Refine3D segfault with CUDA10 and ICC

Issue #33 new
TJ Ragan created an issue

I suspect this is not currently a supported configuration, but decided to offer the bug up anyway. When I compile 5aa7d38 with ICC 2018.3.222 and CUDA10.0 using:

CC=mpiicc CXX=mpiicpc cmake -DCUDA=ON -DCudaTexture=ON -DMKLFFT=ON -DCUDA_ARCH=61 -D CMAKE_C_FLAGS="-O3 -ip -g -debug inline-debug-info -xCORE-AVX2 -restrict " -D CMAKE_CXX_FLAGS="-O3 -ip -g -debug inline-debug-info -xCORE-AVX2 -restrict " -D CMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/net/prog/relion3/5aa7d38-icc-gpu ..

or

CC=mpiicc CXX=mpiicpc cmake -DCUDA=ON -DCudaTexture=ON -DCUDA_ARCH=61 -D CMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/net/prog/relion3/5aa7d38-icc-gpu ..

all steps up to Refine3D (including initial model and 3D classification) run fine with GPU acceleration. Refine3D segfaults after mapping the threads to the GPU devices. Refine3D runs successfully on the same build if GPU acceleration is turned off.

Comments (0)

  1. Log in to comment