Message ID | 1448995276-9599-4-git-send-email-jianbo.liu@linaro.org (mailing list archive) |
---|---|
State | Superseded, archived |
Headers |
Return-Path: <dev-bounces@dpdk.org> X-Original-To: patchwork@dpdk.org Delivered-To: patchwork@dpdk.org Received: from [92.243.14.124] (localhost [IPv6:::1]) by dpdk.org (Postfix) with ESMTP id D7B128DAA; Tue, 1 Dec 2015 11:41:47 +0100 (CET) Received: from mail-wm0-f41.google.com (mail-wm0-f41.google.com [74.125.82.41]) by dpdk.org (Postfix) with ESMTP id 887EA8E6C for <dev@dpdk.org>; Tue, 1 Dec 2015 11:41:46 +0100 (CET) Received: by wmec201 with SMTP id c201so7239325wme.1 for <dev@dpdk.org>; Tue, 01 Dec 2015 02:41:46 -0800 (PST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro-org.20150623.gappssmtp.com; s=20150623; h=from:to:cc:subject:date:message-id:in-reply-to:references; bh=a3yY3uibXXKY2gSckfGsu3RR781otj2yvmMCMJUkKAw=; b=AlynnpEl3utWq9yv9B88cyvV6ITUkHoX5xWXifzaS9gdSphQBq1xMr8qXJypqnVUmK pwH1L+iMZw1QswhwzQz4k3Wo5H4cNbVtn4sqacfws8gPhHmaXMn/zRAfBIVkP8kXEx1m ftuJ2VZ7YQTp5/IGFfXHi6TflmH19IK66JUV/iWdiUMnPoEbAfAGnDxQu7EzpEyI9x+P hUY8sHK5wuYPS33b6mT+2YmteW2dGO5D2FlmyRv9VV1rolF/36WQrOe0Of8PN8qvSwA9 yPmv0hXtoL5JrRu0/1DZB70hMNW4Hu+l4XhZRPf/vySPTnXxIjdl4cDc/44dkNiOEl0e shuw== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20130820; h=x-gm-message-state:from:to:cc:subject:date:message-id:in-reply-to :references; bh=a3yY3uibXXKY2gSckfGsu3RR781otj2yvmMCMJUkKAw=; b=R74oD6c/6QBY9yVG2nKIfPVuVPhL3Sl60OSsUFnYvVfNNehDMJr12rSRXRVm7meULJ zXtaZAucl61poV+juz9UVhXOUBpbGWXcsBnoPgqXOYwBjjxJuoEiMHl2zZiAAjUaoEZq 7k2B3Mf730hAe2+UbdRIQNe5eZyQrwJQGSRK7Oqq2I+h+PUVXeDnf34qtK6wbwGTHnUS tUgLAFvWw0fBAJy8eG/boPbO1kkYiDEVQfuI6LMKGNogaSBlMlmpissxkPKrbHG+yUWK UbRAwelHk61waRCejgPKXf/nB/7PT0P0bha4H9YtpQnXD9Nc49Sb8hq1b8rxxTIylY0R XjLQ== X-Gm-Message-State: ALoCoQmNrp95fjNUN3ETZs8AQzYYKcQrtklbD2zR/KAdCvZNeNVYtssAanphFgP6gcWzVAS3EffY X-Received: by 10.28.45.72 with SMTP id t69mr36356036wmt.32.1448966506385; Tue, 01 Dec 2015 02:41:46 -0800 (PST) Received: from localhost ([112.65.63.41]) by smtp.gmail.com with ESMTPSA id h4sm51056034wjx.41.2015.12.01.02.41.44 (version=TLS1_2 cipher=ECDHE-RSA-AES128-GCM-SHA256 bits=128/128); Tue, 01 Dec 2015 02:41:45 -0800 (PST) From: Jianbo Liu <jianbo.liu@linaro.org> To: dev@dpdk.org Date: Tue, 1 Dec 2015 13:41:15 -0500 Message-Id: <1448995276-9599-4-git-send-email-jianbo.liu@linaro.org> X-Mailer: git-send-email 1.8.3.1 In-Reply-To: <1448995276-9599-1-git-send-email-jianbo.liu@linaro.org> References: <1448995276-9599-1-git-send-email-jianbo.liu@linaro.org> Subject: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.15 Precedence: list List-Id: patches and discussions about DPDK <dev.dpdk.org> List-Unsubscribe: <http://dpdk.org/ml/options/dev>, <mailto:dev-request@dpdk.org?subject=unsubscribe> List-Archive: <http://dpdk.org/ml/archives/dev/> List-Post: <mailto:dev@dpdk.org> List-Help: <mailto:dev-request@dpdk.org?subject=help> List-Subscribe: <http://dpdk.org/ml/listinfo/dev>, <mailto:dev-request@dpdk.org?subject=subscribe> Errors-To: dev-bounces@dpdk.org Sender: "dev" <dev-bounces@dpdk.org> |
Commit Message
Jianbo Liu
Dec. 1, 2015, 6:41 p.m. UTC
Adds ARM NEON support for lpm.
And enables table/pipeline libraries which depend on lpm.
Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org>
---
config/defconfig_arm-armv7a-linuxapp-gcc | 3 -
config/defconfig_arm64-armv8a-linuxapp-gcc | 3 -
lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++
lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++-------
4 files changed, 77 insertions(+), 25 deletions(-)
Comments
On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > Adds ARM NEON support for lpm. > And enables table/pipeline libraries which depend on lpm. I already sent the patch on the same yesterday. We can converge the patches after the discussion. Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > > Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> > --- > config/defconfig_arm-armv7a-linuxapp-gcc | 3 - > config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - > lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ > lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- > 4 files changed, 77 insertions(+), 25 deletions(-) > > diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc > index cbebd64..efffa1f 100644 > --- a/config/defconfig_arm-armv7a-linuxapp-gcc > +++ b/config/defconfig_arm-armv7a-linuxapp-gcc > @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n > CONFIG_RTE_EAL_IGB_UIO=n > > # fails to compile on ARM > -CONFIG_RTE_LIBRTE_LPM=n > -CONFIG_RTE_LIBRTE_TABLE=n > -CONFIG_RTE_LIBRTE_PIPELINE=n > CONFIG_RTE_SCHED_VECTOR=n > > # cannot use those on ARM > diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc > index 504f3ed..57f7941 100644 > --- a/config/defconfig_arm64-armv8a-linuxapp-gcc > +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc > @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n > CONFIG_RTE_LIBRTE_FM10K_PMD=n > CONFIG_RTE_LIBRTE_I40E_PMD=n > > -CONFIG_RTE_LIBRTE_LPM=n > -CONFIG_RTE_LIBRTE_TABLE=n > -CONFIG_RTE_LIBRTE_PIPELINE=n > CONFIG_RTE_SCHED_VECTOR=n > diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h > index a33c054..7437711 100644 > --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h > +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h > @@ -41,6 +41,8 @@ extern "C" { > > typedef int32x4_t xmm_t; > > +typedef int32x4_t __m128i; > + > #define XMM_SIZE (sizeof(xmm_t)) > #define XMM_MASK (XMM_SIZE - 1) > > @@ -53,6 +55,32 @@ typedef union rte_xmm { > double pd[XMM_SIZE / sizeof(double)]; > } __attribute__((aligned(16))) rte_xmm_t; > > +static __inline __m128i > +_mm_set_epi32(int i3, int i2, int i1, int i0) > +{ > + int32_t r[4] = {i0, i1, i2, i3}; > + > + return vld1q_s32(r); > +} > + > +static __inline __m128i > +_mm_loadu_si128(__m128i *p) > +{ > + return vld1q_s32((int32_t *)p); > +} > + > +static __inline __m128i > +_mm_set1_epi32(int i) > +{ > + return vdupq_n_s32(i); > +} > + > +static __inline __m128i > +_mm_and_si128(__m128i a, __m128i b) > +{ > + return vandq_s32(a, b); > +} > + IMO, it makes sense to not emulate the SSE intrinsics with NEON Let's create the rte_vect_* as required. look at the existing patch. > #ifdef RTE_ARCH_ARM > /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ > static __inline uint8x16_t > diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > index c299ce2..c76c07d 100644 > --- a/lib/librte_lpm/rte_lpm.h > +++ b/lib/librte_lpm/rte_lpm.h > @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > /* Mask four results. */ > #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > > +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) Separate out arm implementation to the different header file. Too many ifdef looks odd in the header file and difficult to manage. > +static inline void > +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) > +{ > + uint32x4_t i24; > + uint32_t idx[4]; > + > + /* get 4 indexes for tbl24[]. */ > + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); > + vst1q_u32(idx, i24); > + > + /* extract values from tbl24[] */ > + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; > + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; > + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; > + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; > +} Nice. There is an improvement in this portion code wrt my patch. This is a candidate for convergence. > +#else > +static inline void > +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) > +{ > + __m128i i24; > + uint64_t idx; > + > + /* get 4 indexes for tbl24[]. */ > + i24 = _mm_srli_epi32(ip, CHAR_BIT); > + > + /* extract values from tbl24[] */ > + idx = _mm_cvtsi128_si64(i24); > + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > + > + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > + > + idx = _mm_cvtsi128_si64(i24); > + > + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > +} > +#endif > + > /** > * Lookup four IP addresses in an LPM table. > * > @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > * if lookup would fail. > */ > static inline void > +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], > + uint16_t defv) This would call for change in the change the ABI, IMO, __m128i can be used to represent 128bit vector to avoid ABI chang > +#else separate out arm implementation to the different header file. Too many ifdef looks odd in the header file. Could you rebase your patch based on existing patch and send the improvement portion as separate patch or I can send update patch with your improvements and with your signoff. > rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > uint16_t defv) > +#endif > { > - __m128i i24; > rte_xmm_t i8; > uint16_t tbl[4]; > - uint64_t idx, pt; > - > - const __m128i mask8 = > - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); > + uint64_t pt; > > + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); > /* > * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries > * as one 64-bit value (0x0300030003000300). > @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | > (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); > > - /* get 4 indexes for tbl24[]. */ > - i24 = _mm_srli_epi32(ip, CHAR_BIT); > - > - /* extract values from tbl24[] */ > - idx = _mm_cvtsi128_si64(i24); > - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > - > - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > - > - idx = _mm_cvtsi128_si64(i24); > - > - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > + rte_lpm_tbl24_val4(lpm, ip, tbl); > > /* get 4 indexes for tbl8[]. */ > i8.x = _mm_and_si128(ip, mask8); > -- > 1.8.3.1 >
On Tue, 1 Dec 2015 22:11:42 +0530 Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > > Adds ARM NEON support for lpm. > > And enables table/pipeline libraries which depend on lpm. > > I already sent the patch on the same yesterday. > We can converge the patches after the discussion. > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml I've missed that too. Did you CC me? Jan
On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: >> Adds ARM NEON support for lpm. >> And enables table/pipeline libraries which depend on lpm. > > I already sent the patch on the same yesterday. > We can converge the patches after the discussion. > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > Yes, I have read your patch. But there are many differences, so I sent mine for your reviewing :) > >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> >> --- >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- >> 4 files changed, 77 insertions(+), 25 deletions(-) >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc >> index cbebd64..efffa1f 100644 >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n >> CONFIG_RTE_EAL_IGB_UIO=n >> >> # fails to compile on ARM >> -CONFIG_RTE_LIBRTE_LPM=n >> -CONFIG_RTE_LIBRTE_TABLE=n >> -CONFIG_RTE_LIBRTE_PIPELINE=n >> CONFIG_RTE_SCHED_VECTOR=n >> >> # cannot use those on ARM >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc >> index 504f3ed..57f7941 100644 >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n >> CONFIG_RTE_LIBRTE_FM10K_PMD=n >> CONFIG_RTE_LIBRTE_I40E_PMD=n >> >> -CONFIG_RTE_LIBRTE_LPM=n >> -CONFIG_RTE_LIBRTE_TABLE=n >> -CONFIG_RTE_LIBRTE_PIPELINE=n >> CONFIG_RTE_SCHED_VECTOR=n >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h >> index a33c054..7437711 100644 >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h >> @@ -41,6 +41,8 @@ extern "C" { >> >> typedef int32x4_t xmm_t; >> >> +typedef int32x4_t __m128i; >> + >> #define XMM_SIZE (sizeof(xmm_t)) >> #define XMM_MASK (XMM_SIZE - 1) >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { >> double pd[XMM_SIZE / sizeof(double)]; >> } __attribute__((aligned(16))) rte_xmm_t; >> >> +static __inline __m128i >> +_mm_set_epi32(int i3, int i2, int i1, int i0) >> +{ >> + int32_t r[4] = {i0, i1, i2, i3}; >> + >> + return vld1q_s32(r); >> +} >> + >> +static __inline __m128i >> +_mm_loadu_si128(__m128i *p) >> +{ >> + return vld1q_s32((int32_t *)p); >> +} >> + >> +static __inline __m128i >> +_mm_set1_epi32(int i) >> +{ >> + return vdupq_n_s32(i); >> +} >> + >> +static __inline __m128i >> +_mm_and_si128(__m128i a, __m128i b) >> +{ >> + return vandq_s32(a, b); >> +} >> + > > IMO, it makes sense to not emulate the SSE intrinsics with NEON > Let's create the rte_vect_* as required. look at the existing patch. > I thought of creating a layer of SIMD over all the platforms before. But can't you see it make things complicated, considering there are only few simple intrinsic to implement? If do so, we also need to explain to others how to use these interfaces. Besides, this patch did the smallest changes to the original code, and more likely to be accepted by others. > >> #ifdef RTE_ARCH_ARM >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ >> static __inline uint8x16_t >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h >> index c299ce2..c76c07d 100644 >> --- a/lib/librte_lpm/rte_lpm.h >> +++ b/lib/librte_lpm/rte_lpm.h >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, >> /* Mask four results. */ >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > > Separate out arm implementation to the different header file. > Too many ifdef looks odd in the header file and difficult to manage. > But there are many ifdefs already. And It seems unreasonable to add a new file only for one small function. > >> +static inline void >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) >> +{ >> + uint32x4_t i24; >> + uint32_t idx[4]; >> + >> + /* get 4 indexes for tbl24[]. */ >> + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); >> + vst1q_u32(idx, i24); >> + >> + /* extract values from tbl24[] */ >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; >> +} > > Nice. There is an improvement in this portion code wrt my patch. This is > a candidate for convergence. > > >> +#else >> +static inline void >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) >> +{ >> + __m128i i24; >> + uint64_t idx; >> + >> + /* get 4 indexes for tbl24[]. */ >> + i24 = _mm_srli_epi32(ip, CHAR_BIT); >> + >> + /* extract values from tbl24[] */ >> + idx = _mm_cvtsi128_si64(i24); >> + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); >> + >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> + >> + idx = _mm_cvtsi128_si64(i24); >> + >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> +} >> +#endif >> + >> /** >> * Lookup four IP addresses in an LPM table. >> * >> @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, >> * if lookup would fail. >> */ >> static inline void >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) >> +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], >> + uint16_t defv) > > This would call for change in the change the ABI, > IMO, __m128i can be used to represent 128bit vector to avoid ABI chang > This redefine rte_lpm_lookupx4 is unncessary, I will remove it, so no ABI change. And there only one ifdef for ARM platforms left. > >> +#else > separate out arm implementation to the different header file. Too many > ifdef looks odd in the header file. > > Could you rebase your patch based on existing patch and send the > improvement portion as separate patch or I can send update patch with > your improvements and with your signoff. > > >> rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], >> uint16_t defv) >> +#endif >> { >> - __m128i i24; >> rte_xmm_t i8; >> uint16_t tbl[4]; >> - uint64_t idx, pt; >> - >> - const __m128i mask8 = >> - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); >> + uint64_t pt; >> >> + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); >> /* >> * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries >> * as one 64-bit value (0x0300030003000300). >> @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); >> >> - /* get 4 indexes for tbl24[]. */ >> - i24 = _mm_srli_epi32(ip, CHAR_BIT); >> - >> - /* extract values from tbl24[] */ >> - idx = _mm_cvtsi128_si64(i24); >> - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); >> - >> - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> - >> - idx = _mm_cvtsi128_si64(i24); >> - >> - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> + rte_lpm_tbl24_val4(lpm, ip, tbl); >> >> /* get 4 indexes for tbl8[]. */ >> i8.x = _mm_and_si128(ip, mask8); >> -- >> 1.8.3.1 >>
On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: > On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > >> Adds ARM NEON support for lpm. > >> And enables table/pipeline libraries which depend on lpm. > > > > I already sent the patch on the same yesterday. > > We can converge the patches after the discussion. > > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > > > Yes, I have read your patch. But there are many differences, so I sent > mine for your reviewing :) > > > > >> > >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> > >> --- > >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - > >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - > >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ > >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- > >> 4 files changed, 77 insertions(+), 25 deletions(-) > >> > >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc > >> index cbebd64..efffa1f 100644 > >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc > >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc > >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n > >> CONFIG_RTE_EAL_IGB_UIO=n > >> > >> # fails to compile on ARM > >> -CONFIG_RTE_LIBRTE_LPM=n > >> -CONFIG_RTE_LIBRTE_TABLE=n > >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> CONFIG_RTE_SCHED_VECTOR=n > >> > >> # cannot use those on ARM > >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> index 504f3ed..57f7941 100644 > >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc > >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n > >> CONFIG_RTE_LIBRTE_FM10K_PMD=n > >> CONFIG_RTE_LIBRTE_I40E_PMD=n > >> > >> -CONFIG_RTE_LIBRTE_LPM=n > >> -CONFIG_RTE_LIBRTE_TABLE=n > >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> CONFIG_RTE_SCHED_VECTOR=n > >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> index a33c054..7437711 100644 > >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> @@ -41,6 +41,8 @@ extern "C" { > >> > >> typedef int32x4_t xmm_t; > >> > >> +typedef int32x4_t __m128i; > >> + > >> #define XMM_SIZE (sizeof(xmm_t)) > >> #define XMM_MASK (XMM_SIZE - 1) > >> > >> @@ -53,6 +55,32 @@ typedef union rte_xmm { > >> double pd[XMM_SIZE / sizeof(double)]; > >> } __attribute__((aligned(16))) rte_xmm_t; > >> > >> +static __inline __m128i > >> +_mm_set_epi32(int i3, int i2, int i1, int i0) > >> +{ > >> + int32_t r[4] = {i0, i1, i2, i3}; > >> + > >> + return vld1q_s32(r); > >> +} > >> + > >> +static __inline __m128i > >> +_mm_loadu_si128(__m128i *p) > >> +{ > >> + return vld1q_s32((int32_t *)p); > >> +} > >> + > >> +static __inline __m128i > >> +_mm_set1_epi32(int i) > >> +{ > >> + return vdupq_n_s32(i); > >> +} > >> + > >> +static __inline __m128i > >> +_mm_and_si128(__m128i a, __m128i b) > >> +{ > >> + return vandq_s32(a, b); > >> +} > >> + IMO, it's not always good to emulate GCC defined intrinsics of other architecture. What if a legacy DPDK application has such mappings then BOOM, multiple definition, which one is correct? which one to comment it out? Integration pain starts for DPDK library consumer:-( > > > > IMO, it makes sense to not emulate the SSE intrinsics with NEON > > Let's create the rte_vect_* as required. look at the existing patch. > > > I thought of creating a layer of SIMD over all the platforms before. > But can't you see it make things complicated, considering there are > only few simple intrinsic to implement? Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON implementation if I were to take this approach and emulation comes with the cost. So my take is, lets the each architecture implementation for specific SIMD version of DPDK API in the library should have the freedom to implement the API in NATIVE. And let's create only rte_vect_* abstraction only for using that API/library. Which boils down to have very minimal rte_vect_* abstraction to load, store, set not beyond that. This makes clear "contract" between DPDK library and the applications. and make easy for remaning new architecture porting effort in DPDK. Imagine how your proposed function will look like if new architecture wants to implement "optimized" version of rte_lpm_lookupx4 > If do so, we also need to explain to others how to use these interfaces. > Besides, this patch did the smallest changes to the original code, and > more likely to be accepted by others. other patch makes no changes to IA version of rte_lpm_lookupx4.I thought that make reviewer easy to review the changes in architecture perspective. > > > > >> #ifdef RTE_ARCH_ARM > >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ > >> static __inline uint8x16_t > >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > >> index c299ce2..c76c07d 100644 > >> --- a/lib/librte_lpm/rte_lpm.h > >> +++ b/lib/librte_lpm/rte_lpm.h > >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> /* Mask four results. */ > >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > >> > >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > > > > Separate out arm implementation to the different header file. > > Too many ifdef looks odd in the header file and difficult to manage. > > > But there are many ifdefs already. > And It seems unreasonable to add a new file only for one small function. > small or big, its matter of each architecture to have the freedom for the optimized version for the implementation. What if other architecture demands to write this function in assembly or restructure it for performance improvement? > > > >> +static inline void > >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) > >> +{ > >> + uint32x4_t i24; > >> + uint32_t idx[4]; > >> + > >> + /* get 4 indexes for tbl24[]. */ > >> + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); > >> + vst1q_u32(idx, i24); > >> + > >> + /* extract values from tbl24[] */ > >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; > >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; > >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; > >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; > >> +} > > > > Nice. There is an improvement in this portion code wrt my patch. This is > > a candidate for convergence. > > > > > >> +#else > >> +static inline void > >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) > >> +{ > >> + __m128i i24; > >> + uint64_t idx; > >> + > >> + /* get 4 indexes for tbl24[]. */ > >> + i24 = _mm_srli_epi32(ip, CHAR_BIT); > >> + > >> + /* extract values from tbl24[] */ > >> + idx = _mm_cvtsi128_si64(i24); > >> + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > >> + > >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> + > >> + idx = _mm_cvtsi128_si64(i24); > >> + > >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> +} > >> +#endif > >> + > >> /** > >> * Lookup four IP addresses in an LPM table. > >> * > >> @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> * if lookup would fail. > >> */ > >> static inline void > >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > >> +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], > >> + uint16_t defv) > > > > This would call for change in the change the ABI, > > IMO, __m128i can be used to represent 128bit vector to avoid ABI chang > > > This redefine rte_lpm_lookupx4 is unncessary, I will remove it, so no > ABI change. > And there only one ifdef for ARM platforms left. > > > > >> +#else > > separate out arm implementation to the different header file. Too many > > ifdef looks odd in the header file. > > > > Could you rebase your patch based on existing patch and send the > > improvement portion as separate patch or I can send update patch with > > your improvements and with your signoff. > > > > > >> rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > >> uint16_t defv) > >> +#endif > >> { > >> - __m128i i24; > >> rte_xmm_t i8; > >> uint16_t tbl[4]; > >> - uint64_t idx, pt; > >> - > >> - const __m128i mask8 = > >> - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); > >> + uint64_t pt; > >> > >> + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); > >> /* > >> * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries > >> * as one 64-bit value (0x0300030003000300). > >> @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | > >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); > >> > >> - /* get 4 indexes for tbl24[]. */ > >> - i24 = _mm_srli_epi32(ip, CHAR_BIT); > >> - > >> - /* extract values from tbl24[] */ > >> - idx = _mm_cvtsi128_si64(i24); > >> - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > >> - > >> - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> - > >> - idx = _mm_cvtsi128_si64(i24); > >> - > >> - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> + rte_lpm_tbl24_val4(lpm, ip, tbl); > >> > >> /* get 4 indexes for tbl8[]. */ > >> i8.x = _mm_and_si128(ip, mask8); > >> -- > >> 1.8.3.1 > >>
On 2 December 2015 at 16:03, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: >> On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: >> > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: >> >> Adds ARM NEON support for lpm. >> >> And enables table/pipeline libraries which depend on lpm. >> > >> > I already sent the patch on the same yesterday. >> > We can converge the patches after the discussion. >> > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml >> > >> Yes, I have read your patch. But there are many differences, so I sent >> mine for your reviewing :) >> >> > >> >> >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> >> >> --- >> >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - >> >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - >> >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ >> >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- >> >> 4 files changed, 77 insertions(+), 25 deletions(-) >> >> >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc >> >> index cbebd64..efffa1f 100644 >> >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc >> >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc >> >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n >> >> CONFIG_RTE_EAL_IGB_UIO=n >> >> >> >> # fails to compile on ARM >> >> -CONFIG_RTE_LIBRTE_LPM=n >> >> -CONFIG_RTE_LIBRTE_TABLE=n >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n >> >> CONFIG_RTE_SCHED_VECTOR=n >> >> >> >> # cannot use those on ARM >> >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc >> >> index 504f3ed..57f7941 100644 >> >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc >> >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc >> >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n >> >> CONFIG_RTE_LIBRTE_FM10K_PMD=n >> >> CONFIG_RTE_LIBRTE_I40E_PMD=n >> >> >> >> -CONFIG_RTE_LIBRTE_LPM=n >> >> -CONFIG_RTE_LIBRTE_TABLE=n >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n >> >> CONFIG_RTE_SCHED_VECTOR=n >> >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h >> >> index a33c054..7437711 100644 >> >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h >> >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h >> >> @@ -41,6 +41,8 @@ extern "C" { >> >> >> >> typedef int32x4_t xmm_t; >> >> >> >> +typedef int32x4_t __m128i; >> >> + >> >> #define XMM_SIZE (sizeof(xmm_t)) >> >> #define XMM_MASK (XMM_SIZE - 1) >> >> >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { >> >> double pd[XMM_SIZE / sizeof(double)]; >> >> } __attribute__((aligned(16))) rte_xmm_t; >> >> >> >> +static __inline __m128i >> >> +_mm_set_epi32(int i3, int i2, int i1, int i0) >> >> +{ >> >> + int32_t r[4] = {i0, i1, i2, i3}; >> >> + >> >> + return vld1q_s32(r); >> >> +} >> >> + >> >> +static __inline __m128i >> >> +_mm_loadu_si128(__m128i *p) >> >> +{ >> >> + return vld1q_s32((int32_t *)p); >> >> +} >> >> + >> >> +static __inline __m128i >> >> +_mm_set1_epi32(int i) >> >> +{ >> >> + return vdupq_n_s32(i); >> >> +} >> >> + >> >> +static __inline __m128i >> >> +_mm_and_si128(__m128i a, __m128i b) >> >> +{ >> >> + return vandq_s32(a, b); >> >> +} >> >> + > > IMO, it's not always good to emulate GCC defined intrinsics of > other architecture. What if a legacy DPDK application has such mappings > then BOOM, multiple definition, which one is correct? which one > to comment it out? Integration pain starts for DPDK library consumer:-( > They can include rte_vect.h in build/include directly, which is linked correctly to the one for that ARCH, so there is no need to worry about. >> > >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON >> > Let's create the rte_vect_* as required. look at the existing patch. >> > >> I thought of creating a layer of SIMD over all the platforms before. >> But can't you see it make things complicated, considering there are >> only few simple intrinsic to implement? > > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON > implementation if I were to take this approach and emulation comes with > the cost. > No, I will not re-implement all the intrinsic like that . I only do with the simple intrinsic, such as load/store, as you said below. > So my take is, > lets the each architecture implementation for specific SIMD version of DPDK > API in the library should have the freedom to implement the API in > NATIVE. > > And let's create only rte_vect_* abstraction only for using > that API/library. Which boils down to have very minimal rte_vect_* > abstraction to load, store, set not beyond that. > > This makes clear "contract" between DPDK library and the applications. > and make easy for remaning new architecture porting effort in DPDK. > Agree. But I reuse existing intrinsic names, and you recreate new ones. And I try to do as few changes as possible, and try to avoid any mistaken which may cause code un-compiled. I think it's design level question, we need to hear what others talk about it. > Imagine how your proposed function will look like if new architecture > wants to implement "optimized" version of rte_lpm_lookupx4 > There is no optimization for this (simple) rte_lpm_lookupx4, otherwise you have done that in your patch. If there is for other new platform, defintely they should do like yours, as you did for NEON ACL. > >> If do so, we also need to explain to others how to use these interfaces. >> Besides, this patch did the smallest changes to the original code, and >> more likely to be accepted by others. > > other patch makes no changes to IA version of rte_lpm_lookupx4.I thought > that make reviewer easy to review the changes in architecture > perspective. > As I know, they don't enable LPM for PPC, and ARM is the first one to touch this issue. >> >> > >> >> #ifdef RTE_ARCH_ARM >> >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ >> >> static __inline uint8x16_t >> >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h >> >> index c299ce2..c76c07d 100644 >> >> --- a/lib/librte_lpm/rte_lpm.h >> >> +++ b/lib/librte_lpm/rte_lpm.h >> >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, >> >> /* Mask four results. */ >> >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) >> >> >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) >> > >> > Separate out arm implementation to the different header file. >> > Too many ifdef looks odd in the header file and difficult to manage. >> > >> But there are many ifdefs already. >> And It seems unreasonable to add a new file only for one small function. >> > > small or big, its matter of each architecture to have > the freedom for the optimized version for the implementation. > > What if other architecture demands to write this function in assembly > or restructure it for performance improvement? > If there is such demands, should do like that. But I don't see any restructure in your patch, and you still follow the logic as x86, is it worth adding a new file? > >> > >> >> +static inline void >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) >> >> +{ >> >> + uint32x4_t i24; >> >> + uint32_t idx[4]; >> >> + >> >> + /* get 4 indexes for tbl24[]. */ >> >> + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); >> >> + vst1q_u32(idx, i24); >> >> + >> >> + /* extract values from tbl24[] */ >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; >> >> +} >> > >> > Nice. There is an improvement in this portion code wrt my patch. This is >> > a candidate for convergence. >> > >> > >> >> +#else >> >> +static inline void >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) >> >> +{ >> >> + __m128i i24; >> >> + uint64_t idx; >> >> + >> >> + /* get 4 indexes for tbl24[]. */ >> >> + i24 = _mm_srli_epi32(ip, CHAR_BIT); >> >> + >> >> + /* extract values from tbl24[] */ >> >> + idx = _mm_cvtsi128_si64(i24); >> >> + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); >> >> + >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> + >> >> + idx = _mm_cvtsi128_si64(i24); >> >> + >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> +} >> >> +#endif >> >> + >> >> /** >> >> * Lookup four IP addresses in an LPM table. >> >> * >> >> @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, >> >> * if lookup would fail. >> >> */ >> >> static inline void >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) >> >> +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], >> >> + uint16_t defv) >> > >> > This would call for change in the change the ABI, >> > IMO, __m128i can be used to represent 128bit vector to avoid ABI chang >> > >> This redefine rte_lpm_lookupx4 is unncessary, I will remove it, so no >> ABI change. >> And there only one ifdef for ARM platforms left. >> >> > >> >> +#else >> > separate out arm implementation to the different header file. Too many >> > ifdef looks odd in the header file. >> > >> > Could you rebase your patch based on existing patch and send the >> > improvement portion as separate patch or I can send update patch with >> > your improvements and with your signoff. >> > >> > >> >> rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], >> >> uint16_t defv) >> >> +#endif >> >> { >> >> - __m128i i24; >> >> rte_xmm_t i8; >> >> uint16_t tbl[4]; >> >> - uint64_t idx, pt; >> >> - >> >> - const __m128i mask8 = >> >> - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); >> >> + uint64_t pt; >> >> >> >> + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); >> >> /* >> >> * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries >> >> * as one 64-bit value (0x0300030003000300). >> >> @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); >> >> >> >> - /* get 4 indexes for tbl24[]. */ >> >> - i24 = _mm_srli_epi32(ip, CHAR_BIT); >> >> - >> >> - /* extract values from tbl24[] */ >> >> - idx = _mm_cvtsi128_si64(i24); >> >> - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); >> >> - >> >> - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> - >> >> - idx = _mm_cvtsi128_si64(i24); >> >> - >> >> - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> + rte_lpm_tbl24_val4(lpm, ip, tbl); >> >> >> >> /* get 4 indexes for tbl8[]. */ >> >> i8.x = _mm_and_si128(ip, mask8); >> >> -- >> >> 1.8.3.1 >> >>
Hi everyone, > -----Original Message----- > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jianbo Liu > Sent: Wednesday, December 02, 2015 9:50 AM > To: Jerin Jacob > Cc: dev@dpdk.org > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > On 2 December 2015 at 16:03, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: > >> On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > >> > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > >> >> Adds ARM NEON support for lpm. > >> >> And enables table/pipeline libraries which depend on lpm. > >> > > >> > I already sent the patch on the same yesterday. > >> > We can converge the patches after the discussion. > >> > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > >> > > >> Yes, I have read your patch. But there are many differences, so I sent > >> mine for your reviewing :) > >> > >> > > >> >> > >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> > >> >> --- > >> >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - > >> >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - > >> >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ > >> >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- > >> >> 4 files changed, 77 insertions(+), 25 deletions(-) > >> >> > >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> index cbebd64..efffa1f 100644 > >> >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n > >> >> CONFIG_RTE_EAL_IGB_UIO=n > >> >> > >> >> # fails to compile on ARM > >> >> -CONFIG_RTE_LIBRTE_LPM=n > >> >> -CONFIG_RTE_LIBRTE_TABLE=n > >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> >> CONFIG_RTE_SCHED_VECTOR=n > >> >> > >> >> # cannot use those on ARM > >> >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> index 504f3ed..57f7941 100644 > >> >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n > >> >> CONFIG_RTE_LIBRTE_FM10K_PMD=n > >> >> CONFIG_RTE_LIBRTE_I40E_PMD=n > >> >> > >> >> -CONFIG_RTE_LIBRTE_LPM=n > >> >> -CONFIG_RTE_LIBRTE_TABLE=n > >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> >> CONFIG_RTE_SCHED_VECTOR=n > >> >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> index a33c054..7437711 100644 > >> >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> @@ -41,6 +41,8 @@ extern "C" { > >> >> > >> >> typedef int32x4_t xmm_t; > >> >> > >> >> +typedef int32x4_t __m128i; > >> >> + > >> >> #define XMM_SIZE (sizeof(xmm_t)) > >> >> #define XMM_MASK (XMM_SIZE - 1) > >> >> > >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { > >> >> double pd[XMM_SIZE / sizeof(double)]; > >> >> } __attribute__((aligned(16))) rte_xmm_t; > >> >> > >> >> +static __inline __m128i > >> >> +_mm_set_epi32(int i3, int i2, int i1, int i0) > >> >> +{ > >> >> + int32_t r[4] = {i0, i1, i2, i3}; > >> >> + > >> >> + return vld1q_s32(r); > >> >> +} > >> >> + > >> >> +static __inline __m128i > >> >> +_mm_loadu_si128(__m128i *p) > >> >> +{ > >> >> + return vld1q_s32((int32_t *)p); > >> >> +} > >> >> + > >> >> +static __inline __m128i > >> >> +_mm_set1_epi32(int i) > >> >> +{ > >> >> + return vdupq_n_s32(i); > >> >> +} > >> >> + > >> >> +static __inline __m128i > >> >> +_mm_and_si128(__m128i a, __m128i b) > >> >> +{ > >> >> + return vandq_s32(a, b); > >> >> +} > >> >> + > > > > IMO, it's not always good to emulate GCC defined intrinsics of > > other architecture. What if a legacy DPDK application has such mappings > > then BOOM, multiple definition, which one is correct? which one > > to comment it out? Integration pain starts for DPDK library consumer:-( > > > They can include rte_vect.h in build/include directly, which is linked correctly > to the one for that ARCH, so there is no need to worry about. > > > >> > > >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON > >> > Let's create the rte_vect_* as required. look at the existing patch. > >> > > >> I thought of creating a layer of SIMD over all the platforms before. > >> But can't you see it make things complicated, considering there are > >> only few simple intrinsic to implement? > > > > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON > > implementation if I were to take this approach and emulation comes with > > the cost. > > > No, I will not re-implement all the intrinsic like that . > I only do with the simple intrinsic, such as load/store, as you said below. > > > So my take is, > > lets the each architecture implementation for specific SIMD version of DPDK > > API in the library should have the freedom to implement the API in > > NATIVE. > > > > And let's create only rte_vect_* abstraction only for using > > that API/library. Which boils down to have very minimal rte_vect_* > > abstraction to load, store, set not beyond that. > > > > This makes clear "contract" between DPDK library and the applications. > > and make easy for remaning new architecture porting effort in DPDK. > > > Agree. > But I reuse existing intrinsic names, and you recreate new ones. > And I try to do as few changes as possible, and try to avoid any > mistaken which may cause code un-compiled. > I think it's design level question, we need to hear what others talk about it. > > > Imagine how your proposed function will look like if new architecture > > wants to implement "optimized" version of rte_lpm_lookupx4 > > > There is no optimization for this (simple) rte_lpm_lookupx4, otherwise > you have done that in your patch. > If there is for other new platform, defintely they should do like > yours, as you did for NEON ACL. > > > > >> If do so, we also need to explain to others how to use these interfaces. > >> Besides, this patch did the smallest changes to the original code, and > >> more likely to be accepted by others. > > > > other patch makes no changes to IA version of rte_lpm_lookupx4.I thought > > that make reviewer easy to review the changes in architecture > > perspective. > > > As I know, they don't enable LPM for PPC, and ARM is the first one to > touch this issue. > > >> > >> > > >> >> #ifdef RTE_ARCH_ARM > >> >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ > >> >> static __inline uint8x16_t > >> >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > >> >> index c299ce2..c76c07d 100644 > >> >> --- a/lib/librte_lpm/rte_lpm.h > >> >> +++ b/lib/librte_lpm/rte_lpm.h > >> >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> >> /* Mask four results. */ > >> >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > >> >> > >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > >> > > >> > Separate out arm implementation to the different header file. > >> > Too many ifdef looks odd in the header file and difficult to manage. > >> > > >> But there are many ifdefs already. > >> And It seems unreasonable to add a new file only for one small function. > >> > > > > small or big, its matter of each architecture to have > > the freedom for the optimized version for the implementation. > > > > What if other architecture demands to write this function in assembly > > or restructure it for performance improvement? > > > If there is such demands, should do like that. > But I don't see any restructure in your patch, and you still follow > the logic as x86, is it worth adding a new file? > My preference would also be to put architecture dependent implementation into different files. Might be create lib/librte_lpm/arch/(arm|x86)/... here? Konstantin
On Wed, Dec 02, 2015 at 05:49:41PM +0800, Jianbo Liu wrote: > On 2 December 2015 at 16:03, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: > >> On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > >> > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > >> >> Adds ARM NEON support for lpm. > >> >> And enables table/pipeline libraries which depend on lpm. > >> > > >> > I already sent the patch on the same yesterday. > >> > We can converge the patches after the discussion. > >> > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > >> > > >> Yes, I have read your patch. But there are many differences, so I sent > >> mine for your reviewing :) > >> > >> > > >> >> > >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> > >> >> --- > >> >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - > >> >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - > >> >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ > >> >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- > >> >> 4 files changed, 77 insertions(+), 25 deletions(-) > >> >> > >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> index cbebd64..efffa1f 100644 > >> >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n > >> >> CONFIG_RTE_EAL_IGB_UIO=n > >> >> > >> >> # fails to compile on ARM > >> >> -CONFIG_RTE_LIBRTE_LPM=n > >> >> -CONFIG_RTE_LIBRTE_TABLE=n > >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> >> CONFIG_RTE_SCHED_VECTOR=n > >> >> > >> >> # cannot use those on ARM > >> >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> index 504f3ed..57f7941 100644 > >> >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n > >> >> CONFIG_RTE_LIBRTE_FM10K_PMD=n > >> >> CONFIG_RTE_LIBRTE_I40E_PMD=n > >> >> > >> >> -CONFIG_RTE_LIBRTE_LPM=n > >> >> -CONFIG_RTE_LIBRTE_TABLE=n > >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> >> CONFIG_RTE_SCHED_VECTOR=n > >> >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> index a33c054..7437711 100644 > >> >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> @@ -41,6 +41,8 @@ extern "C" { > >> >> > >> >> typedef int32x4_t xmm_t; > >> >> > >> >> +typedef int32x4_t __m128i; > >> >> + > >> >> #define XMM_SIZE (sizeof(xmm_t)) > >> >> #define XMM_MASK (XMM_SIZE - 1) > >> >> > >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { > >> >> double pd[XMM_SIZE / sizeof(double)]; > >> >> } __attribute__((aligned(16))) rte_xmm_t; > >> >> > >> >> +static __inline __m128i > >> >> +_mm_set_epi32(int i3, int i2, int i1, int i0) > >> >> +{ > >> >> + int32_t r[4] = {i0, i1, i2, i3}; > >> >> + > >> >> + return vld1q_s32(r); > >> >> +} > >> >> + > >> >> +static __inline __m128i > >> >> +_mm_loadu_si128(__m128i *p) > >> >> +{ > >> >> + return vld1q_s32((int32_t *)p); > >> >> +} > >> >> + > >> >> +static __inline __m128i > >> >> +_mm_set1_epi32(int i) > >> >> +{ > >> >> + return vdupq_n_s32(i); > >> >> +} > >> >> + > >> >> +static __inline __m128i > >> >> +_mm_and_si128(__m128i a, __m128i b) > >> >> +{ > >> >> + return vandq_s32(a, b); > >> >> +} > >> >> + > > > > IMO, it's not always good to emulate GCC defined intrinsics of > > other architecture. What if a legacy DPDK application has such mappings > > then BOOM, multiple definition, which one is correct? which one > > to comment it out? Integration pain starts for DPDK library consumer:-( > > > They can include rte_vect.h in build/include directly, which is linked correctly > to the one for that ARCH, so there is no need to worry about. I think you missed the point,I was trying to say that legacy DPDK application and third party stacks uses SSE2NEON kind of libraries for quick integration, for example, something like this https://github.com/jratcliff63367/sse2neon/blob/master/SSE2NEON.h AND they include "rte_lpm.h"(it internally includes rte_vect.h) that lead to multiple definition and its not good. > > > >> > > >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON > >> > Let's create the rte_vect_* as required. look at the existing patch. > >> > > >> I thought of creating a layer of SIMD over all the platforms before. > >> But can't you see it make things complicated, considering there are > >> only few simple intrinsic to implement? > > > > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON > > implementation if I were to take this approach and emulation comes with > > the cost. > > > No, I will not re-implement all the intrinsic like that . > I only do with the simple intrinsic, such as load/store, as you said below. but you forced to add _mm_and_si128 also to the list and emulated _mm_and_si128 intrinsic. Am just saying no emulation. > > > So my take is, > > lets the each architecture implementation for specific SIMD version of DPDK > > API in the library should have the freedom to implement the API in > > NATIVE. > > > > And let's create only rte_vect_* abstraction only for using > > that API/library. Which boils down to have very minimal rte_vect_* > > abstraction to load, store, set not beyond that. > > > > This makes clear "contract" between DPDK library and the applications. > > and make easy for remaning new architecture porting effort in DPDK. > > > Agree. > But I reuse existing intrinsic names, and you recreate new ones. > And I try to do as few changes as possible, and try to avoid any > mistaken which may cause code un-compiled. Its trival to verify. Just compile it > I think it's design level question, we need to hear what others talk about it. > > > Imagine how your proposed function will look like if new architecture > > wants to implement "optimized" version of rte_lpm_lookupx4 > > > There is no optimization for this (simple) rte_lpm_lookupx4, otherwise > you have done that in your patch. > If there is for other new platform, defintely they should do like > yours, as you did for NEON ACL. > > > > >> If do so, we also need to explain to others how to use these interfaces. > >> Besides, this patch did the smallest changes to the original code, and > >> more likely to be accepted by others. > > > > other patch makes no changes to IA version of rte_lpm_lookupx4.I thought > > that make reviewer easy to review the changes in architecture > > perspective. > > > As I know, they don't enable LPM for PPC, and ARM is the first one to > touch this issue. > > >> > >> > > >> >> #ifdef RTE_ARCH_ARM > >> >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ > >> >> static __inline uint8x16_t > >> >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > >> >> index c299ce2..c76c07d 100644 > >> >> --- a/lib/librte_lpm/rte_lpm.h > >> >> +++ b/lib/librte_lpm/rte_lpm.h > >> >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> >> /* Mask four results. */ > >> >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > >> >> > >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > >> > > >> > Separate out arm implementation to the different header file. > >> > Too many ifdef looks odd in the header file and difficult to manage. > >> > > >> But there are many ifdefs already. > >> And It seems unreasonable to add a new file only for one small function. > >> > > > > small or big, its matter of each architecture to have > > the freedom for the optimized version for the implementation. > > > > What if other architecture demands to write this function in assembly > > or restructure it for performance improvement? > > > If there is such demands, should do like that. > But I don't see any restructure in your patch, and you still follow > the logic as x86, is it worth adding a new file? SIMD Logic on getting 4 indexes for tbl24[] is different. /* get 4 indexes for tbl24[]. */ i24 = _mm_srli_epi32(ip, CHAR_BIT); /* extract values from tbl24[] */ idx = _mm_cvtsi128_si64(i24); i24 = _mm_srli_si128(i24, sizeof(uint64_t)); tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; idx = _mm_cvtsi128_si64(i24); tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; VS /* extract values from tbl24[] */ idx = vgetq_lane_u64((uint64x2_t)i24, 0); tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; idx = vgetq_lane_u64((uint64x2_t)i24, 1); tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > > > >> > > >> >> +static inline void > >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) > >> >> +{ > >> >> + uint32x4_t i24; > >> >> + uint32_t idx[4]; > >> >> + > >> >> + /* get 4 indexes for tbl24[]. */ > >> >> + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); > >> >> + vst1q_u32(idx, i24); > >> >> + > >> >> + /* extract values from tbl24[] */ > >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; > >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; > >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; > >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; > >> >> +} > >> > > >> > Nice. There is an improvement in this portion code wrt my patch. This is > >> > a candidate for convergence. > >> > > >> > > >> >> +#else > >> >> +static inline void > >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) > >> >> +{ > >> >> + __m128i i24; > >> >> + uint64_t idx; > >> >> + > >> >> + /* get 4 indexes for tbl24[]. */ > >> >> + i24 = _mm_srli_epi32(ip, CHAR_BIT); > >> >> + > >> >> + /* extract values from tbl24[] */ > >> >> + idx = _mm_cvtsi128_si64(i24); > >> >> + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > >> >> + > >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> + > >> >> + idx = _mm_cvtsi128_si64(i24); > >> >> + > >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> +} > >> >> +#endif > >> >> + > >> >> /** > >> >> * Lookup four IP addresses in an LPM table. > >> >> * > >> >> @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> >> * if lookup would fail. > >> >> */ > >> >> static inline void > >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > >> >> +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], > >> >> + uint16_t defv) > >> > > >> > This would call for change in the change the ABI, > >> > IMO, __m128i can be used to represent 128bit vector to avoid ABI chang > >> > > >> This redefine rte_lpm_lookupx4 is unncessary, I will remove it, so no > >> ABI change. > >> And there only one ifdef for ARM platforms left. > >> > >> > > >> >> +#else > >> > separate out arm implementation to the different header file. Too many > >> > ifdef looks odd in the header file. > >> > > >> > Could you rebase your patch based on existing patch and send the > >> > improvement portion as separate patch or I can send update patch with > >> > your improvements and with your signoff. > >> > > >> > > >> >> rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > >> >> uint16_t defv) > >> >> +#endif > >> >> { > >> >> - __m128i i24; > >> >> rte_xmm_t i8; > >> >> uint16_t tbl[4]; > >> >> - uint64_t idx, pt; > >> >> - > >> >> - const __m128i mask8 = > >> >> - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); > >> >> + uint64_t pt; > >> >> > >> >> + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); > >> >> /* > >> >> * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries > >> >> * as one 64-bit value (0x0300030003000300). > >> >> @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | > >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); > >> >> > >> >> - /* get 4 indexes for tbl24[]. */ > >> >> - i24 = _mm_srli_epi32(ip, CHAR_BIT); > >> >> - > >> >> - /* extract values from tbl24[] */ > >> >> - idx = _mm_cvtsi128_si64(i24); > >> >> - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > >> >> - > >> >> - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> - > >> >> - idx = _mm_cvtsi128_si64(i24); > >> >> - > >> >> - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> + rte_lpm_tbl24_val4(lpm, ip, tbl); > >> >> > >> >> /* get 4 indexes for tbl8[]. */ > >> >> i8.x = _mm_and_si128(ip, mask8); > >> >> -- > >> >> 1.8.3.1 > >> >>
On Wed, Dec 02, 2015 at 10:33:44AM +0000, Ananyev, Konstantin wrote: > Hi everyone, > > > -----Original Message----- > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jianbo Liu > > Sent: Wednesday, December 02, 2015 9:50 AM > > To: Jerin Jacob > > Cc: dev@dpdk.org > > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > > > On 2 December 2015 at 16:03, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: > > >> On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > >> > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > > >> >> Adds ARM NEON support for lpm. > > >> >> And enables table/pipeline libraries which depend on lpm. > > >> > > > >> > I already sent the patch on the same yesterday. > > >> > We can converge the patches after the discussion. > > >> > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > > >> > > > >> Yes, I have read your patch. But there are many differences, so I sent > > >> mine for your reviewing :) > > >> > > >> > > > >> >> > > >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> > > >> >> --- > > >> >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - > > >> >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - > > >> >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ > > >> >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- > > >> >> 4 files changed, 77 insertions(+), 25 deletions(-) > > >> >> > > >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc > > >> >> index cbebd64..efffa1f 100644 > > >> >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc > > >> >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc > > >> >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n > > >> >> CONFIG_RTE_EAL_IGB_UIO=n > > >> >> > > >> >> # fails to compile on ARM > > >> >> -CONFIG_RTE_LIBRTE_LPM=n > > >> >> -CONFIG_RTE_LIBRTE_TABLE=n > > >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > > >> >> CONFIG_RTE_SCHED_VECTOR=n > > >> >> > > >> >> # cannot use those on ARM > > >> >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc > > >> >> index 504f3ed..57f7941 100644 > > >> >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc > > >> >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc > > >> >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n > > >> >> CONFIG_RTE_LIBRTE_FM10K_PMD=n > > >> >> CONFIG_RTE_LIBRTE_I40E_PMD=n > > >> >> > > >> >> -CONFIG_RTE_LIBRTE_LPM=n > > >> >> -CONFIG_RTE_LIBRTE_TABLE=n > > >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > > >> >> CONFIG_RTE_SCHED_VECTOR=n > > >> >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h > > >> >> index a33c054..7437711 100644 > > >> >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h > > >> >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h > > >> >> @@ -41,6 +41,8 @@ extern "C" { > > >> >> > > >> >> typedef int32x4_t xmm_t; > > >> >> > > >> >> +typedef int32x4_t __m128i; > > >> >> + > > >> >> #define XMM_SIZE (sizeof(xmm_t)) > > >> >> #define XMM_MASK (XMM_SIZE - 1) > > >> >> > > >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { > > >> >> double pd[XMM_SIZE / sizeof(double)]; > > >> >> } __attribute__((aligned(16))) rte_xmm_t; > > >> >> > > >> >> +static __inline __m128i > > >> >> +_mm_set_epi32(int i3, int i2, int i1, int i0) > > >> >> +{ > > >> >> + int32_t r[4] = {i0, i1, i2, i3}; > > >> >> + > > >> >> + return vld1q_s32(r); > > >> >> +} > > >> >> + > > >> >> +static __inline __m128i > > >> >> +_mm_loadu_si128(__m128i *p) > > >> >> +{ > > >> >> + return vld1q_s32((int32_t *)p); > > >> >> +} > > >> >> + > > >> >> +static __inline __m128i > > >> >> +_mm_set1_epi32(int i) > > >> >> +{ > > >> >> + return vdupq_n_s32(i); > > >> >> +} > > >> >> + > > >> >> +static __inline __m128i > > >> >> +_mm_and_si128(__m128i a, __m128i b) > > >> >> +{ > > >> >> + return vandq_s32(a, b); > > >> >> +} > > >> >> + > > > > > > IMO, it's not always good to emulate GCC defined intrinsics of > > > other architecture. What if a legacy DPDK application has such mappings > > > then BOOM, multiple definition, which one is correct? which one > > > to comment it out? Integration pain starts for DPDK library consumer:-( > > > > > They can include rte_vect.h in build/include directly, which is linked correctly > > to the one for that ARCH, so there is no need to worry about. > > > > > > >> > > > >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON > > >> > Let's create the rte_vect_* as required. look at the existing patch. > > >> > > > >> I thought of creating a layer of SIMD over all the platforms before. > > >> But can't you see it make things complicated, considering there are > > >> only few simple intrinsic to implement? > > > > > > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON > > > implementation if I were to take this approach and emulation comes with > > > the cost. > > > > > No, I will not re-implement all the intrinsic like that . > > I only do with the simple intrinsic, such as load/store, as you said below. > > > > > So my take is, > > > lets the each architecture implementation for specific SIMD version of DPDK > > > API in the library should have the freedom to implement the API in > > > NATIVE. > > > > > > And let's create only rte_vect_* abstraction only for using > > > that API/library. Which boils down to have very minimal rte_vect_* > > > abstraction to load, store, set not beyond that. > > > > > > This makes clear "contract" between DPDK library and the applications. > > > and make easy for remaning new architecture porting effort in DPDK. > > > > > Agree. > > But I reuse existing intrinsic names, and you recreate new ones. > > And I try to do as few changes as possible, and try to avoid any > > mistaken which may cause code un-compiled. > > I think it's design level question, we need to hear what others talk about it. > > > > > Imagine how your proposed function will look like if new architecture > > > wants to implement "optimized" version of rte_lpm_lookupx4 > > > > > There is no optimization for this (simple) rte_lpm_lookupx4, otherwise > > you have done that in your patch. > > If there is for other new platform, defintely they should do like > > yours, as you did for NEON ACL. > > > > > > > >> If do so, we also need to explain to others how to use these interfaces. > > >> Besides, this patch did the smallest changes to the original code, and > > >> more likely to be accepted by others. > > > > > > other patch makes no changes to IA version of rte_lpm_lookupx4.I thought > > > that make reviewer easy to review the changes in architecture > > > perspective. > > > > > As I know, they don't enable LPM for PPC, and ARM is the first one to > > touch this issue. > > > > >> > > >> > > > >> >> #ifdef RTE_ARCH_ARM > > >> >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ > > >> >> static __inline uint8x16_t > > >> >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > > >> >> index c299ce2..c76c07d 100644 > > >> >> --- a/lib/librte_lpm/rte_lpm.h > > >> >> +++ b/lib/librte_lpm/rte_lpm.h > > >> >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > > >> >> /* Mask four results. */ > > >> >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > > >> >> > > >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > > >> > > > >> > Separate out arm implementation to the different header file. > > >> > Too many ifdef looks odd in the header file and difficult to manage. > > >> > > > >> But there are many ifdefs already. > > >> And It seems unreasonable to add a new file only for one small function. > > >> > > > > > > small or big, its matter of each architecture to have > > > the freedom for the optimized version for the implementation. > > > > > > What if other architecture demands to write this function in assembly > > > or restructure it for performance improvement? > > > > > If there is such demands, should do like that. > > But I don't see any restructure in your patch, and you still follow > > the logic as x86, is it worth adding a new file? > > > > My preference would also be to put architecture dependent implementation > into different files. > Might be create lib/librte_lpm/arch/(arm|x86)/... here? > Konstantin +1 my existing patch creates lib/librte_lpm/rte_lpm_neon.h instead of lib/librte_lpm/arch/arm/rte_lpm_neon.h like lib/librte_hash/rte_cmp_x86.h I am OK for changing the directory structure as proposed in my next revision of patch. Let me know if anyone has any objections/concerns. Jerin
On Wed, 2 Dec 2015 16:09:06 +0530 Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > [snip] > > > IMO, it's not always good to emulate GCC defined intrinsics of > > > other architecture. What if a legacy DPDK application has such mappings > > > then BOOM, multiple definition, which one is correct? which one > > > to comment it out? Integration pain starts for DPDK library consumer:-( > > > > > They can include rte_vect.h in build/include directly, which is linked correctly > > to the one for that ARCH, so there is no need to worry about. > > I think you missed the point,I was trying to say that > legacy DPDK application and third party stacks uses SSE2NEON kind of > libraries > for quick integration, for example, something like this > https://github.com/jratcliff63367/sse2neon/blob/master/SSE2NEON.h > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > that lead to multiple definition and its not good. > > > > > > > >> > > > >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON > > >> > Let's create the rte_vect_* as required. look at the existing patch. > > >> > > > >> I thought of creating a layer of SIMD over all the platforms before. > > >> But can't you see it make things complicated, considering there are > > >> only few simple intrinsic to implement? > > > > > > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON > > > implementation if I were to take this approach and emulation comes with > > > the cost. > > > > > No, I will not re-implement all the intrinsic like that . > > I only do with the simple intrinsic, such as load/store, as you said below. > > but you forced to add _mm_and_si128 also to the list and emulated > _mm_and_si128 intrinsic. Am just saying no emulation. > Guys, do we want emulate x86 on ARM? I hope we don't ;). I think, as more platforms might come into DPDK, there will be a need for a proper abstract vector operations API. Yes, we have to describe this API to people. However, otherwise, the ARM guys must learn SSE and write for ARM platform something that looks quite odd. And if there are some "neon emulations" as shown above, it's definitely an argue to have the API that can hide those approachs. Regards Jan
On Wed, 2 Dec 2015 16:18:13 +0530 Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > [snip] > > > > My preference would also be to put architecture dependent implementation > > into different files. > > Might be create lib/librte_lpm/arch/(arm|x86)/... here? > > Konstantin > > +1 > > my existing patch creates lib/librte_lpm/rte_lpm_neon.h instead > of lib/librte_lpm/arch/arm/rte_lpm_neon.h like > lib/librte_hash/rte_cmp_x86.h > > I am OK for changing the directory structure as proposed in my next revision > of patch. > Let me know if anyone has any objections/concerns. > > Jerin I don't like the idea to have arch/... directory structure inside libraries. I would delay such decision until there are really a big number of different optimized implementations. However, the rte_lpm_neon.h approach is OK from my point of view. Jan > > [snip]
On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > On Wed, Dec 02, 2015 at 05:49:41PM +0800, Jianbo Liu wrote: >> On 2 December 2015 at 16:03, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: >> > On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: >> >> On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: >> >> > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: >> >> >> Adds ARM NEON support for lpm. >> >> >> And enables table/pipeline libraries which depend on lpm. >> >> > >> >> > I already sent the patch on the same yesterday. >> >> > We can converge the patches after the discussion. >> >> > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml >> >> > >> >> Yes, I have read your patch. But there are many differences, so I sent >> >> mine for your reviewing :) >> >> >> >> > >> >> >> >> >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> >> >> >> --- >> >> >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - >> >> >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - >> >> >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ >> >> >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- >> >> >> 4 files changed, 77 insertions(+), 25 deletions(-) >> >> >> >> >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc >> >> >> index cbebd64..efffa1f 100644 >> >> >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc >> >> >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc >> >> >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n >> >> >> CONFIG_RTE_EAL_IGB_UIO=n >> >> >> >> >> >> # fails to compile on ARM >> >> >> -CONFIG_RTE_LIBRTE_LPM=n >> >> >> -CONFIG_RTE_LIBRTE_TABLE=n >> >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n >> >> >> CONFIG_RTE_SCHED_VECTOR=n >> >> >> >> >> >> # cannot use those on ARM >> >> >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc >> >> >> index 504f3ed..57f7941 100644 >> >> >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc >> >> >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc >> >> >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n >> >> >> CONFIG_RTE_LIBRTE_FM10K_PMD=n >> >> >> CONFIG_RTE_LIBRTE_I40E_PMD=n >> >> >> >> >> >> -CONFIG_RTE_LIBRTE_LPM=n >> >> >> -CONFIG_RTE_LIBRTE_TABLE=n >> >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n >> >> >> CONFIG_RTE_SCHED_VECTOR=n >> >> >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h >> >> >> index a33c054..7437711 100644 >> >> >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h >> >> >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h >> >> >> @@ -41,6 +41,8 @@ extern "C" { >> >> >> >> >> >> typedef int32x4_t xmm_t; >> >> >> >> >> >> +typedef int32x4_t __m128i; >> >> >> + >> >> >> #define XMM_SIZE (sizeof(xmm_t)) >> >> >> #define XMM_MASK (XMM_SIZE - 1) >> >> >> >> >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { >> >> >> double pd[XMM_SIZE / sizeof(double)]; >> >> >> } __attribute__((aligned(16))) rte_xmm_t; >> >> >> >> >> >> +static __inline __m128i >> >> >> +_mm_set_epi32(int i3, int i2, int i1, int i0) >> >> >> +{ >> >> >> + int32_t r[4] = {i0, i1, i2, i3}; >> >> >> + >> >> >> + return vld1q_s32(r); >> >> >> +} >> >> >> + >> >> >> +static __inline __m128i >> >> >> +_mm_loadu_si128(__m128i *p) >> >> >> +{ >> >> >> + return vld1q_s32((int32_t *)p); >> >> >> +} >> >> >> + >> >> >> +static __inline __m128i >> >> >> +_mm_set1_epi32(int i) >> >> >> +{ >> >> >> + return vdupq_n_s32(i); >> >> >> +} >> >> >> + >> >> >> +static __inline __m128i >> >> >> +_mm_and_si128(__m128i a, __m128i b) >> >> >> +{ >> >> >> + return vandq_s32(a, b); >> >> >> +} >> >> >> + >> > >> > IMO, it's not always good to emulate GCC defined intrinsics of >> > other architecture. What if a legacy DPDK application has such mappings >> > then BOOM, multiple definition, which one is correct? which one >> > to comment it out? Integration pain starts for DPDK library consumer:-( >> > >> They can include rte_vect.h in build/include directly, which is linked correctly >> to the one for that ARCH, so there is no need to worry about. > > I think you missed the point,I was trying to say that > legacy DPDK application and third party stacks uses SSE2NEON kind of > libraries > for quick integration, for example, something like this > https://github.com/jratcliff63367/sse2neon/blob/master/SSE2NEON.h > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > that lead to multiple definition and its not good. > But you will have similar issue since "typedef int32x4_t __m128i" appears in both your patch and this header file. >> >> >> >> > >> >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON >> >> > Let's create the rte_vect_* as required. look at the existing patch. >> >> > >> >> I thought of creating a layer of SIMD over all the platforms before. >> >> But can't you see it make things complicated, considering there are >> >> only few simple intrinsic to implement? >> > >> > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON >> > implementation if I were to take this approach and emulation comes with >> > the cost. >> > >> No, I will not re-implement all the intrinsic like that . >> I only do with the simple intrinsic, such as load/store, as you said below. > > but you forced to add _mm_and_si128 also to the list and emulated > _mm_and_si128 intrinsic. Am just saying no emulation. > I means simple intrinsic, not load/store only. Depends on how you define emulation. Actually, these simple intrisinic could be only one NEON instruction, and will not bring cost. > >> >> > So my take is, >> > lets the each architecture implementation for specific SIMD version of DPDK >> > API in the library should have the freedom to implement the API in >> > NATIVE. >> > >> > And let's create only rte_vect_* abstraction only for using >> > that API/library. Which boils down to have very minimal rte_vect_* >> > abstraction to load, store, set not beyond that. >> > >> > This makes clear "contract" between DPDK library and the applications. >> > and make easy for remaning new architecture porting effort in DPDK. >> > >> Agree. >> But I reuse existing intrinsic names, and you recreate new ones. >> And I try to do as few changes as possible, and try to avoid any >> mistaken which may cause code un-compiled. > > Its trival to verify. Just compile it > >> I think it's design level question, we need to hear what others talk about it. >> >> > Imagine how your proposed function will look like if new architecture >> > wants to implement "optimized" version of rte_lpm_lookupx4 >> > >> There is no optimization for this (simple) rte_lpm_lookupx4, otherwise >> you have done that in your patch. >> If there is for other new platform, defintely they should do like >> yours, as you did for NEON ACL. >> >> > >> >> If do so, we also need to explain to others how to use these interfaces. >> >> Besides, this patch did the smallest changes to the original code, and >> >> more likely to be accepted by others. >> > >> > other patch makes no changes to IA version of rte_lpm_lookupx4.I thought >> > that make reviewer easy to review the changes in architecture >> > perspective. >> > >> As I know, they don't enable LPM for PPC, and ARM is the first one to >> touch this issue. >> >> >> >> >> > >> >> >> #ifdef RTE_ARCH_ARM >> >> >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ >> >> >> static __inline uint8x16_t >> >> >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h >> >> >> index c299ce2..c76c07d 100644 >> >> >> --- a/lib/librte_lpm/rte_lpm.h >> >> >> +++ b/lib/librte_lpm/rte_lpm.h >> >> >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, >> >> >> /* Mask four results. */ >> >> >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) >> >> >> >> >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) >> >> > >> >> > Separate out arm implementation to the different header file. >> >> > Too many ifdef looks odd in the header file and difficult to manage. >> >> > >> >> But there are many ifdefs already. >> >> And It seems unreasonable to add a new file only for one small function. >> >> >> > >> > small or big, its matter of each architecture to have >> > the freedom for the optimized version for the implementation. >> > >> > What if other architecture demands to write this function in assembly >> > or restructure it for performance improvement? >> > >> If there is such demands, should do like that. >> But I don't see any restructure in your patch, and you still follow >> the logic as x86, is it worth adding a new file? > > SIMD Logic on getting 4 indexes for tbl24[] is different. > > /* get 4 indexes for tbl24[]. */ > i24 = _mm_srli_epi32(ip, CHAR_BIT); > > /* extract values from tbl24[] */ > idx = _mm_cvtsi128_si64(i24); > i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > > tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > idx = _mm_cvtsi128_si64(i24); > > tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > VS > > /* extract values from tbl24[] */ > idx = vgetq_lane_u64((uint64x2_t)i24, 0); > > tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > idx = vgetq_lane_u64((uint64x2_t)i24, 1); > > tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > It's only the optimazation of part of code in that function. I did the similar in my patch. But, looking from the whole, this function is not restructured, and the logic is the same as x86. >> >> > >> >> > >> >> >> +static inline void >> >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) >> >> >> +{ >> >> >> + uint32x4_t i24; >> >> >> + uint32_t idx[4]; >> >> >> + >> >> >> + /* get 4 indexes for tbl24[]. */ >> >> >> + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); >> >> >> + vst1q_u32(idx, i24); >> >> >> + >> >> >> + /* extract values from tbl24[] */ >> >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; >> >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; >> >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; >> >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; >> >> >> +} >> >> > >> >> > Nice. There is an improvement in this portion code wrt my patch. This is >> >> > a candidate for convergence. >> >> > >> >> > >> >> >> +#else >> >> >> +static inline void >> >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) >> >> >> +{ >> >> >> + __m128i i24; >> >> >> + uint64_t idx; >> >> >> + >> >> >> + /* get 4 indexes for tbl24[]. */ >> >> >> + i24 = _mm_srli_epi32(ip, CHAR_BIT); >> >> >> + >> >> >> + /* extract values from tbl24[] */ >> >> >> + idx = _mm_cvtsi128_si64(i24); >> >> >> + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); >> >> >> + >> >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> >> + >> >> >> + idx = _mm_cvtsi128_si64(i24); >> >> >> + >> >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> >> +} >> >> >> +#endif >> >> >> + >> >> >> /** >> >> >> * Lookup four IP addresses in an LPM table. >> >> >> * >> >> >> @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, >> >> >> * if lookup would fail. >> >> >> */ >> >> >> static inline void >> >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) >> >> >> +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], >> >> >> + uint16_t defv) >> >> > >> >> > This would call for change in the change the ABI, >> >> > IMO, __m128i can be used to represent 128bit vector to avoid ABI chang >> >> > >> >> This redefine rte_lpm_lookupx4 is unncessary, I will remove it, so no >> >> ABI change. >> >> And there only one ifdef for ARM platforms left. >> >> >> >> > >> >> >> +#else >> >> > separate out arm implementation to the different header file. Too many >> >> > ifdef looks odd in the header file. >> >> > >> >> > Could you rebase your patch based on existing patch and send the >> >> > improvement portion as separate patch or I can send update patch with >> >> > your improvements and with your signoff. >> >> > >> >> > >> >> >> rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], >> >> >> uint16_t defv) >> >> >> +#endif >> >> >> { >> >> >> - __m128i i24; >> >> >> rte_xmm_t i8; >> >> >> uint16_t tbl[4]; >> >> >> - uint64_t idx, pt; >> >> >> - >> >> >> - const __m128i mask8 = >> >> >> - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); >> >> >> + uint64_t pt; >> >> >> >> >> >> + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); >> >> >> /* >> >> >> * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries >> >> >> * as one 64-bit value (0x0300030003000300). >> >> >> @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], >> >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | >> >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); >> >> >> >> >> >> - /* get 4 indexes for tbl24[]. */ >> >> >> - i24 = _mm_srli_epi32(ip, CHAR_BIT); >> >> >> - >> >> >> - /* extract values from tbl24[] */ >> >> >> - idx = _mm_cvtsi128_si64(i24); >> >> >> - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); >> >> >> - >> >> >> - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> >> - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> >> - >> >> >> - idx = _mm_cvtsi128_si64(i24); >> >> >> - >> >> >> - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; >> >> >> - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; >> >> >> + rte_lpm_tbl24_val4(lpm, ip, tbl); >> >> >> >> >> >> /* get 4 indexes for tbl8[]. */ >> >> >> i8.x = _mm_and_si128(ip, mask8); >> >> >> -- >> >> >> 1.8.3.1 >> >> >>
On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > On Wed, Dec 02, 2015 at 05:49:41PM +0800, Jianbo Liu wrote: > >> On 2 December 2015 at 16:03, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > >> > On Wed, Dec 02, 2015 at 02:54:52PM +0800, Jianbo Liu wrote: > >> >> On 2 December 2015 at 00:41, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > >> >> > On Tue, Dec 01, 2015 at 01:41:15PM -0500, Jianbo Liu wrote: > >> >> >> Adds ARM NEON support for lpm. > >> >> >> And enables table/pipeline libraries which depend on lpm. > >> >> > > >> >> > I already sent the patch on the same yesterday. > >> >> > We can converge the patches after the discussion. > >> >> > Please check "[dpdk-dev] [PATCH 0/3] add lpm support for NEON" on ml > >> >> > > >> >> Yes, I have read your patch. But there are many differences, so I sent > >> >> mine for your reviewing :) > >> >> > >> >> > > >> >> >> > >> >> >> Signed-off-by: Jianbo Liu <jianbo.liu@linaro.org> > >> >> >> --- > >> >> >> config/defconfig_arm-armv7a-linuxapp-gcc | 3 - > >> >> >> config/defconfig_arm64-armv8a-linuxapp-gcc | 3 - > >> >> >> lib/librte_eal/common/include/arch/arm/rte_vect.h | 28 ++++++++++ > >> >> >> lib/librte_lpm/rte_lpm.h | 68 ++++++++++++++++------- > >> >> >> 4 files changed, 77 insertions(+), 25 deletions(-) > >> >> >> > >> >> >> diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> >> index cbebd64..efffa1f 100644 > >> >> >> --- a/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> >> +++ b/config/defconfig_arm-armv7a-linuxapp-gcc > >> >> >> @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n > >> >> >> CONFIG_RTE_EAL_IGB_UIO=n > >> >> >> > >> >> >> # fails to compile on ARM > >> >> >> -CONFIG_RTE_LIBRTE_LPM=n > >> >> >> -CONFIG_RTE_LIBRTE_TABLE=n > >> >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> >> >> CONFIG_RTE_SCHED_VECTOR=n > >> >> >> > >> >> >> # cannot use those on ARM > >> >> >> diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> >> index 504f3ed..57f7941 100644 > >> >> >> --- a/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> >> +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc > >> >> >> @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n > >> >> >> CONFIG_RTE_LIBRTE_FM10K_PMD=n > >> >> >> CONFIG_RTE_LIBRTE_I40E_PMD=n > >> >> >> > >> >> >> -CONFIG_RTE_LIBRTE_LPM=n > >> >> >> -CONFIG_RTE_LIBRTE_TABLE=n > >> >> >> -CONFIG_RTE_LIBRTE_PIPELINE=n > >> >> >> CONFIG_RTE_SCHED_VECTOR=n > >> >> >> diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> >> index a33c054..7437711 100644 > >> >> >> --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> >> +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h > >> >> >> @@ -41,6 +41,8 @@ extern "C" { > >> >> >> > >> >> >> typedef int32x4_t xmm_t; > >> >> >> > >> >> >> +typedef int32x4_t __m128i; > >> >> >> + > >> >> >> #define XMM_SIZE (sizeof(xmm_t)) > >> >> >> #define XMM_MASK (XMM_SIZE - 1) > >> >> >> > >> >> >> @@ -53,6 +55,32 @@ typedef union rte_xmm { > >> >> >> double pd[XMM_SIZE / sizeof(double)]; > >> >> >> } __attribute__((aligned(16))) rte_xmm_t; > >> >> >> > >> >> >> +static __inline __m128i > >> >> >> +_mm_set_epi32(int i3, int i2, int i1, int i0) > >> >> >> +{ > >> >> >> + int32_t r[4] = {i0, i1, i2, i3}; > >> >> >> + > >> >> >> + return vld1q_s32(r); > >> >> >> +} > >> >> >> + > >> >> >> +static __inline __m128i > >> >> >> +_mm_loadu_si128(__m128i *p) > >> >> >> +{ > >> >> >> + return vld1q_s32((int32_t *)p); > >> >> >> +} > >> >> >> + > >> >> >> +static __inline __m128i > >> >> >> +_mm_set1_epi32(int i) > >> >> >> +{ > >> >> >> + return vdupq_n_s32(i); > >> >> >> +} > >> >> >> + > >> >> >> +static __inline __m128i > >> >> >> +_mm_and_si128(__m128i a, __m128i b) > >> >> >> +{ > >> >> >> + return vandq_s32(a, b); > >> >> >> +} > >> >> >> + > >> > > >> > IMO, it's not always good to emulate GCC defined intrinsics of > >> > other architecture. What if a legacy DPDK application has such mappings > >> > then BOOM, multiple definition, which one is correct? which one > >> > to comment it out? Integration pain starts for DPDK library consumer:-( > >> > > >> They can include rte_vect.h in build/include directly, which is linked correctly > >> to the one for that ARCH, so there is no need to worry about. > > > > I think you missed the point,I was trying to say that > > legacy DPDK application and third party stacks uses SSE2NEON kind of > > libraries > > for quick integration, for example, something like this > > https://github.com/jratcliff63367/sse2neon/blob/master/SSE2NEON.h > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > that lead to multiple definition and its not good. > > > But you will have similar issue since "typedef int32x4_t __m128i" > appears in both your patch and this header file. I just tested it, it won't break, back to back "typedef int32x4_t __m128i" is fine(unlike inline function). my intention to keep __m128i "as is" because changing the __m128i to rte_??? something would break the ABI. > > >> > >> > >> >> > > >> >> > IMO, it makes sense to not emulate the SSE intrinsics with NEON > >> >> > Let's create the rte_vect_* as required. look at the existing patch. > >> >> > > >> >> I thought of creating a layer of SIMD over all the platforms before. > >> >> But can't you see it make things complicated, considering there are > >> >> only few simple intrinsic to implement? > >> > > >> > Not true, There were, a lot of SSE intrinsics needs be to emulated for ACL NEON > >> > implementation if I were to take this approach and emulation comes with > >> > the cost. > >> > > >> No, I will not re-implement all the intrinsic like that . > >> I only do with the simple intrinsic, such as load/store, as you said below. > > > > but you forced to add _mm_and_si128 also to the list and emulated > > _mm_and_si128 intrinsic. Am just saying no emulation. > > > I means simple intrinsic, not load/store only. > Depends on how you define emulation. Actually, these simple intrisinic > could be only one NEON instruction, and will not bring cost. > > > > >> > >> > So my take is, > >> > lets the each architecture implementation for specific SIMD version of DPDK > >> > API in the library should have the freedom to implement the API in > >> > NATIVE. > >> > > >> > And let's create only rte_vect_* abstraction only for using > >> > that API/library. Which boils down to have very minimal rte_vect_* > >> > abstraction to load, store, set not beyond that. > >> > > >> > This makes clear "contract" between DPDK library and the applications. > >> > and make easy for remaning new architecture porting effort in DPDK. > >> > > >> Agree. > >> But I reuse existing intrinsic names, and you recreate new ones. > >> And I try to do as few changes as possible, and try to avoid any > >> mistaken which may cause code un-compiled. > > > > Its trival to verify. Just compile it > > > >> I think it's design level question, we need to hear what others talk about it. > >> > >> > Imagine how your proposed function will look like if new architecture > >> > wants to implement "optimized" version of rte_lpm_lookupx4 > >> > > >> There is no optimization for this (simple) rte_lpm_lookupx4, otherwise > >> you have done that in your patch. > >> If there is for other new platform, defintely they should do like > >> yours, as you did for NEON ACL. > >> > >> > > >> >> If do so, we also need to explain to others how to use these interfaces. > >> >> Besides, this patch did the smallest changes to the original code, and > >> >> more likely to be accepted by others. > >> > > >> > other patch makes no changes to IA version of rte_lpm_lookupx4.I thought > >> > that make reviewer easy to review the changes in architecture > >> > perspective. > >> > > >> As I know, they don't enable LPM for PPC, and ARM is the first one to > >> touch this issue. > >> > >> >> > >> >> > > >> >> >> #ifdef RTE_ARCH_ARM > >> >> >> /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ > >> >> >> static __inline uint8x16_t > >> >> >> diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h > >> >> >> index c299ce2..c76c07d 100644 > >> >> >> --- a/lib/librte_lpm/rte_lpm.h > >> >> >> +++ b/lib/librte_lpm/rte_lpm.h > >> >> >> @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> >> >> /* Mask four results. */ > >> >> >> #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) > >> >> >> > >> >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > >> >> > > >> >> > Separate out arm implementation to the different header file. > >> >> > Too many ifdef looks odd in the header file and difficult to manage. > >> >> > > >> >> But there are many ifdefs already. > >> >> And It seems unreasonable to add a new file only for one small function. > >> >> > >> > > >> > small or big, its matter of each architecture to have > >> > the freedom for the optimized version for the implementation. > >> > > >> > What if other architecture demands to write this function in assembly > >> > or restructure it for performance improvement? > >> > > >> If there is such demands, should do like that. > >> But I don't see any restructure in your patch, and you still follow > >> the logic as x86, is it worth adding a new file? > > > > SIMD Logic on getting 4 indexes for tbl24[] is different. > > > > /* get 4 indexes for tbl24[]. */ > > i24 = _mm_srli_epi32(ip, CHAR_BIT); > > > > /* extract values from tbl24[] */ > > idx = _mm_cvtsi128_si64(i24); > > i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > > > > tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > > tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > > > idx = _mm_cvtsi128_si64(i24); > > > > tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > > tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > > > VS > > > > /* extract values from tbl24[] */ > > idx = vgetq_lane_u64((uint64x2_t)i24, 0); > > > > tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > > tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > > > idx = vgetq_lane_u64((uint64x2_t)i24, 1); > > > > tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > > tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > > > It's only the optimazation of part of code in that function. I did the > similar in my patch. > But, looking from the whole, this function is not restructured, and > the logic is the same as x86. > > >> > >> > > >> >> > > >> >> >> +static inline void > >> >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) > >> >> >> +{ > >> >> >> + uint32x4_t i24; > >> >> >> + uint32_t idx[4]; > >> >> >> + > >> >> >> + /* get 4 indexes for tbl24[]. */ > >> >> >> + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); > >> >> >> + vst1q_u32(idx, i24); > >> >> >> + > >> >> >> + /* extract values from tbl24[] */ > >> >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; > >> >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; > >> >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; > >> >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; > >> >> >> +} > >> >> > > >> >> > Nice. There is an improvement in this portion code wrt my patch. This is > >> >> > a candidate for convergence. > >> >> > > >> >> > > >> >> >> +#else > >> >> >> +static inline void > >> >> >> +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) > >> >> >> +{ > >> >> >> + __m128i i24; > >> >> >> + uint64_t idx; > >> >> >> + > >> >> >> + /* get 4 indexes for tbl24[]. */ > >> >> >> + i24 = _mm_srli_epi32(ip, CHAR_BIT); > >> >> >> + > >> >> >> + /* extract values from tbl24[] */ > >> >> >> + idx = _mm_cvtsi128_si64(i24); > >> >> >> + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > >> >> >> + > >> >> >> + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> >> + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> >> + > >> >> >> + idx = _mm_cvtsi128_si64(i24); > >> >> >> + > >> >> >> + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> >> + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> >> +} > >> >> >> +#endif > >> >> >> + > >> >> >> /** > >> >> >> * Lookup four IP addresses in an LPM table. > >> >> >> * > >> >> >> @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, > >> >> >> * if lookup would fail. > >> >> >> */ > >> >> >> static inline void > >> >> >> +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) > >> >> >> +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], > >> >> >> + uint16_t defv) > >> >> > > >> >> > This would call for change in the change the ABI, > >> >> > IMO, __m128i can be used to represent 128bit vector to avoid ABI chang > >> >> > > >> >> This redefine rte_lpm_lookupx4 is unncessary, I will remove it, so no > >> >> ABI change. > >> >> And there only one ifdef for ARM platforms left. > >> >> > >> >> > > >> >> >> +#else > >> >> > separate out arm implementation to the different header file. Too many > >> >> > ifdef looks odd in the header file. > >> >> > > >> >> > Could you rebase your patch based on existing patch and send the > >> >> > improvement portion as separate patch or I can send update patch with > >> >> > your improvements and with your signoff. > >> >> > > >> >> > > >> >> >> rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > >> >> >> uint16_t defv) > >> >> >> +#endif > >> >> >> { > >> >> >> - __m128i i24; > >> >> >> rte_xmm_t i8; > >> >> >> uint16_t tbl[4]; > >> >> >> - uint64_t idx, pt; > >> >> >> - > >> >> >> - const __m128i mask8 = > >> >> >> - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); > >> >> >> + uint64_t pt; > >> >> >> > >> >> >> + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); > >> >> >> /* > >> >> >> * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries > >> >> >> * as one 64-bit value (0x0300030003000300). > >> >> >> @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], > >> >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | > >> >> >> (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); > >> >> >> > >> >> >> - /* get 4 indexes for tbl24[]. */ > >> >> >> - i24 = _mm_srli_epi32(ip, CHAR_BIT); > >> >> >> - > >> >> >> - /* extract values from tbl24[] */ > >> >> >> - idx = _mm_cvtsi128_si64(i24); > >> >> >> - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); > >> >> >> - > >> >> >> - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> >> - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> >> - > >> >> >> - idx = _mm_cvtsi128_si64(i24); > >> >> >> - > >> >> >> - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; > >> >> >> - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; > >> >> >> + rte_lpm_tbl24_val4(lpm, ip, tbl); > >> >> >> > >> >> >> /* get 4 indexes for tbl8[]. */ > >> >> >> i8.x = _mm_and_si128(ip, mask8); > >> >> >> -- > >> >> >> 1.8.3.1 > >> >> >>
2015-12-02 20:04, Jerin Jacob: > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > that lead to multiple definition and its not good. > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > appears in both your patch and this header file. > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > is fine(unlike inline function). > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > something would break the ABI. Isn't it already broken in 2.2?
On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > 2015-12-02 20:04, Jerin Jacob: > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > that lead to multiple definition and its not good. > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > appears in both your patch and this header file. > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > is fine(unlike inline function). > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > something would break the ABI. > > Isn't it already broken in 2.2? Does it mean, You would like to have rte_128i(or similar) kind of abstraction to represent 128bit SIMD variable in DPDK?
2015-12-02 22:23, Jerin Jacob: > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > 2015-12-02 20:04, Jerin Jacob: > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > that lead to multiple definition and its not good. > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > appears in both your patch and this header file. > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > is fine(unlike inline function). > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > something would break the ABI. > > > > Isn't it already broken in 2.2? > > Does it mean, You would like to have rte_128i(or similar) kind of > abstraction to represent 128bit SIMD variable in DPDK? If you are convinced that it is the best way to write a generic code, yes. I think the most important question is to know what is the best solution for performance and maintainability. The API/ABI questions will be considered after. Thanks for your involvement guys.
On Wed, Dec 02, 2015 at 05:57:10PM +0100, Thomas Monjalon wrote: > 2015-12-02 22:23, Jerin Jacob: > > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > > 2015-12-02 20:04, Jerin Jacob: > > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > > that lead to multiple definition and its not good. > > > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > > appears in both your patch and this header file. > > > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > > is fine(unlike inline function). > > > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > > something would break the ABI. > > > > > > Isn't it already broken in 2.2? > > > > Does it mean, You would like to have rte_128i(or similar) kind of > > abstraction to represent 128bit SIMD variable in DPDK? > > If you are convinced that it is the best way to write a generic code, yes. > I think the most important question is to know what is the best solution > for performance and maintainability. The API/ABI questions will be considered IMO, a true portable platform-independent library may need rte_128i kind of abstracttion to represent a 128bit SIMD variable. I can send an RFC patch to see the changes required across the DPDK. > after. > > Thanks for your involvement guys.
On Wed, Dec 02, 2015 at 05:57:10PM +0100, Thomas Monjalon wrote: > 2015-12-02 22:23, Jerin Jacob: > > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > > 2015-12-02 20:04, Jerin Jacob: > > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > > that lead to multiple definition and its not good. > > > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > > appears in both your patch and this header file. > > > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > > is fine(unlike inline function). > > > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > > something would break the ABI. > > > > > > Isn't it already broken in 2.2? > > > > Does it mean, You would like to have rte_128i(or similar) kind of > > abstraction to represent 128bit SIMD variable in DPDK? > > If you are convinced that it is the best way to write a generic code, yes. I grep-ed through DPDK API list to see the dependency with SIMD in API definition.I see only rte_lpm_lookupx4 API has SIMD dependency in API definition. I believe that's the root cause of the problem. IMO, The better way to fix this would be to remove __m128i from API and have more general representation to remove the architecture dependency from API something like this, rte_lpm_lookupx4(const struct rte_lpm *lpm, uint32_t *ip, uint16_t hop[4], uint16_t defv) instead of rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], uint16_t defv) Now I am not sure why this API was created like this, from l3fwd.c example, it looks to accommodate the IPV4 byte swap[1]. If it's true, maybe we can have eal byte swap abstraction for optimized byte swap on memory for 4 IP address in one shot or Have rte_lpm_lookupx4 take an argument for byte swap or not ? or something similar? Thoughts ? [1] const __m128i bswap_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3); /* Byte swap 4 IPV4 addresses. */ dip = _mm_shuffle_epi8(dip, bswap_mask); Jerin > I think the most important question is to know what is the best solution > for performance and maintainability. The API/ABI questions will be considered > after. > > Thanks for your involvement guys.
Hi Jerin, > -----Original Message----- > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jerin Jacob > Sent: Thursday, December 03, 2015 9:34 AM > To: Thomas Monjalon > Cc: dev@dpdk.org > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > On Wed, Dec 02, 2015 at 05:57:10PM +0100, Thomas Monjalon wrote: > > 2015-12-02 22:23, Jerin Jacob: > > > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > > > 2015-12-02 20:04, Jerin Jacob: > > > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > > > that lead to multiple definition and its not good. > > > > > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > > > appears in both your patch and this header file. > > > > > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > > > is fine(unlike inline function). > > > > > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > > > something would break the ABI. > > > > > > > > Isn't it already broken in 2.2? > > > > > > Does it mean, You would like to have rte_128i(or similar) kind of > > > abstraction to represent 128bit SIMD variable in DPDK? > > > > If you are convinced that it is the best way to write a generic code, yes. > > I grep-ed through DPDK API list to see the dependency with SIMD in API > definition.I see only rte_lpm_lookupx4 API has SIMD dependency in API > definition. > > I believe that's the root cause of the problem. IMO, The > better way to fix this would be to remove __m128i from API and have more > general representation to remove the architecture dependency from API > > something like this, > > rte_lpm_lookupx4(const struct rte_lpm *lpm, uint32_t *ip, uint16_t > hop[4], uint16_t defv) > > instead of > > rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t > hop[4], uint16_t defv) The idea for that function was that rte_lpm_lookupx4() accepts 4 IPv4 addresses that are: 1. already in 128bit register 2. 'prepared' - byte swap is already done for them if needed. About ways to fix __m128i dependency: as I can see x86 and arm DPDK code already has xmm_t typedef: $ find lib -type f | xargs grep xmm_t | grep typedef lib/librte_eal/common/include/arch/x86/rte_vect.h:typedef __m128i xmm_t; lib/librte_eal/common/include/arch/arm/rte_vect.h:typedef int32x4_t xmm_t; Why not to change rte_lpm_lookupx4() to accept xmm_t as input parameter. As I understand it would solve the problem, and wouldn't introduce any API/ABI breakage, right? Konstantin > > Now I am not sure why this API was created like this, from l3fwd.c > example, it looks to accommodate the IPV4 byte swap[1]. If it's true, > maybe we can have eal byte swap abstraction for optimized byte swap on > memory for 4 IP address in one shot > > or > > Have rte_lpm_lookupx4 take an argument for byte swap or not ? > > or > > something similar? > > Thoughts ? > > [1] > const __m128i bswap_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, 11, > 4, 5, 6, 7, 0, 1, 2, 3); > /* Byte swap 4 IPV4 addresses. */ > dip = _mm_shuffle_epi8(dip, bswap_mask); > > Jerin > > > I think the most important question is to know what is the best solution > > for performance and maintainability. The API/ABI questions will be considered > > after. > > > > Thanks for your involvement guys.
On Thu, Dec 03, 2015 at 11:02:07AM +0000, Ananyev, Konstantin wrote: Hi Konstantin, > Hi Jerin, > > > -----Original Message----- > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jerin Jacob > > Sent: Thursday, December 03, 2015 9:34 AM > > To: Thomas Monjalon > > Cc: dev@dpdk.org > > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > > > On Wed, Dec 02, 2015 at 05:57:10PM +0100, Thomas Monjalon wrote: > > > 2015-12-02 22:23, Jerin Jacob: > > > > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > > > > 2015-12-02 20:04, Jerin Jacob: > > > > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > > > > that lead to multiple definition and its not good. > > > > > > > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > > > > appears in both your patch and this header file. > > > > > > > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > > > > is fine(unlike inline function). > > > > > > > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > > > > something would break the ABI. > > > > > > > > > > Isn't it already broken in 2.2? > > > > > > > > Does it mean, You would like to have rte_128i(or similar) kind of > > > > abstraction to represent 128bit SIMD variable in DPDK? > > > > > > If you are convinced that it is the best way to write a generic code, yes. > > > > I grep-ed through DPDK API list to see the dependency with SIMD in API > > definition.I see only rte_lpm_lookupx4 API has SIMD dependency in API > > definition. > > > > I believe that's the root cause of the problem. IMO, The > > better way to fix this would be to remove __m128i from API and have more > > general representation to remove the architecture dependency from API > > > > something like this, > > > > rte_lpm_lookupx4(const struct rte_lpm *lpm, uint32_t *ip, uint16_t > > hop[4], uint16_t defv) > > > > instead of > > > > rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t > > hop[4], uint16_t defv) > > The idea for that function was that rte_lpm_lookupx4() accepts 4 IPv4 addresses that are: > 1. already in 128bit register > 2. 'prepared' - byte swap is already done for them if needed. > > About ways to fix __m128i dependency: as I can see x86 and arm DPDK code > already has xmm_t typedef: > > $ find lib -type f | xargs grep xmm_t | grep typedef > lib/librte_eal/common/include/arch/x86/rte_vect.h:typedef __m128i xmm_t; > lib/librte_eal/common/include/arch/arm/rte_vect.h:typedef int32x4_t xmm_t; > > Why not to change rte_lpm_lookupx4() to accept xmm_t as input parameter. > As I understand it would solve the problem, and wouldn't introduce any API/ABI breakage, right? Yes, If we have API/ABI breakage concerns. IMO, Now this would call for some kind of rte_vect_* abstraction load, store, set kind of SIMD operation on xmm_t in common test code to aviod #ifdef's in app/test/test_lpm.c I guess we may not need those abstractions in lib/librte_eal/common/include/arch/ directory. keeping in app/test/xmmt_ops.h should be enough, right? > > Konstantin > > > > > Now I am not sure why this API was created like this, from l3fwd.c > > example, it looks to accommodate the IPV4 byte swap[1]. If it's true, > > maybe we can have eal byte swap abstraction for optimized byte swap on > > memory for 4 IP address in one shot > > > > or > > > > Have rte_lpm_lookupx4 take an argument for byte swap or not ? > > > > or > > > > something similar? > > > > Thoughts ? > > > > [1] > > const __m128i bswap_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, 11, > > 4, 5, 6, 7, 0, 1, 2, 3); > > /* Byte swap 4 IPV4 addresses. */ > > dip = _mm_shuffle_epi8(dip, bswap_mask); > > > > Jerin > > > > > I think the most important question is to know what is the best solution > > > for performance and maintainability. The API/ABI questions will be considered > > > after. > > > > > > Thanks for your involvement guys.
> -----Original Message----- > From: Jerin Jacob [mailto:jerin.jacob@caviumnetworks.com] > Sent: Thursday, December 03, 2015 12:17 PM > To: Ananyev, Konstantin > Cc: Thomas Monjalon; dev@dpdk.org; viktorin@rehivetech.com > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > On Thu, Dec 03, 2015 at 11:02:07AM +0000, Ananyev, Konstantin wrote: > > Hi Konstantin, > > > Hi Jerin, > > > > > -----Original Message----- > > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jerin Jacob > > > Sent: Thursday, December 03, 2015 9:34 AM > > > To: Thomas Monjalon > > > Cc: dev@dpdk.org > > > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > > > > > On Wed, Dec 02, 2015 at 05:57:10PM +0100, Thomas Monjalon wrote: > > > > 2015-12-02 22:23, Jerin Jacob: > > > > > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > > > > > 2015-12-02 20:04, Jerin Jacob: > > > > > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > > > > > that lead to multiple definition and its not good. > > > > > > > > > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > > > > > appears in both your patch and this header file. > > > > > > > > > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > > > > > is fine(unlike inline function). > > > > > > > > > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > > > > > something would break the ABI. > > > > > > > > > > > > Isn't it already broken in 2.2? > > > > > > > > > > Does it mean, You would like to have rte_128i(or similar) kind of > > > > > abstraction to represent 128bit SIMD variable in DPDK? > > > > > > > > If you are convinced that it is the best way to write a generic code, yes. > > > > > > I grep-ed through DPDK API list to see the dependency with SIMD in API > > > definition.I see only rte_lpm_lookupx4 API has SIMD dependency in API > > > definition. > > > > > > I believe that's the root cause of the problem. IMO, The > > > better way to fix this would be to remove __m128i from API and have more > > > general representation to remove the architecture dependency from API > > > > > > something like this, > > > > > > rte_lpm_lookupx4(const struct rte_lpm *lpm, uint32_t *ip, uint16_t > > > hop[4], uint16_t defv) > > > > > > instead of > > > > > > rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t > > > hop[4], uint16_t defv) > > > > The idea for that function was that rte_lpm_lookupx4() accepts 4 IPv4 addresses that are: > > 1. already in 128bit register > > 2. 'prepared' - byte swap is already done for them if needed. > > > > About ways to fix __m128i dependency: as I can see x86 and arm DPDK code > > already has xmm_t typedef: > > > > $ find lib -type f | xargs grep xmm_t | grep typedef > > lib/librte_eal/common/include/arch/x86/rte_vect.h:typedef __m128i xmm_t; > > lib/librte_eal/common/include/arch/arm/rte_vect.h:typedef int32x4_t xmm_t; > > > > Why not to change rte_lpm_lookupx4() to accept xmm_t as input parameter. > > As I understand it would solve the problem, and wouldn't introduce any API/ABI breakage, right? > > Yes, If we have API/ABI breakage concerns. > > IMO, Now this would call for some kind of rte_vect_* abstraction load, > store, set kind of SIMD operation on xmm_t in common test code to > aviod #ifdef's in app/test/test_lpm.c Yes, seems so. > > I guess we may not need those abstractions in > lib/librte_eal/common/include/arch/ directory. > keeping in app/test/xmmt_ops.h should be enough, right? That sounds ok to me. At least for now. For future the more generic question - do we like to have some generic layer abstraction for similar vector instrincts across different archs? From one side it might help people writing/using vector implementation of some stuff, from other side - there would be extra hassle creating/supporting it. Konstantin > > > > > > Konstantin > > > > > > > > Now I am not sure why this API was created like this, from l3fwd.c > > > example, it looks to accommodate the IPV4 byte swap[1]. If it's true, > > > maybe we can have eal byte swap abstraction for optimized byte swap on > > > memory for 4 IP address in one shot > > > > > > or > > > > > > Have rte_lpm_lookupx4 take an argument for byte swap or not ? > > > > > > or > > > > > > something similar? > > > > > > Thoughts ? > > > > > > [1] > > > const __m128i bswap_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, 11, > > > 4, 5, 6, 7, 0, 1, 2, 3); > > > /* Byte swap 4 IPV4 addresses. */ > > > dip = _mm_shuffle_epi8(dip, bswap_mask); > > > > > > Jerin > > > > > > > I think the most important question is to know what is the best solution > > > > for performance and maintainability. The API/ABI questions will be considered > > > > after. > > > > > > > > Thanks for your involvement guys.
On Thu, Dec 03, 2015 at 12:42:13PM +0000, Ananyev, Konstantin wrote: > > > > -----Original Message----- > > From: Jerin Jacob [mailto:jerin.jacob@caviumnetworks.com] > > Sent: Thursday, December 03, 2015 12:17 PM > > To: Ananyev, Konstantin > > Cc: Thomas Monjalon; dev@dpdk.org; viktorin@rehivetech.com > > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > > > On Thu, Dec 03, 2015 at 11:02:07AM +0000, Ananyev, Konstantin wrote: > > > > Hi Konstantin, > > > > > Hi Jerin, > > > > > > > -----Original Message----- > > > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Jerin Jacob > > > > Sent: Thursday, December 03, 2015 9:34 AM > > > > To: Thomas Monjalon > > > > Cc: dev@dpdk.org > > > > Subject: Re: [dpdk-dev] [PATCH 3/4] eal/arm: Enable lpm/table/pipeline libs > > > > > > > > On Wed, Dec 02, 2015 at 05:57:10PM +0100, Thomas Monjalon wrote: > > > > > 2015-12-02 22:23, Jerin Jacob: > > > > > > On Wed, Dec 02, 2015 at 05:40:13PM +0100, Thomas Monjalon wrote: > > > > > > > 2015-12-02 20:04, Jerin Jacob: > > > > > > > > On Wed, Dec 02, 2015 at 09:13:51PM +0800, Jianbo Liu wrote: > > > > > > > > > On 2 December 2015 at 18:39, Jerin Jacob <jerin.jacob@caviumnetworks.com> wrote: > > > > > > > > > > AND they include "rte_lpm.h"(it internally includes rte_vect.h) > > > > > > > > > > that lead to multiple definition and its not good. > > > > > > > > > > > > > > > > > > > But you will have similar issue since "typedef int32x4_t __m128i" > > > > > > > > > appears in both your patch and this header file. > > > > > > > > > > > > > > > > I just tested it, it won't break, back to back "typedef int32x4_t __m128i" > > > > > > > > is fine(unlike inline function). > > > > > > > > > > > > > > > > my intention to keep __m128i "as is" because changing the __m128i to rte_??? > > > > > > > > something would break the ABI. > > > > > > > > > > > > > > Isn't it already broken in 2.2? > > > > > > > > > > > > Does it mean, You would like to have rte_128i(or similar) kind of > > > > > > abstraction to represent 128bit SIMD variable in DPDK? > > > > > > > > > > If you are convinced that it is the best way to write a generic code, yes. > > > > > > > > I grep-ed through DPDK API list to see the dependency with SIMD in API > > > > definition.I see only rte_lpm_lookupx4 API has SIMD dependency in API > > > > definition. > > > > > > > > I believe that's the root cause of the problem. IMO, The > > > > better way to fix this would be to remove __m128i from API and have more > > > > general representation to remove the architecture dependency from API > > > > > > > > something like this, > > > > > > > > rte_lpm_lookupx4(const struct rte_lpm *lpm, uint32_t *ip, uint16_t > > > > hop[4], uint16_t defv) > > > > > > > > instead of > > > > > > > > rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t > > > > hop[4], uint16_t defv) > > > > > > The idea for that function was that rte_lpm_lookupx4() accepts 4 IPv4 addresses that are: > > > 1. already in 128bit register > > > 2. 'prepared' - byte swap is already done for them if needed. > > > > > > About ways to fix __m128i dependency: as I can see x86 and arm DPDK code > > > already has xmm_t typedef: > > > > > > $ find lib -type f | xargs grep xmm_t | grep typedef > > > lib/librte_eal/common/include/arch/x86/rte_vect.h:typedef __m128i xmm_t; > > > lib/librte_eal/common/include/arch/arm/rte_vect.h:typedef int32x4_t xmm_t; > > > > > > Why not to change rte_lpm_lookupx4() to accept xmm_t as input parameter. > > > As I understand it would solve the problem, and wouldn't introduce any API/ABI breakage, right? > > > > Yes, If we have API/ABI breakage concerns. > > > > IMO, Now this would call for some kind of rte_vect_* abstraction load, > > store, set kind of SIMD operation on xmm_t in common test code to > > aviod #ifdef's in app/test/test_lpm.c > > Yes, seems so. > > > > > I guess we may not need those abstractions in > > lib/librte_eal/common/include/arch/ directory. > > keeping in app/test/xmmt_ops.h should be enough, right? > > That sounds ok to me. > At least for now. > For future the more generic question - do we like to have some > generic layer abstraction for similar vector instrincts across different archs? > From one side it might help people writing/using vector implementation of some stuff, > from other side - there would be extra hassle creating/supporting it. There are few such libaries avilable on web. example: NEON -> SSE https://software.intel.com/sites/default/files/managed/cf/f6/NEONvsSSE.h SSE -> NEON https://github.com/jratcliff63367/sse2neon/blob/master/SSE2NEON.h but coming up with common abstraction will be difficult as it's not one to one mapped all the time and performance criteria to choose the instruction on given architecture to realize a certain logic etc Jerin > > Konstantin > > > > > > > > > > > Konstantin > > > > > > > > > > > Now I am not sure why this API was created like this, from l3fwd.c > > > > example, it looks to accommodate the IPV4 byte swap[1]. If it's true, > > > > maybe we can have eal byte swap abstraction for optimized byte swap on > > > > memory for 4 IP address in one shot > > > > > > > > or > > > > > > > > Have rte_lpm_lookupx4 take an argument for byte swap or not ? > > > > > > > > or > > > > > > > > something similar? > > > > > > > > Thoughts ? > > > > > > > > [1] > > > > const __m128i bswap_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, 11, > > > > 4, 5, 6, 7, 0, 1, 2, 3); > > > > /* Byte swap 4 IPV4 addresses. */ > > > > dip = _mm_shuffle_epi8(dip, bswap_mask); > > > > > > > > Jerin > > > > > > > > > I think the most important question is to know what is the best solution > > > > > for performance and maintainability. The API/ABI questions will be considered > > > > > after. > > > > > > > > > > Thanks for your involvement guys.
diff --git a/config/defconfig_arm-armv7a-linuxapp-gcc b/config/defconfig_arm-armv7a-linuxapp-gcc index cbebd64..efffa1f 100644 --- a/config/defconfig_arm-armv7a-linuxapp-gcc +++ b/config/defconfig_arm-armv7a-linuxapp-gcc @@ -53,9 +53,6 @@ CONFIG_RTE_LIBRTE_KNI=n CONFIG_RTE_EAL_IGB_UIO=n # fails to compile on ARM -CONFIG_RTE_LIBRTE_LPM=n -CONFIG_RTE_LIBRTE_TABLE=n -CONFIG_RTE_LIBRTE_PIPELINE=n CONFIG_RTE_SCHED_VECTOR=n # cannot use those on ARM diff --git a/config/defconfig_arm64-armv8a-linuxapp-gcc b/config/defconfig_arm64-armv8a-linuxapp-gcc index 504f3ed..57f7941 100644 --- a/config/defconfig_arm64-armv8a-linuxapp-gcc +++ b/config/defconfig_arm64-armv8a-linuxapp-gcc @@ -51,7 +51,4 @@ CONFIG_RTE_LIBRTE_IVSHMEM=n CONFIG_RTE_LIBRTE_FM10K_PMD=n CONFIG_RTE_LIBRTE_I40E_PMD=n -CONFIG_RTE_LIBRTE_LPM=n -CONFIG_RTE_LIBRTE_TABLE=n -CONFIG_RTE_LIBRTE_PIPELINE=n CONFIG_RTE_SCHED_VECTOR=n diff --git a/lib/librte_eal/common/include/arch/arm/rte_vect.h b/lib/librte_eal/common/include/arch/arm/rte_vect.h index a33c054..7437711 100644 --- a/lib/librte_eal/common/include/arch/arm/rte_vect.h +++ b/lib/librte_eal/common/include/arch/arm/rte_vect.h @@ -41,6 +41,8 @@ extern "C" { typedef int32x4_t xmm_t; +typedef int32x4_t __m128i; + #define XMM_SIZE (sizeof(xmm_t)) #define XMM_MASK (XMM_SIZE - 1) @@ -53,6 +55,32 @@ typedef union rte_xmm { double pd[XMM_SIZE / sizeof(double)]; } __attribute__((aligned(16))) rte_xmm_t; +static __inline __m128i +_mm_set_epi32(int i3, int i2, int i1, int i0) +{ + int32_t r[4] = {i0, i1, i2, i3}; + + return vld1q_s32(r); +} + +static __inline __m128i +_mm_loadu_si128(__m128i *p) +{ + return vld1q_s32((int32_t *)p); +} + +static __inline __m128i +_mm_set1_epi32(int i) +{ + return vdupq_n_s32(i); +} + +static __inline __m128i +_mm_and_si128(__m128i a, __m128i b) +{ + return vandq_s32(a, b); +} + #ifdef RTE_ARCH_ARM /* NEON intrinsic vqtbl1q_u8() is not supported in ARMv7-A(AArch32) */ static __inline uint8x16_t diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h index c299ce2..c76c07d 100644 --- a/lib/librte_lpm/rte_lpm.h +++ b/lib/librte_lpm/rte_lpm.h @@ -361,6 +361,47 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, /* Mask four results. */ #define RTE_LPM_MASKX4_RES UINT64_C(0x00ff00ff00ff00ff) +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) +static inline void +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t tbl[4]) +{ + uint32x4_t i24; + uint32_t idx[4]; + + /* get 4 indexes for tbl24[]. */ + i24 = vshrq_n_u32(vreinterpretq_u32_s32(ip), CHAR_BIT); + vst1q_u32(idx, i24); + + /* extract values from tbl24[] */ + tbl[0] = *(const uint16_t *)&lpm->tbl24[idx[0]]; + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx[1]]; + tbl[2] = *(const uint16_t *)&lpm->tbl24[idx[2]]; + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx[3]]; +} +#else +static inline void +rte_lpm_tbl24_val4(const struct rte_lpm *lpm, __m128i ip, uint16_t tbl[4]) +{ + __m128i i24; + uint64_t idx; + + /* get 4 indexes for tbl24[]. */ + i24 = _mm_srli_epi32(ip, CHAR_BIT); + + /* extract values from tbl24[] */ + idx = _mm_cvtsi128_si64(i24); + i24 = _mm_srli_si128(i24, sizeof(uint64_t)); + + tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; + tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; + + idx = _mm_cvtsi128_si64(i24); + + tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; + tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; +} +#endif + /** * Lookup four IP addresses in an LPM table. * @@ -381,17 +422,19 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const uint32_t * ips, * if lookup would fail. */ static inline void +#if defined(RTE_ARCH_ARM) || defined(RTE_ARCH_ARM64) +rte_lpm_lookupx4(const struct rte_lpm *lpm, int32x4_t ip, uint16_t hop[4], + uint16_t defv) +#else rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], uint16_t defv) +#endif { - __m128i i24; rte_xmm_t i8; uint16_t tbl[4]; - uint64_t idx, pt; - - const __m128i mask8 = - _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX); + uint64_t pt; + const __m128i mask8 = _mm_set1_epi32(UINT8_MAX); /* * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 4 LPM entries * as one 64-bit value (0x0300030003000300). @@ -412,20 +455,7 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip, uint16_t hop[4], (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32 | (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 48); - /* get 4 indexes for tbl24[]. */ - i24 = _mm_srli_epi32(ip, CHAR_BIT); - - /* extract values from tbl24[] */ - idx = _mm_cvtsi128_si64(i24); - i24 = _mm_srli_si128(i24, sizeof(uint64_t)); - - tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; - tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; - - idx = _mm_cvtsi128_si64(i24); - - tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx]; - tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32]; + rte_lpm_tbl24_val4(lpm, ip, tbl); /* get 4 indexes for tbl8[]. */ i8.x = _mm_and_si128(ip, mask8);