[dpdk-dev] [PATCH] fib6: improve lookup performance

[dpdk-dev] [PATCH] fib6: improve lookup performance

[Vladimir Medvedkin]

Improved performance for AVX512 FIB6 lookup by doubling the number
of flows being processed

Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
---
 lib/librte_fib/trie_avx512.c | 230 ++++++++++++++++++++++++++++---------------
 1 file changed, 153 insertions(+), 77 deletions(-)

diff --git a/lib/librte_fib/trie_avx512.c b/lib/librte_fib/trie_avx512.c
index b1c9e4e..d4d70d8 100644
--- a/lib/librte_fib/trie_avx512.c
+++ b/lib/librte_fib/trie_avx512.c
@@ -67,16 +67,22 @@ transpose_x8(uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
 }
 
 static __rte_always_inline void
-trie_vec_lookup_x16(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
+trie_vec_lookup_x16x2(void *p, uint8_t ips[32][RTE_FIB6_IPV6_ADDR_SIZE],
 	uint64_t *next_hops, int size)
 {
 	struct rte_trie_tbl *dp = (struct rte_trie_tbl *)p;
 	const __m512i zero = _mm512_set1_epi32(0);
 	const __m512i lsb = _mm512_set1_epi32(1);
 	const __m512i two_lsb = _mm512_set1_epi32(3);
-	__m512i first, second, third, fourth; /*< IPv6 four byte chunks */
-	__m512i idxes, res, shuf_idxes;
-	__m512i tmp, tmp2, bytes, byte_chunk, base_idxes;
+	/* IPv6 four byte chunks */
+	__m512i first_1, second_1, third_1, fourth_1;
+	__m512i first_2, second_2, third_2, fourth_2;
+	__m512i idxes_1, res_1;
+	__m512i idxes_2, res_2;
+	__m512i shuf_idxes;
+	__m512i tmp_1, tmp2_1, bytes_1, byte_chunk_1;
+	__m512i tmp_2, tmp2_2, bytes_2, byte_chunk_2;
+	__m512i base_idxes;
 	/* used to mask gather values if size is 2 (16 bit next hops) */
 	const __m512i res_msk = _mm512_set1_epi32(UINT16_MAX);
 	const __rte_x86_zmm_t bswap = {
@@ -92,29 +98,41 @@ trie_vec_lookup_x16(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
 	};
 	const __mmask64 k = 0x1111111111111111;
 	int i = 3;
-	__mmask16 msk_ext, new_msk;
+	__mmask16 msk_ext_1, new_msk_1;
+	__mmask16 msk_ext_2, new_msk_2;
 	__mmask16 exp_msk = 0x5555;
 
-	transpose_x16(ips, &first, &second, &third, &fourth);
+	transpose_x16(ips, &first_1, &second_1, &third_1, &fourth_1);
+	transpose_x16(ips + 16, &first_2, &second_2, &third_2, &fourth_2);
 
 	/* get_tbl24_idx() for every 4 byte chunk */
-	idxes = _mm512_shuffle_epi8(first, bswap.z);
+	idxes_1 = _mm512_shuffle_epi8(first_1, bswap.z);
+	idxes_2 = _mm512_shuffle_epi8(first_2, bswap.z);
 
 	/**
 	 * lookup in tbl24
 	 * Put it inside branch to make compiller happy with -O0
 	 */
 	if (size == sizeof(uint16_t)) {
-		res = _mm512_i32gather_epi32(idxes, (const int *)dp->tbl24, 2);
-		res = _mm512_and_epi32(res, res_msk);
-	} else
-		res = _mm512_i32gather_epi32(idxes, (const int *)dp->tbl24, 4);
-
+		res_1 = _mm512_i32gather_epi32(idxes_1,
+				(const int *)dp->tbl24, 2);
+		res_2 = _mm512_i32gather_epi32(idxes_2,
+				(const int *)dp->tbl24, 2);
+		res_1 = _mm512_and_epi32(res_1, res_msk);
+		res_2 = _mm512_and_epi32(res_2, res_msk);
+	} else {
+		res_1 = _mm512_i32gather_epi32(idxes_1,
+				(const int *)dp->tbl24, 4);
+		res_2 = _mm512_i32gather_epi32(idxes_2,
+				(const int *)dp->tbl24, 4);
+	}
 
 	/* get extended entries indexes */
-	msk_ext = _mm512_test_epi32_mask(res, lsb);
+	msk_ext_1 = _mm512_test_epi32_mask(res_1, lsb);
+	msk_ext_2 = _mm512_test_epi32_mask(res_2, lsb);
 
-	tmp = _mm512_srli_epi32(res, 1);
+	tmp_1 = _mm512_srli_epi32(res_1, 1);
+	tmp_2 = _mm512_srli_epi32(res_2, 1);
 
 	/* idxes to retrieve bytes */
 	shuf_idxes = _mm512_setr_epi32(3, 7, 11, 15,
@@ -128,24 +146,44 @@ trie_vec_lookup_x16(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
 				48, 52, 56, 60);
 
 	/* traverse down the trie */
-	while (msk_ext) {
-		idxes = _mm512_maskz_slli_epi32(msk_ext, tmp, 8);
-		byte_chunk = (i < 8) ?
-			((i >= 4) ? second : first) :
-			((i >= 12) ? fourth : third);
-		bytes = _mm512_maskz_shuffle_epi8(k, byte_chunk, shuf_idxes);
-		idxes = _mm512_maskz_add_epi32(msk_ext, idxes, bytes);
+	while (msk_ext_1 || msk_ext_2) {
+		idxes_1 = _mm512_maskz_slli_epi32(msk_ext_1, tmp_1, 8);
+		idxes_2 = _mm512_maskz_slli_epi32(msk_ext_2, tmp_2, 8);
+		byte_chunk_1 = (i < 8) ?
+			((i >= 4) ? second_1 : first_1) :
+			((i >= 12) ? fourth_1 : third_1);
+		byte_chunk_2 = (i < 8) ?
+			((i >= 4) ? second_2 : first_2) :
+			((i >= 12) ? fourth_2 : third_2);
+		bytes_1 = _mm512_maskz_shuffle_epi8(k, byte_chunk_1,
+				shuf_idxes);
+		bytes_2 = _mm512_maskz_shuffle_epi8(k, byte_chunk_2,
+				shuf_idxes);
+		idxes_1 = _mm512_maskz_add_epi32(msk_ext_1, idxes_1, bytes_1);
+		idxes_2 = _mm512_maskz_add_epi32(msk_ext_2, idxes_2, bytes_2);
 		if (size == sizeof(uint16_t)) {
-			tmp = _mm512_mask_i32gather_epi32(zero, msk_ext,
-				idxes, (const int *)dp->tbl8, 2);
-			tmp = _mm512_and_epi32(tmp, res_msk);
-		} else
-			tmp = _mm512_mask_i32gather_epi32(zero, msk_ext,
-				idxes, (const int *)dp->tbl8, 4);
-		new_msk = _mm512_test_epi32_mask(tmp, lsb);
-		res = _mm512_mask_blend_epi32(msk_ext ^ new_msk, res, tmp);
-		tmp = _mm512_srli_epi32(tmp, 1);
-		msk_ext = new_msk;
+			tmp_1 = _mm512_mask_i32gather_epi32(zero, msk_ext_1,
+				idxes_1, (const int *)dp->tbl8, 2);
+			tmp_2 = _mm512_mask_i32gather_epi32(zero, msk_ext_2,
+				idxes_2, (const int *)dp->tbl8, 2);
+			tmp_1 = _mm512_and_epi32(tmp_1, res_msk);
+			tmp_2 = _mm512_and_epi32(tmp_2, res_msk);
+		} else {
+			tmp_1 = _mm512_mask_i32gather_epi32(zero, msk_ext_1,
+				idxes_1, (const int *)dp->tbl8, 4);
+			tmp_2 = _mm512_mask_i32gather_epi32(zero, msk_ext_2,
+				idxes_2, (const int *)dp->tbl8, 4);
+		}
+		new_msk_1 = _mm512_test_epi32_mask(tmp_1, lsb);
+		new_msk_2 = _mm512_test_epi32_mask(tmp_2, lsb);
+		res_1 = _mm512_mask_blend_epi32(msk_ext_1 ^ new_msk_1, res_1,
+				tmp_1);
+		res_2 = _mm512_mask_blend_epi32(msk_ext_2 ^ new_msk_2, res_2,
+				tmp_2);
+		tmp_1 = _mm512_srli_epi32(tmp_1, 1);
+		tmp_2 = _mm512_srli_epi32(tmp_2, 1);
+		msk_ext_1 = new_msk_1;
+		msk_ext_2 = new_msk_2;
 
 		shuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, lsb);
 		shuf_idxes = _mm512_and_epi32(shuf_idxes, two_lsb);
@@ -153,27 +191,43 @@ trie_vec_lookup_x16(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
 		i++;
 	}
 
-	res = _mm512_srli_epi32(res, 1);
-	tmp = _mm512_maskz_expand_epi32(exp_msk, res);
-	__m256i tmp256;
-	tmp256 = _mm512_extracti32x8_epi32(res, 1);
-	tmp2 = _mm512_maskz_expand_epi32(exp_msk,
-		_mm512_castsi256_si512(tmp256));
-	_mm512_storeu_si512(next_hops, tmp);
-	_mm512_storeu_si512(next_hops + 8, tmp2);
+	/* get rid of 1 LSB, now we have HN in every epi32 */
+	res_1 = _mm512_srli_epi32(res_1, 1);
+	res_2 = _mm512_srli_epi32(res_2, 1);
+	/* extract first half of NH's each in epi64 chunk */
+	tmp_1 = _mm512_maskz_expand_epi32(exp_msk, res_1);
+	tmp_2 = _mm512_maskz_expand_epi32(exp_msk, res_2);
+	/* extract second half of NH's */
+	__m256i tmp256_1, tmp256_2;
+	tmp256_1 = _mm512_extracti32x8_epi32(res_1, 1);
+	tmp256_2 = _mm512_extracti32x8_epi32(res_2, 1);
+	tmp2_1 = _mm512_maskz_expand_epi32(exp_msk,
+		_mm512_castsi256_si512(tmp256_1));
+	tmp2_2 = _mm512_maskz_expand_epi32(exp_msk,
+		_mm512_castsi256_si512(tmp256_2));
+	/* return NH's from two sets of registers */
+	_mm512_storeu_si512(next_hops, tmp_1);
+	_mm512_storeu_si512(next_hops + 8, tmp2_1);
+	_mm512_storeu_si512(next_hops + 16, tmp_2);
+	_mm512_storeu_si512(next_hops + 24, tmp2_2);
 }
 
 static void
-trie_vec_lookup_x8_8b(void *p, uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
+trie_vec_lookup_x8x2_8b(void *p, uint8_t ips[16][RTE_FIB6_IPV6_ADDR_SIZE],
 	uint64_t *next_hops)
 {
 	struct rte_trie_tbl *dp = (struct rte_trie_tbl *)p;
 	const __m512i zero = _mm512_set1_epi32(0);
 	const __m512i lsb = _mm512_set1_epi32(1);
 	const __m512i three_lsb = _mm512_set1_epi32(7);
-	__m512i first, second; /*< IPv6 eight byte chunks */
-	__m512i idxes, res, shuf_idxes;
-	__m512i tmp, bytes, byte_chunk, base_idxes;
+	/* IPv6 eight byte chunks */
+	__m512i first_1, second_1;
+	__m512i first_2, second_2;
+	__m512i idxes_1, res_1;
+	__m512i idxes_2, res_2;
+	__m512i shuf_idxes, base_idxes;
+	__m512i tmp_1, bytes_1, byte_chunk_1;
+	__m512i tmp_2, bytes_2, byte_chunk_2;
 	const __rte_x86_zmm_t bswap = {
 		.u8 = { 2, 1, 0, 255, 255, 255, 255, 255,
 			10, 9, 8, 255, 255, 255, 255, 255,
@@ -187,19 +241,25 @@ trie_vec_lookup_x8_8b(void *p, uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
 	};
 	const __mmask64 k = 0x101010101010101;
 	int i = 3;
-	__mmask8 msk_ext, new_msk;
+	__mmask8 msk_ext_1, new_msk_1;
+	__mmask8 msk_ext_2, new_msk_2;
 
-	transpose_x8(ips, &first, &second);
+	transpose_x8(ips, &first_1, &second_1);
+	transpose_x8(ips + 8, &first_2, &second_2);
 
 	/* get_tbl24_idx() for every 4 byte chunk */
-	idxes = _mm512_shuffle_epi8(first, bswap.z);
+	idxes_1 = _mm512_shuffle_epi8(first_1, bswap.z);
+	idxes_2 = _mm512_shuffle_epi8(first_2, bswap.z);
 
 	/* lookup in tbl24 */
-	res = _mm512_i64gather_epi64(idxes, (const void *)dp->tbl24, 8);
+	res_1 = _mm512_i64gather_epi64(idxes_1, (const void *)dp->tbl24, 8);
+	res_2 = _mm512_i64gather_epi64(idxes_2, (const void *)dp->tbl24, 8);
 	/* get extended entries indexes */
-	msk_ext = _mm512_test_epi64_mask(res, lsb);
+	msk_ext_1 = _mm512_test_epi64_mask(res_1, lsb);
+	msk_ext_2 = _mm512_test_epi64_mask(res_2, lsb);
 
-	tmp = _mm512_srli_epi64(res, 1);
+	tmp_1 = _mm512_srli_epi64(res_1, 1);
+	tmp_2 = _mm512_srli_epi64(res_2, 1);
 
 	/* idxes to retrieve bytes */
 	shuf_idxes = _mm512_setr_epi64(3, 11, 19, 27, 35, 43, 51, 59);
@@ -207,17 +267,31 @@ trie_vec_lookup_x8_8b(void *p, uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
 	base_idxes = _mm512_setr_epi64(0, 8, 16, 24, 32, 40, 48, 56);
 
 	/* traverse down the trie */
-	while (msk_ext) {
-		idxes = _mm512_maskz_slli_epi64(msk_ext, tmp, 8);
-		byte_chunk = (i < 8) ? first : second;
-		bytes = _mm512_maskz_shuffle_epi8(k, byte_chunk, shuf_idxes);
-		idxes = _mm512_maskz_add_epi64(msk_ext, idxes, bytes);
-		tmp = _mm512_mask_i64gather_epi64(zero, msk_ext,
-				idxes, (const void *)dp->tbl8, 8);
-		new_msk = _mm512_test_epi64_mask(tmp, lsb);
-		res = _mm512_mask_blend_epi64(msk_ext ^ new_msk, res, tmp);
-		tmp = _mm512_srli_epi64(tmp, 1);
-		msk_ext = new_msk;
+	while (msk_ext_1 || msk_ext_2) {
+		idxes_1 = _mm512_maskz_slli_epi64(msk_ext_1, tmp_1, 8);
+		idxes_2 = _mm512_maskz_slli_epi64(msk_ext_2, tmp_2, 8);
+		byte_chunk_1 = (i < 8) ? first_1 : second_1;
+		byte_chunk_2 = (i < 8) ? first_2 : second_2;
+		bytes_1 = _mm512_maskz_shuffle_epi8(k, byte_chunk_1,
+				shuf_idxes);
+		bytes_2 = _mm512_maskz_shuffle_epi8(k, byte_chunk_2,
+				shuf_idxes);
+		idxes_1 = _mm512_maskz_add_epi64(msk_ext_1, idxes_1, bytes_1);
+		idxes_2 = _mm512_maskz_add_epi64(msk_ext_2, idxes_2, bytes_2);
+		tmp_1 = _mm512_mask_i64gather_epi64(zero, msk_ext_1,
+				idxes_1, (const void *)dp->tbl8, 8);
+		tmp_2 = _mm512_mask_i64gather_epi64(zero, msk_ext_2,
+				idxes_2, (const void *)dp->tbl8, 8);
+		new_msk_1 = _mm512_test_epi64_mask(tmp_1, lsb);
+		new_msk_2 = _mm512_test_epi64_mask(tmp_2, lsb);
+		res_1 = _mm512_mask_blend_epi64(msk_ext_1 ^ new_msk_1, res_1,
+				tmp_1);
+		res_2 = _mm512_mask_blend_epi64(msk_ext_2 ^ new_msk_2, res_2,
+				tmp_2);
+		tmp_1 = _mm512_srli_epi64(tmp_1, 1);
+		tmp_2 = _mm512_srli_epi64(tmp_2, 1);
+		msk_ext_1 = new_msk_1;
+		msk_ext_2 = new_msk_2;
 
 		shuf_idxes = _mm512_maskz_add_epi8(k, shuf_idxes, lsb);
 		shuf_idxes = _mm512_and_epi64(shuf_idxes, three_lsb);
@@ -225,8 +299,10 @@ trie_vec_lookup_x8_8b(void *p, uint8_t ips[8][RTE_FIB6_IPV6_ADDR_SIZE],
 		i++;
 	}
 
-	res = _mm512_srli_epi64(res, 1);
-	_mm512_storeu_si512(next_hops, res);
+	res_1 = _mm512_srli_epi64(res_1, 1);
+	res_2 = _mm512_srli_epi64(res_2, 1);
+	_mm512_storeu_si512(next_hops, res_1);
+	_mm512_storeu_si512(next_hops + 8, res_2);
 }
 
 void
@@ -234,12 +310,12 @@ rte_trie_vec_lookup_bulk_2b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
 	uint64_t *next_hops, const unsigned int n)
 {
 	uint32_t i;
-	for (i = 0; i < (n / 16); i++) {
-		trie_vec_lookup_x16(p, (uint8_t (*)[16])&ips[i * 16][0],
-				next_hops + i * 16, sizeof(uint16_t));
+	for (i = 0; i < (n / 32); i++) {
+		trie_vec_lookup_x16x2(p, (uint8_t (*)[16])&ips[i * 32][0],
+				next_hops + i * 32, sizeof(uint16_t));
 	}
-	rte_trie_lookup_bulk_2b(p, (uint8_t (*)[16])&ips[i * 16][0],
-			next_hops + i * 16, n - i * 16);
+	rte_trie_lookup_bulk_2b(p, (uint8_t (*)[16])&ips[i * 32][0],
+			next_hops + i * 32, n - i * 32);
 }
 
 void
@@ -247,12 +323,12 @@ rte_trie_vec_lookup_bulk_4b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
 	uint64_t *next_hops, const unsigned int n)
 {
 	uint32_t i;
-	for (i = 0; i < (n / 16); i++) {
-		trie_vec_lookup_x16(p, (uint8_t (*)[16])&ips[i * 16][0],
-				next_hops + i * 16, sizeof(uint32_t));
+	for (i = 0; i < (n / 32); i++) {
+		trie_vec_lookup_x16x2(p, (uint8_t (*)[16])&ips[i * 32][0],
+				next_hops + i * 32, sizeof(uint32_t));
 	}
-	rte_trie_lookup_bulk_4b(p, (uint8_t (*)[16])&ips[i * 16][0],
-			next_hops + i * 16, n - i * 16);
+	rte_trie_lookup_bulk_4b(p, (uint8_t (*)[16])&ips[i * 32][0],
+			next_hops + i * 32, n - i * 32);
 }
 
 void
@@ -260,10 +336,10 @@ rte_trie_vec_lookup_bulk_8b(void *p, uint8_t ips[][RTE_FIB6_IPV6_ADDR_SIZE],
 	uint64_t *next_hops, const unsigned int n)
 {
 	uint32_t i;
-	for (i = 0; i < (n / 8); i++) {
-		trie_vec_lookup_x8_8b(p, (uint8_t (*)[16])&ips[i * 8][0],
-				next_hops + i * 8);
+	for (i = 0; i < (n / 16); i++) {
+		trie_vec_lookup_x8x2_8b(p, (uint8_t (*)[16])&ips[i * 16][0],
+				next_hops + i * 16);
 	}
-	rte_trie_lookup_bulk_8b(p, (uint8_t (*)[16])&ips[i * 8][0],
-			next_hops + i * 8, n - i * 8);
+	rte_trie_lookup_bulk_8b(p, (uint8_t (*)[16])&ips[i * 16][0],
+			next_hops + i * 16, n - i * 16);
 }
-- 
2.7.4

Re: [dpdk-dev] [PATCH] fib6: improve lookup performance

[Ananyev, Konstantin]

> Improved performance for AVX512 FIB6 lookup by doubling the number
> of flows being processed
> 
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> ---
>  lib/librte_fib/trie_avx512.c | 230 ++++++++++++++++++++++++++++---------------
>  1 file changed, 153 insertions(+), 77 deletions(-)
> 

Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>

> --
> 2.7.4

Re: [dpdk-dev] [PATCH] fib6: improve lookup performance

[Kinsella, Ray]

On 15/12/2020 18:25, Vladimir Medvedkin wrote:
> Improved performance for AVX512 FIB6 lookup by doubling the number
> of flows being processed
> 
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> ---
Acked-by: Ray Kinsella <mdr@ashroe.eu>

Re: [dpdk-dev] [PATCH] fib6: improve lookup performance

[Thomas Monjalon]

12/01/2021 15:13, Ananyev, Konstantin:
> 
> > Improved performance for AVX512 FIB6 lookup by doubling the number
> > of flows being processed
> > 
> > Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> 
> Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>

Applied, thanks

PS: added "AVX512" in the title