Discussion:
[dpdk-dev] [PATCH 1/4] baseband: enhancement of offload cost test
(too old to reply)
Kamil Chalupnik
2018-12-07 14:31:23 UTC
Permalink
Offload cost test was improved in order to collect
more accurate results.

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/test_bbdev_perf.c | 152 +++++++++++------------
config/common_base | 2 +-
drivers/baseband/turbo_sw/bbdev_turbo_software.c | 70 ++++++++---
lib/librte_bbdev/rte_bbdev.h | 9 +-
4 files changed, 135 insertions(+), 98 deletions(-)

diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index fbe6cc9..bf97edb 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -88,19 +88,19 @@ struct thread_params {
/* Stores time statistics */
struct test_time_stats {
/* Stores software enqueue total working time */
- uint64_t enq_sw_tot_time;
+ uint64_t enq_sw_total_time;
/* Stores minimum value of software enqueue working time */
uint64_t enq_sw_min_time;
/* Stores maximum value of software enqueue working time */
uint64_t enq_sw_max_time;
/* Stores turbo enqueue total working time */
- uint64_t enq_tur_tot_time;
- /* Stores minimum value of turbo enqueue working time */
- uint64_t enq_tur_min_time;
- /* Stores maximum value of turbo enqueue working time */
- uint64_t enq_tur_max_time;
+ uint64_t enq_acc_total_time;
+ /* Stores minimum value of accelerator enqueue working time */
+ uint64_t enq_acc_min_time;
+ /* Stores maximum value of accelerator enqueue working time */
+ uint64_t enq_acc_max_time;
/* Stores dequeue total working time */
- uint64_t deq_tot_time;
+ uint64_t deq_total_time;
/* Stores minimum value of dequeue working time */
uint64_t deq_min_time;
/* Stores maximum value of dequeue working time */
@@ -1200,12 +1200,15 @@ typedef int (test_case_function)(struct active_device *ad,
burst_sz = tp->op_params->burst_sz;
num_to_process = tp->op_params->num_to_process;

- if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id, dec_ops,
burst_sz);
- else
+ rte_bbdev_dec_op_free_bulk(dec_ops, deq);
+ } else {
deq = rte_bbdev_dequeue_enc_ops(dev_id, queue_id, enc_ops,
burst_sz);
+ rte_bbdev_enc_op_free_bulk(enc_ops, deq);
+ }

if (deq < burst_sz) {
printf(
@@ -1316,8 +1319,6 @@ typedef int (test_case_function)(struct active_device *ad,

enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id, ops,
num_to_enq);
-
- rte_bbdev_dec_op_free_bulk(ops, num_to_enq);
}

if (allocs_failed > 0)
@@ -1380,8 +1381,6 @@ typedef int (test_case_function)(struct active_device *ad,

enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id, ops,
num_to_enq);
-
- rte_bbdev_enc_op_free_bulk(ops, num_to_enq);
}

if (allocs_failed > 0)
@@ -1575,13 +1574,14 @@ typedef int (test_case_function)(struct active_device *ad,
RTE_LCORE_FOREACH(lcore_id) {
if (iter++ >= used_cores)
break;
- printf("\tlcore_id: %u, throughput: %.8lg MOPS, %.8lg Mbps\n",
- lcore_id, t_params[lcore_id].mops, t_params[lcore_id].mbps);
+ printf("Throughput for core (%u): %.8lg MOPS, %.8lg Mbps\n",
+ lcore_id, t_params[lcore_id].mops,
+ t_params[lcore_id].mbps);
total_mops += t_params[lcore_id].mops;
total_mbps += t_params[lcore_id].mbps;
}
printf(
- "\n\tTotal stats for %u cores: throughput: %.8lg MOPS, %.8lg Mbps\n",
+ "\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps\n",
used_cores, total_mops, total_mbps);
}

@@ -1882,7 +1882,7 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_ASSERT_NOT_NULL(op_type_str, "Invalid op type: %u", op_type);

printf(
- "Validation/Latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
+ "\nValidation/Latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
info.dev_name, burst_sz, num_to_process, op_type_str);

if (op_type == RTE_BBDEV_OP_TURBO_DEC)
@@ -1899,10 +1899,10 @@ typedef int (test_case_function)(struct active_device *ad,
if (iter <= 0)
return TEST_FAILED;

- printf("\toperation latency:\n"
- "\t\tavg latency: %lg cycles, %lg us\n"
- "\t\tmin latency: %lg cycles, %lg us\n"
- "\t\tmax latency: %lg cycles, %lg us\n",
+ printf("Operation latency:\n"
+ "\tavg latency: %lg cycles, %lg us\n"
+ "\tmin latency: %lg cycles, %lg us\n"
+ "\tmax latency: %lg cycles, %lg us\n",
(double)total_time / (double)iter,
(double)(total_time * 1000000) / (double)iter /
(double)rte_get_tsc_hz(), (double)min_time,
@@ -1930,7 +1930,7 @@ typedef int (test_case_function)(struct active_device *ad,
stats->dequeued_count = q_stats->dequeued_count;
stats->enqueue_err_count = q_stats->enqueue_err_count;
stats->dequeue_err_count = q_stats->dequeue_err_count;
- stats->offload_time = q_stats->offload_time;
+ stats->acc_offload_cycles = q_stats->acc_offload_cycles;

return 0;
}
@@ -1974,18 +1974,18 @@ typedef int (test_case_function)(struct active_device *ad,
queue_id, dev_id);

enq_sw_last_time = rte_rdtsc_precise() - enq_start_time -
- stats.offload_time;
+ stats.acc_offload_cycles;
time_st->enq_sw_max_time = RTE_MAX(time_st->enq_sw_max_time,
enq_sw_last_time);
time_st->enq_sw_min_time = RTE_MIN(time_st->enq_sw_min_time,
enq_sw_last_time);
- time_st->enq_sw_tot_time += enq_sw_last_time;
+ time_st->enq_sw_total_time += enq_sw_last_time;

- time_st->enq_tur_max_time = RTE_MAX(time_st->enq_tur_max_time,
- stats.offload_time);
- time_st->enq_tur_min_time = RTE_MIN(time_st->enq_tur_min_time,
- stats.offload_time);
- time_st->enq_tur_tot_time += stats.offload_time;
+ time_st->enq_acc_max_time = RTE_MAX(time_st->enq_acc_max_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_min_time = RTE_MIN(time_st->enq_acc_min_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_total_time += stats.acc_offload_cycles;

/* ensure enqueue has been completed */
rte_delay_ms(10);
@@ -2003,7 +2003,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time);
time_st->deq_min_time = RTE_MIN(time_st->deq_min_time,
deq_last_time);
- time_st->deq_tot_time += deq_last_time;
+ time_st->deq_total_time += deq_last_time;

/* Dequeue remaining operations if needed*/
while (burst_sz != deq)
@@ -2055,18 +2055,18 @@ typedef int (test_case_function)(struct active_device *ad,
queue_id, dev_id);

enq_sw_last_time = rte_rdtsc_precise() - enq_start_time -
- stats.offload_time;
+ stats.acc_offload_cycles;
time_st->enq_sw_max_time = RTE_MAX(time_st->enq_sw_max_time,
enq_sw_last_time);
time_st->enq_sw_min_time = RTE_MIN(time_st->enq_sw_min_time,
enq_sw_last_time);
- time_st->enq_sw_tot_time += enq_sw_last_time;
+ time_st->enq_sw_total_time += enq_sw_last_time;

- time_st->enq_tur_max_time = RTE_MAX(time_st->enq_tur_max_time,
- stats.offload_time);
- time_st->enq_tur_min_time = RTE_MIN(time_st->enq_tur_min_time,
- stats.offload_time);
- time_st->enq_tur_tot_time += stats.offload_time;
+ time_st->enq_acc_max_time = RTE_MAX(time_st->enq_acc_max_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_min_time = RTE_MIN(time_st->enq_acc_min_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_total_time += stats.acc_offload_cycles;

/* ensure enqueue has been completed */
rte_delay_ms(10);
@@ -2084,7 +2084,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time);
time_st->deq_min_time = RTE_MIN(time_st->deq_min_time,
deq_last_time);
- time_st->deq_tot_time += deq_last_time;
+ time_st->deq_total_time += deq_last_time;

while (burst_sz != deq)
deq += rte_bbdev_dequeue_enc_ops(dev_id, queue_id,
@@ -2121,7 +2121,7 @@ typedef int (test_case_function)(struct active_device *ad,

memset(&time_st, 0, sizeof(struct test_time_stats));
time_st.enq_sw_min_time = UINT64_MAX;
- time_st.enq_tur_min_time = UINT64_MAX;
+ time_st.enq_acc_min_time = UINT64_MAX;
time_st.deq_min_time = UINT64_MAX;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
@@ -2134,7 +2134,7 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_ASSERT_NOT_NULL(op_type_str, "Invalid op type: %u", op_type);

printf(
- "Offload latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
+ "\nOffload latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
info.dev_name, burst_sz, num_to_process, op_type_str);

if (op_type == RTE_BBDEV_OP_TURBO_DEC)
@@ -2149,36 +2149,36 @@ typedef int (test_case_function)(struct active_device *ad,
if (iter <= 0)
return TEST_FAILED;

- printf("\tenq offload cost latency:\n"
- "\t\tsoftware avg %lg cycles, %lg us\n"
- "\t\tsoftware min %lg cycles, %lg us\n"
- "\t\tsoftware max %lg cycles, %lg us\n"
- "\t\tturbo avg %lg cycles, %lg us\n"
- "\t\tturbo min %lg cycles, %lg us\n"
- "\t\tturbo max %lg cycles, %lg us\n",
- (double)time_st.enq_sw_tot_time / (double)iter,
- (double)(time_st.enq_sw_tot_time * 1000000) /
+ printf("Enqueue offload cost latency:\n"
+ "\tDriver offload avg %lg cycles, %lg us\n"
+ "\tDriver offload min %lg cycles, %lg us\n"
+ "\tDriver offload max %lg cycles, %lg us\n"
+ "\tAccelerator offload avg %lg cycles, %lg us\n"
+ "\tAccelerator offload min %lg cycles, %lg us\n"
+ "\tAccelerator offload max %lg cycles, %lg us\n",
+ (double)time_st.enq_sw_total_time / (double)iter,
+ (double)(time_st.enq_sw_total_time * 1000000) /
(double)iter / (double)rte_get_tsc_hz(),
(double)time_st.enq_sw_min_time,
(double)(time_st.enq_sw_min_time * 1000000) /
rte_get_tsc_hz(), (double)time_st.enq_sw_max_time,
(double)(time_st.enq_sw_max_time * 1000000) /
- rte_get_tsc_hz(), (double)time_st.enq_tur_tot_time /
+ rte_get_tsc_hz(), (double)time_st.enq_acc_total_time /
(double)iter,
- (double)(time_st.enq_tur_tot_time * 1000000) /
+ (double)(time_st.enq_acc_total_time * 1000000) /
(double)iter / (double)rte_get_tsc_hz(),
- (double)time_st.enq_tur_min_time,
- (double)(time_st.enq_tur_min_time * 1000000) /
- rte_get_tsc_hz(), (double)time_st.enq_tur_max_time,
- (double)(time_st.enq_tur_max_time * 1000000) /
+ (double)time_st.enq_acc_min_time,
+ (double)(time_st.enq_acc_min_time * 1000000) /
+ rte_get_tsc_hz(), (double)time_st.enq_acc_max_time,
+ (double)(time_st.enq_acc_max_time * 1000000) /
rte_get_tsc_hz());

- printf("\tdeq offload cost latency - one op:\n"
- "\t\tavg %lg cycles, %lg us\n"
- "\t\tmin %lg cycles, %lg us\n"
- "\t\tmax %lg cycles, %lg us\n",
- (double)time_st.deq_tot_time / (double)iter,
- (double)(time_st.deq_tot_time * 1000000) /
+ printf("Dequeue offload cost latency - one op:\n"
+ "\tavg %lg cycles, %lg us\n"
+ "\tmin %lg cycles, %lg us\n"
+ "\tmax %lg cycles, %lg us\n",
+ (double)time_st.deq_total_time / (double)iter,
+ (double)(time_st.deq_total_time * 1000000) /
(double)iter / (double)rte_get_tsc_hz(),
(double)time_st.deq_min_time,
(double)(time_st.deq_min_time * 1000000) /
@@ -2194,7 +2194,7 @@ typedef int (test_case_function)(struct active_device *ad,
static int
offload_latency_empty_q_test_dec(uint16_t dev_id, uint16_t queue_id,
const uint16_t num_to_process, uint16_t burst_sz,
- uint64_t *deq_tot_time, uint64_t *deq_min_time,
+ uint64_t *deq_total_time, uint64_t *deq_min_time,
uint64_t *deq_max_time)
{
int i, deq_total;
@@ -2214,7 +2214,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time = rte_rdtsc_precise() - deq_start_time;
*deq_max_time = RTE_MAX(*deq_max_time, deq_last_time);
*deq_min_time = RTE_MIN(*deq_min_time, deq_last_time);
- *deq_tot_time += deq_last_time;
+ *deq_total_time += deq_last_time;
}

return i;
@@ -2223,7 +2223,7 @@ typedef int (test_case_function)(struct active_device *ad,
static int
offload_latency_empty_q_test_enc(uint16_t dev_id, uint16_t queue_id,
const uint16_t num_to_process, uint16_t burst_sz,
- uint64_t *deq_tot_time, uint64_t *deq_min_time,
+ uint64_t *deq_total_time, uint64_t *deq_min_time,
uint64_t *deq_max_time)
{
int i, deq_total;
@@ -2242,7 +2242,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time = rte_rdtsc_precise() - deq_start_time;
*deq_max_time = RTE_MAX(*deq_max_time, deq_last_time);
*deq_min_time = RTE_MIN(*deq_min_time, deq_last_time);
- *deq_tot_time += deq_last_time;
+ *deq_total_time += deq_last_time;
}

return i;
@@ -2261,7 +2261,7 @@ typedef int (test_case_function)(struct active_device *ad,
return TEST_SKIPPED;
#else
int iter;
- uint64_t deq_tot_time, deq_min_time, deq_max_time;
+ uint64_t deq_total_time, deq_min_time, deq_max_time;
uint16_t burst_sz = op_params->burst_sz;
const uint16_t num_to_process = op_params->num_to_process;
const enum rte_bbdev_op_type op_type = test_vector.op_type;
@@ -2269,7 +2269,7 @@ typedef int (test_case_function)(struct active_device *ad,
struct rte_bbdev_info info;
const char *op_type_str;

- deq_tot_time = deq_max_time = 0;
+ deq_total_time = deq_max_time = 0;
deq_min_time = UINT64_MAX;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
@@ -2281,27 +2281,27 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_ASSERT_NOT_NULL(op_type_str, "Invalid op type: %u", op_type);

printf(
- "Offload latency empty dequeue test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
+ "\nOffload latency empty dequeue test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
info.dev_name, burst_sz, num_to_process, op_type_str);

if (op_type == RTE_BBDEV_OP_TURBO_DEC)
iter = offload_latency_empty_q_test_dec(ad->dev_id, queue_id,
- num_to_process, burst_sz, &deq_tot_time,
+ num_to_process, burst_sz, &deq_total_time,
&deq_min_time, &deq_max_time);
else
iter = offload_latency_empty_q_test_enc(ad->dev_id, queue_id,
- num_to_process, burst_sz, &deq_tot_time,
+ num_to_process, burst_sz, &deq_total_time,
&deq_min_time, &deq_max_time);

if (iter <= 0)
return TEST_FAILED;

- printf("\tempty deq offload\n"
- "\t\tavg. latency: %lg cycles, %lg us\n"
- "\t\tmin. latency: %lg cycles, %lg us\n"
- "\t\tmax. latency: %lg cycles, %lg us\n",
- (double)deq_tot_time / (double)iter,
- (double)(deq_tot_time * 1000000) / (double)iter /
+ printf("Empty dequeue offload\n"
+ "\tavg. latency: %lg cycles, %lg us\n"
+ "\tmin. latency: %lg cycles, %lg us\n"
+ "\tmax. latency: %lg cycles, %lg us\n",
+ (double)deq_total_time / (double)iter,
+ (double)(deq_total_time * 1000000) / (double)iter /
(double)rte_get_tsc_hz(), (double)deq_min_time,
(double)(deq_min_time * 1000000) / rte_get_tsc_hz(),
(double)deq_max_time, (double)(deq_max_time * 1000000) /
diff --git a/config/common_base b/config/common_base
index d12ae98..3ff98bb 100644
--- a/config/common_base
+++ b/config/common_base
@@ -481,7 +481,7 @@ CONFIG_RTE_PMD_PACKET_PREFETCH=y
#
CONFIG_RTE_LIBRTE_BBDEV=y
CONFIG_RTE_BBDEV_MAX_DEVS=128
-CONFIG_RTE_BBDEV_OFFLOAD_COST=n
+CONFIG_RTE_BBDEV_OFFLOAD_COST=y

#
# Compile PMD for NULL bbdev device
diff --git a/drivers/baseband/turbo_sw/bbdev_turbo_software.c b/drivers/baseband/turbo_sw/bbdev_turbo_software.c
index 8ceb276..57f6ba1 100644
--- a/drivers/baseband/turbo_sw/bbdev_turbo_software.c
+++ b/drivers/baseband/turbo_sw/bbdev_turbo_software.c
@@ -510,9 +510,10 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
+ /* CRC24A generation */
bblib_lte_crc24a_gen(&crc_req, &crc_resp);
#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif
} else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
/* CRC24B */
@@ -542,9 +543,10 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
+ /* CRC24B generation */
bblib_lte_crc24b_gen(&crc_req, &crc_resp);
#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif
} else {
ret = is_enc_input_valid(k, k_idx, total_left);
@@ -596,15 +598,14 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
-
+ /* Turbo encoding */
if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
op->status |= 1 << RTE_BBDEV_DRV_ERROR;
rte_bbdev_log(ERR, "Turbo Encoder failed");
return;
}
-
#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif

/* Restore 3 first bytes of next CB if they were overwritten by CRC*/
@@ -671,23 +672,21 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
-
+ /* Rate-Matching */
if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
op->status |= 1 << RTE_BBDEV_DRV_ERROR;
rte_bbdev_log(ERR, "Rate matching failed");
return;
}
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif

/* SW fills an entire last byte even if E%8 != 0. Clear the
* superfluous data bits for consistency with HW device.
*/
mask_id = (e & 7) >> 1;
rm_out[out_len - 1] &= mask_out[mask_id];
-
-#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
-#endif
-
enc->output.length += rm_resp.OutputLen;
} else {
/* Rate matching is bypassed */
@@ -798,7 +797,7 @@ struct turbo_sw_queue {
{
uint16_t i;
#ifdef RTE_BBDEV_OFFLOAD_COST
- queue_stats->offload_time = 0;
+ queue_stats->acc_offload_cycles = 0;
#endif

for (i = 0; i < nb_ops; ++i)
@@ -905,7 +904,8 @@ struct turbo_sw_queue {
process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
- bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left)
+ bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left,
+ struct rte_bbdev_stats *q_stats)
{
int ret;
int32_t k_idx;
@@ -917,6 +917,11 @@ struct turbo_sw_queue {
struct bblib_turbo_decoder_request turbo_req;
struct bblib_turbo_decoder_response turbo_resp;
struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ uint64_t start_time;
+#else
+ RTE_SET_USED(q_stats);
+#endif

k_idx = compute_idx(k);

@@ -942,7 +947,14 @@ struct turbo_sw_queue {
deint_req.pharqbuffer = q->deint_input;
deint_req.ncb = ncb_without_null;
deint_resp.pinteleavebuffer = q->deint_output;
+
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ start_time = rte_rdtsc_precise();
+#endif
bblib_deinterleave_ul(&deint_req, &deint_resp);
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif
} else
move_padding_bytes(in, q->deint_output, k, ncb);

@@ -961,7 +973,15 @@ struct turbo_sw_queue {
adapter_req.ncb = ncb_without_null;
adapter_req.pinteleavebuffer = adapter_input;
adapter_resp.pharqout = q->adapter_output;
+
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ start_time = rte_rdtsc_precise();
+#endif
+ /* Turbo decode adaptation */
bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif

out = (uint8_t *)rte_pktmbuf_append(m_out, ((k - crc24_overlap) >> 3));
if (out == NULL) {
@@ -986,12 +1006,20 @@ struct turbo_sw_queue {
turbo_resp.ag_buf = q->ag;
turbo_resp.cb_buf = q->code_block;
turbo_resp.output = out;
+
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ start_time = rte_rdtsc_precise();
+#endif
+ /* Turbo decode */
iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif
dec->hard_output.length += (k >> 3);

if (iter_cnt > 0) {
/* Temporary solution for returned iter_count from SDK */
- iter_cnt = (iter_cnt - 1) / 2;
+ iter_cnt = (iter_cnt - 1) >> 1;
dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
} else {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
@@ -1001,7 +1029,8 @@ struct turbo_sw_queue {
}

static inline void
-enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op)
+enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
+ struct rte_bbdev_stats *queue_stats)
{
uint8_t c, r = 0;
uint16_t kw, k = 0;
@@ -1053,7 +1082,7 @@ struct turbo_sw_queue {
process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
out_offset, check_bit(dec->op_flags,
RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap,
- total_left);
+ total_left, queue_stats);
/* To keep CRC24 attached to end of Code block, use
* RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it
* removed by default once verified.
@@ -1075,12 +1104,15 @@ struct turbo_sw_queue {

static inline uint16_t
enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
- uint16_t nb_ops)
+ uint16_t nb_ops, struct rte_bbdev_stats *queue_stats)
{
uint16_t i;
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ queue_stats->acc_offload_cycles = 0;
+#endif

for (i = 0; i < nb_ops; ++i)
- enqueue_dec_one_op(q, ops[i]);
+ enqueue_dec_one_op(q, ops[i], queue_stats);

return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
NULL);
@@ -1112,7 +1144,7 @@ struct turbo_sw_queue {
struct turbo_sw_queue *q = queue;
uint16_t nb_enqueued = 0;

- nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops);
+ nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops, &q_data->queue_stats);

q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
q_data->queue_stats.enqueued_count += nb_enqueued;
diff --git a/lib/librte_bbdev/rte_bbdev.h b/lib/librte_bbdev/rte_bbdev.h
index 25ef409..da8cf07 100644
--- a/lib/librte_bbdev/rte_bbdev.h
+++ b/lib/librte_bbdev/rte_bbdev.h
@@ -239,8 +239,13 @@ struct rte_bbdev_stats {
uint64_t enqueue_err_count;
/** Total error count on operations dequeued */
uint64_t dequeue_err_count;
- /** Offload time */
- uint64_t offload_time;
+ /** CPU cycles consumed by the (HW/SW) accelerator device to offload
+ * the enqueue request to its internal queues.
+ * - For a HW device this is the cycles consumed in MMIO write
+ * - For a SW (vdev) device, this is the processing time of the
+ * bbdev operation
+ */
+ uint64_t acc_offload_cycles;
};

/**
--
1.8.3.1
Kamil Chalupnik
2018-12-07 14:31:24 UTC
Permalink
Improvements added to throughput test:
- test is run in loop (number of iterations is specified by
TEST_REPETITIONS define) which ensures more accurate results
- length of input data is calculated based on amount of CBs in TB
- maximum number of decoding iterations is gathered from results
- added new functions responsible for printing results
- small fixes for memory management

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/main.c | 2 -
app/test-bbdev/test_bbdev_perf.c | 471 ++++++++++++++++++++-----------------
app/test-bbdev/test_bbdev_vector.c | 7 +
lib/librte_bbdev/rte_bbdev_op.h | 2 +
4 files changed, 263 insertions(+), 219 deletions(-)

diff --git a/app/test-bbdev/main.c b/app/test-bbdev/main.c
index 41b54bb..7af2522 100644
--- a/app/test-bbdev/main.c
+++ b/app/test-bbdev/main.c
@@ -316,8 +316,6 @@
return 1;
}

- rte_log_set_global_level(RTE_LOG_INFO);
-
/* If no argument provided - run all tests */
if (test_params.num_tests == 0)
return run_all_tests();
diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index bf97edb..a25e3a7 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -23,6 +23,7 @@
#define GET_SOCKET(socket_id) (((socket_id) == SOCKET_ID_ANY) ? 0 : (socket_id))

#define MAX_QUEUES RTE_MAX_LCORE
+#define TEST_REPETITIONS 1000

#define OPS_CACHE_SIZE 256U
#define OPS_POOL_SIZE_MIN 511U /* 0.5K per queue */
@@ -77,8 +78,9 @@ struct thread_params {
uint8_t dev_id;
uint16_t queue_id;
uint64_t start_time;
- double mops;
+ double ops_per_sec;
double mbps;
+ uint8_t iter_count;
rte_atomic16_t nb_dequeued;
rte_atomic16_t processing_status;
struct test_op_params *op_params;
@@ -757,6 +759,8 @@ typedef int (test_case_function)(struct active_device *ad,
turbo_dec->tb_params.c_neg;
ops[i]->turbo_dec.tb_params.cab =
turbo_dec->tb_params.cab;
+ ops[i]->turbo_dec.tb_params.r =
+ turbo_dec->tb_params.r;
} else {
ops[i]->turbo_dec.cb_params.e = turbo_dec->cb_params.e;
ops[i]->turbo_dec.cb_params.k = turbo_dec->cb_params.k;
@@ -884,47 +888,6 @@ typedef int (test_case_function)(struct active_device *ad,
}

static int
-validate_dec_buffers(struct rte_bbdev_dec_op *ref_op, struct test_buffers *bufs,
- const uint16_t num_to_process)
-{
- int i;
-
- struct op_data_entries *hard_data_orig =
- &test_vector.entries[DATA_HARD_OUTPUT];
- struct op_data_entries *soft_data_orig =
- &test_vector.entries[DATA_SOFT_OUTPUT];
-
- for (i = 0; i < num_to_process; i++) {
- TEST_ASSERT_SUCCESS(validate_op_chain(&bufs->hard_outputs[i],
- hard_data_orig),
- "Hard output buffers are not equal");
- if (ref_op->turbo_dec.op_flags &
- RTE_BBDEV_TURBO_SOFT_OUTPUT)
- TEST_ASSERT_SUCCESS(validate_op_chain(
- &bufs->soft_outputs[i],
- soft_data_orig),
- "Soft output buffers are not equal");
- }
-
- return TEST_SUCCESS;
-}
-
-static int
-validate_enc_buffers(struct test_buffers *bufs, const uint16_t num_to_process)
-{
- int i;
-
- struct op_data_entries *hard_data_orig =
- &test_vector.entries[DATA_HARD_OUTPUT];
-
- for (i = 0; i < num_to_process; i++)
- TEST_ASSERT_SUCCESS(validate_op_chain(&bufs->hard_outputs[i],
- hard_data_orig), "");
-
- return TEST_SUCCESS;
-}
-
-static int
validate_dec_op(struct rte_bbdev_dec_op **ops, const uint16_t n,
struct rte_bbdev_dec_op *ref_op, const int vector_mask)
{
@@ -1016,6 +979,44 @@ typedef int (test_case_function)(struct active_device *ad,
entry->segments[i].length;
}

+static uint32_t
+calc_dec_TB_size(struct rte_bbdev_dec_op *op)
+{
+ uint8_t i;
+ uint32_t c, r, tb_size = 0;
+
+ if (op->turbo_dec.code_block_mode) {
+ tb_size = op->turbo_dec.tb_params.k_neg;
+ } else {
+ c = op->turbo_dec.tb_params.c;
+ r = op->turbo_dec.tb_params.r;
+ for (i = 0; i < c-r; i++)
+ tb_size += (r < op->turbo_dec.tb_params.c_neg) ?
+ op->turbo_dec.tb_params.k_neg :
+ op->turbo_dec.tb_params.k_pos;
+ }
+ return tb_size;
+}
+
+static uint32_t
+calc_enc_TB_size(struct rte_bbdev_enc_op *op)
+{
+ uint8_t i;
+ uint32_t c, r, tb_size = 0;
+
+ if (op->turbo_enc.code_block_mode) {
+ tb_size = op->turbo_enc.tb_params.k_neg;
+ } else {
+ c = op->turbo_enc.tb_params.c;
+ r = op->turbo_enc.tb_params.r;
+ for (i = 0; i < c-r; i++)
+ tb_size += (r < op->turbo_enc.tb_params.c_neg) ?
+ op->turbo_enc.tb_params.k_neg :
+ op->turbo_enc.tb_params.k_pos;
+ }
+ return tb_size;
+}
+
static int
init_test_op_params(struct test_op_params *op_params,
enum rte_bbdev_op_type op_type, const int expected_status,
@@ -1163,17 +1164,13 @@ typedef int (test_case_function)(struct active_device *ad,
int ret;
uint16_t i;
uint64_t total_time;
- uint16_t deq, burst_sz, num_to_process;
+ uint16_t deq, burst_sz, num_ops;
uint16_t queue_id = INVALID_QUEUE_ID;
struct rte_bbdev_dec_op *dec_ops[MAX_BURST];
struct rte_bbdev_enc_op *enc_ops[MAX_BURST];
- struct test_buffers *bufs;
struct rte_bbdev_info info;

- /* Input length in bytes, million operations per second,
- * million bits per second.
- */
- double in_len;
+ double tb_len_bits;

struct thread_params *tp = cb_arg;
RTE_SET_USED(ret_param);
@@ -1198,7 +1195,7 @@ typedef int (test_case_function)(struct active_device *ad,
}

burst_sz = tp->op_params->burst_sz;
- num_to_process = tp->op_params->num_to_process;
+ num_ops = tp->op_params->num_to_process;

if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id, dec_ops,
@@ -1218,7 +1215,7 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- if (rte_atomic16_read(&tp->nb_dequeued) + deq < num_to_process) {
+ if (rte_atomic16_read(&tp->nb_dequeued) + deq < num_ops) {
rte_atomic16_add(&tp->nb_dequeued, deq);
return;
}
@@ -1227,14 +1224,18 @@ typedef int (test_case_function)(struct active_device *ad,

rte_bbdev_info_get(dev_id, &info);

- bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];
-
ret = TEST_SUCCESS;
- if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
- ret = validate_dec_buffers(tp->op_params->ref_dec_op, bufs,
- num_to_process);
- else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC)
- ret = validate_enc_buffers(bufs, num_to_process);
+
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
+ struct rte_bbdev_dec_op *ref_op = tp->op_params->ref_dec_op;
+ ret = validate_dec_op(dec_ops, num_ops, ref_op,
+ tp->op_params->vector_mask);
+ rte_bbdev_dec_op_free_bulk(dec_ops, deq);
+ } else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC) {
+ struct rte_bbdev_enc_op *ref_op = tp->op_params->ref_enc_op;
+ ret = validate_enc_op(enc_ops, num_ops, ref_op);
+ rte_bbdev_enc_op_free_bulk(enc_ops, deq);
+ }

if (ret) {
printf("Buffers validation failed\n");
@@ -1243,13 +1244,13 @@ typedef int (test_case_function)(struct active_device *ad,

switch (test_vector.op_type) {
case RTE_BBDEV_OP_TURBO_DEC:
- in_len = tp->op_params->ref_dec_op->turbo_dec.input.length;
+ tb_len_bits = calc_dec_TB_size(tp->op_params->ref_dec_op);
break;
case RTE_BBDEV_OP_TURBO_ENC:
- in_len = tp->op_params->ref_enc_op->turbo_enc.input.length;
+ tb_len_bits = calc_enc_TB_size(tp->op_params->ref_enc_op);
break;
case RTE_BBDEV_OP_NONE:
- in_len = 0.0;
+ tb_len_bits = 0.0;
break;
default:
printf("Unknown op type: %d\n", test_vector.op_type);
@@ -1257,9 +1258,9 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- tp->mops = ((double)num_to_process / 1000000.0) /
+ tp->ops_per_sec = ((double)num_ops) /
((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = ((double)num_to_process * in_len * 8 / 1000000.0) /
+ tp->mbps = (((double)(num_ops * tb_len_bits)) / 1000000.0) /
((double)total_time / (double)rte_get_tsc_hz());

rte_atomic16_add(&tp->nb_dequeued, deq);
@@ -1270,14 +1271,14 @@ typedef int (test_case_function)(struct active_device *ad,
{
struct thread_params *tp = arg;
unsigned int enqueued;
- struct rte_bbdev_dec_op *ops[MAX_BURST];
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
const uint16_t num_to_process = tp->op_params->num_to_process;
+ struct rte_bbdev_dec_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
struct rte_bbdev_info info;
int ret;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1287,6 +1288,11 @@ typedef int (test_case_function)(struct active_device *ad,
tp->dev_id, queue_id);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_to_process > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

rte_atomic16_clear(&tp->processing_status);
@@ -1295,36 +1301,27 @@ typedef int (test_case_function)(struct active_device *ad,
while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

+ ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp, ops,
+ num_to_process);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops",
+ num_to_process);
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_dec_op(ops, num_to_process, 0, bufs->inputs,
+ bufs->hard_outputs, bufs->soft_outputs,
+ tp->op_params->ref_dec_op);
+
tp->start_time = rte_rdtsc_precise();
for (enqueued = 0; enqueued < num_to_process;) {

- uint16_t num_to_enq = burst_sz;
+ num_to_enq = burst_sz;

if (unlikely(num_to_process - enqueued < num_to_enq))
num_to_enq = num_to_process - enqueued;

- ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp, ops,
- num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- continue;
- }
-
- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_dec_op(ops, num_to_enq, enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- bufs->soft_outputs,
- tp->op_params->ref_dec_op);
-
- enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id, ops,
- num_to_enq);
+ enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id,
+ &ops[enqueued], num_to_enq);
}

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
-
return TEST_SUCCESS;
}

@@ -1333,14 +1330,14 @@ typedef int (test_case_function)(struct active_device *ad,
{
struct thread_params *tp = arg;
unsigned int enqueued;
- struct rte_bbdev_enc_op *ops[MAX_BURST];
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
const uint16_t num_to_process = tp->op_params->num_to_process;
+ struct rte_bbdev_enc_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
struct rte_bbdev_info info;
int ret;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1350,6 +1347,11 @@ typedef int (test_case_function)(struct active_device *ad,
tp->dev_id, queue_id);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_to_process > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

rte_atomic16_clear(&tp->processing_status);
@@ -1358,35 +1360,26 @@ typedef int (test_case_function)(struct active_device *ad,
while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

+ ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp, ops,
+ num_to_process);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops",
+ num_to_process);
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_enc_op(ops, num_to_process, 0, bufs->inputs,
+ bufs->hard_outputs, tp->op_params->ref_enc_op);
+
tp->start_time = rte_rdtsc_precise();
for (enqueued = 0; enqueued < num_to_process;) {

- uint16_t num_to_enq = burst_sz;
+ num_to_enq = burst_sz;

if (unlikely(num_to_process - enqueued < num_to_enq))
num_to_enq = num_to_process - enqueued;

- ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp, ops,
- num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- continue;
- }
-
- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_enc_op(ops, num_to_enq, enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- tp->op_params->ref_enc_op);
-
- enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id, ops,
- num_to_enq);
+ enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id,
+ &ops[enqueued], num_to_enq);
}

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
-
return TEST_SUCCESS;
}

@@ -1394,86 +1387,97 @@ typedef int (test_case_function)(struct active_device *ad,
throughput_pmd_lcore_dec(void *arg)
{
struct thread_params *tp = arg;
- unsigned int enqueued, dequeued;
- struct rte_bbdev_dec_op *ops_enq[MAX_BURST], *ops_deq[MAX_BURST];
- uint64_t total_time, start_time;
+ uint16_t enq, deq;
+ uint64_t total_time = 0, start_time;
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
- const uint16_t num_to_process = tp->op_params->num_to_process;
+ const uint16_t num_ops = tp->op_params->num_to_process;
+ struct rte_bbdev_dec_op *ops_enq[num_ops];
+ struct rte_bbdev_dec_op *ops_deq[num_ops];
struct rte_bbdev_dec_op *ref_op = tp->op_params->ref_dec_op;
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
- int ret;
+ int i, j, ret;
struct rte_bbdev_info info;
-
- /* Input length in bytes, million operations per second, million bits
- * per second.
- */
- double in_len;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_ops > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

- start_time = rte_rdtsc_precise();
- for (enqueued = 0, dequeued = 0; dequeued < num_to_process;) {
- uint16_t deq;
+ ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp, ops_enq, num_ops);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops", num_ops);

- if (likely(enqueued < num_to_process)) {
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_dec_op(ops_enq, num_ops, 0, bufs->inputs,
+ bufs->hard_outputs, bufs->soft_outputs, ref_op);

- uint16_t num_to_enq = burst_sz;
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_ops; ++j)
+ ops_enq[j]->opaque_data = (void *)(uintptr_t)j;

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ for (i = 0; i < TEST_REPETITIONS; ++i) {

- ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp,
- ops_enq, num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- goto do_dequeue;
- }
+ for (j = 0; j < num_ops; ++j) {
+ struct rte_bbdev_dec_op *op = ops_enq[j];
+ rte_pktmbuf_reset(op->turbo_dec.hard_output.data);
+ }
+
+ start_time = rte_rdtsc_precise();
+
+ for (enq = 0, deq = 0; enq < num_ops;) {
+ num_to_enq = burst_sz;
+
+ if (unlikely(num_ops - enq < num_to_enq))
+ num_to_enq = num_ops - enq;

- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_dec_op(ops_enq, num_to_enq,
- enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- bufs->soft_outputs,
- ref_op);
+ enq += rte_bbdev_enqueue_dec_ops(tp->dev_id,
+ queue_id, &ops_enq[enq], num_to_enq);

- enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id,
- queue_id, ops_enq, num_to_enq);
+ deq += rte_bbdev_dequeue_dec_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
}
-do_dequeue:
- deq = rte_bbdev_dequeue_dec_ops(tp->dev_id, queue_id, ops_deq,
- burst_sz);
- dequeued += deq;
- rte_bbdev_dec_op_free_bulk(ops_enq, deq);
- }
- total_time = rte_rdtsc_precise() - start_time;

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
+ /* dequeue the remaining */
+ while (deq < enq) {
+ deq += rte_bbdev_dequeue_dec_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
+ }

- TEST_ASSERT(enqueued == dequeued, "enqueued (%u) != dequeued (%u)",
- enqueued, dequeued);
+ total_time += rte_rdtsc_precise() - start_time;
+ }
+
+ tp->iter_count = 0;
+ /* get the max of iter_count for all dequeued ops */
+ for (i = 0; i < num_ops; ++i) {
+ tp->iter_count = RTE_MAX(ops_enq[i]->turbo_dec.iter_count,
+ tp->iter_count);
+ }

if (test_vector.op_type != RTE_BBDEV_OP_NONE) {
- ret = validate_dec_buffers(ref_op, bufs, num_to_process);
- TEST_ASSERT_SUCCESS(ret, "Buffers validation failed");
+ ret = validate_dec_op(ops_deq, num_ops, ref_op,
+ tp->op_params->vector_mask);
+ TEST_ASSERT_SUCCESS(ret, "Validation failed!");
}

- in_len = ref_op->turbo_dec.input.length;
- tp->mops = ((double)num_to_process / 1000000.0) /
- ((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = ((double)num_to_process * in_len * 8 / 1000000.0) /
+ rte_bbdev_dec_op_free_bulk(ops_enq, num_ops);
+
+ double tb_len_bits = calc_dec_TB_size(ref_op);
+
+ tp->ops_per_sec = ((double)num_ops * TEST_REPETITIONS) /
((double)total_time / (double)rte_get_tsc_hz());
+ tp->mbps = (((double)(num_ops * TEST_REPETITIONS * tb_len_bits)) /
+ 1000000.0) / ((double)total_time /
+ (double)rte_get_tsc_hz());

return TEST_SUCCESS;
}
@@ -1482,91 +1486,94 @@ typedef int (test_case_function)(struct active_device *ad,
throughput_pmd_lcore_enc(void *arg)
{
struct thread_params *tp = arg;
- unsigned int enqueued, dequeued;
- struct rte_bbdev_enc_op *ops_enq[MAX_BURST], *ops_deq[MAX_BURST];
- uint64_t total_time, start_time;
+ uint16_t enq, deq;
+ uint64_t total_time = 0, start_time;
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
- const uint16_t num_to_process = tp->op_params->num_to_process;
+ const uint16_t num_ops = tp->op_params->num_to_process;
+ struct rte_bbdev_enc_op *ops_enq[num_ops];
+ struct rte_bbdev_enc_op *ops_deq[num_ops];
struct rte_bbdev_enc_op *ref_op = tp->op_params->ref_enc_op;
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
- int ret;
+ int i, j, ret;
struct rte_bbdev_info info;
-
- /* Input length in bytes, million operations per second, million bits
- * per second.
- */
- double in_len;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_ops > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

- start_time = rte_rdtsc_precise();
- for (enqueued = 0, dequeued = 0; dequeued < num_to_process;) {
- uint16_t deq;
+ ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp, ops_enq,
+ num_ops);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops",
+ num_ops);
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_enc_op(ops_enq, num_ops, 0, bufs->inputs,
+ bufs->hard_outputs, ref_op);

- if (likely(enqueued < num_to_process)) {
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_ops; ++j)
+ ops_enq[j]->opaque_data = (void *)(uintptr_t)j;

- uint16_t num_to_enq = burst_sz;
+ for (i = 0; i < TEST_REPETITIONS; ++i) {

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ for (j = 0; j < num_ops; ++j)
+ rte_pktmbuf_reset(
+ ops_enq[j]->turbo_enc.output.data);

- ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp,
- ops_enq, num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- goto do_dequeue;
- }
+ start_time = rte_rdtsc_precise();
+
+ for (enq = 0, deq = 0; enq < num_ops;) {
+ num_to_enq = burst_sz;

- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_enc_op(ops_enq, num_to_enq,
- enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- ref_op);
+ if (unlikely(num_ops - enq < num_to_enq))
+ num_to_enq = num_ops - enq;

- enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id,
- queue_id, ops_enq, num_to_enq);
+ enq += rte_bbdev_enqueue_enc_ops(tp->dev_id,
+ queue_id, &ops_enq[enq], num_to_enq);
+
+ deq += rte_bbdev_dequeue_enc_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
}
-do_dequeue:
- deq = rte_bbdev_dequeue_enc_ops(tp->dev_id, queue_id, ops_deq,
- burst_sz);
- dequeued += deq;
- rte_bbdev_enc_op_free_bulk(ops_enq, deq);
- }
- total_time = rte_rdtsc_precise() - start_time;

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
+ /* dequeue the remaining */
+ while (deq < enq) {
+ deq += rte_bbdev_dequeue_enc_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
+ }

- TEST_ASSERT(enqueued == dequeued, "enqueued (%u) != dequeued (%u)",
- enqueued, dequeued);
+ total_time += rte_rdtsc_precise() - start_time;
+ }

if (test_vector.op_type != RTE_BBDEV_OP_NONE) {
- ret = validate_enc_buffers(bufs, num_to_process);
- TEST_ASSERT_SUCCESS(ret, "Buffers validation failed");
+ ret = validate_enc_op(ops_deq, num_ops, ref_op);
+ TEST_ASSERT_SUCCESS(ret, "Validation failed!");
}

- in_len = ref_op->turbo_enc.input.length;
+ double tb_len_bits = calc_enc_TB_size(ref_op);

- tp->mops = ((double)num_to_process / 1000000.0) /
- ((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = ((double)num_to_process * in_len * 8 / 1000000.0) /
+ tp->ops_per_sec = ((double)num_ops * TEST_REPETITIONS) /
((double)total_time / (double)rte_get_tsc_hz());
+ tp->mbps = (((double)(num_ops * TEST_REPETITIONS * tb_len_bits))
+ / 1000000.0) / ((double)total_time /
+ (double)rte_get_tsc_hz());

return TEST_SUCCESS;
}
+
static void
-print_throughput(struct thread_params *t_params, unsigned int used_cores)
+print_enc_throughput(struct thread_params *t_params, unsigned int used_cores)
{
unsigned int lcore_id, iter = 0;
double total_mops = 0, total_mbps = 0;
@@ -1574,10 +1581,11 @@ typedef int (test_case_function)(struct active_device *ad,
RTE_LCORE_FOREACH(lcore_id) {
if (iter++ >= used_cores)
break;
- printf("Throughput for core (%u): %.8lg MOPS, %.8lg Mbps\n",
- lcore_id, t_params[lcore_id].mops,
+ printf(
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps\n",
+ lcore_id, t_params[lcore_id].ops_per_sec,
t_params[lcore_id].mbps);
- total_mops += t_params[lcore_id].mops;
+ total_mops += t_params[lcore_id].ops_per_sec;
total_mbps += t_params[lcore_id].mbps;
}
printf(
@@ -1585,6 +1593,30 @@ typedef int (test_case_function)(struct active_device *ad,
used_cores, total_mops, total_mbps);
}

+static void
+print_dec_throughput(struct thread_params *t_params, unsigned int used_cores)
+{
+ unsigned int lcore_id, iter = 0;
+ double total_mops = 0, total_mbps = 0;
+ uint8_t iter_count = 0;
+
+ RTE_LCORE_FOREACH(lcore_id) {
+ if (iter++ >= used_cores)
+ break;
+ printf(
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps @ max %u iterations\n",
+ lcore_id, t_params[lcore_id].ops_per_sec,
+ t_params[lcore_id].mbps,
+ t_params[lcore_id].iter_count);
+ total_mops += t_params[lcore_id].ops_per_sec;
+ total_mbps += t_params[lcore_id].mbps;
+ iter_count = RTE_MAX(iter_count, t_params[lcore_id].iter_count);
+ }
+ printf(
+ "\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps @ max %u iterations\n",
+ used_cores, total_mops, total_mbps, iter_count);
+}
+
/*
* Test function that determines how long an enqueue + dequeue of a burst
* takes on available lcores.
@@ -1677,8 +1709,10 @@ typedef int (test_case_function)(struct active_device *ad,

/* Print throughput if interrupts are disabled and test passed */
if (!intr_enabled) {
- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- print_throughput(t_params, num_lcores);
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ print_dec_throughput(t_params, num_lcores);
+ else
+ print_enc_throughput(t_params, num_lcores);
return ret;
}

@@ -1713,9 +1747,12 @@ typedef int (test_case_function)(struct active_device *ad,
}

/* Print throughput if test passed */
- if (!ret && test_vector.op_type != RTE_BBDEV_OP_NONE)
- print_throughput(t_params, num_lcores);
-
+ if (!ret) {
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ print_dec_throughput(t_params, num_lcores);
+ else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC)
+ print_enc_throughput(t_params, num_lcores);
+ }
return ret;
}

diff --git a/app/test-bbdev/test_bbdev_vector.c b/app/test-bbdev/test_bbdev_vector.c
index 81b8ee7..45fe999 100644
--- a/app/test-bbdev/test_bbdev_vector.c
+++ b/app/test-bbdev/test_bbdev_vector.c
@@ -412,6 +412,10 @@
vector->mask |= TEST_BBDEV_VF_NUM_MAPS;
turbo_dec->num_maps = (uint8_t) strtoul(token, &err, 0);
ret = ((err == NULL) || (*err != '\0')) ? -1 : 0;
+ } else if (!strcmp(key_token, "r")) {
+ vector->mask |= TEST_BBDEV_VF_R;
+ turbo_dec->tb_params.r = (uint8_t) strtoul(token, &err, 0);
+ ret = ((err == NULL) || (*err != '\0')) ? -1 : 0;
} else if (!strcmp(key_token, "code_block_mode")) {
vector->mask |= TEST_BBDEV_VF_CODE_BLOCK_MODE;
turbo_dec->code_block_mode = (uint8_t) strtoul(token, &err, 0);
@@ -714,6 +718,9 @@
if (!(mask & TEST_BBDEV_VF_CAB))
printf(
"WARNING: cab was not specified in vector file and will be set to 0\n");
+ if (!(mask & TEST_BBDEV_VF_R))
+ printf(
+ "WARNING: r was not specified in vector file and will be set to 0\n");
} else {
if (!(mask & TEST_BBDEV_VF_E))
printf(
diff --git a/lib/librte_bbdev/rte_bbdev_op.h b/lib/librte_bbdev/rte_bbdev_op.h
index 83f62c2..962e2ed 100644
--- a/lib/librte_bbdev/rte_bbdev_op.h
+++ b/lib/librte_bbdev/rte_bbdev_op.h
@@ -216,6 +216,8 @@ struct rte_bbdev_op_dec_tb_params {
* operation when r >= cab
*/
uint32_t eb;
+ /**< The index of the first CB in the inbound mbuf data, default is 0 */
+ uint8_t r;
};

/**< Operation structure for Turbo decode.
--
1.8.3.1
Kamil Chalupnik
2018-12-07 14:31:26 UTC
Permalink
Improvements added to interrupt test:
- test is run in loop (number of iterations is specified by
TEST_REPETITIONS define) which ensures more accurate results
- mapping cores to thread parameteres was put in order.
Master core is always set at first index. It fixes problem with
running test for only one core

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/test_bbdev_perf.c | 246 +++++++++++++++++++++++++--------------
1 file changed, 161 insertions(+), 85 deletions(-)

diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index 5bec70d..1c4a645 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -77,13 +77,17 @@ struct test_op_params {
struct thread_params {
uint8_t dev_id;
uint16_t queue_id;
+ uint32_t lcore_id;
uint64_t start_time;
double ops_per_sec;
double mbps;
uint8_t iter_count;
rte_atomic16_t nb_dequeued;
rte_atomic16_t processing_status;
+ rte_atomic16_t burst_sz;
struct test_op_params *op_params;
+ struct rte_bbdev_dec_op *dec_ops[MAX_BURST];
+ struct rte_bbdev_enc_op *enc_ops[MAX_BURST];
};

#ifdef RTE_BBDEV_OFFLOAD_COST
@@ -1206,16 +1210,12 @@ typedef int (test_case_function)(struct active_device *ad,
uint16_t i;
uint64_t total_time;
uint16_t deq, burst_sz, num_ops;
- uint16_t queue_id = INVALID_QUEUE_ID;
- struct rte_bbdev_dec_op *dec_ops[MAX_BURST];
- struct rte_bbdev_enc_op *enc_ops[MAX_BURST];
+ uint16_t queue_id = *(uint16_t *) ret_param;
struct rte_bbdev_info info;

double tb_len_bits;

struct thread_params *tp = cb_arg;
- RTE_SET_USED(ret_param);
- queue_id = tp->queue_id;

/* Find matching thread params using queue_id */
for (i = 0; i < MAX_QUEUES; ++i, ++tp)
@@ -1235,18 +1235,19 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- burst_sz = tp->op_params->burst_sz;
+ burst_sz = rte_atomic16_read(&tp->burst_sz);
num_ops = tp->op_params->num_to_process;

- if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
- deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id, dec_ops,
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id,
+ &tp->dec_ops[
+ rte_atomic16_read(&tp->nb_dequeued)],
burst_sz);
- rte_bbdev_dec_op_free_bulk(dec_ops, deq);
- } else {
- deq = rte_bbdev_dequeue_enc_ops(dev_id, queue_id, enc_ops,
+ else
+ deq = rte_bbdev_dequeue_enc_ops(dev_id, queue_id,
+ &tp->enc_ops[
+ rte_atomic16_read(&tp->nb_dequeued)],
burst_sz);
- rte_bbdev_enc_op_free_bulk(enc_ops, deq);
- }

if (deq < burst_sz) {
printf(
@@ -1269,13 +1270,18 @@ typedef int (test_case_function)(struct active_device *ad,

if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
struct rte_bbdev_dec_op *ref_op = tp->op_params->ref_dec_op;
- ret = validate_dec_op(dec_ops, num_ops, ref_op,
+ ret = validate_dec_op(tp->dec_ops, num_ops, ref_op,
tp->op_params->vector_mask);
- rte_bbdev_dec_op_free_bulk(dec_ops, deq);
+ /* get the max of iter_count for all dequeued ops */
+ for (i = 0; i < num_ops; ++i)
+ tp->iter_count = RTE_MAX(
+ tp->dec_ops[i]->turbo_dec.iter_count,
+ tp->iter_count);
+ rte_bbdev_dec_op_free_bulk(tp->dec_ops, deq);
} else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC) {
struct rte_bbdev_enc_op *ref_op = tp->op_params->ref_enc_op;
- ret = validate_enc_op(enc_ops, num_ops, ref_op);
- rte_bbdev_enc_op_free_bulk(enc_ops, deq);
+ ret = validate_enc_op(tp->enc_ops, num_ops, ref_op);
+ rte_bbdev_enc_op_free_bulk(tp->enc_ops, deq);
}

if (ret) {
@@ -1299,9 +1305,9 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- tp->ops_per_sec = ((double)num_ops) /
+ tp->ops_per_sec += ((double)num_ops) /
((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = (((double)(num_ops * tb_len_bits)) / 1000000.0) /
+ tp->mbps += (((double)(num_ops * tb_len_bits)) / 1000000.0) /
((double)total_time / (double)rte_get_tsc_hz());

rte_atomic16_add(&tp->nb_dequeued, deq);
@@ -1318,8 +1324,8 @@ typedef int (test_case_function)(struct active_device *ad,
struct rte_bbdev_dec_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
struct rte_bbdev_info info;
- int ret;
- uint16_t num_to_enq;
+ int ret, i, j;
+ uint16_t num_to_enq, enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1351,16 +1357,47 @@ typedef int (test_case_function)(struct active_device *ad,
bufs->hard_outputs, bufs->soft_outputs,
tp->op_params->ref_dec_op);

- tp->start_time = rte_rdtsc_precise();
- for (enqueued = 0; enqueued < num_to_process;) {
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_to_process; ++j)
+ ops[j]->opaque_data = (void *)(uintptr_t)j;

- num_to_enq = burst_sz;
+ for (j = 0; j < TEST_REPETITIONS; ++j) {
+ for (i = 0; i < num_to_process; ++i)
+ rte_pktmbuf_reset(ops[i]->turbo_dec.hard_output.data);

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ tp->start_time = rte_rdtsc_precise();
+ for (enqueued = 0; enqueued < num_to_process;) {
+ num_to_enq = burst_sz;
+
+ if (unlikely(num_to_process - enqueued < num_to_enq))
+ num_to_enq = num_to_process - enqueued;
+
+ enq = 0;
+ do {
+ enq += rte_bbdev_enqueue_dec_ops(tp->dev_id,
+ queue_id, &ops[enqueued],
+ num_to_enq);
+ } while (unlikely(num_to_enq != enq));
+ enqueued += enq;
+
+ /* Write to thread burst_sz current number of enqueued
+ * descriptors. It ensures that proper number of
+ * descriptors will be dequeued in callback
+ * function - needed for last batch in case where
+ * the number of operations is not a multiple of
+ * burst size.
+ */
+ rte_atomic16_set(&tp->burst_sz, num_to_enq);

- enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id,
- &ops[enqueued], num_to_enq);
+ /* Wait until processing of previous batch is
+ * completed.
+ */
+ while (rte_atomic16_read(&tp->nb_dequeued) !=
+ (int16_t) enqueued)
+ rte_pause();
+ }
+ if (j != TEST_REPETITIONS - 1)
+ rte_atomic16_clear(&tp->nb_dequeued);
}

return TEST_SUCCESS;
@@ -1377,8 +1414,8 @@ typedef int (test_case_function)(struct active_device *ad,
struct rte_bbdev_enc_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
struct rte_bbdev_info info;
- int ret;
- uint16_t num_to_enq;
+ int ret, i, j;
+ uint16_t num_to_enq, enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1409,16 +1446,47 @@ typedef int (test_case_function)(struct active_device *ad,
copy_reference_enc_op(ops, num_to_process, 0, bufs->inputs,
bufs->hard_outputs, tp->op_params->ref_enc_op);

- tp->start_time = rte_rdtsc_precise();
- for (enqueued = 0; enqueued < num_to_process;) {
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_to_process; ++j)
+ ops[j]->opaque_data = (void *)(uintptr_t)j;
+
+ for (j = 0; j < TEST_REPETITIONS; ++j) {
+ for (i = 0; i < num_to_process; ++i)
+ rte_pktmbuf_reset(ops[i]->turbo_enc.output.data);

- num_to_enq = burst_sz;
+ tp->start_time = rte_rdtsc_precise();
+ for (enqueued = 0; enqueued < num_to_process;) {
+ num_to_enq = burst_sz;

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ if (unlikely(num_to_process - enqueued < num_to_enq))
+ num_to_enq = num_to_process - enqueued;
+
+ enq = 0;
+ do {
+ enq += rte_bbdev_enqueue_enc_ops(tp->dev_id,
+ queue_id, &ops[enqueued],
+ num_to_enq);
+ } while (unlikely(enq != num_to_enq));
+ enqueued += enq;
+
+ /* Write to thread burst_sz current number of enqueued
+ * descriptors. It ensures that proper number of
+ * descriptors will be dequeued in callback
+ * function - needed for last batch in case where
+ * the number of operations is not a multiple of
+ * burst size.
+ */
+ rte_atomic16_set(&tp->burst_sz, num_to_enq);

- enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id,
- &ops[enqueued], num_to_enq);
+ /* Wait until processing of previous batch is
+ * completed.
+ */
+ while (rte_atomic16_read(&tp->nb_dequeued) !=
+ (int16_t) enqueued)
+ rte_pause();
+ }
+ if (j != TEST_REPETITIONS - 1)
+ rte_atomic16_clear(&tp->nb_dequeued);
}

return TEST_SUCCESS;
@@ -1613,18 +1681,16 @@ typedef int (test_case_function)(struct active_device *ad,
static void
print_enc_throughput(struct thread_params *t_params, unsigned int used_cores)
{
- unsigned int lcore_id, iter = 0;
+ unsigned int iter = 0;
double total_mops = 0, total_mbps = 0;

- RTE_LCORE_FOREACH(lcore_id) {
- if (iter++ >= used_cores)
- break;
+ for (iter = 0; iter < used_cores; iter++) {
printf(
- "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps\n",
- lcore_id, t_params[lcore_id].ops_per_sec,
- t_params[lcore_id].mbps);
- total_mops += t_params[lcore_id].ops_per_sec;
- total_mbps += t_params[lcore_id].mbps;
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps\n",
+ t_params[iter].lcore_id, t_params[iter].ops_per_sec,
+ t_params[iter].mbps);
+ total_mops += t_params[iter].ops_per_sec;
+ total_mbps += t_params[iter].mbps;
}
printf(
"\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps\n",
@@ -1634,21 +1700,18 @@ typedef int (test_case_function)(struct active_device *ad,
static void
print_dec_throughput(struct thread_params *t_params, unsigned int used_cores)
{
- unsigned int lcore_id, iter = 0;
+ unsigned int iter = 0;
double total_mops = 0, total_mbps = 0;
uint8_t iter_count = 0;

- RTE_LCORE_FOREACH(lcore_id) {
- if (iter++ >= used_cores)
- break;
+ for (iter = 0; iter < used_cores; iter++) {
printf(
- "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps @ max %u iterations\n",
- lcore_id, t_params[lcore_id].ops_per_sec,
- t_params[lcore_id].mbps,
- t_params[lcore_id].iter_count);
- total_mops += t_params[lcore_id].ops_per_sec;
- total_mbps += t_params[lcore_id].mbps;
- iter_count = RTE_MAX(iter_count, t_params[lcore_id].iter_count);
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps @ max %u iterations\n",
+ t_params[iter].lcore_id, t_params[iter].ops_per_sec,
+ t_params[iter].mbps, t_params[iter].iter_count);
+ total_mops += t_params[iter].ops_per_sec;
+ total_mbps += t_params[iter].mbps;
+ iter_count = RTE_MAX(iter_count, t_params[iter].iter_count);
}
printf(
"\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps @ max %u iterations\n",
@@ -1665,10 +1728,9 @@ typedef int (test_case_function)(struct active_device *ad,
{
int ret;
unsigned int lcore_id, used_cores = 0;
- struct thread_params t_params[MAX_QUEUES];
+ struct thread_params *t_params, *tp;
struct rte_bbdev_info info;
lcore_function_t *throughput_function;
- struct thread_params *tp;
uint16_t num_lcores;
const char *op_type_str;

@@ -1691,6 +1753,13 @@ typedef int (test_case_function)(struct active_device *ad,
? ad->nb_queues
: op_params->num_lcores;

+ /* Allocate memory for thread parameters structure */
+ t_params = rte_zmalloc(NULL, num_lcores * sizeof(struct thread_params),
+ RTE_CACHE_LINE_SIZE);
+ TEST_ASSERT_NOT_NULL(t_params, "Failed to alloc %zuB for t_params",
+ RTE_ALIGN(sizeof(struct thread_params) * num_lcores,
+ RTE_CACHE_LINE_SIZE));
+
if (intr_enabled) {
if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
throughput_function = throughput_intr_lcore_dec;
@@ -1700,9 +1769,11 @@ typedef int (test_case_function)(struct active_device *ad,
/* Dequeue interrupt callback registration */
ret = rte_bbdev_callback_register(ad->dev_id,
RTE_BBDEV_EVENT_DEQUEUE, dequeue_event_callback,
- &t_params);
- if (ret < 0)
+ t_params);
+ if (ret < 0) {
+ rte_free(t_params);
return ret;
+ }
} else {
if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
throughput_function = throughput_pmd_lcore_dec;
@@ -1712,38 +1783,39 @@ typedef int (test_case_function)(struct active_device *ad,

rte_atomic16_set(&op_params->sync, SYNC_WAIT);

- t_params[rte_lcore_id()].dev_id = ad->dev_id;
- t_params[rte_lcore_id()].op_params = op_params;
- t_params[rte_lcore_id()].queue_id =
- ad->queue_ids[used_cores++];
+ /* Master core is set at first entry */
+ t_params[0].dev_id = ad->dev_id;
+ t_params[0].lcore_id = rte_lcore_id();
+ t_params[0].op_params = op_params;
+ t_params[0].queue_id = ad->queue_ids[used_cores++];
+ t_params[0].iter_count = 0;

RTE_LCORE_FOREACH_SLAVE(lcore_id) {
if (used_cores >= num_lcores)
break;

- t_params[lcore_id].dev_id = ad->dev_id;
- t_params[lcore_id].op_params = op_params;
- t_params[lcore_id].queue_id = ad->queue_ids[used_cores++];
+ t_params[used_cores].dev_id = ad->dev_id;
+ t_params[used_cores].lcore_id = lcore_id;
+ t_params[used_cores].op_params = op_params;
+ t_params[used_cores].queue_id = ad->queue_ids[used_cores];
+ t_params[used_cores].iter_count = 0;

- rte_eal_remote_launch(throughput_function, &t_params[lcore_id],
- lcore_id);
+ rte_eal_remote_launch(throughput_function,
+ &t_params[used_cores++], lcore_id);
}

rte_atomic16_set(&op_params->sync, SYNC_START);
- ret = throughput_function(&t_params[rte_lcore_id()]);
+ ret = throughput_function(&t_params[0]);

/* Master core is always used */
- used_cores = 1;
- RTE_LCORE_FOREACH_SLAVE(lcore_id) {
- if (used_cores++ >= num_lcores)
- break;
-
- ret |= rte_eal_wait_lcore(lcore_id);
- }
+ for (used_cores = 1; used_cores < num_lcores; used_cores++)
+ ret |= rte_eal_wait_lcore(t_params[used_cores].lcore_id);

/* Return if test failed */
- if (ret)
+ if (ret) {
+ rte_free(t_params);
return ret;
+ }

/* Print throughput if interrupts are disabled and test passed */
if (!intr_enabled) {
@@ -1751,6 +1823,7 @@ typedef int (test_case_function)(struct active_device *ad,
print_dec_throughput(t_params, num_lcores);
else
print_enc_throughput(t_params, num_lcores);
+ rte_free(t_params);
return ret;
}

@@ -1759,21 +1832,20 @@ typedef int (test_case_function)(struct active_device *ad,
* error using processing_status variable.
* Wait for master lcore operations.
*/
- tp = &t_params[rte_lcore_id()];
+ tp = &t_params[0];
while ((rte_atomic16_read(&tp->nb_dequeued) <
op_params->num_to_process) &&
(rte_atomic16_read(&tp->processing_status) !=
TEST_FAILED))
rte_pause();

+ tp->ops_per_sec /= TEST_REPETITIONS;
+ tp->mbps /= TEST_REPETITIONS;
ret |= rte_atomic16_read(&tp->processing_status);

/* Wait for slave lcores operations */
- used_cores = 1;
- RTE_LCORE_FOREACH_SLAVE(lcore_id) {
- tp = &t_params[lcore_id];
- if (used_cores++ >= num_lcores)
- break;
+ for (used_cores = 1; used_cores < num_lcores; used_cores++) {
+ tp = &t_params[used_cores];

while ((rte_atomic16_read(&tp->nb_dequeued) <
op_params->num_to_process) &&
@@ -1781,6 +1853,8 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_FAILED))
rte_pause();

+ tp->ops_per_sec /= TEST_REPETITIONS;
+ tp->mbps /= TEST_REPETITIONS;
ret |= rte_atomic16_read(&tp->processing_status);
}

@@ -1791,6 +1865,8 @@ typedef int (test_case_function)(struct active_device *ad,
else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC)
print_enc_throughput(t_params, num_lcores);
}
+
+ rte_free(t_params);
return ret;
}
--
1.8.3.1
Kamil Chalupnik
2018-12-07 15:15:31 UTC
Permalink
Offload cost test was improved in order to collect
more accurate results.

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/test_bbdev_perf.c | 152 +++++++++++------------
config/common_base | 2 +-
drivers/baseband/turbo_sw/bbdev_turbo_software.c | 70 ++++++++---
lib/librte_bbdev/rte_bbdev.h | 9 +-
4 files changed, 135 insertions(+), 98 deletions(-)

diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index fbe6cc9..bf97edb 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -88,19 +88,19 @@ struct thread_params {
/* Stores time statistics */
struct test_time_stats {
/* Stores software enqueue total working time */
- uint64_t enq_sw_tot_time;
+ uint64_t enq_sw_total_time;
/* Stores minimum value of software enqueue working time */
uint64_t enq_sw_min_time;
/* Stores maximum value of software enqueue working time */
uint64_t enq_sw_max_time;
/* Stores turbo enqueue total working time */
- uint64_t enq_tur_tot_time;
- /* Stores minimum value of turbo enqueue working time */
- uint64_t enq_tur_min_time;
- /* Stores maximum value of turbo enqueue working time */
- uint64_t enq_tur_max_time;
+ uint64_t enq_acc_total_time;
+ /* Stores minimum value of accelerator enqueue working time */
+ uint64_t enq_acc_min_time;
+ /* Stores maximum value of accelerator enqueue working time */
+ uint64_t enq_acc_max_time;
/* Stores dequeue total working time */
- uint64_t deq_tot_time;
+ uint64_t deq_total_time;
/* Stores minimum value of dequeue working time */
uint64_t deq_min_time;
/* Stores maximum value of dequeue working time */
@@ -1200,12 +1200,15 @@ typedef int (test_case_function)(struct active_device *ad,
burst_sz = tp->op_params->burst_sz;
num_to_process = tp->op_params->num_to_process;

- if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id, dec_ops,
burst_sz);
- else
+ rte_bbdev_dec_op_free_bulk(dec_ops, deq);
+ } else {
deq = rte_bbdev_dequeue_enc_ops(dev_id, queue_id, enc_ops,
burst_sz);
+ rte_bbdev_enc_op_free_bulk(enc_ops, deq);
+ }

if (deq < burst_sz) {
printf(
@@ -1316,8 +1319,6 @@ typedef int (test_case_function)(struct active_device *ad,

enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id, ops,
num_to_enq);
-
- rte_bbdev_dec_op_free_bulk(ops, num_to_enq);
}

if (allocs_failed > 0)
@@ -1380,8 +1381,6 @@ typedef int (test_case_function)(struct active_device *ad,

enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id, ops,
num_to_enq);
-
- rte_bbdev_enc_op_free_bulk(ops, num_to_enq);
}

if (allocs_failed > 0)
@@ -1575,13 +1574,14 @@ typedef int (test_case_function)(struct active_device *ad,
RTE_LCORE_FOREACH(lcore_id) {
if (iter++ >= used_cores)
break;
- printf("\tlcore_id: %u, throughput: %.8lg MOPS, %.8lg Mbps\n",
- lcore_id, t_params[lcore_id].mops, t_params[lcore_id].mbps);
+ printf("Throughput for core (%u): %.8lg MOPS, %.8lg Mbps\n",
+ lcore_id, t_params[lcore_id].mops,
+ t_params[lcore_id].mbps);
total_mops += t_params[lcore_id].mops;
total_mbps += t_params[lcore_id].mbps;
}
printf(
- "\n\tTotal stats for %u cores: throughput: %.8lg MOPS, %.8lg Mbps\n",
+ "\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps\n",
used_cores, total_mops, total_mbps);
}

@@ -1882,7 +1882,7 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_ASSERT_NOT_NULL(op_type_str, "Invalid op type: %u", op_type);

printf(
- "Validation/Latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
+ "\nValidation/Latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
info.dev_name, burst_sz, num_to_process, op_type_str);

if (op_type == RTE_BBDEV_OP_TURBO_DEC)
@@ -1899,10 +1899,10 @@ typedef int (test_case_function)(struct active_device *ad,
if (iter <= 0)
return TEST_FAILED;

- printf("\toperation latency:\n"
- "\t\tavg latency: %lg cycles, %lg us\n"
- "\t\tmin latency: %lg cycles, %lg us\n"
- "\t\tmax latency: %lg cycles, %lg us\n",
+ printf("Operation latency:\n"
+ "\tavg latency: %lg cycles, %lg us\n"
+ "\tmin latency: %lg cycles, %lg us\n"
+ "\tmax latency: %lg cycles, %lg us\n",
(double)total_time / (double)iter,
(double)(total_time * 1000000) / (double)iter /
(double)rte_get_tsc_hz(), (double)min_time,
@@ -1930,7 +1930,7 @@ typedef int (test_case_function)(struct active_device *ad,
stats->dequeued_count = q_stats->dequeued_count;
stats->enqueue_err_count = q_stats->enqueue_err_count;
stats->dequeue_err_count = q_stats->dequeue_err_count;
- stats->offload_time = q_stats->offload_time;
+ stats->acc_offload_cycles = q_stats->acc_offload_cycles;

return 0;
}
@@ -1974,18 +1974,18 @@ typedef int (test_case_function)(struct active_device *ad,
queue_id, dev_id);

enq_sw_last_time = rte_rdtsc_precise() - enq_start_time -
- stats.offload_time;
+ stats.acc_offload_cycles;
time_st->enq_sw_max_time = RTE_MAX(time_st->enq_sw_max_time,
enq_sw_last_time);
time_st->enq_sw_min_time = RTE_MIN(time_st->enq_sw_min_time,
enq_sw_last_time);
- time_st->enq_sw_tot_time += enq_sw_last_time;
+ time_st->enq_sw_total_time += enq_sw_last_time;

- time_st->enq_tur_max_time = RTE_MAX(time_st->enq_tur_max_time,
- stats.offload_time);
- time_st->enq_tur_min_time = RTE_MIN(time_st->enq_tur_min_time,
- stats.offload_time);
- time_st->enq_tur_tot_time += stats.offload_time;
+ time_st->enq_acc_max_time = RTE_MAX(time_st->enq_acc_max_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_min_time = RTE_MIN(time_st->enq_acc_min_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_total_time += stats.acc_offload_cycles;

/* ensure enqueue has been completed */
rte_delay_ms(10);
@@ -2003,7 +2003,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time);
time_st->deq_min_time = RTE_MIN(time_st->deq_min_time,
deq_last_time);
- time_st->deq_tot_time += deq_last_time;
+ time_st->deq_total_time += deq_last_time;

/* Dequeue remaining operations if needed*/
while (burst_sz != deq)
@@ -2055,18 +2055,18 @@ typedef int (test_case_function)(struct active_device *ad,
queue_id, dev_id);

enq_sw_last_time = rte_rdtsc_precise() - enq_start_time -
- stats.offload_time;
+ stats.acc_offload_cycles;
time_st->enq_sw_max_time = RTE_MAX(time_st->enq_sw_max_time,
enq_sw_last_time);
time_st->enq_sw_min_time = RTE_MIN(time_st->enq_sw_min_time,
enq_sw_last_time);
- time_st->enq_sw_tot_time += enq_sw_last_time;
+ time_st->enq_sw_total_time += enq_sw_last_time;

- time_st->enq_tur_max_time = RTE_MAX(time_st->enq_tur_max_time,
- stats.offload_time);
- time_st->enq_tur_min_time = RTE_MIN(time_st->enq_tur_min_time,
- stats.offload_time);
- time_st->enq_tur_tot_time += stats.offload_time;
+ time_st->enq_acc_max_time = RTE_MAX(time_st->enq_acc_max_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_min_time = RTE_MIN(time_st->enq_acc_min_time,
+ stats.acc_offload_cycles);
+ time_st->enq_acc_total_time += stats.acc_offload_cycles;

/* ensure enqueue has been completed */
rte_delay_ms(10);
@@ -2084,7 +2084,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time);
time_st->deq_min_time = RTE_MIN(time_st->deq_min_time,
deq_last_time);
- time_st->deq_tot_time += deq_last_time;
+ time_st->deq_total_time += deq_last_time;

while (burst_sz != deq)
deq += rte_bbdev_dequeue_enc_ops(dev_id, queue_id,
@@ -2121,7 +2121,7 @@ typedef int (test_case_function)(struct active_device *ad,

memset(&time_st, 0, sizeof(struct test_time_stats));
time_st.enq_sw_min_time = UINT64_MAX;
- time_st.enq_tur_min_time = UINT64_MAX;
+ time_st.enq_acc_min_time = UINT64_MAX;
time_st.deq_min_time = UINT64_MAX;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
@@ -2134,7 +2134,7 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_ASSERT_NOT_NULL(op_type_str, "Invalid op type: %u", op_type);

printf(
- "Offload latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
+ "\nOffload latency test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
info.dev_name, burst_sz, num_to_process, op_type_str);

if (op_type == RTE_BBDEV_OP_TURBO_DEC)
@@ -2149,36 +2149,36 @@ typedef int (test_case_function)(struct active_device *ad,
if (iter <= 0)
return TEST_FAILED;

- printf("\tenq offload cost latency:\n"
- "\t\tsoftware avg %lg cycles, %lg us\n"
- "\t\tsoftware min %lg cycles, %lg us\n"
- "\t\tsoftware max %lg cycles, %lg us\n"
- "\t\tturbo avg %lg cycles, %lg us\n"
- "\t\tturbo min %lg cycles, %lg us\n"
- "\t\tturbo max %lg cycles, %lg us\n",
- (double)time_st.enq_sw_tot_time / (double)iter,
- (double)(time_st.enq_sw_tot_time * 1000000) /
+ printf("Enqueue offload cost latency:\n"
+ "\tDriver offload avg %lg cycles, %lg us\n"
+ "\tDriver offload min %lg cycles, %lg us\n"
+ "\tDriver offload max %lg cycles, %lg us\n"
+ "\tAccelerator offload avg %lg cycles, %lg us\n"
+ "\tAccelerator offload min %lg cycles, %lg us\n"
+ "\tAccelerator offload max %lg cycles, %lg us\n",
+ (double)time_st.enq_sw_total_time / (double)iter,
+ (double)(time_st.enq_sw_total_time * 1000000) /
(double)iter / (double)rte_get_tsc_hz(),
(double)time_st.enq_sw_min_time,
(double)(time_st.enq_sw_min_time * 1000000) /
rte_get_tsc_hz(), (double)time_st.enq_sw_max_time,
(double)(time_st.enq_sw_max_time * 1000000) /
- rte_get_tsc_hz(), (double)time_st.enq_tur_tot_time /
+ rte_get_tsc_hz(), (double)time_st.enq_acc_total_time /
(double)iter,
- (double)(time_st.enq_tur_tot_time * 1000000) /
+ (double)(time_st.enq_acc_total_time * 1000000) /
(double)iter / (double)rte_get_tsc_hz(),
- (double)time_st.enq_tur_min_time,
- (double)(time_st.enq_tur_min_time * 1000000) /
- rte_get_tsc_hz(), (double)time_st.enq_tur_max_time,
- (double)(time_st.enq_tur_max_time * 1000000) /
+ (double)time_st.enq_acc_min_time,
+ (double)(time_st.enq_acc_min_time * 1000000) /
+ rte_get_tsc_hz(), (double)time_st.enq_acc_max_time,
+ (double)(time_st.enq_acc_max_time * 1000000) /
rte_get_tsc_hz());

- printf("\tdeq offload cost latency - one op:\n"
- "\t\tavg %lg cycles, %lg us\n"
- "\t\tmin %lg cycles, %lg us\n"
- "\t\tmax %lg cycles, %lg us\n",
- (double)time_st.deq_tot_time / (double)iter,
- (double)(time_st.deq_tot_time * 1000000) /
+ printf("Dequeue offload cost latency - one op:\n"
+ "\tavg %lg cycles, %lg us\n"
+ "\tmin %lg cycles, %lg us\n"
+ "\tmax %lg cycles, %lg us\n",
+ (double)time_st.deq_total_time / (double)iter,
+ (double)(time_st.deq_total_time * 1000000) /
(double)iter / (double)rte_get_tsc_hz(),
(double)time_st.deq_min_time,
(double)(time_st.deq_min_time * 1000000) /
@@ -2194,7 +2194,7 @@ typedef int (test_case_function)(struct active_device *ad,
static int
offload_latency_empty_q_test_dec(uint16_t dev_id, uint16_t queue_id,
const uint16_t num_to_process, uint16_t burst_sz,
- uint64_t *deq_tot_time, uint64_t *deq_min_time,
+ uint64_t *deq_total_time, uint64_t *deq_min_time,
uint64_t *deq_max_time)
{
int i, deq_total;
@@ -2214,7 +2214,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time = rte_rdtsc_precise() - deq_start_time;
*deq_max_time = RTE_MAX(*deq_max_time, deq_last_time);
*deq_min_time = RTE_MIN(*deq_min_time, deq_last_time);
- *deq_tot_time += deq_last_time;
+ *deq_total_time += deq_last_time;
}

return i;
@@ -2223,7 +2223,7 @@ typedef int (test_case_function)(struct active_device *ad,
static int
offload_latency_empty_q_test_enc(uint16_t dev_id, uint16_t queue_id,
const uint16_t num_to_process, uint16_t burst_sz,
- uint64_t *deq_tot_time, uint64_t *deq_min_time,
+ uint64_t *deq_total_time, uint64_t *deq_min_time,
uint64_t *deq_max_time)
{
int i, deq_total;
@@ -2242,7 +2242,7 @@ typedef int (test_case_function)(struct active_device *ad,
deq_last_time = rte_rdtsc_precise() - deq_start_time;
*deq_max_time = RTE_MAX(*deq_max_time, deq_last_time);
*deq_min_time = RTE_MIN(*deq_min_time, deq_last_time);
- *deq_tot_time += deq_last_time;
+ *deq_total_time += deq_last_time;
}

return i;
@@ -2261,7 +2261,7 @@ typedef int (test_case_function)(struct active_device *ad,
return TEST_SKIPPED;
#else
int iter;
- uint64_t deq_tot_time, deq_min_time, deq_max_time;
+ uint64_t deq_total_time, deq_min_time, deq_max_time;
uint16_t burst_sz = op_params->burst_sz;
const uint16_t num_to_process = op_params->num_to_process;
const enum rte_bbdev_op_type op_type = test_vector.op_type;
@@ -2269,7 +2269,7 @@ typedef int (test_case_function)(struct active_device *ad,
struct rte_bbdev_info info;
const char *op_type_str;

- deq_tot_time = deq_max_time = 0;
+ deq_total_time = deq_max_time = 0;
deq_min_time = UINT64_MAX;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
@@ -2281,27 +2281,27 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_ASSERT_NOT_NULL(op_type_str, "Invalid op type: %u", op_type);

printf(
- "Offload latency empty dequeue test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
+ "\nOffload latency empty dequeue test: dev: %s, burst size: %u, num ops: %u, op type: %s\n",
info.dev_name, burst_sz, num_to_process, op_type_str);

if (op_type == RTE_BBDEV_OP_TURBO_DEC)
iter = offload_latency_empty_q_test_dec(ad->dev_id, queue_id,
- num_to_process, burst_sz, &deq_tot_time,
+ num_to_process, burst_sz, &deq_total_time,
&deq_min_time, &deq_max_time);
else
iter = offload_latency_empty_q_test_enc(ad->dev_id, queue_id,
- num_to_process, burst_sz, &deq_tot_time,
+ num_to_process, burst_sz, &deq_total_time,
&deq_min_time, &deq_max_time);

if (iter <= 0)
return TEST_FAILED;

- printf("\tempty deq offload\n"
- "\t\tavg. latency: %lg cycles, %lg us\n"
- "\t\tmin. latency: %lg cycles, %lg us\n"
- "\t\tmax. latency: %lg cycles, %lg us\n",
- (double)deq_tot_time / (double)iter,
- (double)(deq_tot_time * 1000000) / (double)iter /
+ printf("Empty dequeue offload\n"
+ "\tavg. latency: %lg cycles, %lg us\n"
+ "\tmin. latency: %lg cycles, %lg us\n"
+ "\tmax. latency: %lg cycles, %lg us\n",
+ (double)deq_total_time / (double)iter,
+ (double)(deq_total_time * 1000000) / (double)iter /
(double)rte_get_tsc_hz(), (double)deq_min_time,
(double)(deq_min_time * 1000000) / rte_get_tsc_hz(),
(double)deq_max_time, (double)(deq_max_time * 1000000) /
diff --git a/config/common_base b/config/common_base
index d12ae98..3ff98bb 100644
--- a/config/common_base
+++ b/config/common_base
@@ -481,7 +481,7 @@ CONFIG_RTE_PMD_PACKET_PREFETCH=y
#
CONFIG_RTE_LIBRTE_BBDEV=y
CONFIG_RTE_BBDEV_MAX_DEVS=128
-CONFIG_RTE_BBDEV_OFFLOAD_COST=n
+CONFIG_RTE_BBDEV_OFFLOAD_COST=y

#
# Compile PMD for NULL bbdev device
diff --git a/drivers/baseband/turbo_sw/bbdev_turbo_software.c b/drivers/baseband/turbo_sw/bbdev_turbo_software.c
index 8ceb276..57f6ba1 100644
--- a/drivers/baseband/turbo_sw/bbdev_turbo_software.c
+++ b/drivers/baseband/turbo_sw/bbdev_turbo_software.c
@@ -510,9 +510,10 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
+ /* CRC24A generation */
bblib_lte_crc24a_gen(&crc_req, &crc_resp);
#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif
} else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
/* CRC24B */
@@ -542,9 +543,10 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
+ /* CRC24B generation */
bblib_lte_crc24b_gen(&crc_req, &crc_resp);
#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif
} else {
ret = is_enc_input_valid(k, k_idx, total_left);
@@ -596,15 +598,14 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
-
+ /* Turbo encoding */
if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
op->status |= 1 << RTE_BBDEV_DRV_ERROR;
rte_bbdev_log(ERR, "Turbo Encoder failed");
return;
}
-
#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif

/* Restore 3 first bytes of next CB if they were overwritten by CRC*/
@@ -671,23 +672,21 @@ struct turbo_sw_queue {
#ifdef RTE_BBDEV_OFFLOAD_COST
start_time = rte_rdtsc_precise();
#endif
-
+ /* Rate-Matching */
if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
op->status |= 1 << RTE_BBDEV_DRV_ERROR;
rte_bbdev_log(ERR, "Rate matching failed");
return;
}
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif

/* SW fills an entire last byte even if E%8 != 0. Clear the
* superfluous data bits for consistency with HW device.
*/
mask_id = (e & 7) >> 1;
rm_out[out_len - 1] &= mask_out[mask_id];
-
-#ifdef RTE_BBDEV_OFFLOAD_COST
- q_stats->offload_time += rte_rdtsc_precise() - start_time;
-#endif
-
enc->output.length += rm_resp.OutputLen;
} else {
/* Rate matching is bypassed */
@@ -798,7 +797,7 @@ struct turbo_sw_queue {
{
uint16_t i;
#ifdef RTE_BBDEV_OFFLOAD_COST
- queue_stats->offload_time = 0;
+ queue_stats->acc_offload_cycles = 0;
#endif

for (i = 0; i < nb_ops; ++i)
@@ -905,7 +904,8 @@ struct turbo_sw_queue {
process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
- bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left)
+ bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left,
+ struct rte_bbdev_stats *q_stats)
{
int ret;
int32_t k_idx;
@@ -917,6 +917,11 @@ struct turbo_sw_queue {
struct bblib_turbo_decoder_request turbo_req;
struct bblib_turbo_decoder_response turbo_resp;
struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ uint64_t start_time;
+#else
+ RTE_SET_USED(q_stats);
+#endif

k_idx = compute_idx(k);

@@ -942,7 +947,14 @@ struct turbo_sw_queue {
deint_req.pharqbuffer = q->deint_input;
deint_req.ncb = ncb_without_null;
deint_resp.pinteleavebuffer = q->deint_output;
+
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ start_time = rte_rdtsc_precise();
+#endif
bblib_deinterleave_ul(&deint_req, &deint_resp);
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif
} else
move_padding_bytes(in, q->deint_output, k, ncb);

@@ -961,7 +973,15 @@ struct turbo_sw_queue {
adapter_req.ncb = ncb_without_null;
adapter_req.pinteleavebuffer = adapter_input;
adapter_resp.pharqout = q->adapter_output;
+
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ start_time = rte_rdtsc_precise();
+#endif
+ /* Turbo decode adaptation */
bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif

out = (uint8_t *)rte_pktmbuf_append(m_out, ((k - crc24_overlap) >> 3));
if (out == NULL) {
@@ -986,12 +1006,20 @@ struct turbo_sw_queue {
turbo_resp.ag_buf = q->ag;
turbo_resp.cb_buf = q->code_block;
turbo_resp.output = out;
+
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ start_time = rte_rdtsc_precise();
+#endif
+ /* Turbo decode */
iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
+#endif
dec->hard_output.length += (k >> 3);

if (iter_cnt > 0) {
/* Temporary solution for returned iter_count from SDK */
- iter_cnt = (iter_cnt - 1) / 2;
+ iter_cnt = (iter_cnt - 1) >> 1;
dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
} else {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
@@ -1001,7 +1029,8 @@ struct turbo_sw_queue {
}

static inline void
-enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op)
+enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
+ struct rte_bbdev_stats *queue_stats)
{
uint8_t c, r = 0;
uint16_t kw, k = 0;
@@ -1053,7 +1082,7 @@ struct turbo_sw_queue {
process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
out_offset, check_bit(dec->op_flags,
RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap,
- total_left);
+ total_left, queue_stats);
/* To keep CRC24 attached to end of Code block, use
* RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it
* removed by default once verified.
@@ -1075,12 +1104,15 @@ struct turbo_sw_queue {

static inline uint16_t
enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
- uint16_t nb_ops)
+ uint16_t nb_ops, struct rte_bbdev_stats *queue_stats)
{
uint16_t i;
+#ifdef RTE_BBDEV_OFFLOAD_COST
+ queue_stats->acc_offload_cycles = 0;
+#endif

for (i = 0; i < nb_ops; ++i)
- enqueue_dec_one_op(q, ops[i]);
+ enqueue_dec_one_op(q, ops[i], queue_stats);

return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
NULL);
@@ -1112,7 +1144,7 @@ struct turbo_sw_queue {
struct turbo_sw_queue *q = queue;
uint16_t nb_enqueued = 0;

- nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops);
+ nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops, &q_data->queue_stats);

q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
q_data->queue_stats.enqueued_count += nb_enqueued;
diff --git a/lib/librte_bbdev/rte_bbdev.h b/lib/librte_bbdev/rte_bbdev.h
index 25ef409..da8cf07 100644
--- a/lib/librte_bbdev/rte_bbdev.h
+++ b/lib/librte_bbdev/rte_bbdev.h
@@ -239,8 +239,13 @@ struct rte_bbdev_stats {
uint64_t enqueue_err_count;
/** Total error count on operations dequeued */
uint64_t dequeue_err_count;
- /** Offload time */
- uint64_t offload_time;
+ /** CPU cycles consumed by the (HW/SW) accelerator device to offload
+ * the enqueue request to its internal queues.
+ * - For a HW device this is the cycles consumed in MMIO write
+ * - For a SW (vdev) device, this is the processing time of the
+ * bbdev operation
+ */
+ uint64_t acc_offload_cycles;
};

/**
--
1.8.3.1
Kamil Chalupnik
2018-12-07 15:15:32 UTC
Permalink
Improvements added to throughput test:
- test is run in loop (number of iterations is specified by
TEST_REPETITIONS define) which ensures more accurate results
- length of input data is calculated based on amount of CBs in TB
- maximum number of decoding iterations is gathered from results
- added new functions responsible for printing results
- small fixes for memory management

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/main.c | 2 -
app/test-bbdev/test_bbdev_perf.c | 471 ++++++++++++++++++++-----------------
app/test-bbdev/test_bbdev_vector.c | 7 +
lib/librte_bbdev/rte_bbdev_op.h | 2 +
4 files changed, 263 insertions(+), 219 deletions(-)

diff --git a/app/test-bbdev/main.c b/app/test-bbdev/main.c
index 41b54bb..7af2522 100644
--- a/app/test-bbdev/main.c
+++ b/app/test-bbdev/main.c
@@ -316,8 +316,6 @@
return 1;
}

- rte_log_set_global_level(RTE_LOG_INFO);
-
/* If no argument provided - run all tests */
if (test_params.num_tests == 0)
return run_all_tests();
diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index bf97edb..a25e3a7 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -23,6 +23,7 @@
#define GET_SOCKET(socket_id) (((socket_id) == SOCKET_ID_ANY) ? 0 : (socket_id))

#define MAX_QUEUES RTE_MAX_LCORE
+#define TEST_REPETITIONS 1000

#define OPS_CACHE_SIZE 256U
#define OPS_POOL_SIZE_MIN 511U /* 0.5K per queue */
@@ -77,8 +78,9 @@ struct thread_params {
uint8_t dev_id;
uint16_t queue_id;
uint64_t start_time;
- double mops;
+ double ops_per_sec;
double mbps;
+ uint8_t iter_count;
rte_atomic16_t nb_dequeued;
rte_atomic16_t processing_status;
struct test_op_params *op_params;
@@ -757,6 +759,8 @@ typedef int (test_case_function)(struct active_device *ad,
turbo_dec->tb_params.c_neg;
ops[i]->turbo_dec.tb_params.cab =
turbo_dec->tb_params.cab;
+ ops[i]->turbo_dec.tb_params.r =
+ turbo_dec->tb_params.r;
} else {
ops[i]->turbo_dec.cb_params.e = turbo_dec->cb_params.e;
ops[i]->turbo_dec.cb_params.k = turbo_dec->cb_params.k;
@@ -884,47 +888,6 @@ typedef int (test_case_function)(struct active_device *ad,
}

static int
-validate_dec_buffers(struct rte_bbdev_dec_op *ref_op, struct test_buffers *bufs,
- const uint16_t num_to_process)
-{
- int i;
-
- struct op_data_entries *hard_data_orig =
- &test_vector.entries[DATA_HARD_OUTPUT];
- struct op_data_entries *soft_data_orig =
- &test_vector.entries[DATA_SOFT_OUTPUT];
-
- for (i = 0; i < num_to_process; i++) {
- TEST_ASSERT_SUCCESS(validate_op_chain(&bufs->hard_outputs[i],
- hard_data_orig),
- "Hard output buffers are not equal");
- if (ref_op->turbo_dec.op_flags &
- RTE_BBDEV_TURBO_SOFT_OUTPUT)
- TEST_ASSERT_SUCCESS(validate_op_chain(
- &bufs->soft_outputs[i],
- soft_data_orig),
- "Soft output buffers are not equal");
- }
-
- return TEST_SUCCESS;
-}
-
-static int
-validate_enc_buffers(struct test_buffers *bufs, const uint16_t num_to_process)
-{
- int i;
-
- struct op_data_entries *hard_data_orig =
- &test_vector.entries[DATA_HARD_OUTPUT];
-
- for (i = 0; i < num_to_process; i++)
- TEST_ASSERT_SUCCESS(validate_op_chain(&bufs->hard_outputs[i],
- hard_data_orig), "");
-
- return TEST_SUCCESS;
-}
-
-static int
validate_dec_op(struct rte_bbdev_dec_op **ops, const uint16_t n,
struct rte_bbdev_dec_op *ref_op, const int vector_mask)
{
@@ -1016,6 +979,44 @@ typedef int (test_case_function)(struct active_device *ad,
entry->segments[i].length;
}

+static uint32_t
+calc_dec_TB_size(struct rte_bbdev_dec_op *op)
+{
+ uint8_t i;
+ uint32_t c, r, tb_size = 0;
+
+ if (op->turbo_dec.code_block_mode) {
+ tb_size = op->turbo_dec.tb_params.k_neg;
+ } else {
+ c = op->turbo_dec.tb_params.c;
+ r = op->turbo_dec.tb_params.r;
+ for (i = 0; i < c-r; i++)
+ tb_size += (r < op->turbo_dec.tb_params.c_neg) ?
+ op->turbo_dec.tb_params.k_neg :
+ op->turbo_dec.tb_params.k_pos;
+ }
+ return tb_size;
+}
+
+static uint32_t
+calc_enc_TB_size(struct rte_bbdev_enc_op *op)
+{
+ uint8_t i;
+ uint32_t c, r, tb_size = 0;
+
+ if (op->turbo_enc.code_block_mode) {
+ tb_size = op->turbo_enc.tb_params.k_neg;
+ } else {
+ c = op->turbo_enc.tb_params.c;
+ r = op->turbo_enc.tb_params.r;
+ for (i = 0; i < c-r; i++)
+ tb_size += (r < op->turbo_enc.tb_params.c_neg) ?
+ op->turbo_enc.tb_params.k_neg :
+ op->turbo_enc.tb_params.k_pos;
+ }
+ return tb_size;
+}
+
static int
init_test_op_params(struct test_op_params *op_params,
enum rte_bbdev_op_type op_type, const int expected_status,
@@ -1163,17 +1164,13 @@ typedef int (test_case_function)(struct active_device *ad,
int ret;
uint16_t i;
uint64_t total_time;
- uint16_t deq, burst_sz, num_to_process;
+ uint16_t deq, burst_sz, num_ops;
uint16_t queue_id = INVALID_QUEUE_ID;
struct rte_bbdev_dec_op *dec_ops[MAX_BURST];
struct rte_bbdev_enc_op *enc_ops[MAX_BURST];
- struct test_buffers *bufs;
struct rte_bbdev_info info;

- /* Input length in bytes, million operations per second,
- * million bits per second.
- */
- double in_len;
+ double tb_len_bits;

struct thread_params *tp = cb_arg;
RTE_SET_USED(ret_param);
@@ -1198,7 +1195,7 @@ typedef int (test_case_function)(struct active_device *ad,
}

burst_sz = tp->op_params->burst_sz;
- num_to_process = tp->op_params->num_to_process;
+ num_ops = tp->op_params->num_to_process;

if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id, dec_ops,
@@ -1218,7 +1215,7 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- if (rte_atomic16_read(&tp->nb_dequeued) + deq < num_to_process) {
+ if (rte_atomic16_read(&tp->nb_dequeued) + deq < num_ops) {
rte_atomic16_add(&tp->nb_dequeued, deq);
return;
}
@@ -1227,14 +1224,18 @@ typedef int (test_case_function)(struct active_device *ad,

rte_bbdev_info_get(dev_id, &info);

- bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];
-
ret = TEST_SUCCESS;
- if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
- ret = validate_dec_buffers(tp->op_params->ref_dec_op, bufs,
- num_to_process);
- else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC)
- ret = validate_enc_buffers(bufs, num_to_process);
+
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
+ struct rte_bbdev_dec_op *ref_op = tp->op_params->ref_dec_op;
+ ret = validate_dec_op(dec_ops, num_ops, ref_op,
+ tp->op_params->vector_mask);
+ rte_bbdev_dec_op_free_bulk(dec_ops, deq);
+ } else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC) {
+ struct rte_bbdev_enc_op *ref_op = tp->op_params->ref_enc_op;
+ ret = validate_enc_op(enc_ops, num_ops, ref_op);
+ rte_bbdev_enc_op_free_bulk(enc_ops, deq);
+ }

if (ret) {
printf("Buffers validation failed\n");
@@ -1243,13 +1244,13 @@ typedef int (test_case_function)(struct active_device *ad,

switch (test_vector.op_type) {
case RTE_BBDEV_OP_TURBO_DEC:
- in_len = tp->op_params->ref_dec_op->turbo_dec.input.length;
+ tb_len_bits = calc_dec_TB_size(tp->op_params->ref_dec_op);
break;
case RTE_BBDEV_OP_TURBO_ENC:
- in_len = tp->op_params->ref_enc_op->turbo_enc.input.length;
+ tb_len_bits = calc_enc_TB_size(tp->op_params->ref_enc_op);
break;
case RTE_BBDEV_OP_NONE:
- in_len = 0.0;
+ tb_len_bits = 0.0;
break;
default:
printf("Unknown op type: %d\n", test_vector.op_type);
@@ -1257,9 +1258,9 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- tp->mops = ((double)num_to_process / 1000000.0) /
+ tp->ops_per_sec = ((double)num_ops) /
((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = ((double)num_to_process * in_len * 8 / 1000000.0) /
+ tp->mbps = (((double)(num_ops * tb_len_bits)) / 1000000.0) /
((double)total_time / (double)rte_get_tsc_hz());

rte_atomic16_add(&tp->nb_dequeued, deq);
@@ -1270,14 +1271,14 @@ typedef int (test_case_function)(struct active_device *ad,
{
struct thread_params *tp = arg;
unsigned int enqueued;
- struct rte_bbdev_dec_op *ops[MAX_BURST];
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
const uint16_t num_to_process = tp->op_params->num_to_process;
+ struct rte_bbdev_dec_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
struct rte_bbdev_info info;
int ret;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1287,6 +1288,11 @@ typedef int (test_case_function)(struct active_device *ad,
tp->dev_id, queue_id);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_to_process > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

rte_atomic16_clear(&tp->processing_status);
@@ -1295,36 +1301,27 @@ typedef int (test_case_function)(struct active_device *ad,
while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

+ ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp, ops,
+ num_to_process);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops",
+ num_to_process);
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_dec_op(ops, num_to_process, 0, bufs->inputs,
+ bufs->hard_outputs, bufs->soft_outputs,
+ tp->op_params->ref_dec_op);
+
tp->start_time = rte_rdtsc_precise();
for (enqueued = 0; enqueued < num_to_process;) {

- uint16_t num_to_enq = burst_sz;
+ num_to_enq = burst_sz;

if (unlikely(num_to_process - enqueued < num_to_enq))
num_to_enq = num_to_process - enqueued;

- ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp, ops,
- num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- continue;
- }
-
- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_dec_op(ops, num_to_enq, enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- bufs->soft_outputs,
- tp->op_params->ref_dec_op);
-
- enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id, ops,
- num_to_enq);
+ enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id,
+ &ops[enqueued], num_to_enq);
}

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
-
return TEST_SUCCESS;
}

@@ -1333,14 +1330,14 @@ typedef int (test_case_function)(struct active_device *ad,
{
struct thread_params *tp = arg;
unsigned int enqueued;
- struct rte_bbdev_enc_op *ops[MAX_BURST];
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
const uint16_t num_to_process = tp->op_params->num_to_process;
+ struct rte_bbdev_enc_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
struct rte_bbdev_info info;
int ret;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1350,6 +1347,11 @@ typedef int (test_case_function)(struct active_device *ad,
tp->dev_id, queue_id);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_to_process > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

rte_atomic16_clear(&tp->processing_status);
@@ -1358,35 +1360,26 @@ typedef int (test_case_function)(struct active_device *ad,
while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

+ ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp, ops,
+ num_to_process);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops",
+ num_to_process);
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_enc_op(ops, num_to_process, 0, bufs->inputs,
+ bufs->hard_outputs, tp->op_params->ref_enc_op);
+
tp->start_time = rte_rdtsc_precise();
for (enqueued = 0; enqueued < num_to_process;) {

- uint16_t num_to_enq = burst_sz;
+ num_to_enq = burst_sz;

if (unlikely(num_to_process - enqueued < num_to_enq))
num_to_enq = num_to_process - enqueued;

- ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp, ops,
- num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- continue;
- }
-
- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_enc_op(ops, num_to_enq, enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- tp->op_params->ref_enc_op);
-
- enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id, ops,
- num_to_enq);
+ enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id,
+ &ops[enqueued], num_to_enq);
}

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
-
return TEST_SUCCESS;
}

@@ -1394,86 +1387,97 @@ typedef int (test_case_function)(struct active_device *ad,
throughput_pmd_lcore_dec(void *arg)
{
struct thread_params *tp = arg;
- unsigned int enqueued, dequeued;
- struct rte_bbdev_dec_op *ops_enq[MAX_BURST], *ops_deq[MAX_BURST];
- uint64_t total_time, start_time;
+ uint16_t enq, deq;
+ uint64_t total_time = 0, start_time;
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
- const uint16_t num_to_process = tp->op_params->num_to_process;
+ const uint16_t num_ops = tp->op_params->num_to_process;
+ struct rte_bbdev_dec_op *ops_enq[num_ops];
+ struct rte_bbdev_dec_op *ops_deq[num_ops];
struct rte_bbdev_dec_op *ref_op = tp->op_params->ref_dec_op;
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
- int ret;
+ int i, j, ret;
struct rte_bbdev_info info;
-
- /* Input length in bytes, million operations per second, million bits
- * per second.
- */
- double in_len;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_ops > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

- start_time = rte_rdtsc_precise();
- for (enqueued = 0, dequeued = 0; dequeued < num_to_process;) {
- uint16_t deq;
+ ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp, ops_enq, num_ops);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops", num_ops);

- if (likely(enqueued < num_to_process)) {
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_dec_op(ops_enq, num_ops, 0, bufs->inputs,
+ bufs->hard_outputs, bufs->soft_outputs, ref_op);

- uint16_t num_to_enq = burst_sz;
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_ops; ++j)
+ ops_enq[j]->opaque_data = (void *)(uintptr_t)j;

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ for (i = 0; i < TEST_REPETITIONS; ++i) {

- ret = rte_bbdev_dec_op_alloc_bulk(tp->op_params->mp,
- ops_enq, num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- goto do_dequeue;
- }
+ for (j = 0; j < num_ops; ++j) {
+ struct rte_bbdev_dec_op *op = ops_enq[j];
+ rte_pktmbuf_reset(op->turbo_dec.hard_output.data);
+ }
+
+ start_time = rte_rdtsc_precise();
+
+ for (enq = 0, deq = 0; enq < num_ops;) {
+ num_to_enq = burst_sz;
+
+ if (unlikely(num_ops - enq < num_to_enq))
+ num_to_enq = num_ops - enq;

- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_dec_op(ops_enq, num_to_enq,
- enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- bufs->soft_outputs,
- ref_op);
+ enq += rte_bbdev_enqueue_dec_ops(tp->dev_id,
+ queue_id, &ops_enq[enq], num_to_enq);

- enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id,
- queue_id, ops_enq, num_to_enq);
+ deq += rte_bbdev_dequeue_dec_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
}
-do_dequeue:
- deq = rte_bbdev_dequeue_dec_ops(tp->dev_id, queue_id, ops_deq,
- burst_sz);
- dequeued += deq;
- rte_bbdev_dec_op_free_bulk(ops_enq, deq);
- }
- total_time = rte_rdtsc_precise() - start_time;

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
+ /* dequeue the remaining */
+ while (deq < enq) {
+ deq += rte_bbdev_dequeue_dec_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
+ }

- TEST_ASSERT(enqueued == dequeued, "enqueued (%u) != dequeued (%u)",
- enqueued, dequeued);
+ total_time += rte_rdtsc_precise() - start_time;
+ }
+
+ tp->iter_count = 0;
+ /* get the max of iter_count for all dequeued ops */
+ for (i = 0; i < num_ops; ++i) {
+ tp->iter_count = RTE_MAX(ops_enq[i]->turbo_dec.iter_count,
+ tp->iter_count);
+ }

if (test_vector.op_type != RTE_BBDEV_OP_NONE) {
- ret = validate_dec_buffers(ref_op, bufs, num_to_process);
- TEST_ASSERT_SUCCESS(ret, "Buffers validation failed");
+ ret = validate_dec_op(ops_deq, num_ops, ref_op,
+ tp->op_params->vector_mask);
+ TEST_ASSERT_SUCCESS(ret, "Validation failed!");
}

- in_len = ref_op->turbo_dec.input.length;
- tp->mops = ((double)num_to_process / 1000000.0) /
- ((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = ((double)num_to_process * in_len * 8 / 1000000.0) /
+ rte_bbdev_dec_op_free_bulk(ops_enq, num_ops);
+
+ double tb_len_bits = calc_dec_TB_size(ref_op);
+
+ tp->ops_per_sec = ((double)num_ops * TEST_REPETITIONS) /
((double)total_time / (double)rte_get_tsc_hz());
+ tp->mbps = (((double)(num_ops * TEST_REPETITIONS * tb_len_bits)) /
+ 1000000.0) / ((double)total_time /
+ (double)rte_get_tsc_hz());

return TEST_SUCCESS;
}
@@ -1482,91 +1486,94 @@ typedef int (test_case_function)(struct active_device *ad,
throughput_pmd_lcore_enc(void *arg)
{
struct thread_params *tp = arg;
- unsigned int enqueued, dequeued;
- struct rte_bbdev_enc_op *ops_enq[MAX_BURST], *ops_deq[MAX_BURST];
- uint64_t total_time, start_time;
+ uint16_t enq, deq;
+ uint64_t total_time = 0, start_time;
const uint16_t queue_id = tp->queue_id;
const uint16_t burst_sz = tp->op_params->burst_sz;
- const uint16_t num_to_process = tp->op_params->num_to_process;
+ const uint16_t num_ops = tp->op_params->num_to_process;
+ struct rte_bbdev_enc_op *ops_enq[num_ops];
+ struct rte_bbdev_enc_op *ops_deq[num_ops];
struct rte_bbdev_enc_op *ref_op = tp->op_params->ref_enc_op;
struct test_buffers *bufs = NULL;
- unsigned int allocs_failed = 0;
- int ret;
+ int i, j, ret;
struct rte_bbdev_info info;
-
- /* Input length in bytes, million operations per second, million bits
- * per second.
- */
- double in_len;
+ uint16_t num_to_enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);

rte_bbdev_info_get(tp->dev_id, &info);
+
+ TEST_ASSERT_SUCCESS((num_ops > info.drv.queue_size_lim),
+ "NUM_OPS cannot exceed %u for this device",
+ info.drv.queue_size_lim);
+
bufs = &tp->op_params->q_bufs[GET_SOCKET(info.socket_id)][queue_id];

while (rte_atomic16_read(&tp->op_params->sync) == SYNC_WAIT)
rte_pause();

- start_time = rte_rdtsc_precise();
- for (enqueued = 0, dequeued = 0; dequeued < num_to_process;) {
- uint16_t deq;
+ ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp, ops_enq,
+ num_ops);
+ TEST_ASSERT_SUCCESS(ret, "Allocation failed for %d ops",
+ num_ops);
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ copy_reference_enc_op(ops_enq, num_ops, 0, bufs->inputs,
+ bufs->hard_outputs, ref_op);

- if (likely(enqueued < num_to_process)) {
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_ops; ++j)
+ ops_enq[j]->opaque_data = (void *)(uintptr_t)j;

- uint16_t num_to_enq = burst_sz;
+ for (i = 0; i < TEST_REPETITIONS; ++i) {

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ if (test_vector.op_type != RTE_BBDEV_OP_NONE)
+ for (j = 0; j < num_ops; ++j)
+ rte_pktmbuf_reset(
+ ops_enq[j]->turbo_enc.output.data);

- ret = rte_bbdev_enc_op_alloc_bulk(tp->op_params->mp,
- ops_enq, num_to_enq);
- if (ret != 0) {
- allocs_failed++;
- goto do_dequeue;
- }
+ start_time = rte_rdtsc_precise();
+
+ for (enq = 0, deq = 0; enq < num_ops;) {
+ num_to_enq = burst_sz;

- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- copy_reference_enc_op(ops_enq, num_to_enq,
- enqueued,
- bufs->inputs,
- bufs->hard_outputs,
- ref_op);
+ if (unlikely(num_ops - enq < num_to_enq))
+ num_to_enq = num_ops - enq;

- enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id,
- queue_id, ops_enq, num_to_enq);
+ enq += rte_bbdev_enqueue_enc_ops(tp->dev_id,
+ queue_id, &ops_enq[enq], num_to_enq);
+
+ deq += rte_bbdev_dequeue_enc_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
}
-do_dequeue:
- deq = rte_bbdev_dequeue_enc_ops(tp->dev_id, queue_id, ops_deq,
- burst_sz);
- dequeued += deq;
- rte_bbdev_enc_op_free_bulk(ops_enq, deq);
- }
- total_time = rte_rdtsc_precise() - start_time;

- if (allocs_failed > 0)
- printf("WARNING: op allocations failed: %u times\n",
- allocs_failed);
+ /* dequeue the remaining */
+ while (deq < enq) {
+ deq += rte_bbdev_dequeue_enc_ops(tp->dev_id,
+ queue_id, &ops_deq[deq], enq - deq);
+ }

- TEST_ASSERT(enqueued == dequeued, "enqueued (%u) != dequeued (%u)",
- enqueued, dequeued);
+ total_time += rte_rdtsc_precise() - start_time;
+ }

if (test_vector.op_type != RTE_BBDEV_OP_NONE) {
- ret = validate_enc_buffers(bufs, num_to_process);
- TEST_ASSERT_SUCCESS(ret, "Buffers validation failed");
+ ret = validate_enc_op(ops_deq, num_ops, ref_op);
+ TEST_ASSERT_SUCCESS(ret, "Validation failed!");
}

- in_len = ref_op->turbo_enc.input.length;
+ double tb_len_bits = calc_enc_TB_size(ref_op);

- tp->mops = ((double)num_to_process / 1000000.0) /
- ((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = ((double)num_to_process * in_len * 8 / 1000000.0) /
+ tp->ops_per_sec = ((double)num_ops * TEST_REPETITIONS) /
((double)total_time / (double)rte_get_tsc_hz());
+ tp->mbps = (((double)(num_ops * TEST_REPETITIONS * tb_len_bits))
+ / 1000000.0) / ((double)total_time /
+ (double)rte_get_tsc_hz());

return TEST_SUCCESS;
}
+
static void
-print_throughput(struct thread_params *t_params, unsigned int used_cores)
+print_enc_throughput(struct thread_params *t_params, unsigned int used_cores)
{
unsigned int lcore_id, iter = 0;
double total_mops = 0, total_mbps = 0;
@@ -1574,10 +1581,11 @@ typedef int (test_case_function)(struct active_device *ad,
RTE_LCORE_FOREACH(lcore_id) {
if (iter++ >= used_cores)
break;
- printf("Throughput for core (%u): %.8lg MOPS, %.8lg Mbps\n",
- lcore_id, t_params[lcore_id].mops,
+ printf(
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps\n",
+ lcore_id, t_params[lcore_id].ops_per_sec,
t_params[lcore_id].mbps);
- total_mops += t_params[lcore_id].mops;
+ total_mops += t_params[lcore_id].ops_per_sec;
total_mbps += t_params[lcore_id].mbps;
}
printf(
@@ -1585,6 +1593,30 @@ typedef int (test_case_function)(struct active_device *ad,
used_cores, total_mops, total_mbps);
}

+static void
+print_dec_throughput(struct thread_params *t_params, unsigned int used_cores)
+{
+ unsigned int lcore_id, iter = 0;
+ double total_mops = 0, total_mbps = 0;
+ uint8_t iter_count = 0;
+
+ RTE_LCORE_FOREACH(lcore_id) {
+ if (iter++ >= used_cores)
+ break;
+ printf(
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps @ max %u iterations\n",
+ lcore_id, t_params[lcore_id].ops_per_sec,
+ t_params[lcore_id].mbps,
+ t_params[lcore_id].iter_count);
+ total_mops += t_params[lcore_id].ops_per_sec;
+ total_mbps += t_params[lcore_id].mbps;
+ iter_count = RTE_MAX(iter_count, t_params[lcore_id].iter_count);
+ }
+ printf(
+ "\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps @ max %u iterations\n",
+ used_cores, total_mops, total_mbps, iter_count);
+}
+
/*
* Test function that determines how long an enqueue + dequeue of a burst
* takes on available lcores.
@@ -1677,8 +1709,10 @@ typedef int (test_case_function)(struct active_device *ad,

/* Print throughput if interrupts are disabled and test passed */
if (!intr_enabled) {
- if (test_vector.op_type != RTE_BBDEV_OP_NONE)
- print_throughput(t_params, num_lcores);
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ print_dec_throughput(t_params, num_lcores);
+ else
+ print_enc_throughput(t_params, num_lcores);
return ret;
}

@@ -1713,9 +1747,12 @@ typedef int (test_case_function)(struct active_device *ad,
}

/* Print throughput if test passed */
- if (!ret && test_vector.op_type != RTE_BBDEV_OP_NONE)
- print_throughput(t_params, num_lcores);
-
+ if (!ret) {
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ print_dec_throughput(t_params, num_lcores);
+ else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC)
+ print_enc_throughput(t_params, num_lcores);
+ }
return ret;
}

diff --git a/app/test-bbdev/test_bbdev_vector.c b/app/test-bbdev/test_bbdev_vector.c
index 81b8ee7..45fe999 100644
--- a/app/test-bbdev/test_bbdev_vector.c
+++ b/app/test-bbdev/test_bbdev_vector.c
@@ -412,6 +412,10 @@
vector->mask |= TEST_BBDEV_VF_NUM_MAPS;
turbo_dec->num_maps = (uint8_t) strtoul(token, &err, 0);
ret = ((err == NULL) || (*err != '\0')) ? -1 : 0;
+ } else if (!strcmp(key_token, "r")) {
+ vector->mask |= TEST_BBDEV_VF_R;
+ turbo_dec->tb_params.r = (uint8_t) strtoul(token, &err, 0);
+ ret = ((err == NULL) || (*err != '\0')) ? -1 : 0;
} else if (!strcmp(key_token, "code_block_mode")) {
vector->mask |= TEST_BBDEV_VF_CODE_BLOCK_MODE;
turbo_dec->code_block_mode = (uint8_t) strtoul(token, &err, 0);
@@ -714,6 +718,9 @@
if (!(mask & TEST_BBDEV_VF_CAB))
printf(
"WARNING: cab was not specified in vector file and will be set to 0\n");
+ if (!(mask & TEST_BBDEV_VF_R))
+ printf(
+ "WARNING: r was not specified in vector file and will be set to 0\n");
} else {
if (!(mask & TEST_BBDEV_VF_E))
printf(
diff --git a/lib/librte_bbdev/rte_bbdev_op.h b/lib/librte_bbdev/rte_bbdev_op.h
index 83f62c2..962e2ed 100644
--- a/lib/librte_bbdev/rte_bbdev_op.h
+++ b/lib/librte_bbdev/rte_bbdev_op.h
@@ -216,6 +216,8 @@ struct rte_bbdev_op_dec_tb_params {
* operation when r >= cab
*/
uint32_t eb;
+ /**< The index of the first CB in the inbound mbuf data, default is 0 */
+ uint8_t r;
};

/**< Operation structure for Turbo decode.
--
1.8.3.1
Kamil Chalupnik
2018-12-07 15:15:33 UTC
Permalink
Test application and Turbo Software driver were adapted
to support chained-mbuf for bigger TB sizes.

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/test_bbdev_perf.c | 60 +++++++++---
drivers/baseband/turbo_sw/bbdev_turbo_software.c | 111 ++++++++++++++++-------
2 files changed, 126 insertions(+), 45 deletions(-)

diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index a25e3a7..5bec70d 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -114,6 +114,17 @@ typedef int (test_case_function)(struct active_device *ad,
struct test_op_params *op_params);

static inline void
+mbuf_reset(struct rte_mbuf *m)
+{
+ m->pkt_len = 0;
+
+ do {
+ m->data_len = 0;
+ m = m->next;
+ } while (m != NULL);
+}
+
+static inline void
set_avail_op(struct active_device *ad, enum rte_bbdev_op_type op_type)
{
ad->supported_ops |= (1 << op_type);
@@ -573,6 +584,10 @@ typedef int (test_case_function)(struct active_device *ad,
op_type, n * ref_entries->nb_segments,
mbuf_pool->size);

+ TEST_ASSERT_SUCCESS(((seg->length + RTE_PKTMBUF_HEADROOM) >
+ (uint32_t)UINT16_MAX),
+ "Given data is bigger than allowed mbuf segment size");
+
bufs[i].data = m_head;
bufs[i].offset = 0;
bufs[i].length = 0;
@@ -589,7 +604,6 @@ typedef int (test_case_function)(struct active_device *ad,
rte_memcpy(data, seg->addr, seg->length);
bufs[i].length += seg->length;

-
for (j = 1; j < ref_entries->nb_segments; ++j) {
struct rte_mbuf *m_tail =
rte_pktmbuf_alloc(mbuf_pool);
@@ -617,6 +631,24 @@ typedef int (test_case_function)(struct active_device *ad,
"Couldn't chain mbufs from %d data type mbuf pool",
op_type);
}
+
+ } else {
+
+ /* allocate chained-mbuf for output buffer */
+ for (j = 1; j < ref_entries->nb_segments; ++j) {
+ struct rte_mbuf *m_tail =
+ rte_pktmbuf_alloc(mbuf_pool);
+ TEST_ASSERT_NOT_NULL(m_tail,
+ "Not enough mbufs in %d data type mbuf pool (needed %u, available %u)",
+ op_type,
+ n * ref_entries->nb_segments,
+ mbuf_pool->size);
+
+ ret = rte_pktmbuf_chain(m_head, m_tail);
+ TEST_ASSERT_SUCCESS(ret,
+ "Couldn't chain mbufs from %d data type mbuf pool",
+ op_type);
+ }
}
}

@@ -655,7 +687,7 @@ typedef int (test_case_function)(struct active_device *ad,
while (m != NULL) {
int8_t *llr = rte_pktmbuf_mtod_offset(m, int8_t *,
input_ops[i].offset);
- for (byte_idx = 0; byte_idx < input_ops[i].length;
+ for (byte_idx = 0; byte_idx < rte_pktmbuf_data_len(m);
++byte_idx)
llr[byte_idx] = round((double)max_llr_modulus *
llr[byte_idx] / INT8_MAX);
@@ -864,15 +896,18 @@ typedef int (test_case_function)(struct active_device *ad,
uint8_t i;
struct rte_mbuf *m = op->data;
uint8_t nb_dst_segments = orig_op->nb_segments;
+ uint32_t total_data_size = 0;

TEST_ASSERT(nb_dst_segments == m->nb_segs,
"Number of segments differ in original (%u) and filled (%u) op",
nb_dst_segments, m->nb_segs);

+ /* Validate each mbuf segment length */
for (i = 0; i < nb_dst_segments; ++i) {
/* Apply offset to the first mbuf segment */
uint16_t offset = (i == 0) ? op->offset : 0;
- uint16_t data_len = m->data_len - offset;
+ uint16_t data_len = rte_pktmbuf_data_len(m) - offset;
+ total_data_size += orig_op->segments[i].length;

TEST_ASSERT(orig_op->segments[i].length == data_len,
"Length of segment differ in original (%u) and filled (%u) op",
@@ -884,6 +919,12 @@ typedef int (test_case_function)(struct active_device *ad,
m = m->next;
}

+ /* Validate total mbuf pkt length */
+ uint32_t pkt_len = rte_pktmbuf_pkt_len(op->data) - op->offset;
+ TEST_ASSERT(total_data_size == pkt_len,
+ "Length of data differ in original (%u) and filled (%u) op",
+ total_data_size, pkt_len);
+
return TEST_SUCCESS;
}

@@ -1427,10 +1468,8 @@ typedef int (test_case_function)(struct active_device *ad,

for (i = 0; i < TEST_REPETITIONS; ++i) {

- for (j = 0; j < num_ops; ++j) {
- struct rte_bbdev_dec_op *op = ops_enq[j];
- rte_pktmbuf_reset(op->turbo_dec.hard_output.data);
- }
+ for (j = 0; j < num_ops; ++j)
+ mbuf_reset(ops_enq[j]->turbo_dec.hard_output.data);

start_time = rte_rdtsc_precise();

@@ -1529,8 +1568,7 @@ typedef int (test_case_function)(struct active_device *ad,

if (test_vector.op_type != RTE_BBDEV_OP_NONE)
for (j = 0; j < num_ops; ++j)
- rte_pktmbuf_reset(
- ops_enq[j]->turbo_enc.output.data);
+ mbuf_reset(ops_enq[j]->turbo_enc.output.data);

start_time = rte_rdtsc_precise();

@@ -2025,7 +2063,7 @@ typedef int (test_case_function)(struct active_device *ad,
time_st->enq_acc_total_time += stats.acc_offload_cycles;

/* ensure enqueue has been completed */
- rte_delay_ms(10);
+ rte_delay_us(200);

/* Start time meas for dequeue function offload latency */
deq_start_time = rte_rdtsc_precise();
@@ -2106,7 +2144,7 @@ typedef int (test_case_function)(struct active_device *ad,
time_st->enq_acc_total_time += stats.acc_offload_cycles;

/* ensure enqueue has been completed */
- rte_delay_ms(10);
+ rte_delay_us(200);

/* Start time meas for dequeue function offload latency */
deq_start_time = rte_rdtsc_precise();
diff --git a/drivers/baseband/turbo_sw/bbdev_turbo_software.c b/drivers/baseband/turbo_sw/bbdev_turbo_software.c
index 57f6ba1..19fbb55 100644
--- a/drivers/baseband/turbo_sw/bbdev_turbo_software.c
+++ b/drivers/baseband/turbo_sw/bbdev_turbo_software.c
@@ -83,6 +83,18 @@ struct turbo_sw_queue {
enum rte_bbdev_op_type type;
} __rte_cache_aligned;

+static inline char *
+mbuf_append(struct rte_mbuf *m_head, struct rte_mbuf *m, uint16_t len)
+{
+ if (unlikely(len > rte_pktmbuf_tailroom(m)))
+ return NULL;
+
+ char *tail = (char *)m->buf_addr + m->data_off + m->data_len;
+ m->data_len = (uint16_t)(m->data_len + len);
+ m_head->pkt_len = (m_head->pkt_len + len);
+ return tail;
+}
+
/* Calculate index based on Table 5.1.3-3 from TS34.212 */
static inline int32_t
compute_idx(uint16_t k)
@@ -437,7 +449,7 @@ struct turbo_sw_queue {
return -1;
}

- if (in_length - kw < 0) {
+ if (in_length < kw) {
rte_bbdev_log(ERR,
"Mismatch between input length (%u) and kw (%u)",
in_length, kw);
@@ -456,9 +468,9 @@ struct turbo_sw_queue {
static inline void
process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
uint8_t r, uint8_t c, uint16_t k, uint16_t ncb,
- uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out,
- uint16_t in_offset, uint16_t out_offset, uint16_t total_left,
- struct rte_bbdev_stats *q_stats)
+ uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out_head,
+ struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
+ uint16_t in_length, struct rte_bbdev_stats *q_stats)
{
int ret;
int16_t k_idx;
@@ -484,7 +496,7 @@ struct turbo_sw_queue {
/* CRC24A (for TB) */
if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) &&
(enc->code_block_mode == 1)) {
- ret = is_enc_input_valid(k - 24, k_idx, total_left);
+ ret = is_enc_input_valid(k - 24, k_idx, in_length);
if (ret != 0) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
return;
@@ -494,7 +506,7 @@ struct turbo_sw_queue {
/* Check if there is a room for CRC bits if not use
* the temporary buffer.
*/
- if (rte_pktmbuf_append(m_in, 3) == NULL) {
+ if (mbuf_append(m_in, m_in, 3) == NULL) {
rte_memcpy(q->enc_in, in, (k - 24) >> 3);
in = q->enc_in;
} else {
@@ -517,7 +529,7 @@ struct turbo_sw_queue {
#endif
} else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
/* CRC24B */
- ret = is_enc_input_valid(k - 24, k_idx, total_left);
+ ret = is_enc_input_valid(k - 24, k_idx, in_length);
if (ret != 0) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
return;
@@ -527,7 +539,7 @@ struct turbo_sw_queue {
/* Check if there is a room for CRC bits if this is the last
* CB in TB. If not use temporary buffer.
*/
- if ((c - r == 1) && (rte_pktmbuf_append(m_in, 3) == NULL)) {
+ if ((c - r == 1) && (mbuf_append(m_in, m_in, 3) == NULL)) {
rte_memcpy(q->enc_in, in, (k - 24) >> 3);
in = q->enc_in;
} else if (c - r > 1) {
@@ -549,7 +561,7 @@ struct turbo_sw_queue {
q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif
} else {
- ret = is_enc_input_valid(k, k_idx, total_left);
+ ret = is_enc_input_valid(k, k_idx, in_length);
if (ret != 0) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
return;
@@ -570,7 +582,8 @@ struct turbo_sw_queue {
out1 = RTE_PTR_ADD(out0, (k >> 3) + 1);
out2 = RTE_PTR_ADD(out1, (k >> 3) + 1);
} else {
- out0 = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3) * 3 + 2);
+ out0 = (uint8_t *)mbuf_append(m_out_head, m_out,
+ (k >> 3) * 3 + 2);
if (out0 == NULL) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
rte_bbdev_log(ERR,
@@ -623,7 +636,7 @@ struct turbo_sw_queue {
const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC};

/* get output data starting address */
- rm_out = (uint8_t *)rte_pktmbuf_append(m_out, out_len);
+ rm_out = (uint8_t *)mbuf_append(m_out_head, m_out, out_len);
if (rm_out == NULL) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
rte_bbdev_log(ERR,
@@ -725,14 +738,16 @@ struct turbo_sw_queue {
uint16_t out_offset = enc->output.offset;
struct rte_mbuf *m_in = enc->input.data;
struct rte_mbuf *m_out = enc->output.data;
- uint16_t total_left = enc->input.length;
+ struct rte_mbuf *m_out_head = enc->output.data;
+ uint32_t in_length, mbuf_total_left = enc->input.length;
+ uint16_t seg_total_left;

/* Clear op status */
op->status = 0;

- if (total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) {
+ if (mbuf_total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) {
rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d",
- total_left, RTE_BBDEV_MAX_TB_SIZE);
+ mbuf_total_left, RTE_BBDEV_MAX_TB_SIZE);
op->status = 1 << RTE_BBDEV_DATA_ERROR;
return;
}
@@ -755,7 +770,10 @@ struct turbo_sw_queue {
r = 0;
}

- while (total_left > 0 && r < c) {
+ while (mbuf_total_left > 0 && r < c) {
+
+ seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset;
+
if (enc->code_block_mode == 0) {
k = (r < enc->tb_params.c_neg) ?
enc->tb_params.k_neg : enc->tb_params.k_pos;
@@ -769,22 +787,32 @@ struct turbo_sw_queue {
e = enc->cb_params.e;
}

- process_enc_cb(q, op, r, c, k, ncb, e, m_in,
- m_out, in_offset, out_offset, total_left,
+ process_enc_cb(q, op, r, c, k, ncb, e, m_in, m_out_head,
+ m_out, in_offset, out_offset, seg_total_left,
queue_stats);
/* Update total_left */
- total_left -= (k - crc24_bits) >> 3;
+ in_length = ((k - crc24_bits) >> 3);
+ mbuf_total_left -= in_length;
/* Update offsets for next CBs (if exist) */
in_offset += (k - crc24_bits) >> 3;
if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH)
out_offset += e >> 3;
else
out_offset += (k >> 3) * 3 + 2;
+
+ /* Update offsets */
+ if (seg_total_left == in_length) {
+ /* Go to the next mbuf */
+ m_in = m_in->next;
+ m_out = m_out->next;
+ in_offset = 0;
+ out_offset = 0;
+ }
r++;
}

/* check if all input data was processed */
- if (total_left != 0) {
+ if (mbuf_total_left != 0) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
rte_bbdev_log(ERR,
"Mismatch between mbuf length and included CBs sizes");
@@ -903,8 +931,9 @@ struct turbo_sw_queue {
static inline void
process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
- struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
- bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left,
+ struct rte_mbuf *m_out_head, struct rte_mbuf *m_out,
+ uint16_t in_offset, uint16_t out_offset, bool check_crc_24b,
+ uint16_t crc24_overlap, uint16_t in_length,
struct rte_bbdev_stats *q_stats)
{
int ret;
@@ -925,7 +954,7 @@ struct turbo_sw_queue {

k_idx = compute_idx(k);

- ret = is_dec_input_valid(k_idx, kw, total_left);
+ ret = is_dec_input_valid(k_idx, kw, in_length);
if (ret != 0) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
return;
@@ -983,7 +1012,8 @@ struct turbo_sw_queue {
q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
#endif

- out = (uint8_t *)rte_pktmbuf_append(m_out, ((k - crc24_overlap) >> 3));
+ out = (uint8_t *)mbuf_append(m_out_head, m_out,
+ ((k - crc24_overlap) >> 3));
if (out == NULL) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
rte_bbdev_log(ERR, "Too little space in output mbuf");
@@ -1038,9 +1068,11 @@ struct turbo_sw_queue {
struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
struct rte_mbuf *m_in = dec->input.data;
struct rte_mbuf *m_out = dec->hard_output.data;
+ struct rte_mbuf *m_out_head = dec->hard_output.data;
uint16_t in_offset = dec->input.offset;
- uint16_t total_left = dec->input.length;
uint16_t out_offset = dec->hard_output.offset;
+ uint32_t mbuf_total_left = dec->input.length;
+ uint16_t seg_total_left;

/* Clear op status */
op->status = 0;
@@ -1062,11 +1094,13 @@ struct turbo_sw_queue {
RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP))
crc24_overlap = 24;

- while (total_left > 0) {
+ while (mbuf_total_left > 0) {
if (dec->code_block_mode == 0)
k = (r < dec->tb_params.c_neg) ?
dec->tb_params.k_neg : dec->tb_params.k_pos;

+ seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset;
+
/* Calculates circular buffer size (Kw).
* According to 3gpp 36.212 section 5.1.4.2
* Kw = 3 * Kpi,
@@ -1079,23 +1113,32 @@ struct turbo_sw_queue {
*/
kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_C_SUBBLOCK) * 3;

- process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
- out_offset, check_bit(dec->op_flags,
+ process_dec_cb(q, op, c, k, kw, m_in, m_out_head, m_out,
+ in_offset, out_offset, check_bit(dec->op_flags,
RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap,
- total_left, queue_stats);
+ seg_total_left, queue_stats);
/* To keep CRC24 attached to end of Code block, use
* RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it
* removed by default once verified.
*/

- /* Update total_left */
- total_left -= kw;
- /* Update offsets for next CBs (if exist) */
- in_offset += kw;
- out_offset += ((k - crc24_overlap) >> 3);
+ mbuf_total_left -= kw;
+
+ /* Update offsets */
+ if (seg_total_left == kw) {
+ /* Go to the next mbuf */
+ m_in = m_in->next;
+ m_out = m_out->next;
+ in_offset = 0;
+ out_offset = 0;
+ } else {
+ /* Update offsets for next CBs (if exist) */
+ in_offset += kw;
+ out_offset += ((k - crc24_overlap) >> 3);
+ }
r++;
}
- if (total_left != 0) {
+ if (mbuf_total_left != 0) {
op->status |= 1 << RTE_BBDEV_DATA_ERROR;
rte_bbdev_log(ERR,
"Mismatch between mbuf length and included Circular buffer sizes");
--
1.8.3.1
Kamil Chalupnik
2018-12-07 15:15:34 UTC
Permalink
Improvements added to interrupt test:
- test is run in loop (number of iterations is specified by
TEST_REPETITIONS define) which ensures more accurate results
- mapping cores to thread parameteres was put in order.
Master core is always set at first index. It fixes problem with
running test for only one core

Signed-off-by: Kamil Chalupnik <***@intel.com>
---
app/test-bbdev/test_bbdev_perf.c | 246 +++++++++++++++++++++++++--------------
1 file changed, 161 insertions(+), 85 deletions(-)

diff --git a/app/test-bbdev/test_bbdev_perf.c b/app/test-bbdev/test_bbdev_perf.c
index 5bec70d..1c4a645 100644
--- a/app/test-bbdev/test_bbdev_perf.c
+++ b/app/test-bbdev/test_bbdev_perf.c
@@ -77,13 +77,17 @@ struct test_op_params {
struct thread_params {
uint8_t dev_id;
uint16_t queue_id;
+ uint32_t lcore_id;
uint64_t start_time;
double ops_per_sec;
double mbps;
uint8_t iter_count;
rte_atomic16_t nb_dequeued;
rte_atomic16_t processing_status;
+ rte_atomic16_t burst_sz;
struct test_op_params *op_params;
+ struct rte_bbdev_dec_op *dec_ops[MAX_BURST];
+ struct rte_bbdev_enc_op *enc_ops[MAX_BURST];
};

#ifdef RTE_BBDEV_OFFLOAD_COST
@@ -1206,16 +1210,12 @@ typedef int (test_case_function)(struct active_device *ad,
uint16_t i;
uint64_t total_time;
uint16_t deq, burst_sz, num_ops;
- uint16_t queue_id = INVALID_QUEUE_ID;
- struct rte_bbdev_dec_op *dec_ops[MAX_BURST];
- struct rte_bbdev_enc_op *enc_ops[MAX_BURST];
+ uint16_t queue_id = *(uint16_t *) ret_param;
struct rte_bbdev_info info;

double tb_len_bits;

struct thread_params *tp = cb_arg;
- RTE_SET_USED(ret_param);
- queue_id = tp->queue_id;

/* Find matching thread params using queue_id */
for (i = 0; i < MAX_QUEUES; ++i, ++tp)
@@ -1235,18 +1235,19 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- burst_sz = tp->op_params->burst_sz;
+ burst_sz = rte_atomic16_read(&tp->burst_sz);
num_ops = tp->op_params->num_to_process;

- if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
- deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id, dec_ops,
+ if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
+ deq = rte_bbdev_dequeue_dec_ops(dev_id, queue_id,
+ &tp->dec_ops[
+ rte_atomic16_read(&tp->nb_dequeued)],
burst_sz);
- rte_bbdev_dec_op_free_bulk(dec_ops, deq);
- } else {
- deq = rte_bbdev_dequeue_enc_ops(dev_id, queue_id, enc_ops,
+ else
+ deq = rte_bbdev_dequeue_enc_ops(dev_id, queue_id,
+ &tp->enc_ops[
+ rte_atomic16_read(&tp->nb_dequeued)],
burst_sz);
- rte_bbdev_enc_op_free_bulk(enc_ops, deq);
- }

if (deq < burst_sz) {
printf(
@@ -1269,13 +1270,18 @@ typedef int (test_case_function)(struct active_device *ad,

if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) {
struct rte_bbdev_dec_op *ref_op = tp->op_params->ref_dec_op;
- ret = validate_dec_op(dec_ops, num_ops, ref_op,
+ ret = validate_dec_op(tp->dec_ops, num_ops, ref_op,
tp->op_params->vector_mask);
- rte_bbdev_dec_op_free_bulk(dec_ops, deq);
+ /* get the max of iter_count for all dequeued ops */
+ for (i = 0; i < num_ops; ++i)
+ tp->iter_count = RTE_MAX(
+ tp->dec_ops[i]->turbo_dec.iter_count,
+ tp->iter_count);
+ rte_bbdev_dec_op_free_bulk(tp->dec_ops, deq);
} else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC) {
struct rte_bbdev_enc_op *ref_op = tp->op_params->ref_enc_op;
- ret = validate_enc_op(enc_ops, num_ops, ref_op);
- rte_bbdev_enc_op_free_bulk(enc_ops, deq);
+ ret = validate_enc_op(tp->enc_ops, num_ops, ref_op);
+ rte_bbdev_enc_op_free_bulk(tp->enc_ops, deq);
}

if (ret) {
@@ -1299,9 +1305,9 @@ typedef int (test_case_function)(struct active_device *ad,
return;
}

- tp->ops_per_sec = ((double)num_ops) /
+ tp->ops_per_sec += ((double)num_ops) /
((double)total_time / (double)rte_get_tsc_hz());
- tp->mbps = (((double)(num_ops * tb_len_bits)) / 1000000.0) /
+ tp->mbps += (((double)(num_ops * tb_len_bits)) / 1000000.0) /
((double)total_time / (double)rte_get_tsc_hz());

rte_atomic16_add(&tp->nb_dequeued, deq);
@@ -1318,8 +1324,8 @@ typedef int (test_case_function)(struct active_device *ad,
struct rte_bbdev_dec_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
struct rte_bbdev_info info;
- int ret;
- uint16_t num_to_enq;
+ int ret, i, j;
+ uint16_t num_to_enq, enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1351,16 +1357,47 @@ typedef int (test_case_function)(struct active_device *ad,
bufs->hard_outputs, bufs->soft_outputs,
tp->op_params->ref_dec_op);

- tp->start_time = rte_rdtsc_precise();
- for (enqueued = 0; enqueued < num_to_process;) {
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_to_process; ++j)
+ ops[j]->opaque_data = (void *)(uintptr_t)j;

- num_to_enq = burst_sz;
+ for (j = 0; j < TEST_REPETITIONS; ++j) {
+ for (i = 0; i < num_to_process; ++i)
+ rte_pktmbuf_reset(ops[i]->turbo_dec.hard_output.data);

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ tp->start_time = rte_rdtsc_precise();
+ for (enqueued = 0; enqueued < num_to_process;) {
+ num_to_enq = burst_sz;
+
+ if (unlikely(num_to_process - enqueued < num_to_enq))
+ num_to_enq = num_to_process - enqueued;
+
+ enq = 0;
+ do {
+ enq += rte_bbdev_enqueue_dec_ops(tp->dev_id,
+ queue_id, &ops[enqueued],
+ num_to_enq);
+ } while (unlikely(num_to_enq != enq));
+ enqueued += enq;
+
+ /* Write to thread burst_sz current number of enqueued
+ * descriptors. It ensures that proper number of
+ * descriptors will be dequeued in callback
+ * function - needed for last batch in case where
+ * the number of operations is not a multiple of
+ * burst size.
+ */
+ rte_atomic16_set(&tp->burst_sz, num_to_enq);

- enqueued += rte_bbdev_enqueue_dec_ops(tp->dev_id, queue_id,
- &ops[enqueued], num_to_enq);
+ /* Wait until processing of previous batch is
+ * completed.
+ */
+ while (rte_atomic16_read(&tp->nb_dequeued) !=
+ (int16_t) enqueued)
+ rte_pause();
+ }
+ if (j != TEST_REPETITIONS - 1)
+ rte_atomic16_clear(&tp->nb_dequeued);
}

return TEST_SUCCESS;
@@ -1377,8 +1414,8 @@ typedef int (test_case_function)(struct active_device *ad,
struct rte_bbdev_enc_op *ops[num_to_process];
struct test_buffers *bufs = NULL;
struct rte_bbdev_info info;
- int ret;
- uint16_t num_to_enq;
+ int ret, i, j;
+ uint16_t num_to_enq, enq;

TEST_ASSERT_SUCCESS((burst_sz > MAX_BURST),
"BURST_SIZE should be <= %u", MAX_BURST);
@@ -1409,16 +1446,47 @@ typedef int (test_case_function)(struct active_device *ad,
copy_reference_enc_op(ops, num_to_process, 0, bufs->inputs,
bufs->hard_outputs, tp->op_params->ref_enc_op);

- tp->start_time = rte_rdtsc_precise();
- for (enqueued = 0; enqueued < num_to_process;) {
+ /* Set counter to validate the ordering */
+ for (j = 0; j < num_to_process; ++j)
+ ops[j]->opaque_data = (void *)(uintptr_t)j;
+
+ for (j = 0; j < TEST_REPETITIONS; ++j) {
+ for (i = 0; i < num_to_process; ++i)
+ rte_pktmbuf_reset(ops[i]->turbo_enc.output.data);

- num_to_enq = burst_sz;
+ tp->start_time = rte_rdtsc_precise();
+ for (enqueued = 0; enqueued < num_to_process;) {
+ num_to_enq = burst_sz;

- if (unlikely(num_to_process - enqueued < num_to_enq))
- num_to_enq = num_to_process - enqueued;
+ if (unlikely(num_to_process - enqueued < num_to_enq))
+ num_to_enq = num_to_process - enqueued;
+
+ enq = 0;
+ do {
+ enq += rte_bbdev_enqueue_enc_ops(tp->dev_id,
+ queue_id, &ops[enqueued],
+ num_to_enq);
+ } while (unlikely(enq != num_to_enq));
+ enqueued += enq;
+
+ /* Write to thread burst_sz current number of enqueued
+ * descriptors. It ensures that proper number of
+ * descriptors will be dequeued in callback
+ * function - needed for last batch in case where
+ * the number of operations is not a multiple of
+ * burst size.
+ */
+ rte_atomic16_set(&tp->burst_sz, num_to_enq);

- enqueued += rte_bbdev_enqueue_enc_ops(tp->dev_id, queue_id,
- &ops[enqueued], num_to_enq);
+ /* Wait until processing of previous batch is
+ * completed.
+ */
+ while (rte_atomic16_read(&tp->nb_dequeued) !=
+ (int16_t) enqueued)
+ rte_pause();
+ }
+ if (j != TEST_REPETITIONS - 1)
+ rte_atomic16_clear(&tp->nb_dequeued);
}

return TEST_SUCCESS;
@@ -1613,18 +1681,16 @@ typedef int (test_case_function)(struct active_device *ad,
static void
print_enc_throughput(struct thread_params *t_params, unsigned int used_cores)
{
- unsigned int lcore_id, iter = 0;
+ unsigned int iter = 0;
double total_mops = 0, total_mbps = 0;

- RTE_LCORE_FOREACH(lcore_id) {
- if (iter++ >= used_cores)
- break;
+ for (iter = 0; iter < used_cores; iter++) {
printf(
- "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps\n",
- lcore_id, t_params[lcore_id].ops_per_sec,
- t_params[lcore_id].mbps);
- total_mops += t_params[lcore_id].ops_per_sec;
- total_mbps += t_params[lcore_id].mbps;
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps\n",
+ t_params[iter].lcore_id, t_params[iter].ops_per_sec,
+ t_params[iter].mbps);
+ total_mops += t_params[iter].ops_per_sec;
+ total_mbps += t_params[iter].mbps;
}
printf(
"\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps\n",
@@ -1634,21 +1700,18 @@ typedef int (test_case_function)(struct active_device *ad,
static void
print_dec_throughput(struct thread_params *t_params, unsigned int used_cores)
{
- unsigned int lcore_id, iter = 0;
+ unsigned int iter = 0;
double total_mops = 0, total_mbps = 0;
uint8_t iter_count = 0;

- RTE_LCORE_FOREACH(lcore_id) {
- if (iter++ >= used_cores)
- break;
+ for (iter = 0; iter < used_cores; iter++) {
printf(
- "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps @ max %u iterations\n",
- lcore_id, t_params[lcore_id].ops_per_sec,
- t_params[lcore_id].mbps,
- t_params[lcore_id].iter_count);
- total_mops += t_params[lcore_id].ops_per_sec;
- total_mbps += t_params[lcore_id].mbps;
- iter_count = RTE_MAX(iter_count, t_params[lcore_id].iter_count);
+ "Throughput for core (%u): %.8lg Ops/s, %.8lg Mbps @ max %u iterations\n",
+ t_params[iter].lcore_id, t_params[iter].ops_per_sec,
+ t_params[iter].mbps, t_params[iter].iter_count);
+ total_mops += t_params[iter].ops_per_sec;
+ total_mbps += t_params[iter].mbps;
+ iter_count = RTE_MAX(iter_count, t_params[iter].iter_count);
}
printf(
"\nTotal throughput for %u cores: %.8lg MOPS, %.8lg Mbps @ max %u iterations\n",
@@ -1665,10 +1728,9 @@ typedef int (test_case_function)(struct active_device *ad,
{
int ret;
unsigned int lcore_id, used_cores = 0;
- struct thread_params t_params[MAX_QUEUES];
+ struct thread_params *t_params, *tp;
struct rte_bbdev_info info;
lcore_function_t *throughput_function;
- struct thread_params *tp;
uint16_t num_lcores;
const char *op_type_str;

@@ -1691,6 +1753,13 @@ typedef int (test_case_function)(struct active_device *ad,
? ad->nb_queues
: op_params->num_lcores;

+ /* Allocate memory for thread parameters structure */
+ t_params = rte_zmalloc(NULL, num_lcores * sizeof(struct thread_params),
+ RTE_CACHE_LINE_SIZE);
+ TEST_ASSERT_NOT_NULL(t_params, "Failed to alloc %zuB for t_params",
+ RTE_ALIGN(sizeof(struct thread_params) * num_lcores,
+ RTE_CACHE_LINE_SIZE));
+
if (intr_enabled) {
if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
throughput_function = throughput_intr_lcore_dec;
@@ -1700,9 +1769,11 @@ typedef int (test_case_function)(struct active_device *ad,
/* Dequeue interrupt callback registration */
ret = rte_bbdev_callback_register(ad->dev_id,
RTE_BBDEV_EVENT_DEQUEUE, dequeue_event_callback,
- &t_params);
- if (ret < 0)
+ t_params);
+ if (ret < 0) {
+ rte_free(t_params);
return ret;
+ }
} else {
if (test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC)
throughput_function = throughput_pmd_lcore_dec;
@@ -1712,38 +1783,39 @@ typedef int (test_case_function)(struct active_device *ad,

rte_atomic16_set(&op_params->sync, SYNC_WAIT);

- t_params[rte_lcore_id()].dev_id = ad->dev_id;
- t_params[rte_lcore_id()].op_params = op_params;
- t_params[rte_lcore_id()].queue_id =
- ad->queue_ids[used_cores++];
+ /* Master core is set at first entry */
+ t_params[0].dev_id = ad->dev_id;
+ t_params[0].lcore_id = rte_lcore_id();
+ t_params[0].op_params = op_params;
+ t_params[0].queue_id = ad->queue_ids[used_cores++];
+ t_params[0].iter_count = 0;

RTE_LCORE_FOREACH_SLAVE(lcore_id) {
if (used_cores >= num_lcores)
break;

- t_params[lcore_id].dev_id = ad->dev_id;
- t_params[lcore_id].op_params = op_params;
- t_params[lcore_id].queue_id = ad->queue_ids[used_cores++];
+ t_params[used_cores].dev_id = ad->dev_id;
+ t_params[used_cores].lcore_id = lcore_id;
+ t_params[used_cores].op_params = op_params;
+ t_params[used_cores].queue_id = ad->queue_ids[used_cores];
+ t_params[used_cores].iter_count = 0;

- rte_eal_remote_launch(throughput_function, &t_params[lcore_id],
- lcore_id);
+ rte_eal_remote_launch(throughput_function,
+ &t_params[used_cores++], lcore_id);
}

rte_atomic16_set(&op_params->sync, SYNC_START);
- ret = throughput_function(&t_params[rte_lcore_id()]);
+ ret = throughput_function(&t_params[0]);

/* Master core is always used */
- used_cores = 1;
- RTE_LCORE_FOREACH_SLAVE(lcore_id) {
- if (used_cores++ >= num_lcores)
- break;
-
- ret |= rte_eal_wait_lcore(lcore_id);
- }
+ for (used_cores = 1; used_cores < num_lcores; used_cores++)
+ ret |= rte_eal_wait_lcore(t_params[used_cores].lcore_id);

/* Return if test failed */
- if (ret)
+ if (ret) {
+ rte_free(t_params);
return ret;
+ }

/* Print throughput if interrupts are disabled and test passed */
if (!intr_enabled) {
@@ -1751,6 +1823,7 @@ typedef int (test_case_function)(struct active_device *ad,
print_dec_throughput(t_params, num_lcores);
else
print_enc_throughput(t_params, num_lcores);
+ rte_free(t_params);
return ret;
}

@@ -1759,21 +1832,20 @@ typedef int (test_case_function)(struct active_device *ad,
* error using processing_status variable.
* Wait for master lcore operations.
*/
- tp = &t_params[rte_lcore_id()];
+ tp = &t_params[0];
while ((rte_atomic16_read(&tp->nb_dequeued) <
op_params->num_to_process) &&
(rte_atomic16_read(&tp->processing_status) !=
TEST_FAILED))
rte_pause();

+ tp->ops_per_sec /= TEST_REPETITIONS;
+ tp->mbps /= TEST_REPETITIONS;
ret |= rte_atomic16_read(&tp->processing_status);

/* Wait for slave lcores operations */
- used_cores = 1;
- RTE_LCORE_FOREACH_SLAVE(lcore_id) {
- tp = &t_params[lcore_id];
- if (used_cores++ >= num_lcores)
- break;
+ for (used_cores = 1; used_cores < num_lcores; used_cores++) {
+ tp = &t_params[used_cores];

while ((rte_atomic16_read(&tp->nb_dequeued) <
op_params->num_to_process) &&
@@ -1781,6 +1853,8 @@ typedef int (test_case_function)(struct active_device *ad,
TEST_FAILED))
rte_pause();

+ tp->ops_per_sec /= TEST_REPETITIONS;
+ tp->mbps /= TEST_REPETITIONS;
ret |= rte_atomic16_read(&tp->processing_status);
}

@@ -1791,6 +1865,8 @@ typedef int (test_case_function)(struct active_device *ad,
else if (test_vector.op_type == RTE_BBDEV_OP_TURBO_ENC)
print_enc_throughput(t_params, num_lcores);
}
+
+ rte_free(t_params);
return ret;
}
--
1.8.3.1
Loading...