From b2a0a0bd3a6224f4381ea14fb996681366f8fa5b Mon Sep 17 00:00:00 2001 From: Nusrat Islam Date: Mon, 17 Jul 2023 21:45:52 -0500 Subject: [PATCH] device: fine-tune RCCL send-recv --- src/collectives/device/sendrecv.h | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/src/collectives/device/sendrecv.h b/src/collectives/device/sendrecv.h index 030b1af7c4..647378f2a3 100644 --- a/src/collectives/device/sendrecv.h +++ b/src/collectives/device/sendrecv.h @@ -202,13 +202,21 @@ struct RunWork { if (args->proto == NCCL_PROTO_LL) { runRecv(tid, nthreads, group, args); } else { +#if defined(__gfx90a__) + runRecv>(tid, nthreads, group, args); +#else runRecv>(tid, nthreads, group, args); +#endif } } else { if (args->proto == NCCL_PROTO_LL) { runSend(tid, nthreads, group, args); } else { +#if defined(__gfx90a__) + runSend>(tid, nthreads, group, args); +#else runSend>(tid, nthreads, group, args); +#endif } } }