HolyLow
HolyLow
@pramodk I ran a huge network with 100000 neurons, on 8 top-tired A100 GPUs. Here is my output log: ``` $ mpiexec -np 8 ./profile_gpu_install/bin/special-core --tstop 1000 --datpath ./networks/100000Sim/RoundRobin-core-8 --mpi...
@pramodk I added the export PGI_ACC_SYNCHRONOUS=1 and reran the program, the log is as below: ``` $ mpiexec -np 8 ./profile_gpu_install/bin/special-core --tstop 1000 --datpath ./networks/100000Sim/RoundRobin-core-8 --mpi --gpu --multisend --ms-phases 1...
@pramodk The log without multisend is as below: ``` $ mpiexec -np 8 ./profile_gpu_install/bin/special-core --tstop 1000 --datpath ./networks/100000Sim/RoundRobin-core-8 --mpi --gpu libibverbs: Warning: couldn't open config directory '/usr/etc/libibverbs.d'. libibverbs: Warning: no...
@pramodk @nrnhines So could you please show me some hints on the following questions? 1. If I want to let the netcons that is going to fire in a timestep...
@nrnhines > These are generally SelfEvents from NET_RECEIVE net_send(...) calls (often 0 delay) and not NetCon events which can only be 0 delay if src and target are in same...
Same problem, still not solved yet @ @a0u @fkan-ec
@chihminchao I tried the #485 , but still `build git:(master) ✗ spike --extension=dummy_rocc $RISCV/riscv64-unknown-elf/bin/pk a.out bbl loader z 0000000000000000 ra 0000000000010102 sp 000000007f7e9b00 gp 0000000000015ab8 tp 0000000000000000 t0 0000000000010454 t1...
@chihminchao I changed the dummy_rocc behavior in my local repository so that the instruction 0x00c5f50b won't fall into an illegal instruction. But it still doesn't work, and I think the...
@chihminchao I solved my problem! It was the issue of my pk, I should set csv to enable xcustom instructions in my used pk, which is described by the patch...
Seems this bug still exists. I ran into the same coredump when using the folly-2024.04.01. My environment is AliOS-7 (based on centos I believe) with libc 2.32 : ``` $...