Added support to create multiple queue-pairs per device to
enqueue and dequeue inference requests. Number of queue pairs
to be created can be specified through "--queue_pairs" option.
Support is also enabled to control the number of descriptors
per each queue pair through "--queue_size" option. Inference
requests for a model are distributed across all available
queue-pairs.
Signed-off-by: Srikanth Yalavarthi <syalavar...@marvell.com>
---
app/test-mldev/ml_options.c | 40 ++++++++++---
app/test-mldev/ml_options.h | 4 ++
app/test-mldev/test_common.c | 2 +-
app/test-mldev/test_inference_common.c | 79 +++++++++++++++++++++-----
app/test-mldev/test_inference_common.h | 1 +
5 files changed, 102 insertions(+), 24 deletions(-)
diff --git a/app/test-mldev/ml_options.c b/app/test-mldev/ml_options.c
index 7fbf9c5678..c81dec6e30 100644
--- a/app/test-mldev/ml_options.c
+++ b/app/test-mldev/ml_options.c
@@ -25,6 +25,8 @@ ml_options_default(struct ml_options *opt)
opt->nb_filelist = 0;
opt->repetitions = 1;
opt->burst_size = 1;
+ opt->queue_pairs = 1;
+ opt->queue_size = 1;
opt->debug = false;
}
@@ -152,11 +154,30 @@ ml_parse_burst_size(struct ml_options *opt, const char
*arg)
return parser_read_uint16(&opt->burst_size, arg);
}
+static int
+ml_parse_queue_pairs(struct ml_options *opt, const char *arg)
+{
+ int ret;
+
+ ret = parser_read_uint16(&opt->queue_pairs, arg);
+
+ return ret;
+}
+
+static int
+ml_parse_queue_size(struct ml_options *opt, const char *arg)
+{
+ return parser_read_uint16(&opt->queue_size, arg);
+}
+
static void
ml_dump_test_options(const char *testname)
{
- if (strcmp(testname, "device_ops") == 0)
+ if (strcmp(testname, "device_ops") == 0) {
+ printf("\t\t--queue_pairs : number of queue pairs to
create\n"
+ "\t\t--queue_size : size fo queue-pair\n");
printf("\n");
+ }
if (strcmp(testname, "model_ops") == 0) {
printf("\t\t--models : comma separated list of
models\n");
@@ -167,7 +188,9 @@ ml_dump_test_options(const char *testname)
(strcmp(testname, "inference_interleave") == 0)) {
printf("\t\t--filelist : comma separated list of model,
input and output\n"
"\t\t--repetitions : number of inference
repetitions\n"
- "\t\t--burst_size : inference burst size\n");
+ "\t\t--burst_size : inference burst size\n"
+ "\t\t--queue_pairs : number of queue pairs to
create\n"
+ "\t\t--queue_size : size fo queue-pair\n");
printf("\n");
}
}
@@ -187,11 +210,11 @@ print_usage(char *program)
ml_test_dump_names(ml_dump_test_options);
}
-static struct option lgopts[] = {{ML_TEST, 1, 0, 0}, {ML_DEVICE_ID, 1, 0,
0},
- {ML_SOCKET_ID, 1, 0, 0}, {ML_MODELS, 1, 0, 0},
- {ML_FILELIST, 1, 0, 0}, {ML_REPETITIONS, 1,
0, 0},
- {ML_BURST_SIZE, 1, 0, 0}, {ML_DEBUG, 0, 0, 0},
- {ML_HELP, 0, 0, 0}, {NULL, 0, 0, 0}};
+static struct option lgopts[] = {
+ {ML_TEST, 1, 0, 0}, {ML_DEVICE_ID, 1, 0, 0}, {ML_SOCKET_ID, 1,
0, 0},
+ {ML_MODELS, 1, 0, 0}, {ML_FILELIST, 1, 0, 0}, {ML_REPETITIONS,
1, 0, 0},
+ {ML_BURST_SIZE, 1, 0, 0}, {ML_QUEUE_PAIRS, 1, 0, 0}, {ML_QUEUE_SIZE, 1,
0, 0},
+ {ML_DEBUG, 0, 0, 0}, {ML_HELP, 0, 0, 0}, {NULL, 0, 0, 0}};
static int
ml_opts_parse_long(int opt_idx, struct ml_options *opt)
@@ -202,7 +225,8 @@ ml_opts_parse_long(int opt_idx, struct ml_options *opt)
{ML_TEST, ml_parse_test_name}, {ML_DEVICE_ID,
ml_parse_dev_id},
{ML_SOCKET_ID, ml_parse_socket_id}, {ML_MODELS,
ml_parse_models},
{ML_FILELIST, ml_parse_filelist}, {ML_REPETITIONS,
ml_parse_repetitions},
- {ML_BURST_SIZE, ml_parse_burst_size},
+ {ML_BURST_SIZE, ml_parse_burst_size}, {ML_QUEUE_PAIRS,
ml_parse_queue_pairs},
+ {ML_QUEUE_SIZE, ml_parse_queue_size},
};
for (i = 0; i < RTE_DIM(parsermap); i++) {
diff --git a/app/test-mldev/ml_options.h b/app/test-mldev/ml_options.h
index 00342d8a0c..c4018ee9d1 100644
--- a/app/test-mldev/ml_options.h
+++ b/app/test-mldev/ml_options.h
@@ -19,6 +19,8 @@
#define ML_FILELIST ("filelist")
#define ML_REPETITIONS ("repetitions")
#define ML_BURST_SIZE ("burst_size")
+#define ML_QUEUE_PAIRS ("queue_pairs")
+#define ML_QUEUE_SIZE ("queue_size")
#define ML_DEBUG ("debug")
#define ML_HELP ("help")
@@ -36,6 +38,8 @@ struct ml_options {
uint8_t nb_filelist;
uint64_t repetitions;
uint16_t burst_size;
+ uint16_t queue_pairs;
+ uint16_t queue_size;
bool debug;
};
diff --git a/app/test-mldev/test_common.c b/app/test-mldev/test_common.c
index 8c4da4609a..016b31c6ba 100644
--- a/app/test-mldev/test_common.c
+++ b/app/test-mldev/test_common.c
@@ -75,7 +75,7 @@ ml_test_device_configure(struct ml_test *test, struct
ml_options *opt)
/* configure device */
dev_config.socket_id = opt->socket_id;
dev_config.nb_models = t->dev_info.max_models;
- dev_config.nb_queue_pairs = t->dev_info.max_queue_pairs;
+ dev_config.nb_queue_pairs = opt->queue_pairs;
ret = rte_ml_dev_configure(opt->dev_id, &dev_config);
if (ret != 0) {
ml_err("Failed to configure ml device, dev_id = %d\n",
opt->dev_id);
diff --git a/app/test-mldev/test_inference_common.c
b/app/test-mldev/test_inference_common.c
index 530872def6..7637d0833e 100644
--- a/app/test-mldev/test_inference_common.c
+++ b/app/test-mldev/test_inference_common.c
@@ -66,7 +66,7 @@ ml_enqueue_single(void *arg)
req->fid = fid;
enqueue_req:
- burst_enq = rte_ml_enqueue_burst(t->cmn.opt->dev_id, 0, &op, 1);
+ burst_enq = rte_ml_enqueue_burst(t->cmn.opt->dev_id, args->qp_id, &op,
1);
if (burst_enq == 0)
goto enqueue_req;
@@ -103,7 +103,7 @@ ml_dequeue_single(void *arg)
return 0;
dequeue_req:
- burst_deq = rte_ml_dequeue_burst(t->cmn.opt->dev_id, 0, &op, 1);
+ burst_deq = rte_ml_dequeue_burst(t->cmn.opt->dev_id, args->qp_id, &op,
1);
if (likely(burst_deq == 1)) {
total_deq += burst_deq;
@@ -182,7 +182,8 @@ ml_enqueue_burst(void *arg)
pending = ops_count;
enqueue_reqs:
- burst_enq = rte_ml_enqueue_burst(t->cmn.opt->dev_id, 0,
&args->enq_ops[idx], pending);
+ burst_enq =
+ rte_ml_enqueue_burst(t->cmn.opt->dev_id, args->qp_id,
&args->enq_ops[idx], pending);
pending = pending - burst_enq;
if (pending > 0) {
@@ -223,8 +224,8 @@ ml_dequeue_burst(void *arg)
return 0;
dequeue_burst:
- burst_deq =
- rte_ml_dequeue_burst(t->cmn.opt->dev_id, 0, args->deq_ops,
t->cmn.opt->burst_size);
+ burst_deq = rte_ml_dequeue_burst(t->cmn.opt->dev_id, args->qp_id,
args->deq_ops,
+ t->cmn.opt->burst_size);
if (likely(burst_deq > 0)) {
total_deq += burst_deq;
@@ -257,6 +258,19 @@ test_inference_cap_check(struct ml_options *opt)
return false;
rte_ml_dev_info_get(opt->dev_id, &dev_info);
+
+ if (opt->queue_pairs > dev_info.max_queue_pairs) {
+ ml_err("Insufficient capabilities: queue_pairs = %u,
max_queue_pairs = %u",
+ opt->queue_pairs, dev_info.max_queue_pairs);
+ return false;
+ }
+
+ if (opt->queue_size > dev_info.max_desc) {
+ ml_err("Insufficient capabilities: queue_size = %u, max_desc =
%u", opt->queue_size,
+ dev_info.max_desc);
+ return false;
+ }
+
if (opt->nb_filelist > dev_info.max_models) {
ml_err("Insufficient capabilities: Filelist count exceeded
device limit, count = %u (max limit = %u)",
opt->nb_filelist, dev_info.max_models);
@@ -308,10 +322,21 @@ test_inference_opt_check(struct ml_options *opt)
return -EINVAL;
}
+ if (opt->queue_pairs == 0) {
+ ml_err("Invalid option, queue_pairs = %u\n", opt->queue_pairs);
+ return -EINVAL;
+ }
+
+ if (opt->queue_size == 0) {
+ ml_err("Invalid option, queue_size = %u\n", opt->queue_size);
+ return -EINVAL;
+ }
+
/* check number of available lcores. */
- if (rte_lcore_count() < 3) {
+ if (rte_lcore_count() < (uint32_t)(opt->queue_pairs * 2 + 1)) {
ml_err("Insufficient lcores = %u\n", rte_lcore_count());
- ml_err("Minimum lcores required to create %u queue-pairs =
%u\n", 1, 3);
+ ml_err("Minimum lcores required to create %u queue-pairs =
%u\n", opt->queue_pairs,
+ (opt->queue_pairs * 2 + 1));
return -EINVAL;
}
@@ -329,6 +354,8 @@ test_inference_opt_dump(struct ml_options *opt)
/* dump test opts */
ml_dump("repetitions", "%" PRIu64, opt->repetitions);
ml_dump("burst_size", "%u", opt->burst_size);
+ ml_dump("queue_pairs", "%u", opt->queue_pairs);
+ ml_dump("queue_size", "%u", opt->queue_size);
ml_dump_begin("filelist");
for (i = 0; i < opt->nb_filelist; i++) {
@@ -419,23 +446,31 @@ ml_inference_mldev_setup(struct ml_test *test, struct
ml_options *opt)
{
struct rte_ml_dev_qp_conf qp_conf;
struct test_inference *t;
+ uint16_t qp_id;
int ret;
t = ml_test_priv(test);
+ RTE_SET_USED(t);
+
ret = ml_test_device_configure(test, opt);
if (ret != 0)
return ret;
/* setup queue pairs */
- qp_conf.nb_desc = t->cmn.dev_info.max_desc;
+ qp_conf.nb_desc = opt->queue_size;
qp_conf.cb = NULL;
- ret = rte_ml_dev_queue_pair_setup(opt->dev_id, 0, &qp_conf,
opt->socket_id);
- if (ret != 0) {
- ml_err("Failed to setup ml device queue-pair, dev_id = %d,
qp_id = %u\n",
- opt->dev_id, 0);
- goto error;
+ for (qp_id = 0; qp_id < opt->queue_pairs; qp_id++) {
+ qp_conf.nb_desc = opt->queue_size;
+ qp_conf.cb = NULL;
+
+ ret = rte_ml_dev_queue_pair_setup(opt->dev_id, qp_id, &qp_conf,
opt->socket_id);
+ if (ret != 0) {
+ ml_err("Failed to setup ml device queue-pair, dev_id =
%d, qp_id = %u\n",
+ opt->dev_id, qp_id);
+ return ret;
+ }
}
ret = ml_test_device_start(test, opt);
@@ -692,14 +727,28 @@ ml_inference_launch_cores(struct ml_test *test, struct
ml_options *opt, uint16_t
{
struct test_inference *t = ml_test_priv(test);
uint32_t lcore_id;
+ uint32_t nb_reqs;
uint32_t id = 0;
+ uint32_t qp_id;
+
+ nb_reqs = opt->repetitions / opt->queue_pairs;
RTE_LCORE_FOREACH_WORKER(lcore_id)
{
- if (id == 2)
+ if (id >= opt->queue_pairs * 2)
break;
- t->args[lcore_id].nb_reqs = opt->repetitions;
+ qp_id = id / 2;
+ t->args[lcore_id].qp_id = qp_id;
+ t->args[lcore_id].nb_reqs = nb_reqs;
+ if (qp_id == 0)
+ t->args[lcore_id].nb_reqs += opt->repetitions - nb_reqs
* opt->queue_pairs;
+
+ if (t->args[lcore_id].nb_reqs == 0) {
+ id++;
+ break;
+ }
+
t->args[lcore_id].start_fid = start_fid;
t->args[lcore_id].end_fid = end_fid;
diff --git a/app/test-mldev/test_inference_common.h
b/app/test-mldev/test_inference_common.h
index d0307a6432..b85fb464b1 100644
--- a/app/test-mldev/test_inference_common.h
+++ b/app/test-mldev/test_inference_common.h
@@ -22,6 +22,7 @@ struct ml_core_args {
uint64_t nb_reqs;
uint16_t start_fid;
uint16_t end_fid;
+ uint32_t qp_id;
struct rte_ml_op **enq_ops;
struct rte_ml_op **deq_ops;
--
2.17.1
