DPDK patches and discussions
 help / color / mirror / Atom feed
From: "Stanisław Kardach" <kda@semihalf.com>
To: "Medvedkin, Vladimir" <vladimir.medvedkin@intel.com>
Cc: Michal Mazurek <maz@semihalf.com>, dev <dev@dpdk.org>,
	 Frank Zhao <Frank.Zhao@starfivetech.com>,
	Sam Grove <sam.grove@sifive.com>,
	 Marcin Wojtas <mw@semihalf.com>,
	upstream@semihalf.com,
	 Bruce Richardson <bruce.richardson@intel.com>
Subject: Re: [PATCH 1/1] lpm: add a scalar version of lookupx4 function
Date: Tue, 24 May 2022 18:28:02 +0200	[thread overview]
Message-ID: <CALVGJWJjf7NX31-0rzWD6tWXC4sgeCT-iKwuU3XAocMBHCGOsQ@mail.gmail.com> (raw)
In-Reply-To: <d60986b2-8b8c-c12e-e960-007b83e85e1d@intel.com>

On Thu, May 19, 2022 at 7:04 PM Medvedkin, Vladimir
<vladimir.medvedkin@intel.com> wrote:
>
> Hi Stanislaw, Michal,
>
> As far as I can see, this implementation almost completely repeats other
> lookupx4() implementations, except for the use of vector instructions.
>
> On my board (x86_64) in lpm_perf_autotest your implementation takes about:
> LPM LookupX4: 29.5 cycles (fails = 12.5%)
>
> replacing this code with a simple loop with rte_lpm_lookup():
>
> uint32_t nh;
> int i, ret;
>
> for (i = 0; i < 4; i++) {
>    ret = rte_lpm_lookup((struct rte_lpm *)lpm, ((rte_xmm_t)ip).u32[i], &nh);
>    hop[i] = (ret == 0) ? nh : defv;
> }
>
> works faster:
> LPM LookupX4: 22.2 cycles (fails = 12.5%)
>
> I'm wondering if this will work faster on your board (I assume it it
> RISC-V arch)?
Hi Vladimir,

On my HiFive Unmatched RISC-V board there is a marginal difference (~ -1.56%):
  Our version: 210.5 cycles (fails = 12.5%)
  rte_lpm_lookup version: 213.8 cycles (fails = 12.5%)
Given that x86 is faster with rte_lpm_lookup, I'll change to this
implementation in the next version.

That said I wonder why do we have different const requirements for
rte_lpm_lookup() and rte_lpm_lookupx4():
  static inline int rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip,
uint32_t *next_hop)
  static inline void rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t
ip, uint32_t hop[4], uint32_t defv);
I think both should be const.


>
> Thanks!
>
> On 10/05/2022 12:58, Stanislaw Kardach wrote:
> > From: Michal Mazurek <maz@semihalf.com>
> >
> > Add an implementation of the rte_lpm_lookupx4() function for platforms
> > without support for vector operations.
> >
> > This will be useful in the upcoming RISC-V port as well as any platform
> > which may want to start with a basic level of LPM support.
> >
> > Signed-off-by: Michal Mazurek <maz@semihalf.com>
> > Signed-off-by: Stanislaw Kardach <kda@semihalf.com>
> > ---
> >   doc/guides/rel_notes/release_22_07.rst |   5 +
> >   lib/lpm/meson.build                    |   1 +
> >   lib/lpm/rte_lpm.h                      |   4 +-
> >   lib/lpm/rte_lpm_scalar.h               | 122 +++++++++++++++++++++++++
> >   4 files changed, 131 insertions(+), 1 deletion(-)
> >   create mode 100644 lib/lpm/rte_lpm_scalar.h
> >
> > diff --git a/doc/guides/rel_notes/release_22_07.rst b/doc/guides/rel_notes/release_22_07.rst
> > index 4ae91dd94d..73e8d632f2 100644
> > --- a/doc/guides/rel_notes/release_22_07.rst
> > +++ b/doc/guides/rel_notes/release_22_07.rst
> > @@ -70,6 +70,11 @@ New Features
> >     * Added AH mode support in lookaside protocol (IPsec) for CN9K & CN10K.
> >     * Added AES-GMAC support in lookaside protocol (IPsec) for CN9K & CN10K.
> >
> > +* **Added scalar version of the LPM library.**
> > +
> > +  * Added scalar implementation of ``rte_lpm_lookupx4``. This is a fall-back
> > +    implementation for platforms that don't support vector operations.
> > +
> >
> >   Removed Items
> >   -------------
> > diff --git a/lib/lpm/meson.build b/lib/lpm/meson.build
> > index 78d91d3421..6b47361fce 100644
> > --- a/lib/lpm/meson.build
> > +++ b/lib/lpm/meson.build
> > @@ -14,6 +14,7 @@ headers = files('rte_lpm.h', 'rte_lpm6.h')
> >   indirect_headers += files(
> >           'rte_lpm_altivec.h',
> >           'rte_lpm_neon.h',
> > +        'rte_lpm_scalar.h',
> >           'rte_lpm_sse.h',
> >           'rte_lpm_sve.h',
> >   )
> > diff --git a/lib/lpm/rte_lpm.h b/lib/lpm/rte_lpm.h
> > index eb91960e81..b5db6a353a 100644
> > --- a/lib/lpm/rte_lpm.h
> > +++ b/lib/lpm/rte_lpm.h
> > @@ -405,8 +405,10 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4],
> >   #endif
> >   #elif defined(RTE_ARCH_PPC_64)
> >   #include "rte_lpm_altivec.h"
> > -#else
> > +#elif defined(RTE_ARCH_X86)
> >   #include "rte_lpm_sse.h"
> > +#else
> > +#include "rte_lpm_scalar.h"
> >   #endif
> >
> >   #ifdef __cplusplus
> > diff --git a/lib/lpm/rte_lpm_scalar.h b/lib/lpm/rte_lpm_scalar.h
> > new file mode 100644
> > index 0000000000..991b94e687
> > --- /dev/null
> > +++ b/lib/lpm/rte_lpm_scalar.h
> > @@ -0,0 +1,122 @@
> > +/* SPDX-License-Identifier: BSD-3-Clause
> > + * Copyright(c) 2022 StarFive
> > + * Copyright(c) 2022 SiFive
> > + * Copyright(c) 2022 Semihalf
> > + */
> > +
> > +#ifndef _RTE_LPM_SCALAR_H_
> > +#define _RTE_LPM_SCALAR_H_
> > +
> > +#include <rte_branch_prediction.h>
> > +#include <rte_byteorder.h>
> > +#include <rte_common.h>
> > +#include <rte_vect.h>
> > +
> > +#ifdef __cplusplus
> > +extern "C" {
> > +#endif
> > +
> > +static inline void
> > +rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4],
> > +             uint32_t defv)
> > +{
> > +     rte_xmm_t i24;
> > +     rte_xmm_t i8;
> > +     uint32_t tbl[4];
> > +     uint64_t pt, pt2;
> > +     const uint32_t *ptbl;
> > +
> > +     const rte_xmm_t mask8 = {
> > +             .u32 = {UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX}};
> > +
> > +     /*
> > +      * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 2 LPM entries
> > +      * as one 64-bit value (0x0300000003000000).
> > +      */
> > +     const uint64_t mask_xv =
> > +             ((uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK |
> > +             (uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK << 32);
> > +
> > +     /*
> > +      * RTE_LPM_LOOKUP_SUCCESS for 2 LPM entries
> > +      * as one 64-bit value (0x0100000001000000).
> > +      */
> > +     const uint64_t mask_v =
> > +             ((uint64_t)RTE_LPM_LOOKUP_SUCCESS |
> > +             (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32);
> > +
> > +     /* get 4 indexes for tbl24[]. */
> > +     i24.x = ip;
> > +     i24.u32[0] >>= CHAR_BIT;
> > +     i24.u32[1] >>= CHAR_BIT;
> > +     i24.u32[2] >>= CHAR_BIT;
> > +     i24.u32[3] >>= CHAR_BIT;
> > +
> > +     /* extract values from tbl24[] */
> > +     ptbl = (const uint32_t *)&lpm->tbl24[i24.u32[0]];
> > +     tbl[0] = *ptbl;
> > +     ptbl = (const uint32_t *)&lpm->tbl24[i24.u32[1]];
> > +     tbl[1] = *ptbl;
> > +     ptbl = (const uint32_t *)&lpm->tbl24[i24.u32[2]];
> > +     tbl[2] = *ptbl;
> > +     ptbl = (const uint32_t *)&lpm->tbl24[i24.u32[3]];
> > +     tbl[3] = *ptbl;
> > +
> > +     /* get 4 indexes for tbl8[]. */
> > +     i8.x = ip;
> > +     i8.u64[0] &= mask8.u64[0];
> > +     i8.u64[1] &= mask8.u64[1];
> > +
> > +     pt = (uint64_t)tbl[0] |
> > +             (uint64_t)tbl[1] << 32;
> > +     pt2 = (uint64_t)tbl[2] |
> > +             (uint64_t)tbl[3] << 32;
> > +
> > +     /* search successfully finished for all 4 IP addresses. */
> > +     if (likely((pt & mask_xv) == mask_v) &&
> > +                     likely((pt2 & mask_xv) == mask_v)) {
> > +             *(uint64_t *)hop = pt & RTE_LPM_MASKX4_RES;
> > +             *(uint64_t *)(hop + 2) = pt2 & RTE_LPM_MASKX4_RES;
> > +             return;
> > +     }
> > +
> > +     if (unlikely((pt & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
> > +                     RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
> > +             i8.u32[0] = i8.u32[0] +
> > +                     (tbl[0] & 0x00FFFFFF) * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
> > +             ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[0]];
> > +             tbl[0] = *ptbl;
> > +     }
> > +     if (unlikely((pt >> 32 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
> > +                     RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
> > +             i8.u32[1] = i8.u32[1] +
> > +                     (tbl[1] & 0x00FFFFFF) * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
> > +             ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[1]];
> > +             tbl[1] = *ptbl;
> > +     }
> > +     if (unlikely((pt2 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
> > +                     RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
> > +             i8.u32[2] = i8.u32[2] +
> > +                     (tbl[2] & 0x00FFFFFF) * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
> > +             ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[2]];
> > +             tbl[2] = *ptbl;
> > +     }
> > +     if (unlikely((pt2 >> 32 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
> > +                     RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
> > +             i8.u32[3] = i8.u32[3] +
> > +                     (tbl[3] & 0x00FFFFFF) * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
> > +             ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[3]];
> > +             tbl[3] = *ptbl;
> > +     }
> > +
> > +     hop[0] = (tbl[0] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[0] & 0x00FFFFFF : defv;
> > +     hop[1] = (tbl[1] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[1] & 0x00FFFFFF : defv;
> > +     hop[2] = (tbl[2] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[2] & 0x00FFFFFF : defv;
> > +     hop[3] = (tbl[3] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[3] & 0x00FFFFFF : defv;
> > +}
> > +
> > +#ifdef __cplusplus
> > +}
> > +#endif
> > +
> > +#endif /* _RTE_LPM_SCALAR_H_ */
>
> --
> Regards,
> Vladimir

  reply	other threads:[~2022-05-24 16:28 UTC|newest]

Thread overview: 19+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-05-10 11:58 Stanislaw Kardach
2022-05-19 17:02 ` Medvedkin, Vladimir
2022-05-24 16:28   ` Stanisław Kardach [this message]
2022-05-27 11:16     ` Stanisław Kardach
2022-05-27 13:16       ` Medvedkin, Vladimir
2022-05-27 18:18 ` [PATCH v2 1/2] lpm: add const to lpm arg of rte_lpm_lookup Stanislaw Kardach
2022-05-27 18:18   ` [PATCH v2 2/2] lpm: add a scalar version of lookupx4 function Stanislaw Kardach
2022-05-27 20:15     ` Stephen Hemminger
2022-05-30  7:52       ` Bruce Richardson
2022-05-30  8:00         ` Morten Brørup
2022-05-30 10:42           ` Bruce Richardson
2022-05-30 11:20             ` Stanisław Kardach
2022-05-30 12:46               ` Bruce Richardson
2022-05-30 18:24   ` [PATCH v3 1/2] lpm: add const to lpm arg of rte_lpm_lookup Stanislaw Kardach
2022-05-30 18:24     ` [PATCH v3 2/2] lpm: add a scalar version of lookupx4 function Stanislaw Kardach
2022-06-01  9:41       ` Medvedkin, Vladimir
2022-06-01 10:32         ` Stanisław Kardach
2022-05-30 20:38     ` [PATCH v3 1/2] lpm: add const to lpm arg of rte_lpm_lookup Stephen Hemminger
2022-06-01  9:35     ` Medvedkin, Vladimir

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CALVGJWJjf7NX31-0rzWD6tWXC4sgeCT-iKwuU3XAocMBHCGOsQ@mail.gmail.com \
    --to=kda@semihalf.com \
    --cc=Frank.Zhao@starfivetech.com \
    --cc=bruce.richardson@intel.com \
    --cc=dev@dpdk.org \
    --cc=maz@semihalf.com \
    --cc=mw@semihalf.com \
    --cc=sam.grove@sifive.com \
    --cc=upstream@semihalf.com \
    --cc=vladimir.medvedkin@intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).