gpu: Fix SMP process to GPU mapping 21/5021/4
authorMichael Robson <mprobson@illinois.edu>
Thu, 14 Mar 2019 19:10:32 +0000 (12:10 -0700)
committerRaghavendra K <raghavendra066@gmail.com>
Tue, 19 Mar 2019 16:17:25 +0000 (11:17 -0500)
commit0004b7629080904432ef980f5743d612d4a2a52c
treea6030ff929b985e64395488773feae6456082849
parent7234587ca7d815a2a30a46d9acff8727e58ace77
gpu: Fix SMP process to GPU mapping

This patch preserves the round robin ordering of non-smp mode
while ensuring that separate SMP processes are mapped to distinct
GPUs, if they're available. Further work needs to be done to enable
topology and locality aware mapping and user sepcification.

Change-Id: Ief8d89d37bdffbbeae20608f75ec29afe3027ca8
src/arch/cuda/hybridAPI/hapi_impl.cpp