mfakto: an OpenCL program for Mersenne prefactoring

[Bdot]

Originally Posted by kracker

Yeah, I believe mine has 600 stream processors on it

It's 400 (@600 MHz) in the HD 6550D / A8 3850, so mfakto guessed it right

[kracker]

Okay, I just got my 7770 today, for some reason (maybe I'm doing it wrong) performance is only a little better compared to my integrated 6550D... (and yes gpu usage was 85%~)

mfakto 0.11-Win (64bit build)


Runtime options
  Inifile                   mfakto.ini
  SievePrimesMin            10000
  SievePrimesMax            200000
  SievePrimes               25000
  SievePrimesAdjust         1
  NumStreams                5
  GridSize                  4
  WorkFile                  worktodo.txt
  ResultsFile               results.txt
  Checkpoints               enabled
  CheckpointDelay           300s
  Stages                    enabled
  StopAfterFactor           class
  PrintMode                 full
  V5UserID                  none
  ComputerID                none
  AllowSleep                yes
  TimeStampInResults        no
  VectorSize                4
  PreferKernel              mfakto_cl_barrett79
  SieveOnGPU                no
  SmallExp                  no
Compiletime options
  SIEVE_SIZE_LIMIT          64kiB
  SIEVE_SIZE                482885bits
  SIEVE_SPLIT               250
  MORE_CLASSES              enabled
Select device - Get device info - Compiling kernels ..........

OpenCL device info
  name                      Capeverde (Advanced Micro Devices, Inc.)
  device (driver) version   OpenCL 1.2 AMD-APP (937.2) (CAL 1.4.1734 (VM))
  maximum threads per block 256
  maximum threads per grid  16777216
  number of multiprocessors 10 (800 compute elements (estimate for ATI GPUs))
  clock rate                1000MHz

Automatic parameters
  threads per grid          2097152

running a simple selftest ...
########## testcase 1/17 (#2) ##########
########## testcase 2/17 (#25) ##########
########## testcase 3/17 (#39) ##########
########## testcase 4/17 (#57) ##########
########## testcase 5/17 (#70) ##########
########## testcase 6/17 (#72) ##########
########## testcase 7/17 (#73) ##########
########## testcase 8/17 (#82) ##########
########## testcase 9/17 (#88) ##########
########## testcase 10/17 (#106) ##########
########## testcase 11/17 (#355) ##########
########## testcase 12/17 (#358) ##########
########## testcase 13/17 (#666) ##########
########## testcase 14/17 (#1547) ##########
########## testcase 15/17 (#1552) ##########
########## testcase 16/17 (#1556) ##########
########## testcase 17/17 (#1557) ##########
Selftest statistics                          
  number of tests           44
  successful tests          44

selftest PASSED!

got assignment: exp=61275631 bit_min=70 bit_max=71
Starting trial factoring M61275631 from 2^70 to 2^71 (3.90GHz-days)
  k_min = 9633451350600 - k_max = 19266902705863
Using GPU kernel "barrett15_75"

found a valid checkpoint file!
  last finished class was: 989
  found 0 factors already

   done |    ETA |     GHz |time/class|    #FCs | avg. rate | SieveP. |CPU idle
 21.67% |  1h54m |   38.43 |   9.139s | 448.79M |  49.11M/s |   25000 |  53.57%
 21.77% |  1h52m |   39.14 |   8.973s | 444.60M |  49.55M/s |   28125 |  51.99%

mfakto will exit once the current class is finished.
press ^C again to exit immediately
 21.88% |  1h53m |   38.52 |   9.118s | 440.40M |  48.30M/s |   31640 |  51.31%

[Bdot]

Originally Posted by kracker

Okay, I just got my 7770 today, for some reason (maybe I'm doing it wrong) performance is only a little better compared to my integrated 6550D... (and yes gpu usage was 85%~)

It's probably not you doing anything wrong, it's mfakto. Could you please send me ~ one minute of 'mfakto-pi -st' output?

I still have to teach mfakto what is good on GCN and what is bad. The kernel it chose is extremely bad. I can probably send you a GCN-optimized version tomorrow, which should have doubled these performance figures.

[kracker]

Originally Posted by Bdot

It's probably not you doing anything wrong, it's mfakto. Could you please send me ~ one minute of 'mfakto-pi -st' output?

I still have to teach mfakto what is good on GCN and what is bad. The kernel it chose is extremely bad. I can probably send you a GCN-optimized version tomorrow, which should have doubled these performance figures.

Will do.

P.S.: Oh Crap. forgot to self-test

[Bdot]

Thanks, kracker for your tests!

It turns out that the GCN-based cards have less registers at their disposal. Already at VectorSize=4, the very register-intensive 15-bit kernel has bad register-spilling into slow scratchpad memory, crippling performance.

On the other hand, GCN can schedule threads more flexibly, so that big vector sizes are no longer needed for high performance - even the more register-efficient kernels run fastest with VectorSize=2.

Therefore, to maximize performance with mfakto 0.11 on HD77xx-79xx, set the following in mfakto.ini:

Stages=1
PreferKernel=mfakto_cl_71
VectorSize=2

The next mfakto version will do this automatically ...

Projecting the test results of the 7770 to 7970, 400M/s should easily be surpassed - dbaugh, do you still/again have an operational 7970?

7770 is now even 7% faster than 5770, even though it has less GFLOPS (1280 vs. 1360) - which indicates improved hardware efficiency.

[kracker]

Thanks Bdot, for everything

Performance is much better, from ~50M to ~135M for my Radeon HD 7770
I'll send some benchmarks (the right way) over to you-know-who

[Dubslow]

I'm about to steal the file locking code
I must say I never would have done such an elaborate system, I can see that I still have much to learn

(PS, compare lines 21 and 25 )

[Bdot]

Originally Posted by kracker

Thanks Bdot, for everything

Performance is much better, from ~50M to ~135M for my Radeon HD 7770
I'll send some benchmarks (the right way) over to you-know-who

exactly the right thing to do

Originally Posted by Dubslow

I'm about to steal the file locking code
I must say I never would have done such an elaborate system, I can see that I still have much to learn

go ahead (steal and learn )

Originally Posted by Dubslow

(PS, compare lines 21 and 25 )

That's just to be sure it really is included

Publishing the source code starts to show added value ... (fixed).

[dbaugh]

I replaced the card with an identical card with the exception that this one works great! I have been running it 24/7 with three instances to get the GPU% to 99 and with the clocks maxed out at 1125/1575. Except for the system pulling a steady 500 watts, all is good. I'll run the benchmark soon. Is there optimized code for the 7970?

[kracker]

Originally Posted by Bdot

Therefore, to maximize performance with mfakto 0.11 on HD77xx-79xx, set the following in mfakto.ini:

Stages=1
PreferKernel=mfakto_cl_71
VectorSize=2

The next mfakto version will do this automatically ...

There, if you haven't already

[Bdot]

Originally Posted by dbaugh

I replaced the card with an identical card with the exception that this one works great! I have been running it 24/7 with three instances to get the GPU% to 99 and with the clocks maxed out at 1125/1575. Except for the system pulling a steady 500 watts, all is good. I'll run the benchmark soon. Is there optimized code for the 7970?

Originally Posted by kracker

There, if you haven't already

Yes, these settings should do. If possible, I'd like to see a mfakto-pi run of your card with VectorSize=2 - this should show the highest mfakto rate I've seen so far.

Regarding the (over)clocks you mention: I've noticed on my slower cards, that apart from more heat there is no measurable difference when increasing the memory clock. Core clock has a direct and linear impact. I could set the core clock to a higher value when lowering the memory clock to the lowest possible setting.