!)
It's possible to redirect both into the same file, although chronological sequence (phasing between stderr and stdout)is not preserved, in my experience. Or each to different files. https://stackoverflow.com/questions/...a-file#7526988
https://stackoverflow.com/questions/...-a-single-file

Tee in Windows 7 Powershell lacks append capability; Windows 10 has it as an option.

[DrobinsonPE]

Originally Posted by kriesel

Is there a pattern to which are not found? Certain bit levels?
I had one with issues above ~81 bits. It's usable lower.

Self test results only logged up to "testcase 2110/34071" and stopped. There were 35 errors logged. The errors are all in the 60 to 64 range, examples of the errors are below.

no factor for M800007823 from 2^60 to 2^61 [mfakto 0.15pre7-MGW cl_barrett15_69_gs_2]
ERROR: selftest failed for M800007823 (cl_barrett15_69_gs)

no factor for M332193203 from 2^61 to 2^62 [mfakto 0.15pre7-MGW cl_barrett15_69_gs_2]
ERROR: selftest failed for M332193203 (cl_barrett15_69_gs)

no factor for M60008387 from 2^62 to 2^63 [mfakto 0.15pre7-MGW cl_barrett15_69_gs_2]
ERROR: selftest failed for M60008387 (cl_barrett15_69_gs)

no factor for M57078799 from 2^63 to 2^64 [mfakto 0.15pre7-MGW cl_barrett15_69_gs_2]
ERROR: selftest failed for M57078799 (cl_barrett15_69_gs)

I just will not run exponents in the 60-69 bit range.

[aheeffer]

Originally Posted by kriesel

Mfakto device callout is a little different. Read the readme or https://www.mersenneforum.org/showpo...58&postcount=1

Thanks again! After some trial and error '-d 12' worked for me, 1 being the platform number for AMD and 2 being the third device. The APU of the Ryzen 5 3400 G now burns 208 Ghz-d/days with mfacto-0.15.

[kriesel]

Originally Posted by DrobinsonPE

For the last almost 2 weeks I have been running the windows compiled version of mfakto 0.15pre7 on the iGPU of a Ryzen 3 3200G. it passed multiple runs of mfacto -st so I started feeding it a few GPU72 TF assignments and a few PrimeNet manual assignments.

Depending on the exponent I have seen between 126 and 132 GHz-d/day. I have not tried tuning any of the ini file settings yet. I also installed the windows compiled version of gpuowl and got it working but it is slow so I stopped it. I was seeing 18.1 ms/itt for a 111M exponent.

The above numbers are with the CPU idling. I just loaded Prime 95 and started running it but do not have any firm numbers yet.

These numbers are pretty good for an integrated graphics processor. Gpuowl timing is more than half the speed of an RX480, and 2.7 times as fast as an RX550. TF throughput ~125% of an RX550.

[DrobinsonPE]

Originally Posted by kriesel

These numbers are pretty good for an integrated graphics processor. Gpuowl timing is more than half the speed of an RX480, and 2.7 times as fast as an RX550. TF throughput ~125% of an RX550.

Is there any other testing you want me to try? For the last month I have been running with Prime 95 and mfakto but have not used GPUOWL since the initial testing.

Here are my notes on running Prime 95 and mfakto together:

Prime 95 running alone - 6.09ms/Itt @57M
Prime 95 with mfakto - 6.36ms/Itt @57M, 11.7ms/itt @98.3M, (4% drop in output when mfakto starts)

Mfakto running alone - TF level 74-76, 126 to 134 GHz-d/day, TF level 77, 122 to 123GHz-d/day
Mfakto with prime 95 - TF level 74-76, 104 to 110 GHz-d/day, TF level 77, 100-102 GHz-d/day (17% drop in output when Prime 95 starts)

In a little while I am going to swap in the Linux hard drive and try again to get mfakto and GPUOWL working on Linux. Before that I still plan to try tuning mfakto settings to see if any settings change the speed. I also want to get more data on GPUOWL.

[kriesel]

Your spreadsheet summary was impressive in its detail. Plain text is good too, more people see that.
I suggest close comparisons of performance, Windows vs. linux for same or very similar inputs, on same hardware.

Tune mfakto on each OS and compare performance and tune values.

[DrobinsonPE]

Originally Posted by kriesel

Tune mfakto on each OS and compare performance and tune values.

Here is the tuning information for mfakto on Windows 10. I will run this for a week or two and then I will swap hard drives and try to get mfakto working on Linux.

mfakto tuning.
exp=101129087 bit 75 to 76

Step 1, very GPUSieveProcessSize 
# Possible values: 8, 16, 24, 32
# Also must divide GPUSieveSize * 1024
# Default: GPUSieveProcessSize=24

GPUSieveProcessSize=8      125.92GHz-day
GPUSieveProcessSize=16    127.74GHz-day
GPUSieveProcessSize=24    129.15GHz-day
GPUSieveProcessSize=32    129.36GHz-day  *

Step 2: very GPUSieveSize with GPUSieveProcessSize=32
# Minimum: GPUSieveSize=4
# Maximum: GPUSieveSize=128
# Default: GPUSieveSize=96

GPUSieveSize=32     129.16GHz-day
GPUSieveSize=64     129.27GHz-day
GPUSieveSize=96     129.36GHz-day  *
GPUSieveSize=128   129.34GHz-day

Step 3: very GPUSievePrimes with GPUSieveSize=96, GPUSieveProcessSize=32
# Minimum: GPUSievePrimes=54
# Maximum: GPUSievePrimes=1075766
# Default: GPUSievePrimes=81157

GPUSievePrimes=21814      119.08GHz-day
GPUSievePrimes=67894      128.29GHz-day
GPUSievePrimes=81157      129.36GHz-day
GPUSievePrimes=99894      130.30GHz-day
GPUSievePrimes=120374    130.75GHz-day
GPUSievePrimes=139830    131.02GHz-day  *
GPUSievePrimes=160310    131.00GHz-day

[LaurV]

Nice job DrobinsonPE!
Even better would be if you specify what toy are you testing, in each post, beside of the OS, so the mortals like us who didn't follow the discussion from the start won't need to scroll 4 pages (but it was a nice reading through all those posts, very well documented!).

[DrobinsonPE]

Originally Posted by LaurV

Nice job DrobinsonPE!
Even better would be if you specify what toy are you testing, in each post, beside of the OS

Thank you. It has been a fun learning experience. Good suggestion on putting the computer information on every post. This computers particulars are: Asrock DeskMini A300, AMD Ryzen 3 3200G, 16GB RAM, 256GB SSD.

I think I have done as much testing as I can do on Windows. After tuning mfakto running alone, I ran some TFs in mfakto with prime95 also running. With prime 95 running, mfakto went from 106-107GHz-day before tuning to 108-109GHz-day after tuning. I am currently running a 58M LL-D with gpuowl that will take about 10 days with prime 95 also running. I have completed many TF assignments and a few PM-1 assignments but with all the mfakto/gpuowl testing I never actually completed a LL/PRP assignment on the iGPU. I posted all of my gpuowl results for this computer in the gpuowl thread.

I have another computer with a GTX 1650 Super in it and I just spent some time tuning mfaktc 0.21. I will be posting those results in the mfaktc thread with the computer information included in the post.

[DrobinsonPE]

For everyone waiting for someone to post RTX3080 results....This is not the post you are looking for. I got bored this morning and tuned mfakto for UHD 600 graphics. Here is the short version of the results.

J4105, UHD 600 IGPU, 4GB ram, SSD, Windows 10, mfakto 0.15pre7
exp=210286903 bit 73 to 74
Initial settings and speed.
GPUSieveProcessSize=24, GPUSieveSize=96, GPUSievePrimes=81157, 3.70GHz-day
Final settings and speed.
GPUSieveProcessSize=32, GPUSieveSize=128, GPUSievePrimes=81157, 3.74GHz-day
Speed with prime95 also running, 3.77GHz-day