Re: [AMBER] Fixed the problem of AMBER20 run with pmemd.cuda with CUDA11.1 and 3080 GPU card

From: Mohd Farid Ismail <mohd.farid.ismail.yandex.com>
Date: Sat, 02 Jan 2021 08:54:56 +0800

   + kailiu.

   Â

   Hi,

   Â

   Can you post some numbers comparing the sm80 vs sm86 or vs Turing/Pascal
   cards? Â I'm interested to see the actual gain from using sm86. Â Also, did
   you compare the energies/properties produced using the sm86 codes to make
   sure they're 'OK'?

   Â

   --Â Farid

   Â

   Â

   Â

   Â

   31.12.2020, 21:34, "liu kai" <[1]kailiu.outlook.com>:

     Dear All,
     I have a problem with amber20 with CUDA11.1, when I used Nvidia 3080 GPU
     card.
     The error message is âcudaMemcpyToSymbol: SetSim copy to cSim failed
     invalid device symbolâ, maybe connected to
     [2]http://archive.ambermd.org/202012/0096.html
     I think there is a potential problem in the configure2 file, like my case.
     For details:
     Â Â 1. I have updated to the newest version including update.8 and
     update.12 for Amber20 and Ambertool20, respectively.
     Â Â 2. I used the legacy way to install. Because the error of âunsupported
     cuda 11.1â, solved by: [3]http://archive.ambermd.org/202011/0157.html
     Â Â 3. After installation, I used the same executable file:pmemd.cuda but
     different machine for my test.
     Â Â 4. The test file is already equilibrated, I originally wanted to run
     the production step for efficiency comparison.
     Â Â 5. Thus, It is a repeated calculation but with different ig value.
     CUDA 11.1 + Nvidia 3080 calculation failed, with above error message.
     CUDA 11.1 + Nvidia 2080Ti OK
     CUDA 10.1 + Nvidia 2080Ti OK
     Solution:
     The modification in step2 is not enough. Because the gencode is only
     support to sm75.
     While The 3080 card requires sm86.
     So, I added two lines, highlighted as red as below.
     Original:
     Â Â Â if [ "$cudaversion" = "11.0"]; then
     Â Â Â Â Â Â # Implement the standard, not warp synchronous, compilation
     Â Â Â Â Â Â sm75flags='-gencode arch=compute_75,code=sm_75'
     Â Â Â Â Â Â sm70flags='-gencode arch=compute_70,code=sm_70'
     Â Â Â Â Â Â echo "CUDA Version $cudaversion detected"
     Â Â Â Â Â Â echo "Configuring for SM5.0, SM5.2, SM5.3, SM6.0, SM6.1,
     SM7.0, SM7.5 and SM8.0"
     Â Â Â Â Â Â nvccflags="$sm50flags $sm52flags $sm53flags $sm60flags
     $sm61flags $sm70flags $sm75flags $sm80flags $sm86flags
     -Wno-deprecated-declarations"
     changed to:
     Â Â Â if [ "$cudaversion" = "11.0" -o "$cudaversion" = "11.1" ]; then
     Â Â Â Â Â Â # Implement the standard, not warp synchronous, compilation
     Â Â Â Â Â Â sm86flags='-gencode arch=compute_86,code=sm_86'
     Â Â Â Â Â Â sm80flags='-gencode arch=compute_80,code=sm_80'
     Â Â Â Â Â Â sm75flags='-gencode arch=compute_75,code=sm_75'
     Â Â Â Â Â Â sm70flags='-gencode arch=compute_70,code=sm_70'
     Â Â Â Â Â Â echo "CUDA Version $cudaversion detected"
     Â Â Â Â Â Â echo "Configuring for SM5.0, SM5.2, SM5.3, SM6.0, SM6.1,
     SM7.0, SM7.5 and SM8.0"
     Â Â Â Â Â Â nvccflags="$sm50flags $sm52flags $sm53flags $sm60flags
     $sm61flags $sm70flags $sm75flags $sm80flags $sm86flags
     -Wno-deprecated-declarations"
     After that, my calculation was terminated normally by CUDA 11.1 + Nvidia
     3080.
     I have not carefully tested other cases.
     But I hope this would provide some hits for someone has similar problem.
     Yours truly,
     Kai
     _______________________________________________
     AMBER mailing list
     [4]AMBER.ambermd.org
     [5]http://lists.ambermd.org/mailman/listinfo/amber

References

   1. mailto:kailiu.outlook.com
   2. http://archive.ambermd.org/202012/0096.html
   3. http://archive.ambermd.org/202011/0157.html
   4. mailto:AMBER.ambermd.org
   5. http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Jan 01 2021 - 17:00:02 PST