CUDA - Class problems. Factor divisible by 2, 3, 5, 7, or 11

[TheJudger]

Hi Luigi,

yes, verbose mode is an old feature, I'm not sure about exact verbosity levels but at least increasing "-v <level>" increases the amount of output (similar to mfaktc).

Oliver

[TheJudger]

Hi Serge,

Originally Posted by Batalov

It's been years since I last built it, but I remember that there are messages about register usage etc. Check at compile time that they are not containing complaints about 'out of registers' (iirc).



There are a couple "#define"s in the header that allow to unroll loops less aggressively (again, iirc, some of them are at 3, at 4; just guessing that the new architectures may have higher throughput but less registers, or something like that).

because I wasn't able to reproduce on Linux using the same code and I *assume* that GPU code has no difference on Linux and Windows I didn't take a deeper look at this. There should be some minor differences on CPU code (check src/compatibility.h) but not in GPU code.

Oliver

[ET_]

Originally Posted by TheJudger

Hi Serge,




because I wasn't able to reproduce on Linux using the same code and I *assume* that GPU code has no difference on Linux and Windows I didn't take a deeper look at this. There should be some minor differences on CPU code (check src/compatibility.h) but not in GPU code.

Oliver

Tried to compile mmff on Linux (Ubuntu 18.04 LTS) and CUDA 10.0 for a RTX 2060 with either cc=6.1 or cc=7.0(BTW, what's its CC? 7.0?).

The error appears at linking:

gcc -fPIC -L/usr/local/cuda/lib64/ timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -lcudart -lm -o ../mmff.exe
tf_barrett96_gs.o:(.data.rel.local.DW.ref.__gxx_personality_v0[DW.ref.__gxx_personality_v0]+0x0): riferimento non definito a "__gxx_personality_v0" (= reference not defined to ...)
collect2: error: ld returned 1 exit status
Makefile:38: recipe for target '../mmff.exe' failed
make: *** [../mmff.exe] Error 1

[Mark Rose]

Originally Posted by ET_

BTW, what's its CC? 7.0?

7.5, like the other Turing chips.

[ET_]

Originally Posted by Mark Rose

7.5, like the other Turing chips.

Still failing linkediting.

[TheJudger]

Hi Luigi,

Originally Posted by ET_

Tried to compile mmff on Linux (Ubuntu 18.04 LTS) and CUDA 10.0 for a RTX 2060 with either cc=6.1 or cc=7.0(BTW, what's its CC? 7.0?).

The error appears at linking:

gcc -fPIC -L/usr/local/cuda/lib64/ timer.o parse.o read_config.o mfaktc.o checkpoint.o signal_handler.o output.o tf_barrett96_gs.o gpusieve.o -lcudart -lm -o ../mmff.exe
tf_barrett96_gs.o:(.data.rel.local.DW.ref.__gxx_personality_v0[DW.ref.__gxx_personality_v0]+0x0): riferimento non definito a "__gxx_personality_v0" (= reference not defined to ...)
collect2: error: ld returned 1 exit status
Makefile:38: recipe for target '../mmff.exe' failed
make: *** [../mmff.exe] Error 1

That one is an easy one, just add

-lstdc++

to the linker flags (host code, not GPU code).
Turing is CC 7.5 but can run CC 7.0 (Volta) code easily.

Oliver

[ET_]

Originally Posted by TheJudger

Hi Luigi,



That one is an easy one, just add

-lstdc++

to the linker flags (host code, not GPU code).
Turing is CC 7.5 but can run CC 7.0 (Volta) code easily.

Oliver

[TheJudger]

Luigi,

so on Ubuntu mmff 0.28 seems to work, right?
Can you try the worktodo from above and run it with -v 3 or or so? Does it show the issue with failing to calculate % 1, too?
@all: did earlier version (on older HW) show this behaviour/issue, too?

Oliver

[Dylan14]

Originally Posted by TheJudger

@all: did earlier version (on older HW) show this behaviour/issue, too?

Oliver

I compiled the other version available on mersenne.ca's download mirror (version .26) and ran the worktodo file that you mentioned previously (on a GTX 1050 ti, I don't have any other compatible hardware to test this on). Unfortunately, this version does not like the exponents tested, since I get messages like:

WARNING: Exponents >= 160 are not supported in Fermat factoring!
         Ignoring TF exponent 167 from 2^197 to 2^198!

which makes sense, since some kernels that exist in v0.28 don't exist in v0.26. When I try the worktodo.txt file included in the v0.26 source (attached), I get a line like

Verifying (2^(2^31)) % 0 = 0
ERROR: Class problems.  Factor divisible by 2, 3, 5, 7, or 11

All relevant output is included in the zip file.

[ET_]

Originally Posted by TheJudger

Luigi,

so on Ubuntu mmff 0.28 seems to work, right?
Can you try the worktodo from above and run it with -v 3 or or so? Does it show the issue with failing to calculate % 1, too?

Oliver

It looks like it does...

[TheJudger]

Originally Posted by ET_

It looks like it does...

It doesn't mean that this is a (big) issue, just want to know if this happens for you, too. And would be good to know if this occurs in earlier builts (or versions) aswell.

Speaking for mfaktc in debug mode (compiletime option) there are lots of checks on (interim) results. Sometimes the results are not within the expected range but this is OK because the issue was already on the input, e.g. the kernel is designed to handle factors abot 2⁶⁴ and on input are numbers slightly below 2⁶⁴. Mfaktc checks the input before writing (error-)messages (ignoring issues when input was not within the designed limits). This could be the case here, too.

Oliver