static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret) {
fprintf(stderr, "sync_test failed!\n");
goto out;
}
ft_start();
if (opts.dst_addr) {
for (i = 0; i < opts.iterations; i++) {
ret = ft_tx(ep, remote_fi_addr, opts.transfer_size, &tx_ctx);
if (ret)
goto out;
}
} else {
ret = wait_for_recv_completion(opts.iterations);
if (ret)
goto out;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations,
&start, &end, 1, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations,
&start, &end, 1);
out:
return ret;
}
static int run_test(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
ft_start();
for (i = 0; i < opts.iterations; i++) {
ret = ft_rma(opts.rma_op, ep, opts.transfer_size, &remote, ep);
if (ret)
return ret;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
1, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations,
&start, &end, 1);
return 0;
}
static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret)
return ret;
clock_gettime(CLOCK_MONOTONIC, &start);
for (i = 0; i < opts.iterations; i++) {
ret = opts.dst_addr ? send_xfer(opts.transfer_size) :
recv_xfer(opts.transfer_size);
if (ret)
return ret;
ret = opts.dst_addr ? recv_xfer(opts.transfer_size) :
send_xfer(opts.transfer_size);
if (ret)
return ret;
}
clock_gettime(CLOCK_MONOTONIC, &end);
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 2, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations, &start, &end, 2);
return 0;
}
static int run_test(void)
{
int ret, i, t;
ret = sync_test();
if (ret)
goto out;
gettimeofday(&start, NULL);
for (i = 0; i < iterations; i++) {
for (t = 0; t < transfer_count; t++) {
ret = dst_addr ? send_xfer(transfer_size) :
recv_xfer(transfer_size);
if (ret)
goto out;
}
for (t = 0; t < transfer_count; t++) {
ret = dst_addr ? recv_xfer(transfer_size) :
send_xfer(transfer_size);
if (ret)
goto out;
}
}
gettimeofday(&end, NULL);
show_perf();
ret = 0;
out:
return ret;
}
int pingpong(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
ft_start();
for (i = 0; i < opts.iterations; i++) {
ret = opts.dst_addr ?
ft_tx(opts.transfer_size) : ft_rx(opts.transfer_size);
if (ret)
return ret;
ret = opts.dst_addr ?
ft_rx(opts.transfer_size) : ft_tx(opts.transfer_size);
if (ret)
return ret;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 2, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations, &start, &end, 2);
return 0;
}
static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret)
goto out;
clock_gettime(CLOCK_MONOTONIC, &start);
for (i = 0; i < iterations; i++) {
ret = dst_addr ? send_xfer(transfer_size) :
recv_xfer(transfer_size);
if (ret)
goto out;
ret = dst_addr ? recv_xfer(transfer_size) :
send_xfer(transfer_size);
if (ret)
goto out;
}
clock_gettime(CLOCK_MONOTONIC, &end);
show_perf(test_name, transfer_size, iterations, &start, &end, 2);
ret = 0;
out:
return ret;
}
static int run_test(void)
{
int ret = 0, i;
if ((ret = sync_test())) {
goto out;
}
if (bidir || client) {
gettimeofday(&start, NULL);
for (i = 0; i < iterations; i++) {
if ((ret = write_xfer(transfer_size))) {
goto out;
}
}
if ((ret = poll_all_sends())) {
goto out;
}
gettimeofday(&end, NULL);
show_perf();
}
if ((ret = sync_test())) {
goto out;
}
out:
return ret;
}
static int stream(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
ft_start();
for (i = 0; i < opts.iterations; i++) {
ret = opts.dst_addr ?
ft_tx(ep, remote_fi_addr, opts.transfer_size, &tx_ctx) : ft_rx(ep, opts.transfer_size);
if (ret)
return ret;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 1,
opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end, 1);
return 0;
}
static int run_op(void)
{
int ret, i;
count = (size_t *) malloc(sizeof(size_t));
sync_test();
clock_gettime(CLOCK_MONOTONIC, &start);
switch (op_type) {
case FI_MIN:
case FI_MAX:
case FI_ATOMIC_READ:
case FI_ATOMIC_WRITE:
ret = is_valid_base_atomic_op(op_type);
if (ret > 0) {
for (i = 0; i < iterations; i++) {
ret = execute_base_atomic_op(op_type);
if (ret)
break;
}
}
ret = is_valid_fetch_atomic_op(op_type);
if (ret > 0) {
for (i = 0; i < iterations; i++) {
ret = execute_fetch_atomic_op(op_type);
if (ret)
break;
}
}
break;
case FI_CSWAP:
ret = is_valid_compare_atomic_op(op_type);
if (ret > 0) {
ret = execute_compare_atomic_op(op_type);
} else {
goto out;
}
break;
default:
ret = -EINVAL;
goto out;
}
clock_gettime(CLOCK_MONOTONIC, &end);
if (ret)
goto out;
show_perf(test_name, transfer_size, iterations, &start, &end, 2);
ret = 0;
out:
free(count);
return ret;
}
static int run_test(void)
{
int ret, i, t;
off_t offset;
uint8_t marker = 0;
poll_byte = buf + transfer_size - 1;
*poll_byte = -1;
offset = riomap(rs, buf, transfer_size, PROT_WRITE, 0, 0);
if (offset == -1) {
perror("riomap");
ret = -1;
goto out;
}
ret = sync_test();
if (ret)
goto out;
gettimeofday(&start, NULL);
for (i = 0; i < iterations; i++) {
if (dst_addr) {
for (t = 0; t < transfer_count - 1; t++) {
ret = send_xfer(transfer_size);
if (ret)
goto out;
}
*poll_byte = (uint8_t) marker++;
ret = send_xfer(transfer_size);
if (ret)
goto out;
ret = recv_xfer(transfer_size, marker++);
} else {
ret = recv_xfer(transfer_size, marker++);
if (ret)
goto out;
for (t = 0; t < transfer_count - 1; t++) {
ret = send_xfer(transfer_size);
if (ret)
goto out;
}
*poll_byte = (uint8_t) marker++;
ret = send_xfer(transfer_size);
}
if (ret)
goto out;
}
gettimeofday(&end, NULL);
show_perf();
ret = riounmap(rs, buf, transfer_size);
out:
return ret;
}
/*--------------------------------------------------------*/
void AzsSvrg::_train_test()
{
if (rseed > 0) {
srand(rseed); /* initialize the random seed */
}
/*--- initialization ---*/
int dim = m_trn_x->rowNum();
reset_weights(dim);
/*--- iterate ... ---*/
AzTimeLog::print("--- Training begins ... ", log_out);
AzsSvrgData_fast prev_fast;
AzsSvrgData_compact prev_compact;
int ite;
for (ite = 0; ite < ite_num; ++ite) {
if (do_show_timing) AzTimeLog::print("--- iteration#", ite+1, log_out);
if (doing_svrg(ite) && (ite-sgd_ite) % svrg_interval == 0) {
if (do_show_timing) AzTimeLog::print("Computing gradient average ... ", log_out);
if (do_compact) get_avg_gradient_compact(&prev_compact);
else get_avg_gradient_fast(&prev_fast);
}
if (do_show_timing) AzTimeLog::print("Updating weights ... ", log_out);
AzIntArr ia_dxs;
const int *dxs = gen_seq(dataSize(), ia_dxs);
int ix;
for (ix = 0; ix < dataSize(); ++ix) {
int dx = dxs[ix]; /* data point index */
AzDvect v_deriv(class_num);
get_deriv(dx, &v_deriv); /* compute the derivatives */
if (doing_svrg(ite)) {
if (do_compact) updateDelta_svrg_compact(dx, &v_deriv, prev_compact);
else updateDelta_svrg_fast(dx, &v_deriv, prev_fast);
}
else {
updateDelta_sgd(dx, &v_deriv);
}
flushDelta();
}
show_perf(ite);
}
if (do_show_timing) AzTimeLog::print("--- End of training ... ", log_out);
/*--- write predictions to a file if requested ---*/
if (s_pred_fn.length() > 0) {
AzTimeLog::print("Writing predictions to ", s_pred_fn.c_str(), log_out);
write_pred(m_tst_x, s_pred_fn.c_str());
}
}
int pingpong(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
if (opts.dst_addr) {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_tx(opts.transfer_size);
if (ret)
return ret;
ret = ft_rx(opts.transfer_size);
if (ret)
return ret;
}
} else {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
ret = ft_rx(opts.transfer_size);
if (ret)
return ret;
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_tx(opts.transfer_size);
if (ret)
return ret;
}
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 2, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end, 2);
return 0;
}
static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret)
return ret;
clock_gettime(CLOCK_MONOTONIC, &start);
for (i = 0; i < opts.iterations; i++) {
switch (op_type) {
case FT_RMA_WRITE:
ret = write_data(opts.transfer_size);
break;
case FT_RMA_WRITEDATA:
ret = write_data_with_cq_data(opts.transfer_size);
if (ret)
return ret;
ret = wait_remote_writedata_completion();
break;
case FT_RMA_READ:
ret = read_data(opts.transfer_size);
break;
}
if (ret)
return ret;
ret = ft_wait_for_comp(txcq, 1);
if (ret)
return ret;
}
clock_gettime(CLOCK_MONOTONIC, &end);
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
1, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations,
&start, &end, 1);
return 0;
}
static int ft_run_bandwidth(void)
{
size_t recv_cnt;
int ret, i;
for (i = 0; i < ft_ctrl.size_cnt; i += ft_ctrl.inc_step) {
ft_tx_ctrl.msg_size = ft_ctrl.size_array[i];
if (ft_tx_ctrl.msg_size > fabric_info->ep_attr->max_msg_size)
break;
if (((test_info.class_function == FT_FUNC_INJECT) ||
(test_info.class_function == FT_FUNC_INJECTDATA)) &&
(ft_tx_ctrl.msg_size > fabric_info->tx_attr->inject_size))
break;
ft_ctrl.xfer_iter = test_info.test_flags & FT_FLAG_QUICKTEST ?
5 : size_to_count(ft_tx_ctrl.msg_size);
recv_cnt = ft_ctrl.xfer_iter;
ret = ft_sync_test(0);
if (ret)
return ret;
ret = ft_post_recv_bufs();
if (ret)
return ret;
clock_gettime(CLOCK_MONOTONIC, &start);
ret = (test_info.ep_type == FI_EP_DGRAM) ?
ft_bw_dgram(&recv_cnt) : ft_bw();
clock_gettime(CLOCK_MONOTONIC, &end);
if (ret) {
FT_PRINTERR("bw test failed!", ret);
return ret;
}
show_perf("bw", ft_tx_ctrl.msg_size, recv_cnt, &start, &end, 1);
}
return 0;
}
int bandwidth(void)
{
int ret, i, j;
ret = ft_sync();
if (ret)
return ret;
/* The loop structured allows for the possibility that the sender
* immediately overruns the receiving side on the first transfer (or
* the entire window). This could result in exercising parts of the
* provider's implementation of FI_RM_ENABLED. For better or worse,
* some MPI-level benchmarks tend to use this type of loop for measuring
* bandwidth. */
if (opts.dst_addr) {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
for(j = 0; j < opts.window_size; j++) {
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_post_tx(opts.transfer_size);
if (ret)
return ret;
}
ret = ft_get_tx_comp(tx_seq);
if (ret)
return ret;
ret = ft_rx(4);
if (ret)
return ret;
}
} else {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
for(j = 0; j < opts.window_size; j++) {
ret = ft_post_rx(opts.transfer_size);
if (ret)
return ret;
}
ret = ft_get_rx_comp(rx_seq-1); /* rx_seq is always one ahead */
if (ret)
return ret;
ret = ft_tx(4);
if (ret)
return ret;
}
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
opts.window_size, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end,
opts.window_size);
return 0;
}
