Re: [dpdk-dev] [PATCH v2 1/3] lpm: make rte_lpm_lookupx4 API definition architecture agnostic

[Jerin Jacob <jerin.jacob@caviumnetworks.com>]

On Mon, Dec 07, 2015 at 02:15:28PM +0800, Jianbo Liu wrote:
> On 4 December 2015 at 23:14, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote:
> > -Used architecture agnostic xmm_t to represent 128 bit SIMD variable
> >
> > -Introduced vect_* API abstraction in app/test to test rte_lpm_lookupx4
> > API in  architecture agnostic way
> >
> > -Moved rte_lpm_lookupx4 SSE implementation to architecture specific
> > rte_lpm_sse.h file to accommodate new rte_lpm_lookupx4 implementation
> > for a different architecture.
> >
> > Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
> > ---
> >  app/test/test_lpm.c          |  21 ++++---
> >  app/test/test_xmmt_ops.h     |  47 ++++++++++++++
> >  lib/librte_lpm/Makefile      |   2 +
> >  lib/librte_lpm/rte_lpm.h     |  93 +---------------------------
> >  lib/librte_lpm/rte_lpm_sse.h | 143 +++++++++++++++++++++++++++++++++++++++++++
> >  5 files changed, 206 insertions(+), 100 deletions(-)
> >  create mode 100644 app/test/test_xmmt_ops.h
> >  create mode 100644 lib/librte_lpm/rte_lpm_sse.h
> >
> > diff --git a/app/test/test_lpm.c b/app/test/test_lpm.c
> > index 8b4ded9..59674f1 100644
> > --- a/app/test/test_lpm.c
> > +++ b/app/test/test_lpm.c
> > @@ -49,6 +49,7 @@
> >
> >  #include "rte_lpm.h"
> >  #include "test_lpm_routes.h"
> > +#include "test_xmmt_ops.h"
> >
> >  #define TEST_LPM_ASSERT(cond) do {                                            \
> >         if (!(cond)) {                                                        \
> > @@ -308,7 +309,7 @@ test6(void)
> >  int32_t
> >  test7(void)
> >  {
> > -       __m128i ipx4;
> > +       xmm_t ipx4;
> >         uint16_t hop[4];
> >         struct rte_lpm *lpm = NULL;
> >         uint32_t ip = IPv4(0, 0, 0, 0);
> > @@ -324,7 +325,7 @@ test7(void)
> >         status = rte_lpm_lookup(lpm, ip, &next_hop_return);
> >         TEST_LPM_ASSERT((status == 0) && (next_hop_return == next_hop_add));
> >
> > -       ipx4 = _mm_set_epi32(ip, ip + 0x100, ip - 0x100, ip);
> > +       ipx4 = vect_set_epi32(ip, ip + 0x100, ip - 0x100, ip);
> >         rte_lpm_lookupx4(lpm, ipx4, hop, UINT16_MAX);
> >         TEST_LPM_ASSERT(hop[0] == next_hop_add);
> >         TEST_LPM_ASSERT(hop[1] == UINT16_MAX);
> > @@ -354,7 +355,7 @@ test7(void)
> >  int32_t
> >  test8(void)
> >  {
> > -       __m128i ipx4;
> > +       xmm_t ipx4;
> >         uint16_t hop[4];
> >         struct rte_lpm *lpm = NULL;
> >         uint32_t ip1 = IPv4(127, 255, 255, 255), ip2 = IPv4(128, 0, 0, 0);
> > @@ -380,7 +381,7 @@ test8(void)
> >                 TEST_LPM_ASSERT((status == 0) &&
> >                         (next_hop_return == next_hop_add));
> >
> > -               ipx4 = _mm_set_epi32(ip2, ip1, ip2, ip1);
> > +               ipx4 = vect_set_epi32(ip2, ip1, ip2, ip1);
> >                 rte_lpm_lookupx4(lpm, ipx4, hop, UINT16_MAX);
> >                 TEST_LPM_ASSERT(hop[0] == UINT16_MAX);
> >                 TEST_LPM_ASSERT(hop[1] == next_hop_add);
> > @@ -408,7 +409,7 @@ test8(void)
> >                 status = rte_lpm_lookup(lpm, ip1, &next_hop_return);
> >                 TEST_LPM_ASSERT(status == -ENOENT);
> >
> > -               ipx4 = _mm_set_epi32(ip1, ip1, ip2, ip2);
> > +               ipx4 = vect_set_epi32(ip1, ip1, ip2, ip2);
> >                 rte_lpm_lookupx4(lpm, ipx4, hop, UINT16_MAX);
> >                 if (depth != 1) {
> >                         TEST_LPM_ASSERT(hop[0] == next_hop_add);
> > @@ -850,7 +851,7 @@ test11(void)
> >  int32_t
> >  test12(void)
> >  {
> > -       __m128i ipx4;
> > +       xmm_t ipx4;
> >         uint16_t hop[4];
> >         struct rte_lpm *lpm = NULL;
> >         uint32_t ip, i;
> > @@ -872,7 +873,7 @@ test12(void)
> >                 TEST_LPM_ASSERT((status == 0) &&
> >                                 (next_hop_return == next_hop_add));
> >
> > -               ipx4 = _mm_set_epi32(ip, ip + 1, ip, ip - 1);
> > +               ipx4 = vect_set_epi32(ip, ip + 1, ip, ip - 1);
> >                 rte_lpm_lookupx4(lpm, ipx4, hop, UINT16_MAX);
> >                 TEST_LPM_ASSERT(hop[0] == UINT16_MAX);
> >                 TEST_LPM_ASSERT(hop[1] == next_hop_add);
> > @@ -1289,10 +1290,10 @@ perf_test(void)
> >                 begin = rte_rdtsc();
> >                 for (j = 0; j < BATCH_SIZE; j += RTE_DIM(next_hops)) {
> >                         unsigned k;
> > -                       __m128i ipx4;
> > +                       xmm_t ipx4;
> >
> > -                       ipx4 = _mm_loadu_si128((__m128i *)(ip_batch + j));
> > -                       ipx4 = *(__m128i *)(ip_batch + j);
> > +                       ipx4 = vect_loadu_sil128((xmm_t *)(ip_batch + j));
> > +                       ipx4 = *(xmm_t *)(ip_batch + j);
> >                         rte_lpm_lookupx4(lpm, ipx4, next_hops, UINT16_MAX);
> >                         for (k = 0; k < RTE_DIM(next_hops); k++)
> >                                 if (unlikely(next_hops[k] == UINT16_MAX))
> > diff --git a/app/test/test_xmmt_ops.h b/app/test/test_xmmt_ops.h
> > new file mode 100644
> > index 0000000..c055912
> > --- /dev/null
> > +++ b/app/test/test_xmmt_ops.h
> Why add this new file under app/test, which is only for test app?
> Should vect_loadu_sil128/vect_set_epi32 be in each ARCH's rte_vect.h?
>

V1 was like that, I thought of moving the file under app/test because

1) all the ARCH can't have the implementation for vector primitives if
architecture doesn't support it like ppc64 and tile so moving EAL may
not be a good idea
2) scope of vector abstraction only for using the API(i.e test app), NOT
for
implementing the library. So its boils down to load/store/set should
not be beyond that.
and I am afraid that if we opening up EAL abstraction that will change the scope
and which will have performance implication to use emulating the logic in
library
3) It's been discussed, There was no disagreement on this
http://dpdk.org/ml/archives/dev/2015-December/029404.html

Thanks,
Jerin

[snip]