static ucs_status_t uct_perf_test_check_capabilities(ucx_perf_params_t *params,
uct_iface_h iface)
{
uct_iface_attr_t attr;
ucs_status_t status;
uint64_t required_flags;
size_t min_size, max_size, max_iov, message_size;
status = uct_iface_query(iface, &attr);
if (status != UCS_OK) {
return status;
}
min_size = 0;
max_iov = 1;
message_size = ucx_perf_get_message_size(params);
switch (params->command) {
case UCX_PERF_CMD_AM:
required_flags = __get_flag(params->uct.data_layout, UCT_IFACE_FLAG_AM_SHORT,
UCT_IFACE_FLAG_AM_BCOPY, UCT_IFACE_FLAG_AM_ZCOPY);
required_flags |= UCT_IFACE_FLAG_CB_SYNC;
min_size = __get_max_size(params->uct.data_layout, 0, 0,
attr.cap.am.min_zcopy);
max_size = __get_max_size(params->uct.data_layout, attr.cap.am.max_short,
attr.cap.am.max_bcopy, attr.cap.am.max_zcopy);
max_iov = attr.cap.am.max_iov;
break;
case UCX_PERF_CMD_PUT:
required_flags = __get_flag(params->uct.data_layout, UCT_IFACE_FLAG_PUT_SHORT,
UCT_IFACE_FLAG_PUT_BCOPY, UCT_IFACE_FLAG_PUT_ZCOPY);
min_size = __get_max_size(params->uct.data_layout, 0, 0,
attr.cap.put.min_zcopy);
max_size = __get_max_size(params->uct.data_layout, attr.cap.put.max_short,
attr.cap.put.max_bcopy, attr.cap.put.max_zcopy);
max_iov = attr.cap.put.max_iov;
break;
case UCX_PERF_CMD_GET:
required_flags = __get_flag(params->uct.data_layout, 0,
UCT_IFACE_FLAG_GET_BCOPY, UCT_IFACE_FLAG_GET_ZCOPY);
min_size = __get_max_size(params->uct.data_layout, 0, 0,
attr.cap.get.min_zcopy);
max_size = __get_max_size(params->uct.data_layout, 0,
attr.cap.get.max_bcopy, attr.cap.get.max_zcopy);
max_iov = attr.cap.get.max_iov;
break;
case UCX_PERF_CMD_ADD:
required_flags = __get_atomic_flag(message_size, UCT_IFACE_FLAG_ATOMIC_ADD32,
UCT_IFACE_FLAG_ATOMIC_ADD64);
max_size = 8;
break;
case UCX_PERF_CMD_FADD:
required_flags = __get_atomic_flag(message_size, UCT_IFACE_FLAG_ATOMIC_FADD32,
UCT_IFACE_FLAG_ATOMIC_FADD64);
max_size = 8;
break;
case UCX_PERF_CMD_SWAP:
required_flags = __get_atomic_flag(message_size, UCT_IFACE_FLAG_ATOMIC_SWAP32,
UCT_IFACE_FLAG_ATOMIC_SWAP64);
max_size = 8;
break;
case UCX_PERF_CMD_CSWAP:
required_flags = __get_atomic_flag(message_size, UCT_IFACE_FLAG_ATOMIC_CSWAP32,
UCT_IFACE_FLAG_ATOMIC_CSWAP64);
max_size = 8;
break;
default:
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Invalid test command");
}
return UCS_ERR_INVALID_PARAM;
}
status = ucx_perf_test_check_params(params);
if (status != UCS_OK) {
return status;
}
if (!ucs_test_all_flags(attr.cap.flags, required_flags) || !required_flags) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Device does not support required operation");
}
return UCS_ERR_UNSUPPORTED;
}
if (message_size < min_size) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Message size too small");
}
return UCS_ERR_UNSUPPORTED;
}
if (message_size > max_size) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Message size too big");
}
return UCS_ERR_UNSUPPORTED;
}
if (params->command == UCX_PERF_CMD_AM) {
if ((params->uct.data_layout == UCT_PERF_DATA_LAYOUT_SHORT) &&
(params->am_hdr_size != sizeof(uint64_t)))
{
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Short AM header size must be 8 bytes");
}
return UCS_ERR_INVALID_PARAM;
}
if ((params->uct.data_layout == UCT_PERF_DATA_LAYOUT_ZCOPY) &&
(params->am_hdr_size > attr.cap.am.max_hdr))
{
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM header size too big");
}
return UCS_ERR_UNSUPPORTED;
}
if (params->am_hdr_size > message_size) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM header size larger than message size");
}
return UCS_ERR_INVALID_PARAM;
}
if (params->uct.fc_window > UCT_PERF_TEST_MAX_FC_WINDOW) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM flow-control window too large (should be <= %d)",
UCT_PERF_TEST_MAX_FC_WINDOW);
}
return UCS_ERR_INVALID_PARAM;
}
if ((params->flags & UCX_PERF_TEST_FLAG_ONE_SIDED) &&
(params->flags & UCX_PERF_TEST_FLAG_VERBOSE))
{
ucs_warn("Running active-message test with on-sided progress");
}
}
if (UCT_PERF_DATA_LAYOUT_ZCOPY == params->uct.data_layout) {
if (params->msg_size_cnt > max_iov) {
if ((params->flags & UCX_PERF_TEST_FLAG_VERBOSE) ||
!params->msg_size_cnt) {
ucs_error("Wrong number of IOV entries. Requested is %lu, "
"should be in the range 1...%lu", params->msg_size_cnt,
max_iov);
}
return UCS_ERR_UNSUPPORTED;
}
/* if msg_size_cnt == 1 the message size checked above */
if ((UCX_PERF_CMD_AM == params->command) && (params->msg_size_cnt > 1)) {
if (params->am_hdr_size > params->msg_size_list[0]) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM header size (%lu) larger than the first IOV "
"message size (%lu)", params->am_hdr_size,
params->msg_size_list[0]);
}
return UCS_ERR_INVALID_PARAM;
}
}
}
return UCS_OK;
}
static ucs_status_t uct_perf_test_check_capabilities(ucx_perf_params_t *params,
uct_iface_h iface)
{
uct_iface_attr_t attr;
ucs_status_t status;
uint64_t required_flags;
size_t max_size;
status = uct_iface_query(iface, &attr);
if (status != UCS_OK) {
return status;
}
switch (params->command) {
case UCX_PERF_CMD_AM:
required_flags = __get_flag(params->uct.data_layout, UCT_IFACE_FLAG_AM_SHORT,
UCT_IFACE_FLAG_AM_BCOPY, UCT_IFACE_FLAG_AM_ZCOPY);
max_size = __get_max_size(params->uct.data_layout, attr.cap.am.max_short,
attr.cap.am.max_bcopy, attr.cap.am.max_zcopy);
break;
case UCX_PERF_CMD_PUT:
required_flags = __get_flag(params->uct.data_layout, UCT_IFACE_FLAG_PUT_SHORT,
UCT_IFACE_FLAG_PUT_BCOPY, UCT_IFACE_FLAG_PUT_ZCOPY);
max_size = __get_max_size(params->uct.data_layout, attr.cap.put.max_short,
attr.cap.put.max_bcopy, attr.cap.put.max_zcopy);
break;
case UCX_PERF_CMD_GET:
required_flags = __get_flag(params->uct.data_layout, 0,
UCT_IFACE_FLAG_GET_BCOPY, UCT_IFACE_FLAG_GET_ZCOPY);
max_size = __get_max_size(params->uct.data_layout, 0,
attr.cap.get.max_bcopy, attr.cap.get.max_zcopy);
break;
case UCX_PERF_CMD_ADD:
required_flags = __get_atomic_flag(params->message_size, UCT_IFACE_FLAG_ATOMIC_ADD32,
UCT_IFACE_FLAG_ATOMIC_ADD64);
max_size = 8;
break;
case UCX_PERF_CMD_FADD:
required_flags = __get_atomic_flag(params->message_size, UCT_IFACE_FLAG_ATOMIC_FADD32,
UCT_IFACE_FLAG_ATOMIC_FADD64);
max_size = 8;
break;
case UCX_PERF_CMD_SWAP:
required_flags = __get_atomic_flag(params->message_size, UCT_IFACE_FLAG_ATOMIC_SWAP32,
UCT_IFACE_FLAG_ATOMIC_SWAP64);
max_size = 8;
break;
case UCX_PERF_CMD_CSWAP:
required_flags = __get_atomic_flag(params->message_size, UCT_IFACE_FLAG_ATOMIC_CSWAP32,
UCT_IFACE_FLAG_ATOMIC_CSWAP64);
max_size = 8;
break;
default:
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Invalid test command");
}
return UCS_ERR_INVALID_PARAM;
}
status = ucx_perf_test_check_params(params);
if (status != UCS_OK) {
return status;
}
if ((attr.cap.flags & required_flags) == 0) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Device does not support required operation");
}
return UCS_ERR_UNSUPPORTED;
}
if (params->message_size > max_size) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Message size too big");
}
return UCS_ERR_INVALID_PARAM;
}
if (params->command == UCX_PERF_CMD_AM) {
if ((params->uct.data_layout == UCT_PERF_DATA_LAYOUT_SHORT) &&
(params->am_hdr_size != sizeof(uint64_t)))
{
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("Short AM header size must be 8 bytes");
}
return UCS_ERR_INVALID_PARAM;
}
if ((params->uct.data_layout == UCT_PERF_DATA_LAYOUT_ZCOPY) &&
(params->am_hdr_size > attr.cap.am.max_hdr))
{
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM header size too big");
}
return UCS_ERR_INVALID_PARAM;
}
if (params->am_hdr_size > params->message_size) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM header size larger than message size");
}
return UCS_ERR_INVALID_PARAM;
}
if (params->uct.fc_window > UCT_PERF_TEST_MAX_FC_WINDOW) {
if (params->flags & UCX_PERF_TEST_FLAG_VERBOSE) {
ucs_error("AM flow-control window too large (should be <= %d)",
UCT_PERF_TEST_MAX_FC_WINDOW);
}
return UCS_ERR_INVALID_PARAM;
}
if ((params->flags & UCX_PERF_TEST_FLAG_ONE_SIDED) &&
(params->flags & UCX_PERF_TEST_FLAG_VERBOSE))
{
ucs_warn("Running active-message test with on-sided progress");
}
}
return UCS_OK;
}
