static size_t ucp_tag_rndv_pack_rkey(ucp_request_t *sreq,
ucp_rndv_rts_hdr_t *rndv_rts_hdr)
{
ucp_ep_h ep = sreq->send.ep;
size_t packed_rkey = 0;
ucs_status_t status;
ucs_assert(UCP_DT_IS_CONTIG(sreq->send.datatype));
/* Check if the sender needs to register the send buffer -
* is its datatype contiguous and does the receive side need it */
if ((ucp_ep_is_rndv_lane_present(ep)) &&
(ucp_ep_rndv_md_flags(ep) & UCT_MD_FLAG_NEED_RKEY)) {
status = ucp_request_send_buffer_reg(sreq, ucp_ep_get_rndv_get_lane(ep));
ucs_assert_always(status == UCS_OK);
/* if the send buffer was registered, send the rkey */
UCS_PROFILE_CALL(uct_md_mkey_pack, ucp_ep_md(ep, ucp_ep_get_rndv_get_lane(ep)),
sreq->send.state.dt.contig.memh, rndv_rts_hdr + 1);
rndv_rts_hdr->flags |= UCP_RNDV_RTS_FLAG_PACKED_RKEY;
packed_rkey = ucp_ep_md_attr(ep, ucp_ep_get_rndv_get_lane(ep))->rkey_packed_size;
}
return packed_rkey;
}
static ucs_status_t ucp_tag_req_start_contig(ucp_request_t *req, size_t count,
ssize_t max_short, size_t zcopy_thresh,
size_t rndv_thresh,
const ucp_proto_t *proto)
{
ucp_ep_config_t *config = ucp_ep_config(req->send.ep);
size_t only_hdr_size = proto->only_hdr_size;
ucs_status_t status;
size_t max_zcopy;
ssize_t length;
length = ucp_contig_dt_length(req->send.datatype, count);
req->send.length = length;
if (length <= max_short) {
/* short */
req->send.uct.func = proto->contig_short;
} else if (length >= rndv_thresh) {
/* rendezvous */
status = ucp_tag_send_start_rndv(req);
if (status != UCS_OK) {
return status;
}
} else if (length < zcopy_thresh) {
/* bcopy */
if (req->send.length <= config->max_am_bcopy - only_hdr_size) {
req->send.uct.func = proto->bcopy_single;
} else {
req->send.uct.func = proto->bcopy_multi;
}
} else {
/* eager zcopy */
status = ucp_request_send_buffer_reg(req, ucp_ep_get_am_lane(req->send.ep));
if (status != UCS_OK) {
return status;
}
req->send.uct_comp.func = proto->contig_zcopy_completion;
max_zcopy = config->max_am_zcopy;
if (req->send.length <= max_zcopy - only_hdr_size) {
req->send.uct_comp.count = 1;
req->send.uct.func = proto->contig_zcopy_single;
} else {
/* calculate number of zcopy fragments */
req->send.uct_comp.count = 1 +
(length + proto->first_hdr_size - proto->mid_hdr_size - 1) /
(max_zcopy - proto->mid_hdr_size);
req->send.uct.func = proto->contig_zcopy_multi;
}
}
return UCS_OK;
}
