Revisiting msieve GPU with Colab

[EdH]

<long listing of undefined references>
. . .
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0x9): undefined reference to `__cudaRegisterFatBinary'
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0x3d): undefined reference to `__cudaRegisterFunction'
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0x6b): undefined reference to `__cudaRegisterFunction'
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0x95): undefined reference to `__cudaRegisterVar'
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0xbd): undefined reference to `__cudaRegisterVar'
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0xe5): undefined reference to `__cudaRegisterVar'
tmpxft_00001db1_00000000-5_cudakernel.compute_75.cudafe1.cpp:(.text.startup+0x10e): undefined reference to `__cudaRegisterVar'
collect2: error: ld returned 1 exit status
Makefile:297: recipe for target 'all' failed
make: *** [all] Error 1

I'm lost! (Not an uncommon event. . .)


At the risk of showing even more of my ignorance, Is this an architecture issue and I can simply add some lines of reference to the Makefiles?

[RichD]

Which GPU did you get for your instance? K80, P4, P100, T4?

[jasonp]

You will need a version of CUDA that supports compute capability 7.5 and a version of Visual Studio that that version of CUDA is compatible with. If you have those, both the Msieve makefile and cub/makefile list out the compute capabilities that they will compile for; you can try adding '75' to the list but I don't know if the generated code will work, or wheither the Msieve binary will select the right PTX at runtime.

You've gone from using Msieve on hardware that's too old to using Msieve on hardware that's too new. The result is probably the same :)

[EdH]

Originally Posted by RichD

Which GPU did you get for your instance? K80, P4, P100, T4?

This happened with both a K80 and a P100.

Originally Posted by jasonp

You will need a version of CUDA that supports compute capability 7.5 and a version of Visual Studio that that version of CUDA is compatible with. If you have those, both the Msieve makefile and cub/makefile list out the compute capabilities that they will compile for; you can try adding '75' to the list but I don't know if the generated code will work, or wheither the Msieve binary will select the right PTX at runtime.

You've gone from using Msieve on hardware that's too old to using Msieve on hardware that's too new. The result is probably the same :)

I'm still playing with old, but the Colab instances give me a playground with the new. I've no familiarity at all with Visual Studio. The Colab CUDA is 10.1 ATM, with 10.0 still instaled. I have succeeded in getting the ECM-GPU running, in the Colab sessions and, in fact, I have been compiling it with my msieve attempts, although I don't know if there's any advantage when compiling msieve with ECM=1. I'm mostly doing it to try to prevent the Colab complaint that I'm not using the GPU, while trying to compile msieve.

Thanks for all the help. Off to play. . .

[harrymelling]

Download a version of CUDA that supports compute capability 7.5 and a compatible version of Visual Studio.