fix build on CUDA 10.1: - Githubissues

h2oai / h2o4gpu

H2Oai GPU Edition

Apache License 2.0

460 stars 95 forks source link

fix build on CUDA 10.1: #768

Closed sh1ng closed 5 years ago

sh1ng commented 5 years ago

use __shfl_down_sync
add kmeans copy overload methods for device_allocator
linking with libcublasLt_static.a

Besides the fact that we're not going to support CUDA-10.1 for now(see https://github.com/dmlc/xgboost/issues/4264). It's better to be compatible with it.

sh1ng commented 5 years ago

It's not linking on CUDA-10.1 because NVidia made extra changes on libcublas. https://devtalk.nvidia.com/default/topic/1048658/gpu-accelerated-libraries/linker-problem-with-cuda-10-1-105-on-cento-7-undefined-references/ https://github.com/tensorflow/tensorflow/issues/26150