static UCS_F_ALWAYS_INLINE void uct_dc_mlx5_iface_atomic_post(uct_dc_mlx5_iface_t *iface, uct_dc_mlx5_ep_t *ep, unsigned opcode, uct_rc_iface_send_desc_t *desc, unsigned length, uint64_t remote_addr, uct_rkey_t rkey, uint64_t compare_mask, uint64_t compare, uint64_t swap_mask, uint64_t swap_add) { uint32_t ib_rkey = uct_ib_resolve_atomic_rkey(rkey, ep->atomic_mr_offset, &remote_addr); UCT_DC_MLX5_TXQP_DECL(txqp, txwq); UCT_DC_MLX5_IFACE_TXQP_GET(iface, ep, txqp, txwq); desc->super.sn = txwq->sw_pi; uct_rc_mlx5_txqp_dptr_post(&iface->super, UCT_IB_QPT_DCI, txqp, txwq, opcode, desc + 1, length, &desc->lkey, remote_addr, ib_rkey, compare_mask, compare, swap_mask, swap_add, &ep->av, uct_dc_mlx5_ep_get_grh(ep), uct_ib_mlx5_wqe_av_size(&ep->av), MLX5_WQE_CTRL_CQ_UPDATE, 0, INT_MAX, NULL); UCT_TL_EP_STAT_ATOMIC(&ep->super); uct_rc_txqp_add_send_op(txqp, &desc->super); }
static UCS_F_ALWAYS_INLINE void uct_rc_verbs_ep_atomic_post(uct_rc_verbs_ep_t *ep, int opcode, uint64_t compare_add, uint64_t swap, uint64_t remote_addr, uct_rkey_t rkey, uct_rc_iface_send_desc_t *desc, int force_sig) { uint32_t ib_rkey = uct_ib_resolve_atomic_rkey(rkey, ep->super.atomic_mr_offset, &remote_addr); struct ibv_send_wr wr; struct ibv_sge sge; UCT_RC_VERBS_FILL_ATOMIC_WR(wr, wr.opcode, sge, opcode, compare_add, swap, remote_addr, ib_rkey); UCT_TL_EP_STAT_ATOMIC(&ep->super.super); uct_rc_verbs_ep_post_send_desc(ep, &wr, desc, force_sig); }
static UCS_F_ALWAYS_INLINE void uct_rc_mlx5_ep_atomic_post(uct_rc_mlx5_ep_t *ep, unsigned opcode, uct_rc_iface_send_desc_t *desc, unsigned length, uint64_t remote_addr, uct_rkey_t rkey, uint64_t compare_mask, uint64_t compare, uint64_t swap_add, int signal) { uct_rc_iface_t *iface = ucs_derived_of(ep->super.super.super.iface, uct_rc_iface_t); uint32_t ib_rkey = uct_ib_resolve_atomic_rkey(rkey, ep->super.atomic_mr_offset, &remote_addr); desc->super.sn = ep->tx.wq.sw_pi; uct_rc_mlx5_txqp_dptr_post(iface, IBV_QPT_RC, &ep->super.txqp, &ep->tx.wq, opcode, desc + 1, length, &desc->lkey, 0, NULL, 0, remote_addr, ib_rkey, compare_mask, compare, swap_add, NULL, 0, signal); UCT_TL_EP_STAT_ATOMIC(&ep->super.super); uct_rc_txqp_add_send_op(&ep->super.txqp, &desc->super); }