static inline void rdtsc_prof_start(struct rdtsc_prof *p)
{
p->clk_start = rte_rdtsc_precise();
}
/* run benchmark per burst size */
static inline int
pmd_cyclecount_bench_burst_sz(
struct pmd_cyclecount_state *state, uint16_t test_burst_size)
{
uint64_t tsc_start;
uint64_t tsc_end;
uint64_t tsc_op;
uint64_t tsc_enq;
uint64_t tsc_deq;
uint32_t cur_op;
/* reset all counters */
tsc_enq = 0;
tsc_deq = 0;
state->ops_enqd = 0;
state->ops_enq_retries = 0;
state->ops_deqd = 0;
state->ops_deq_retries = 0;
/*
* Benchmark crypto op alloc-build-free separately.
*/
tsc_start = rte_rdtsc_precise();
for (cur_op = 0; cur_op < state->opts->total_ops;
cur_op += state->opts->nb_descriptors) {
if (unlikely(pmd_cyclecount_bench_ops(
state, cur_op, test_burst_size)))
return -1;
}
tsc_end = rte_rdtsc_precise();
tsc_op = tsc_end - tsc_start;
/*
* Hardware acceleration cyclecount benchmarking loop.
*
* We're benchmarking raw enq/deq performance by filling up the device
* queue, so we never get any failed enqs unless the driver won't accept
* the exact number of descriptors we requested, or the driver won't
* wrap around the end of the TX ring. However, since we're only
* dequeueing once we've filled up the queue, we have to benchmark it
* piecemeal and then average out the results.
*/
cur_op = 0;
while (cur_op < state->opts->total_ops) {
uint32_t iter_ops_left = state->opts->total_ops - cur_op;
uint32_t iter_ops_needed = RTE_MIN(
state->opts->nb_descriptors, iter_ops_left);
uint32_t iter_ops_allocd = iter_ops_needed;
/* allocate and build ops */
if (unlikely(pmd_cyclecount_build_ops(state, iter_ops_needed,
test_burst_size)))
return -1;
tsc_start = rte_rdtsc_precise();
/* fill up TX ring */
iter_ops_needed = pmd_cyclecount_bench_enq(state,
iter_ops_needed, test_burst_size);
tsc_end = rte_rdtsc_precise();
tsc_enq += tsc_end - tsc_start;
/* allow for HW to catch up */
if (state->delay)
rte_delay_us_block(state->delay);
tsc_start = rte_rdtsc_precise();
/* drain RX ring */
pmd_cyclecount_bench_deq(state, iter_ops_needed,
test_burst_size);
tsc_end = rte_rdtsc_precise();
tsc_deq += tsc_end - tsc_start;
cur_op += iter_ops_needed;
/*
* we may not have processed all ops that we allocated, so
* free everything we've allocated.
*/
rte_mempool_put_bulk(state->ctx->pool,
(void **)state->ctx->ops, iter_ops_allocd);
}
state->cycles_per_build = (double)tsc_op / state->opts->total_ops;
state->cycles_per_enq = (double)tsc_enq / state->ops_enqd;
state->cycles_per_deq = (double)tsc_deq / state->ops_deqd;
return 0;
}
