@@ -24,6 +24,7 @@ ml_options_default(struct ml_options *opt)
opt->socket_id = SOCKET_ID_ANY;
opt->nb_filelist = 0;
opt->repetitions = 1;
+ opt->burst_size = 1;
opt->debug = false;
}
@@ -145,6 +146,12 @@ ml_parse_repetitions(struct ml_options *opt, const char *arg)
return parser_read_uint64(&opt->repetitions, arg);
}
+static int
+ml_parse_burst_size(struct ml_options *opt, const char *arg)
+{
+ return parser_read_uint16(&opt->burst_size, arg);
+}
+
static void
ml_dump_test_options(const char *testname)
{
@@ -159,7 +166,8 @@ ml_dump_test_options(const char *testname)
if ((strcmp(testname, "inference_ordered") == 0) ||
(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--repetitions : number of inference repetitions\n"
+ "\t\t--burst_size : inference burst size\n");
printf("\n");
}
}
@@ -186,6 +194,7 @@ static struct option lgopts[] = {
{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}};
@@ -202,6 +211,7 @@ ml_opts_parse_long(int opt_idx, struct ml_options *opt)
{ML_MODELS, ml_parse_models},
{ML_FILELIST, ml_parse_filelist},
{ML_REPETITIONS, ml_parse_repetitions},
+ {ML_BURST_SIZE, ml_parse_burst_size},
};
for (i = 0; i < RTE_DIM(parsermap); i++) {
@@ -18,6 +18,7 @@
#define ML_MODELS ("models")
#define ML_FILELIST ("filelist")
#define ML_REPETITIONS ("repetitions")
+#define ML_BURST_SIZE ("burst_size")
#define ML_DEBUG ("debug")
#define ML_HELP ("help")
@@ -34,6 +35,7 @@ struct ml_options {
struct ml_filelist filelist[ML_TEST_MAX_MODELS];
uint8_t nb_filelist;
uint64_t repetitions;
+ uint16_t burst_size;
bool debug;
};
@@ -124,6 +124,132 @@ ml_dequeue_single(void *arg)
return 0;
}
+/* Enqueue inference requests with burst size greater than 1 */
+static int
+ml_enqueue_burst(void *arg)
+{
+ struct test_inference *t = ml_test_priv((struct ml_test *)arg);
+ struct ml_core_args *args;
+ uint16_t ops_count;
+ uint64_t model_enq;
+ uint16_t burst_enq;
+ uint32_t lcore_id;
+ uint16_t pending;
+ uint16_t idx;
+ uint16_t fid;
+ uint16_t i;
+ int ret;
+
+ lcore_id = rte_lcore_id();
+ args = &t->args[lcore_id];
+ model_enq = 0;
+
+ if (args->nb_reqs == 0)
+ return 0;
+
+next_rep:
+ fid = args->start_fid;
+
+next_model:
+ ops_count = RTE_MIN(t->cmn.opt->burst_size, args->nb_reqs - model_enq);
+ ret = rte_mempool_get_bulk(t->op_pool, (void **)args->enq_ops, ops_count);
+ if (ret != 0)
+ goto next_model;
+
+retry:
+ ret = rte_mempool_get_bulk(t->model[fid].io_pool, (void **)args->reqs, ops_count);
+ if (ret != 0)
+ goto retry;
+
+ for (i = 0; i < ops_count; i++) {
+ args->enq_ops[i]->model_id = t->model[fid].id;
+ args->enq_ops[i]->nb_batches = t->model[fid].info.batch_size;
+ args->enq_ops[i]->mempool = t->op_pool;
+
+ args->enq_ops[i]->input.addr = args->reqs[i]->input;
+ args->enq_ops[i]->input.length = t->model[fid].inp_qsize;
+ args->enq_ops[i]->input.next = NULL;
+
+ args->enq_ops[i]->output.addr = args->reqs[i]->output;
+ args->enq_ops[i]->output.length = t->model[fid].out_qsize;
+ args->enq_ops[i]->output.next = NULL;
+
+ args->enq_ops[i]->user_ptr = args->reqs[i];
+ args->reqs[i]->niters++;
+ args->reqs[i]->fid = fid;
+ }
+
+ idx = 0;
+ pending = ops_count;
+
+enqueue_reqs:
+ burst_enq = rte_ml_enqueue_burst(t->cmn.opt->dev_id, 0, &args->enq_ops[idx], pending);
+ pending = pending - burst_enq;
+
+ if (pending > 0) {
+ idx = idx + burst_enq;
+ goto enqueue_reqs;
+ }
+
+ fid++;
+ if (fid <= args->end_fid)
+ goto next_model;
+
+ model_enq = model_enq + ops_count;
+ if (model_enq < args->nb_reqs)
+ goto next_rep;
+
+ return 0;
+}
+
+/* Dequeue inference requests with burst size greater than 1 */
+static int
+ml_dequeue_burst(void *arg)
+{
+ struct test_inference *t = ml_test_priv((struct ml_test *)arg);
+ struct rte_ml_op_error error;
+ struct ml_core_args *args;
+ struct ml_request *req;
+ uint64_t total_deq = 0;
+ uint16_t burst_deq = 0;
+ uint8_t nb_filelist;
+ uint32_t lcore_id;
+ uint32_t i;
+
+ lcore_id = rte_lcore_id();
+ args = &t->args[lcore_id];
+ nb_filelist = args->end_fid - args->start_fid + 1;
+
+ if (args->nb_reqs == 0)
+ return 0;
+
+dequeue_burst:
+ burst_deq =
+ rte_ml_dequeue_burst(t->cmn.opt->dev_id, 0, args->deq_ops, t->cmn.opt->burst_size);
+
+ if (likely(burst_deq > 0)) {
+ total_deq += burst_deq;
+
+ for (i = 0; i < burst_deq; i++) {
+ if (unlikely(args->deq_ops[i]->status == RTE_ML_OP_STATUS_ERROR)) {
+ rte_ml_op_error_get(t->cmn.opt->dev_id, args->deq_ops[i], &error);
+ ml_err("error_code = 0x%" PRIx64 ", error_message = %s\n",
+ error.errcode, error.message);
+ t->error_count[lcore_id]++;
+ }
+ req = (struct ml_request *)args->deq_ops[i]->user_ptr;
+ if (req != NULL)
+ rte_mempool_put(t->model[req->fid].io_pool, req);
+ }
+ rte_mempool_put_bulk(t->op_pool, (void *)args->deq_ops, burst_deq);
+ }
+
+ if (total_deq < args->nb_reqs * nb_filelist)
+ goto dequeue_burst;
+
+ return 0;
+}
+
bool
test_inference_cap_check(struct ml_options *opt)
{
@@ -173,6 +299,17 @@ test_inference_opt_check(struct ml_options *opt)
return -EINVAL;
}
+ if (opt->burst_size == 0) {
+ ml_err("Invalid option, burst_size = %u\n", opt->burst_size);
+ return -EINVAL;
+ }
+
+ if (opt->burst_size > ML_TEST_MAX_POOL_SIZE) {
+ ml_err("Invalid option, burst_size = %u (> max supported = %d)\n", opt->burst_size,
+ ML_TEST_MAX_POOL_SIZE);
+ return -EINVAL;
+ }
+
/* check number of available lcores. */
if (rte_lcore_count() < 3) {
ml_err("Insufficient lcores = %u\n", rte_lcore_count());
@@ -193,6 +330,7 @@ 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_begin("filelist");
for (i = 0; i < opt->nb_filelist; i++) {
@@ -208,6 +346,7 @@ test_inference_setup(struct ml_test *test, struct ml_options *opt)
{
struct test_inference *t;
void *test_inference;
+ uint32_t lcore_id;
int ret = 0;
uint32_t i;
@@ -233,13 +372,30 @@ test_inference_setup(struct ml_test *test, struct ml_options *opt)
goto error;
}
- t->enqueue = ml_enqueue_single;
- t->dequeue = ml_dequeue_single;
+ if (opt->burst_size == 1) {
+ t->enqueue = ml_enqueue_single;
+ t->dequeue = ml_dequeue_single;
+ } else {
+ t->enqueue = ml_enqueue_burst;
+ t->dequeue = ml_dequeue_burst;
+ }
/* set model initial state */
for (i = 0; i < opt->nb_filelist; i++)
t->model[i].state = MODEL_INITIAL;
+ for (lcore_id = 0; lcore_id < RTE_MAX_LCORE; lcore_id++) {
+ t->args[lcore_id].enq_ops = rte_zmalloc_socket(
+ "ml_test_enq_ops", opt->burst_size * sizeof(struct rte_ml_op *),
+ RTE_CACHE_LINE_SIZE, opt->socket_id);
+ t->args[lcore_id].deq_ops = rte_zmalloc_socket(
+ "ml_test_deq_ops", opt->burst_size * sizeof(struct rte_ml_op *),
+ RTE_CACHE_LINE_SIZE, opt->socket_id);
+ t->args[lcore_id].reqs = rte_zmalloc_socket(
+ "ml_test_requests", opt->burst_size * sizeof(struct ml_request *),
+ RTE_CACHE_LINE_SIZE, opt->socket_id);
+ }
+
return 0;
error:
@@ -22,6 +22,10 @@ struct ml_core_args {
uint64_t nb_reqs;
uint16_t start_fid;
uint16_t end_fid;
+
+ struct rte_ml_op **enq_ops;
+ struct rte_ml_op **deq_ops;
+ struct ml_request **reqs;
};
struct test_inference {
@@ -91,6 +91,10 @@ The following are the command-line options supported by the test application.
Set the number of inference repetitions to be executed in the test per each model. Default
value is `1`.
+* ``--burst_size <n>``
+
+ Set the burst size to be used when enqueuing / dequeuing inferences. Default value is `1`.
+
* ``--debug``
Enable the tests to run in debug mode.
@@ -236,6 +240,7 @@ Supported command line options for inference tests are following::
--socket_id
--filelist
--repetitions
+ --burst_size
List of files to be used for the inference tests can be specified through the option
@@ -244,6 +249,9 @@ List of files to be used for the inference tests can be specified through the op
list of files required to test with a single model. Multiple filelist entries are supported by
the test, one entry per model. Maximum number of file entries supported by the test is `8`.
+When ``--burst_size <num>`` option is specified for the test, enqueue and dequeue burst would
+try to enqueue or dequeue ``num`` number of inferences per each call respectively.
+
.. Note::
* The ``--filelist <file_list>`` is a mandatory option for running inference tests.
@@ -280,6 +288,14 @@ Example command to run inference_ordered test:
sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
--test=inference_ordered --filelist model.bin,input.bin,output.bin
+Example command to run inference_ordered test with a specific burst size:
+
+.. code-block:: console
+
+ sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
+ --test=inference_ordered --filelist model.bin,input.bin,output.bin \
+ --burst_size 12
+
INFERENCE_INTERLEAVE Test
~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -320,6 +336,14 @@ Example command to run inference_interleave test with multiple models:
--test=inference_interleave --filelist model_A.bin,input_A.bin,output_A.bin \
--filelist model_B.bin,input_B.bin,output_B.bin
+Example command to run inference_interleave test with a specific burst size:
+
+.. code-block:: console
+
+ sudo <build_dir>/app/dpdk-test-mldev -c 0xf -a <PCI_ID> -- \
+ --test=inference_interleave --filelist model.bin,input.bin,output.bin \
+ --burst_size 16
+
Debug mode
----------