From 6d8cadeab8d98681f4fc22e8a66855c4593f9404 Mon Sep 17 00:00:00 2001
From: Juan ESCOBAR <juan.escobar@aero.obs-mip.fr>
Date: Wed, 1 Feb 2023 11:42:03 +0100
Subject: [PATCH] Juan 01/02/2023:set_rocm_bind_device, new script for better
 bing GPU & CPU on ADASTRA

---
 bin/set_rocm_bind_device | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)
 create mode 100755 bin/set_rocm_bind_device

diff --git a/bin/set_rocm_bind_device b/bin/set_rocm_bind_device
new file mode 100755
index 000000000..cea4b811a
--- /dev/null
+++ b/bin/set_rocm_bind_device
@@ -0,0 +1,20 @@
+#!/bin/bash
+#set -x
+
+GPUSID=(4 5 2 3 6 7 0 1) 
+
+Gpuinfo='rocm-smi -i'
+NB_DEVICE=${NB_DEVICE:-$( ${Gpuinfo} | grep GPU | wc -l )}
+[ ${NB_DEVICE} -eq 0 ] && NB_DEVICE=1
+
+export LIP=${OMPI_COMM_WORLD_LOCAL_RANK:-${SLURM_LOCALID}}
+export IP=${OMPI_COMM_WORLD_RANK:-${SLURM_PROCID}}
+export NP=${OMPI_COMM_WORLD_SIZE:-${SLURM_NTASKS}}
+export NN=${OMPI_MCA_orte_num_nodes:-${SLURM_NNODES}}
+export NPN=$(( 1 + (NP-1)/ NN ))
+export HN=$( hostname )
+export IG=$(( LIP / ( 1 + (NPN-1) / NB_DEVICE ) )) 
+export ROCR_VISIBLE_DEVICES=${GPUSID[${IG}]}
+echo LIP=${LIP} IP=${IP} NP=${NP} NN=${NN} NPN=${NPN} NG=${NB_DEVICE} IG=${IG} GPU=${ROCR_VISIBLE_DEVICES} ${HN} `taskset -pc $$`
+
+exec $*
-- 
GitLab