Hi Konstantin, On Tue, 27 Oct 2015 15:31:44 +0000 "Ananyev, Konstantin" <konstantin.ananyev at intel.com> wrote:
> Hi Jan, > > > -----Original Message----- > > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Jan Viktorin > > Sent: Monday, October 26, 2015 4:38 PM > > To: Thomas Monjalon; Hunt, David; dev at dpdk.org > > Cc: Vlastimil Kosar > > Subject: [dpdk-dev] [PATCH v2 15/16] lpm/arm: implement rte_lpm_lookupx4 > > using rte_lpm_lookup_bulk on for-x86 > > > > From: Vlastimil Kosar <kosar at rehivetech.com> > > > > LPM function rte_lpm_lookupx4() uses i686/x86_64 SIMD intrinsics. Therefore, > > the function is reimplemented using non-vector operations for non-x86 > > architectures. In the future, each architecture should have vectorized code. > > This patch includes rudimentary emulation of intrinsic functions > > _mm_set_epi32(), > > _mm_loadu_si128() and _mm_load_si128() for easy portability of existing > > applications. > > > > LPM builds now when on ARM. > > > > FIXME: to be reworked > > > > Signed-off-by: Vlastimil Kosar <kosar at rehivetech.com> > > Signed-off-by: Jan Viktorin <viktorin at rehivetech.com> > > --- > > config/defconfig_arm-armv7-a-linuxapp-gcc | 1 - > > lib/librte_lpm/rte_lpm.h | 71 > > +++++++++++++++++++++++++++++++ > > 2 files changed, 71 insertions(+), 1 deletion(-) > > > > diff --git a/config/defconfig_arm-armv7-a-linuxapp-gcc > > b/config/defconfig_arm-armv7-a-linuxapp-gcc > > index 5b582a8..33afb33 100644 > > --- a/config/defconfig_arm-armv7-a-linuxapp-gcc > > +++ b/config/defconfig_arm-armv7-a-linuxapp-gcc > > @@ -58,7 +58,6 @@ CONFIG_XMM_SIZE=16 > > > > # fails to compile on ARM > > CONFIG_RTE_LIBRTE_ACL=n > > -CONFIG_RTE_LIBRTE_LPM=n > > > > # cannot use those on ARM > > CONFIG_RTE_KNI_KMOD=n > > diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > > index c299ce2..4619992 100644 > > --- a/lib/librte_lpm/rte_lpm.h > > +++ b/lib/librte_lpm/rte_lpm.h > > @@ -47,7 +47,9 @@ > > #include <rte_byteorder.h> > > #include <rte_memory.h> > > #include <rte_common.h> > > +#if defined(RTE_ARCH_X86_64) || defined(RTE_ARCH_I686) > > #include <rte_vect.h> > > +#endif > > > > #ifdef __cplusplus > > extern "C" { > > @@ -358,6 +360,7 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, > > const uint32_t * ips, > > return 0; > > } > > > > +#if defined(RTE_ARCH_X86_64) || defined(RTE_ARCH_I686) > > /* Mask four results. */ > > #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > > > > @@ -472,6 +475,74 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i > > ip, uint16_t hop[4], > > hop[2] = (tbl[2] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[2] : defv; > > hop[3] = (tbl[3] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[3] : defv; > > } > > +#else > > Probably better to create an > lib/librte_eal/common/include/arch/arm/rte_vect.h, > and move all these x86 vector support emulation there? > Konstantin Sure. This patch is terribly wrong and it's not to be merged. It is a question whether to make it this way (with the refactoring as you suggested) or to make some general abstraction of the SSE calls in DPDK. Jan > > > +// TODO: this code should be reworked. > > + > > +typedef struct { > > + union uint128 { > > + uint8_t uint8[16]; > > + uint32_t uint32[4]; > > + } val; > > +} __m128i; > > + > > +static inline __m128i > > +_mm_set_epi32(uint32_t v0, uint32_t v1, uint32_t v2, uint32_t v3) > > +{ > > + __m128i res; > > + res.val.uint32[0] = v0; > > + res.val.uint32[1] = v1; > > + res.val.uint32[2] = v2; > > + res.val.uint32[3] = v3; > > + return res; > > +} > > + > > +static inline __m128i > > +_mm_loadu_si128(__m128i * v) > > +{ > > + __m128i res; > > + res = *v; > > + return res; > > +} > > + > > +static inline __m128i > > +_mm_load_si128(__m128i * v) > > +{ > > + __m128i res; > > + res = *v; > > + return res; > > +} > > + > > +/** > > + * Lookup four IP addresses in an LPM table. > > + * > > + * @param lpm > > + * LPM object handle > > + * @param ip > > + * Four IPs to be looked up in the LPM table > > + * @param hop > > + * Next hop of the most specific rule found for IP (valid on lookup hit > > only). > > + * This is an 4 elements array of two byte values. > > + * If the lookup was succesfull for the given IP, then least significant > > byte > > + * of the corresponding element is the actual next hop and the most > > + * significant byte is zero. > > + * If the lookup for the given IP failed, then corresponding element > > would > > + * contain default value, see description of then next parameter. > > + * @param defv > > + * Default value to populate into corresponding element of hop[] array, > > + * if lookup would fail. > > + */ > > +static inline void > > +rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > > + uint16_t defv) > > +{ > > + rte_lpm_lookup_bulk(lpm, ip.val.uint32, hop, 4); > > + > > + hop[0] = (hop[0] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)hop[0] : defv; > > + hop[1] = (hop[1] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)hop[1] : defv; > > + hop[2] = (hop[2] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)hop[2] : defv; > > + hop[3] = (hop[3] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)hop[3] : defv; > > +} > > +#endif > > > > #ifdef __cplusplus > > } > > -- > > 2.6.1 > -- Jan Viktorin E-mail: Viktorin at RehiveTech.com System Architect Web: www.RehiveTech.com RehiveTech Brno, Czech Republic
