On 24/03/2020 09:40, Pavan Nikhilesh Bhagavatula wrote:
> Hi Ray,
>
> I have tried to avoid hand unrolling loops and found the following
> observations.
>
> 1. Although it decreases LOC it also takes away readability too.
> Example:
> Avoiding unrolled code below
[SNIP]
> Which is kind of unreadable.
I am confused - isn't it exactly the same code?
You still haven't completely unrolled the loop either?
I don't know how one is readable and the other is not.
>
> 2. Not all compilers are made equal. I found that most of the compilers don’t
> Unroll the loop above even when compiled with `-funroll-all-loops`.
> I have checked with following compilers:
> GCC 9.2.0
> Clang 9.0.1
> Aarch64 GCC 7.3.0
> Aarch64 GCC 9.2.0
Compilers have been unrolling fixed length loops for as long time - this isn't
new technology.
If the compiler isn't unrolling you are doing something that makes it think it
is a bad idea.
Hand unrolling the loop isn't the solution, understanding what the compiler is
doing is a better idea.
In front of your for loop insert, to indicate to the compiler what you want to
do.
#pragma unroll BUF_PER_LOOP
With clang you can ask it why it is not unrolling the loop with the following
switches.
(output is verbose, but the reason is in there).
-Rpass=loop-unroll -Rpass-missed=loop-unroll
>
> 3. Performance wise I see a lot of degradation on our platform at least 13%.
Is the loop being unrolled?
> On IA with a Broadwell(Xeon E5-2690) and i40e the performance remain same
> w.r.t Rx/Tx since the
> hotspot is in the Tx path of the driver which limits the per core
> capability.
> But the performance difference in number of cycles per node can be seen
> below:
>
> Hand unrolling:
> +-------------------------------+---------------+---------------+---------------+---------------+---------------+-----------+
> |Node |calls |objs
> |realloc_count |objs/call |objs/sec(10E6) |cycles/call|
> +-------------------------------+---------------+---------------+---------------+---------------+---------------+-----------+
> |ip4_lookup |7765918 |248509344 |1
> |32.000 |27.725408 |779.0000 |
> |ip4_rewrite |7765925 |248509568 |1
> |32.000 |27.725408 |425.0000 |
> |ethdev_tx-1 |7765927 |204056223 |1
> |26.000 |22.762720 |597.0000 |
> |pkt_drop |1389170 |44453409 |1
> |32.000 |4.962688 |298.0000 |
> |ethdev_rx-0-0 |63604111 |248509792 |2
> |32.000 |27.725408 |982.0000 |
> +-------------------------------+---------------+---------------+---------------+---------------+---------------+-----------+
>
> W/o unrolling:
>
> +-------------------------------+---------------+---------------+---------------+---------------+---------------+-----------+
> |Node |calls |objs
> |realloc_count |objs/call |objs/sec(10E6) |cycles/call|
> +-------------------------------+---------------+---------------+---------------+---------------+---------------+-----------+
> |ip4_lookup |18864640 |603668448 |1
> |32.000 |26.051328 |828.0000 |
> |ip4_rewrite |18864646 |603668640 |1
> |32.000 |26.051328 |534.0000 |
> |ethdev_tx-1 |18864648 |527874175 |1
> |27.000 |22.780256 |633.0000 |
> |pkt_drop |2368580 |75794529 |1
> |32.000 |3.271072 |286.0000 |
> |ethdev_rx-0-0 |282058226 |603668864 |2
> |32.000 |26.051328 |994.0000 |
> +-------------------------------+---------------+---------------+---------------+---------------+---------------+-----------+
>
> Considering the above findings we would like to continue unrolling the loops
> by hand.
>
> Regards,
> Pavan.
>
>> -----Original Message-----
>> From: Ray Kinsella <m...@ashroe.eu>
>> Sent: Friday, March 20, 2020 2:44 PM
>> To: Pavan Nikhilesh Bhagavatula <pbhagavat...@marvell.com>; Jerin
>> Jacob Kollanukkaran <jer...@marvell.com>; Nithin Kumar Dabilpuram
>> <ndabilpu...@marvell.com>
>> Cc: dev@dpdk.org; tho...@monjalon.net;
>> david.march...@redhat.com; mattias.ronnb...@ericsson.com; Kiran
>> Kumar Kokkilagadda <kirankum...@marvell.com>
>> Subject: Re: [EXT] Re: [dpdk-dev] [PATCH v1 20/26] node: ipv4 lookup
>> for x86
>>
>>
>>
>> On 19/03/2020 16:13, Pavan Nikhilesh Bhagavatula wrote:
>>>
>>>
>>>> -----Original Message-----
>>>> From: Ray Kinsella <m...@ashroe.eu>
>>>> Sent: Thursday, March 19, 2020 9:21 PM
>>>> To: Pavan Nikhilesh Bhagavatula <pbhagavat...@marvell.com>;
>> Jerin
>>>> Jacob Kollanukkaran <jer...@marvell.com>; Nithin Kumar
>> Dabilpuram
>>>> <ndabilpu...@marvell.com>
>>>> Cc: dev@dpdk.org; tho...@monjalon.net;
>>>> david.march...@redhat.com; mattias.ronnb...@ericsson.com;
>> Kiran
>>>> Kumar Kokkilagadda <kirankum...@marvell.com>
>>>> Subject: Re: [EXT] Re: [dpdk-dev] [PATCH v1 20/26] node: ipv4
>> lookup
>>>> for x86
>>>>
>>>>
>>>>
>>>> On 19/03/2020 14:22, Pavan Nikhilesh Bhagavatula wrote:
>>>>>> On 18/03/2020 21:35, jer...@marvell.com wrote:
>>>>>>> From: Pavan Nikhilesh <pbhagavat...@marvell.com>
>>>>>>>
>>>>>>> Add IPv4 lookup process function for ip4_lookup
>>>>>>> rte_node. This node performs LPM lookup using x86_64
>>>>>>> vector supported RTE_LPM API on every packet received
>>>>>>> and forwards it to a next node that is identified by
>>>>>>> lookup result.
>>>>>>>
>>>>>>> Signed-off-by: Pavan Nikhilesh <pbhagavat...@marvell.com>
>>>>>>> Signed-off-by: Nithin Dabilpuram <ndabilpu...@marvell.com>
>>>>>>> Signed-off-by: Kiran Kumar K <kirankum...@marvell.com>
>>>>>>> ---
>>>>>>> lib/librte_node/ip4_lookup.c | 245
>>>>>> +++++++++++++++++++++++++++++++++++
>>>>>>> 1 file changed, 245 insertions(+)
>>>>>>>
>>>>>>> diff --git a/lib/librte_node/ip4_lookup.c
>>>>>> b/lib/librte_node/ip4_lookup.c
>>>>>>> index d7fcd1158..c003e9c91 100644
>>>>>>> --- a/lib/librte_node/ip4_lookup.c
>>>>>>> +++ b/lib/librte_node/ip4_lookup.c
>>>>>>> @@ -264,6 +264,251 @@ ip4_lookup_node_process(struct
>>>> rte_graph
>>>>>> *graph, struct rte_node *node,
>>>>>>> return nb_objs;
>>>>>>> }
>>>>>>>
>>>>>>> +#elif defined(RTE_ARCH_X86)
>>>>>>> +
>>>>>>> +/* X86 SSE */
>>>>>>> +static uint16_t
>>>>>>> +ip4_lookup_node_process(struct rte_graph *graph, struct
>>>> rte_node
>>>>>> *node,
>>>>>>> + void **objs, uint16_t nb_objs)
>>>>>>> +{
>>>>>>> + struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3,
>> **pkts;
>>>>>>> + rte_edge_t next0, next1, next2, next3, next_index;
>>>>>>> + struct rte_ipv4_hdr *ipv4_hdr;
>>>>>>> + struct rte_ether_hdr *eth_hdr;
>>>>>>> + uint32_t ip0, ip1, ip2, ip3;
>>>>>>> + void **to_next, **from;
>>>>>>> + uint16_t last_spec = 0;
>>>>>>> + uint16_t n_left_from;
>>>>>>> + struct rte_lpm *lpm;
>>>>>>> + uint16_t held = 0;
>>>>>>> + uint32_t drop_nh;
>>>>>>> + rte_xmm_t dst;
>>>>>>> + __m128i dip; /* SSE register */
>>>>>>> + int rc, i;
>>>>>>> +
>>>>>>> + /* Speculative next */
>>>>>>> + next_index =
>> RTE_NODE_IP4_LOOKUP_NEXT_REWRITE;
>>>>>>> + /* Drop node */
>>>>>>> + drop_nh =
>>>>>> ((uint32_t)RTE_NODE_IP4_LOOKUP_NEXT_PKT_DROP) << 16;
>>>>>>> +
>>>>>>> + /* Get socket specific LPM from ctx */
>>>>>>> + lpm = *((struct rte_lpm **)node->ctx);
>>>>>>> +
>>>>>>> + pkts = (struct rte_mbuf **)objs;
>>>>>>> + from = objs;
>>>>>>> + n_left_from = nb_objs;
>>>>>>
>>>>>> I doubt this initial prefetch of the first 4 packets has any benefit.
>>>>>
>>>>> Ack will remove in v2 for x86.
>>>>>
>>>>>>
>>>>>>> + if (n_left_from >= 4) {
>>>>>>> + for (i = 0; i < 4; i++) {
>>>>>>> +
>> rte_prefetch0(rte_pktmbuf_mtod(pkts[i],
>>>>>>> + struct
>> rte_ether_hdr
>>>>>> *) +
>>>>>>> + 1);
>>>>>>> + }
>>>>>>> + }
>>>>>>> +
>>>>>>> + /* Get stream for the speculated next node */
>>>>>>> + to_next = rte_node_next_stream_get(graph, node,
>>>>>> next_index, nb_objs);
>>>>>>
>>>>>> Suggest you don't reuse the hand-unrolling optimization from
>> FD.io
>>>>>> VPP.
>>>>>> I have never found any performance benefit from them, and they
>>>>>> make the code unnecessarily verbose.
>>>>>>
>>>>>
>>>>> How would be take the benefit of rte_lpm_lookupx4 without
>>>> unrolling the loop?.
>>>>> Also, in future if we are using rte_rib and fib with a CPU supporting
>>>> wider SIMD we might
>>>>> need to unroll them further (AVX256 AND 512 currently
>>>> rte_lpm_lookup uses only 128bit
>>>>> since it is only uses SSE extension).
>>>>
>>>> Let the compiler do it for you, but using a constant vector length.
>>>> for (int i=0; i < 4; ++i) { ... }
>>>>
>>>
>>> Ok, I think I misunderstood the previous comment.
>>> It was only for the prefetches in the loop right?
>>
>>
>> no, it was for all the needless repetition.
>> hand-unrolling loops serve no purpose but to add verbosity.
>>
>>>
>>>>>
>>>>>>
>>>>>>> + while (n_left_from >= 4) {
>>>>>>> + /* Prefetch next-next mbufs */
>>>>>>> + if (likely(n_left_from >= 11)) {
>>>>>>> + rte_prefetch0(pkts[8]);
>>>>>>> + rte_prefetch0(pkts[9]);
>>>>>>> + rte_prefetch0(pkts[10]);
>>>>>>> + rte_prefetch0(pkts[11]);
>>>>>>> + }
>>>>>>> +
>>>>>>> + /* Prefetch next mbuf data */
>>>>>>> + if (likely(n_left_from >= 7)) {
>>>>>>> +
>> rte_prefetch0(rte_pktmbuf_mtod(pkts[4],
>>>>>>> + struct
>> rte_ether_hdr
>>>>>> *) +
>>>>>>> + 1);
>>>>>>> +
>> rte_prefetch0(rte_pktmbuf_mtod(pkts[5],
>>>>>>> + struct
>> rte_ether_hdr
>>>>>> *) +
>>>>>>> + 1);
>>>>>>> +
>> rte_prefetch0(rte_pktmbuf_mtod(pkts[6],
>>>>>>> + struct
>> rte_ether_hdr
>>>>>> *) +
>>>>>>> + 1);
>>>>>>> +
>> rte_prefetch0(rte_pktmbuf_mtod(pkts[7],
>>>>>>> + struct
>> rte_ether_hdr
>>>>>> *) +
>>>>>>> + 1);
>>>>>>> + }
>>>>>>> +
>>>>>>> + mbuf0 = pkts[0];
>>>>>>> + mbuf1 = pkts[1];
>>>>>>> + mbuf2 = pkts[2];
>>>>>>> + mbuf3 = pkts[3];
>>>>>>> +
>>>>>>> + pkts += 4;
>>>>>>> + n_left_from -= 4;
>>>>>>> +
>>>>>>> + /* Extract DIP of mbuf0 */
>>>>>>> + eth_hdr = rte_pktmbuf_mtod(mbuf0, struct
>>>>>> rte_ether_hdr *);
>>>>>>> + ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
>>>>>>> + ip0 = ipv4_hdr->dst_addr;
>>>>>>> + /* Extract cksum, ttl as ipv4 hdr is in cache */
>>>>>>> + rte_node_mbuf_priv1(mbuf0)->cksum =
>> ipv4_hdr-
>>>>>>> hdr_checksum;
>>>>>>> + rte_node_mbuf_priv1(mbuf0)->ttl = ipv4_hdr-
>>>>>>> time_to_live;
>>>>>>> +
>>>>>>> + /* Extract DIP of mbuf1 */
>>>>>>> + eth_hdr = rte_pktmbuf_mtod(mbuf1, struct
>>>>>> rte_ether_hdr *);
>>>>>>> + ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
>>>>>>> + ip1 = ipv4_hdr->dst_addr;
>>>>>>> + /* Extract cksum, ttl as ipv4 hdr is in cache */
>>>>>>> + rte_node_mbuf_priv1(mbuf1)->cksum =
>> ipv4_hdr-
>>>>>>> hdr_checksum;
>>>>>>> + rte_node_mbuf_priv1(mbuf1)->ttl = ipv4_hdr-
>>>>>>> time_to_live;
>>>>>>> +
>>>>>>> + /* Extract DIP of mbuf2 */
>>>>>>> + eth_hdr = rte_pktmbuf_mtod(mbuf2, struct
>>>>>> rte_ether_hdr *);
>>>>>>> + ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
>>>>>>> + ip2 = ipv4_hdr->dst_addr;
>>>>>>> + /* Extract cksum, ttl as ipv4 hdr is in cache */
>>>>>>> + rte_node_mbuf_priv1(mbuf2)->cksum =
>> ipv4_hdr-
>>>>>>> hdr_checksum;
>>>>>>> + rte_node_mbuf_priv1(mbuf2)->ttl = ipv4_hdr-
>>>>>>> time_to_live;
>>>>>>> +
>>>>>>> + /* Extract DIP of mbuf3 */
>>>>>>> + eth_hdr = rte_pktmbuf_mtod(mbuf3, struct
>>>>>> rte_ether_hdr *);
>>>>>>> + ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
>>>>>>> + ip3 = ipv4_hdr->dst_addr;
>>>>>>> +
>>>>>>> + /* Prepare for lookup x4 */
>>>>>>> + dip = _mm_set_epi32(ip3, ip2, ip1, ip0);
>>>>>>> +
>>>>>>> + /* Byte swap 4 IPV4 addresses. */
>>>>>>> + const __m128i bswap_mask = _mm_set_epi8(
>>>>>>> + 12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1,
>> 2, 3);
>>>>>>> + dip = _mm_shuffle_epi8(dip, bswap_mask);
>>>>>>> +
>>>>>>> + /* Extract cksum, ttl as ipv4 hdr is in cache */
>>>>>>> + rte_node_mbuf_priv1(mbuf3)->cksum =
>> ipv4_hdr-
>>>>>>> hdr_checksum;
>>>>>>> + rte_node_mbuf_priv1(mbuf3)->ttl = ipv4_hdr-
>>>>>>> time_to_live;
>>>>>>> +
>>>>>>> + /* Perform LPM lookup to get NH and next
>> node */
>>>>>>> + rte_lpm_lookupx4(lpm, dip, dst.u32, drop_nh);
>>>>>>> +
>>>>>>> + /* Extract next node id and NH */
>>>>>>> + rte_node_mbuf_priv1(mbuf0)->nh = dst.u32[0]
>> &
>>>>>> 0xFFFF;
>>>>>>> + next0 = (dst.u32[0] >> 16);
>>>>>>> +
>>>>>>> + rte_node_mbuf_priv1(mbuf1)->nh = dst.u32[1]
>> &
>>>>>> 0xFFFF;
>>>>>>> + next1 = (dst.u32[1] >> 16);
>>>>>>> +
>>>>>>> + rte_node_mbuf_priv1(mbuf2)->nh = dst.u32[2]
>> &
>>>>>> 0xFFFF;
>>>>>>> + next2 = (dst.u32[2] >> 16);
>>>>>>> +
>>>>>>> + rte_node_mbuf_priv1(mbuf3)->nh = dst.u32[3]
>> &
>>>>>> 0xFFFF;
>>>>>>> + next3 = (dst.u32[3] >> 16);
>>>>>>> +
>>>>>>> + /* Enqueue four to next node */
>>>>>>> + rte_edge_t fix_spec =
>>>>>>> + (next_index ^ next0) | (next_index ^
>> next1) |
>>>>>>> + (next_index ^ next2) | (next_index ^
>> next3);
>>>>>>> +
>>>>>>> + if (unlikely(fix_spec)) {
>>>>>>> + /* Copy things successfully speculated
>> till now
>>>>>> */
>>>>>>> + rte_memcpy(to_next, from, last_spec *
>>>>>> sizeof(from[0]));
>>>>>>> + from += last_spec;
>>>>>>> + to_next += last_spec;
>>>>>>> + held += last_spec;
>>>>>>> + last_spec = 0;
>>>>>>> +
>>>>>>> + /* Next0 */
>>>>>>> + if (next_index == next0) {
>>>>>>> + to_next[0] = from[0];
>>>>>>> + to_next++;
>>>>>>> + held++;
>>>>>>> + } else {
>>>>>>> + rte_node_enqueue_x1(graph,
>> node,
>>>>>> next0,
>>>>>>> + from[0]);
>>>>>>> + }
>>>>>>> +
>>>>>>> + /* Next1 */
>>>>>>> + if (next_index == next1) {
>>>>>>> + to_next[0] = from[1];
>>>>>>> + to_next++;
>>>>>>> + held++;
>>>>>>> + } else {
>>>>>>> + rte_node_enqueue_x1(graph,
>> node,
>>>>>> next1,
>>>>>>> + from[1]);
>>>>>>> + }
>>>>>>> +
>>>>>>> + /* Next2 */
>>>>>>> + if (next_index == next2) {
>>>>>>> + to_next[0] = from[2];
>>>>>>> + to_next++;
>>>>>>> + held++;
>>>>>>> + } else {
>>>>>>> + rte_node_enqueue_x1(graph,
>> node,
>>>>>> next2,
>>>>>>> + from[2]);
>>>>>>> + }
>>>>>>> +
>>>>>>> + /* Next3 */
>>>>>>> + if (next_index == next3) {
>>>>>>> + to_next[0] = from[3];
>>>>>>> + to_next++;
>>>>>>> + held++;
>>>>>>> + } else {
>>>>>>> + rte_node_enqueue_x1(graph,
>> node,
>>>>>> next3,
>>>>>>> + from[3]);
>>>>>>> + }
>>>>>>> +
>>>>>>> + from += 4;
>>>>>>> +
>>>>>>> + } else {
>>>>>>> + last_spec += 4;
>>>>>>> + }
>>>>>>> + }
>>>>>>> +
>>>>>>> + while (n_left_from > 0) {
>>>>>>> + uint32_t next_hop;
>>>>>>> +
>>>>>>> + mbuf0 = pkts[0];
>>>>>>> +
>>>>>>> + pkts += 1;
>>>>>>> + n_left_from -= 1;
>>>>>>> +
>>>>>>> + /* Extract DIP of mbuf0 */
>>>>>>> + eth_hdr = rte_pktmbuf_mtod(mbuf0, struct
>>>>>> rte_ether_hdr *);
>>>>>>> + ipv4_hdr = (struct rte_ipv4_hdr *)(eth_hdr + 1);
>>>>>>> + /* Extract cksum, ttl as ipv4 hdr is in cache */
>>>>>>> + rte_node_mbuf_priv1(mbuf0)->cksum =
>> ipv4_hdr-
>>>>>>> hdr_checksum;
>>>>>>> + rte_node_mbuf_priv1(mbuf0)->ttl = ipv4_hdr-
>>>>>>> time_to_live;
>>>>>>> +
>>>>>>> + rc = rte_lpm_lookup(lpm,
>> rte_be_to_cpu_32(ipv4_hdr-
>>>>>>> dst_addr),
>>>>>>> + &next_hop);
>>>>>>> + next_hop = (rc == 0) ? next_hop : drop_nh;
>>>>>>> +
>>>>>>> + rte_node_mbuf_priv1(mbuf0)->nh = next_hop
>> &
>>>>>> 0xFFFF;
>>>>>>> + next0 = (next_hop >> 16);
>>>>>>> +
>>>>>>> + if (unlikely(next_index ^ next0)) {
>>>>>>> + /* Copy things successfully speculated
>> till now
>>>>>> */
>>>>>>> + rte_memcpy(to_next, from, last_spec *
>>>>>> sizeof(from[0]));
>>>>>>> + from += last_spec;
>>>>>>> + to_next += last_spec;
>>>>>>> + held += last_spec;
>>>>>>> + last_spec = 0;
>>>>>>> +
>>>>>>> + rte_node_enqueue_x1(graph, node,
>> next0,
>>>>>> from[0]);
>>>>>>> + from += 1;
>>>>>>> + } else {
>>>>>>> + last_spec += 1;
>>>>>>> + }
>>>>>>> + }
>>>>>>> +
>>>>>>> + /* !!! Home run !!! */
>>>>>>> + if (likely(last_spec == nb_objs)) {
>>>>>>> + rte_node_next_stream_move(graph, node,
>>>>>> next_index);
>>>>>>> + return nb_objs;
>>>>>>> + }
>>>>>>> +
>>>>>>> + held += last_spec;
>>>>>>> + /* Copy things successfully speculated till now */
>>>>>>> + rte_memcpy(to_next, from, last_spec *
>> sizeof(from[0]));
>>>>>>> + rte_node_next_stream_put(graph, node, next_index,
>> held);
>>>>>>> +
>>>>>>> + return nb_objs;
>>>>>>> +}
>>>>>>> +
>>>>>>> #else
>>>>>>>
>>>>>>> static uint16_t
>>>>>>>