Re: [PATCH v1 07/13] test/bbdev: add support for BLER for 4G

[Maxime Coquelin <maxime.coquelin@redhat.com>]

On 2/13/23 21:59, Vargas, Hernan wrote:
> 
> 
>> -----Original Message-----
>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
>> Sent: Tuesday, January 31, 2023 4:20 AM
>> To: Vargas, Hernan <hernan.vargas@intel.com>; dev@dpdk.org;
>> gakhil@marvell.com; Rix, Tom <trix@redhat.com>
>> Cc: Chautru, Nicolas <nicolas.chautru@intel.com>; Zhang, Qi Z
>> <qi.z.zhang@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@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

> 
>>> +
>>> +			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;
>>>
>