[PATCH v1 07/13] test/bbdev: add support for BLER for 4G
Vargas, Hernan
hernan.vargas at intel.com
Wed Feb 22 22:55:06 CET 2023
Hi Maxime,
> -----Original Message-----
> From: Maxime Coquelin <maxime.coquelin at redhat.com>
> Sent: Monday, February 20, 2023 9:44 AM
> To: Vargas, Hernan <hernan.vargas at intel.com>; dev at dpdk.org;
> gakhil at marvell.com; Rix, Tom <trix at redhat.com>
> Cc: Chautru, Nicolas <nicolas.chautru at intel.com>; Zhang, Qi Z
> <qi.z.zhang at intel.com>
> Subject: Re: [PATCH v1 07/13] test/bbdev: add support for BLER for 4G
>
>
>
> On 2/13/23 21:59, Vargas, Hernan wrote:
> >
> >
> >> -----Original Message-----
> >> From: Maxime Coquelin <maxime.coquelin at redhat.com>
> >> Sent: Tuesday, January 31, 2023 4:20 AM
> >> To: Vargas, Hernan <hernan.vargas at intel.com>; dev at dpdk.org;
> >> gakhil at marvell.com; Rix, Tom <trix at redhat.com>
> >> Cc: Chautru, Nicolas <nicolas.chautru at intel.com>; Zhang, Qi Z
> >> <qi.z.zhang at intel.com>
> >> Subject: Re: [PATCH v1 07/13] test/bbdev: add support for BLER for 4G
> >>
> >>
> >>
> >> On 1/17/23 17:50, Hernan Vargas wrote:
> >>> New feature to add BLER support for 4G in bbdev-test.
> >>>
> >>> Signed-off-by: Hernan Vargas <hernan.vargas at intel.com>
> >>> ---
> >>> app/test-bbdev/test_bbdev_perf.c | 184
> >> ++++++++++++++++++++++++++++++-
> >>> 1 file changed, 183 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/app/test-bbdev/test_bbdev_perf.c
> >>> b/app/test-bbdev/test_bbdev_perf.c
> >>> index ffb2de0604..69b86cdeb1 100644
> >>> --- a/app/test-bbdev/test_bbdev_perf.c
> >>> +++ b/app/test-bbdev/test_bbdev_perf.c
> >>> @@ -1756,6 +1756,30 @@ gen_qm2_llr(int8_t *llrs, uint32_t j, double
> >>> N0,
> >> double llr_max)
> >>> llrs[j] = (int8_t) b;
> >>> }
> >>>
> >>> +/* Simple LLR generation assuming AWGN and QPSK */ static void
> >>> +gen_turbo_llr(int8_t *llrs, uint32_t j, double N0, double llr_max) {
> >>> + double b, b1, n;
> >>> + double coeff = 2.0 * sqrt(N0);
> >>> +
> >>> + /* Ignore in vectors null LLRs not to be saturated */
> >>> + if (llrs[j] == 0)
> >>> + return;
> >>> +
> >>> + /* Note don't change sign here */
> >>> + n = randn(j % 2);
> >>> + b1 = ((llrs[j] > 0 ? 2.0 : -2.0)
> >>> + + coeff * n) / N0;
> >>> + b = b1 * (1 << 4);
> >>> + b = round(b);
> >>> + if (b > llr_max)
> >>> + b = llr_max;
> >>> + if (b < -llr_max)
> >>> + b = -llr_max;
> >>> + llrs[j] = (int8_t) b;
> >>> +}
> >>> +
> >>> /* Generate LLR for a given SNR */
> >>> static void
> >>> generate_llr_input(uint16_t n, struct rte_bbdev_op_data *inputs,
> >>> @@
> >>> -1791,6 +1815,27 @@ generate_llr_input(uint16_t n, struct
> >> rte_bbdev_op_data *inputs,
> >>> }
> >>> }
> >>>
> >>> +/* Generate LLR for turbo decoder for a given SNR */ static void
> >>> +generate_turbo_llr_input(uint16_t n, struct rte_bbdev_op_data
> *inputs,
> >>> + struct rte_bbdev_dec_op *ref_op)
> >>> +{
> >>> + struct rte_mbuf *m;
> >>> + uint32_t i, j, range;
> >>> + double N0, llr_max;
> >>> +
> >>> + llr_max = 127;
> >>> + range = ref_op->turbo_dec.input.length;
> >>> + N0 = 1.0 / pow(10.0, get_snr() / 10.0);
> >>> +
> >>> + for (i = 0; i < n; ++i) {
> >>> + m = inputs[i].data;
> >>> + int8_t *llrs = rte_pktmbuf_mtod_offset(m, int8_t *, 0);
> >>> + for (j = 0; j < range; ++j)
> >>> + gen_turbo_llr(llrs, j, N0, llr_max);
> >>> + }
> >>> +}
> >>> +
> >>> static void
> >>> copy_reference_ldpc_dec_op(struct rte_bbdev_dec_op **ops,
> >>> unsigned
> >> int n,
> >>> unsigned int start_idx,
> >>> @@ -2301,6 +2346,31 @@ validate_ldpc_bler(struct rte_bbdev_dec_op
> >> **ops, const uint16_t n)
> >>> return errors;
> >>> }
> >>>
> >>> +/* Check Number of code blocks errors */ static int
> >>> +validate_turbo_bler(struct rte_bbdev_dec_op **ops, const uint16_t
> >>> +n) {
> >>> + unsigned int i;
> >>> + struct op_data_entries *hard_data_orig =
> >>> + &test_vector.entries[DATA_HARD_OUTPUT];
> >>> + struct rte_bbdev_op_turbo_dec *ops_td;
> >>> + struct rte_bbdev_op_data *hard_output;
> >>> + int errors = 0;
> >>> + struct rte_mbuf *m;
> >>> +
> >>> + for (i = 0; i < n; ++i) {
> >>> + ops_td = &ops[i]->turbo_dec;
> >>> + hard_output = &ops_td->hard_output;
> >>> + m = hard_output->data;
> >>> + if (memcmp(rte_pktmbuf_mtod_offset(m, uint32_t *, 0),
> >>> + hard_data_orig->segments[0].addr,
> >>> + hard_data_orig->segments[0].length))
> >>
> >> No need to check mbuf is at least as large as segment 0 length?
> >>
> >>> + errors++;
> >>> + }
> >>> + return errors;
> >>> +}
> >>> +
> >>> +
> >>> static int
> >>> validate_ldpc_dec_op(struct rte_bbdev_dec_op **ops, const uint16_t
> n,
> >>> struct rte_bbdev_dec_op *ref_op, const int vector_mask)
> >> @@ -3736,6
> >>> +3806,114 @@ bler_pmd_lcore_ldpc_dec(void *arg)
> >>> return TEST_SUCCESS;
> >>> }
> >>>
> >>> +
> >>> +static int
> >>> +bler_pmd_lcore_turbo_dec(void *arg) {
> >>> + struct thread_params *tp = arg;
> >>> + 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_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;
> >>> + int i, j, ret;
> >>> + struct rte_bbdev_info info;
> >>> + 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];
> >>> +
> >>> + rte_wait_until_equal_16(&tp->op_params->sync, SYNC_START,
> >>> +__ATOMIC_RELAXED);
> >>> +
> >>> + 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);
> >>> +
> >>> + /* For BLER tests we need to enable early termination */
> >>> + if (!check_bit(ref_op->turbo_dec.op_flags,
> >>> + RTE_BBDEV_TURBO_EARLY_TERMINATION))
> >>> + ref_op->turbo_dec.op_flags +=
> >>> + RTE_BBDEV_TURBO_EARLY_TERMINATION;
> >>> + ref_op->turbo_dec.iter_max = get_iter_max();
> >>> + ref_op->turbo_dec.iter_count = ref_op->turbo_dec.iter_max;
> >>> +
> >>> + 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);
> >>> + generate_turbo_llr_input(num_ops, bufs->inputs, ref_op);
> >>> +
> >>> + /* Set counter to validate the ordering */
> >>> + for (j = 0; j < num_ops; ++j)
> >>> + ops_enq[j]->opaque_data = (void *)(uintptr_t)j;
> >>> +
> >>> + for (i = 0; i < 1; ++i) { /* Could add more iterations */
> >>> + for (j = 0; j < num_ops; ++j) {
> >>> + mbuf_reset(
> >>> + ops_enq[j]->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;
> >>> +
> >>> + enq += rte_bbdev_enqueue_dec_ops(tp->dev_id,
> >>> + queue_id, &ops_enq[enq],
> >> num_to_enq);
> >>
> >> If for some reason the operation fails (looks like it can fail in at
> >> least ACC200 I checked), it ends up doing an endless loop as enc won't get
> incremented.
> > The python wrapper script used to invoke test-bbdev has a configurable
> timeout (default 600 seconds) that would stop the test in case it ends up in
> an infinite loop.
>
> Ok, but having a timeout in the test script does not prevent to do a proprer
> error propagation/handling in the application.
>
> It is happens in when running in a CI, having errors properly checked would
> provide useful information to investigate without having to reproduce, or
> when failing to reproduce.
>
> Regards,
> Maxime
>
This is a good point for future improvement but not related to this series.
We can implement it in the future. What do you think?
Thanks,
Hernan
> >
> >>> +
> >>> + deq += rte_bbdev_dequeue_dec_ops(tp->dev_id,
> >>> + queue_id, &ops_deq[deq], enq -
> >> deq);
> >>> + }
> >>> +
> >>> + /* dequeue the remaining */
> >>> + while (deq < enq) {
> >>> + deq += rte_bbdev_dequeue_dec_ops(tp->dev_id,
> >>> + queue_id, &ops_deq[deq], enq -
> >> deq);
> >>> + }
> >>> +
> >>> + total_time += rte_rdtsc_precise() - start_time;
> >>> + }
> >>> +
> >>> + tp->iter_count = 0;
> >>> + tp->iter_average = 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);
> >>> + tp->iter_average += (double) ops_enq[i]-
> >>> turbo_dec.iter_count;
> >>> + }
> >>> +
> >>> + tp->iter_average /= num_ops;
> >>> + tp->bler = (double) validate_turbo_bler(ops_deq, num_ops) /
> >> num_ops;
> >>> +
> >>> + 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 * 1) /
> >>> + ((double)total_time / (double)rte_get_tsc_hz());
> >>> + tp->mbps = (((double)(num_ops * 1 * tb_len_bits)) /
> >>> + 1000000.0) / ((double)total_time /
> >>> + (double)rte_get_tsc_hz());
> >>> + printf("TBS %.0f Time %.0f\n", tb_len_bits, 1000000.0 *
> >>> + ((double)total_time / (double)rte_get_tsc_hz()));
> >>> +
> >>> + return TEST_SUCCESS;
> >>> +}
> >>> +
> >>> static int
> >>> throughput_pmd_lcore_ldpc_dec(void *arg)
> >>> {
> >>> @@ -4193,7 +4371,7 @@ print_dec_bler(struct thread_params
> *t_params,
> >> unsigned int used_cores)
> >>> total_bler /= used_cores;
> >>> total_iter /= used_cores;
> >>>
> >>> - printf("SNR %.2f BLER %.1f %% - Iterations %.1f %d - Tp %.1f Mbps
> >> %s\n",
> >>> + printf("SNR %.2f BLER %.1f %% - Iterations %.1f %d - Tp %.3f Mbps
> >>> +%s\n",
> >>> snr, total_bler * 100, total_iter, get_iter_max(),
> >>> total_mbps, get_vector_filename());
> >>> }
> >>> @@ -4245,6 +4423,10 @@ bler_test(struct active_device *ad,
> >>> && !check_bit(test_vector.ldpc_dec.op_flags,
> >>> RTE_BBDEV_LDPC_LLR_COMPRESSION))
> >>> bler_function = bler_pmd_lcore_ldpc_dec;
> >>> + else if ((test_vector.op_type == RTE_BBDEV_OP_TURBO_DEC) &&
> >>> + !check_bit(test_vector.turbo_dec.op_flags,
> >>> + RTE_BBDEV_TURBO_SOFT_OUTPUT))
> >>> + bler_function = bler_pmd_lcore_turbo_dec;
> >>> else
> >>> return TEST_SKIPPED;
> >>>
> >
More information about the dev
mailing list