To enhance the per-core performance, this patch adds some AVX512 instructions to the data path to handle the legacy RX descriptors.
Signed-off-by: Wenzhuo Lu <wenzhuo...@intel.com> Signed-off-by: Bruce Richardson <bruce.richard...@intel.com> Signed-off-by: Leyi Rong <leyi.r...@intel.com> --- drivers/net/iavf/iavf_rxtx.c | 28 +- drivers/net/iavf/iavf_rxtx.h | 5 + drivers/net/iavf/iavf_rxtx_vec_avx512.c | 691 ++++++++++++++++++++++++++++++++ drivers/net/iavf/meson.build | 17 + 4 files changed, 737 insertions(+), 4 deletions(-) create mode 100644 drivers/net/iavf/iavf_rxtx_vec_avx512.c diff --git a/drivers/net/iavf/iavf_rxtx.c b/drivers/net/iavf/iavf_rxtx.c index edb2dc3..0067b64 100644 --- a/drivers/net/iavf/iavf_rxtx.c +++ b/drivers/net/iavf/iavf_rxtx.c @@ -2104,6 +2104,9 @@ struct iavf_rx_queue *rxq; int i; bool use_avx2 = false; +#ifdef CC_AVX512_SUPPORT + bool use_avx512 = false; +#endif if (!iavf_rx_vec_dev_check(dev) && rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_128) { @@ -2116,6 +2119,11 @@ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) && rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_256) use_avx2 = true; +#ifdef CC_AVX512_SUPPORT + if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1 && + rte_vect_get_max_simd_bitwidth() >= RTE_VECT_SIMD_512) + use_avx512 = true; +#endif if (dev->data->scattered_rx) { PMD_DRV_LOG(DEBUG, @@ -2123,27 +2131,39 @@ use_avx2 ? "avx2 " : "", dev->data->port_id); if (vf->vf_res->vf_cap_flags & - VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) + VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) { dev->rx_pkt_burst = use_avx2 ? iavf_recv_scattered_pkts_vec_avx2_flex_rxd : iavf_recv_scattered_pkts_vec_flex_rxd; - else + } else { dev->rx_pkt_burst = use_avx2 ? iavf_recv_scattered_pkts_vec_avx2 : iavf_recv_scattered_pkts_vec; +#ifdef CC_AVX512_SUPPORT + if (use_avx512) + dev->rx_pkt_burst = + iavf_recv_scattered_pkts_vec_avx512; +#endif + } } else { PMD_DRV_LOG(DEBUG, "Using %sVector Rx (port %d).", use_avx2 ? "avx2 " : "", dev->data->port_id); if (vf->vf_res->vf_cap_flags & - VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) + VIRTCHNL_VF_OFFLOAD_RX_FLEX_DESC) { dev->rx_pkt_burst = use_avx2 ? iavf_recv_pkts_vec_avx2_flex_rxd : iavf_recv_pkts_vec_flex_rxd; - else + } else { dev->rx_pkt_burst = use_avx2 ? iavf_recv_pkts_vec_avx2 : iavf_recv_pkts_vec; +#ifdef CC_AVX512_SUPPORT + if (use_avx512) + dev->rx_pkt_burst = + iavf_recv_pkts_vec_avx512; +#endif + } } return; diff --git a/drivers/net/iavf/iavf_rxtx.h b/drivers/net/iavf/iavf_rxtx.h index 3d02c65..7c1f05f 100644 --- a/drivers/net/iavf/iavf_rxtx.h +++ b/drivers/net/iavf/iavf_rxtx.h @@ -438,6 +438,11 @@ uint16_t iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, int iavf_tx_vec_dev_check(struct rte_eth_dev *dev); int iavf_rxq_vec_setup(struct iavf_rx_queue *rxq); int iavf_txq_vec_setup(struct iavf_tx_queue *txq); +uint16_t iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts); +uint16_t iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts); const uint32_t *iavf_get_default_ptype_table(void); diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx512.c b/drivers/net/iavf/iavf_rxtx_vec_avx512.c new file mode 100644 index 0000000..959067c --- /dev/null +++ b/drivers/net/iavf/iavf_rxtx_vec_avx512.c @@ -0,0 +1,691 @@ +/* SPDX-License-Identifier: BSD-3-Clause + * Copyright(c) 2020 Intel Corporation + */ + +#include "iavf_rxtx_vec_common.h" + +#include <x86intrin.h> + +#ifndef __INTEL_COMPILER +#pragma GCC diagnostic ignored "-Wcast-qual" +#endif + +#define IAVF_DESCS_PER_LOOP_AVX 8 +#define PKTLEN_SHIFT 10 + +static inline void +iavf_rxq_rearm(struct iavf_rx_queue *rxq) +{ + int i; + uint16_t rx_id; + volatile union iavf_rx_desc *rxdp; + struct rte_mempool_cache *cache = + rte_mempool_default_cache(rxq->mp, rte_lcore_id()); + struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start]; + + rxdp = rxq->rx_ring + rxq->rxrearm_start; + + /* We need to pull 'n' more MBUFs into the software ring from mempool + * We inline the mempool function here, so we can vectorize the copy + * from the cache into the shadow ring. + */ + + /* Can this be satisfied from the cache? */ + if (cache->len < IAVF_RXQ_REARM_THRESH) { + /* No. Backfill the cache first, and then fill from it */ + uint32_t req = IAVF_RXQ_REARM_THRESH + (cache->size - + cache->len); + + /* How many do we require i.e. number to fill the cache + the request */ + int ret = rte_mempool_ops_dequeue_bulk + (rxq->mp, &cache->objs[cache->len], req); + if (ret == 0) { + cache->len += req; + } else { + if (rxq->rxrearm_nb + IAVF_RXQ_REARM_THRESH >= + rxq->nb_rx_desc) { + __m128i dma_addr0; + + dma_addr0 = _mm_setzero_si128(); + for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) { + rxp[i] = &rxq->fake_mbuf; + _mm_storeu_si128((__m128i *)&rxdp[i].read, + dma_addr0); + } + } + rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed += + IAVF_RXQ_REARM_THRESH; + return; + } + } + + const __m512i iova_offsets = _mm512_set1_epi64(offsetof + (struct rte_mbuf, buf_iova)); + const __m512i headroom = _mm512_set1_epi64(RTE_PKTMBUF_HEADROOM); + +#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC + /* to shuffle the addresses to correct slots. Values 4-7 will contain + * zeros, so use 7 for a zero-value. + */ + const __m512i permute_idx = _mm512_set_epi64(7, 7, 3, 1, 7, 7, 2, 0); +#else + const __m512i permute_idx = _mm512_set_epi64(7, 3, 6, 2, 5, 1, 4, 0); +#endif + + /* Initialize the mbufs in vector, process 8 mbufs in one loop, taking + * from mempool cache and populating both shadow and HW rings + */ + for (i = 0; i < IAVF_RXQ_REARM_THRESH / IAVF_DESCS_PER_LOOP_AVX; i++) { + const __m512i mbuf_ptrs = _mm512_loadu_si512 + (&cache->objs[cache->len - IAVF_DESCS_PER_LOOP_AVX]); + _mm512_storeu_si512(rxp, mbuf_ptrs); + + const __m512i iova_base_addrs = _mm512_i64gather_epi64 + (_mm512_add_epi64(mbuf_ptrs, iova_offsets), + 0, /* base */ + 1 /* scale */); + const __m512i iova_addrs = _mm512_add_epi64(iova_base_addrs, + headroom); +#ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC + const __m512i iovas0 = _mm512_castsi256_si512 + (_mm512_extracti64x4_epi64(iova_addrs, 0)); + const __m512i iovas1 = _mm512_castsi256_si512 + (_mm512_extracti64x4_epi64(iova_addrs, 1)); + + /* permute leaves desc 2-3 addresses in header address slots 0-1 + * but these are ignored by driver since header split not + * enabled. Similarly for desc 6 & 7. + */ + const __m512i desc0_1 = _mm512_permutexvar_epi64 + (permute_idx, + iovas0); + const __m512i desc2_3 = _mm512_bsrli_epi128(desc0_1, 8); + + const __m512i desc4_5 = _mm512_permutexvar_epi64 + (permute_idx, + iovas1); + const __m512i desc6_7 = _mm512_bsrli_epi128(desc4_5, 8); + + _mm512_storeu_si512((void *)rxdp, desc0_1); + _mm512_storeu_si512((void *)(rxdp + 2), desc2_3); + _mm512_storeu_si512((void *)(rxdp + 4), desc4_5); + _mm512_storeu_si512((void *)(rxdp + 6), desc6_7); +#else + /* permute leaves desc 4-7 addresses in header address slots 0-3 + * but these are ignored by driver since header split not + * enabled. + */ + const __m512i desc0_3 = _mm512_permutexvar_epi64(permute_idx, + iova_addrs); + const __m512i desc4_7 = _mm512_bsrli_epi128(desc0_3, 8); + + _mm512_storeu_si512((void *)rxdp, desc0_3); + _mm512_storeu_si512((void *)(rxdp + 4), desc4_7); +#endif + rxp += IAVF_DESCS_PER_LOOP_AVX; + rxdp += IAVF_DESCS_PER_LOOP_AVX; + cache->len -= IAVF_DESCS_PER_LOOP_AVX; + } + + rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH; + if (rxq->rxrearm_start >= rxq->nb_rx_desc) + rxq->rxrearm_start = 0; + + rxq->rxrearm_nb -= IAVF_RXQ_REARM_THRESH; + + rx_id = (uint16_t)((rxq->rxrearm_start == 0) ? + (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1)); + + /* Update the tail pointer on the NIC */ + IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id); +} + +#define IAVF_RX_LEN_MASK 0x80808080 +static inline uint16_t +_iavf_recv_raw_pkts_vec_avx512(struct iavf_rx_queue *rxq, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts, uint8_t *split_packet) +{ + const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl; + + const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0, + rxq->mbuf_initializer); + struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail]; + volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail; + + rte_prefetch0(rxdp); + + /* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */ + nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_DESCS_PER_LOOP_AVX); + + /* See if we need to rearm the RX queue - gives the prefetch a bit + * of time to act + */ + if (rxq->rxrearm_nb > IAVF_RXQ_REARM_THRESH) + iavf_rxq_rearm(rxq); + + /* Before we start moving massive data around, check to see if + * there is actually a packet available + */ + if (!(rxdp->wb.qword1.status_error_len & + rte_cpu_to_le_32(1 << IAVF_RX_DESC_STATUS_DD_SHIFT))) + return 0; + + /* constants used in processing loop */ + const __m512i crc_adjust = + _mm512_set_epi32 + (/* 1st descriptor */ + 0, /* ignore non-length fields */ + -rxq->crc_len, /* sub crc on data_len */ + -rxq->crc_len, /* sub crc on pkt_len */ + 0, /* ignore pkt_type field */ + /* 2nd descriptor */ + 0, /* ignore non-length fields */ + -rxq->crc_len, /* sub crc on data_len */ + -rxq->crc_len, /* sub crc on pkt_len */ + 0, /* ignore pkt_type field */ + /* 3rd descriptor */ + 0, /* ignore non-length fields */ + -rxq->crc_len, /* sub crc on data_len */ + -rxq->crc_len, /* sub crc on pkt_len */ + 0, /* ignore pkt_type field */ + /* 4th descriptor */ + 0, /* ignore non-length fields */ + -rxq->crc_len, /* sub crc on data_len */ + -rxq->crc_len, /* sub crc on pkt_len */ + 0 /* ignore pkt_type field */ + ); + + /* 8 packets DD mask, LSB in each 32-bit value */ + const __m256i dd_check = _mm256_set1_epi32(1); + + /* 8 packets EOP mask, second-LSB in each 32-bit value */ + const __m256i eop_check = _mm256_slli_epi32(dd_check, + IAVF_RX_DESC_STATUS_EOF_SHIFT); + + /* mask to shuffle from desc. to mbuf (4 descriptors)*/ + const __m512i shuf_msk = + _mm512_set_epi32 + (/* 1st descriptor */ + 0x07060504, /* octet 4~7, 32bits rss */ + 0x03020F0E, /* octet 2~3, low 16 bits vlan_macip */ + /* octet 15~14, 16 bits data_len */ + 0xFFFF0F0E, /* skip high 16 bits pkt_len, zero out */ + /* octet 15~14, low 16 bits pkt_len */ + 0xFFFFFFFF, /* pkt_type set as unknown */ + /* 2nd descriptor */ + 0x07060504, /* octet 4~7, 32bits rss */ + 0x03020F0E, /* octet 2~3, low 16 bits vlan_macip */ + /* octet 15~14, 16 bits data_len */ + 0xFFFF0F0E, /* skip high 16 bits pkt_len, zero out */ + /* octet 15~14, low 16 bits pkt_len */ + 0xFFFFFFFF, /* pkt_type set as unknown */ + /* 3rd descriptor */ + 0x07060504, /* octet 4~7, 32bits rss */ + 0x03020F0E, /* octet 2~3, low 16 bits vlan_macip */ + /* octet 15~14, 16 bits data_len */ + 0xFFFF0F0E, /* skip high 16 bits pkt_len, zero out */ + /* octet 15~14, low 16 bits pkt_len */ + 0xFFFFFFFF, /* pkt_type set as unknown */ + /* 4th descriptor */ + 0x07060504, /* octet 4~7, 32bits rss */ + 0x03020F0E, /* octet 2~3, low 16 bits vlan_macip */ + /* octet 15~14, 16 bits data_len */ + 0xFFFF0F0E, /* skip high 16 bits pkt_len, zero out */ + /* octet 15~14, low 16 bits pkt_len */ + 0xFFFFFFFF /* pkt_type set as unknown */ + ); + /** + * compile-time check the above crc and shuffle layout is correct. + * NOTE: the first field (lowest address) is given last in set_epi + * calls above. + */ + RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) != + offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4); + RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) != + offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8); + RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) != + offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10); + RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) != + offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12); + + /* Status/Error flag masks */ + /** + * mask everything except RSS, flow director and VLAN flags + * bit2 is for VLAN tag, bit11 for flow director indication + * bit13:12 for RSS indication. Bits 3-5 of error + * field (bits 22-24) are for IP/L4 checksum errors + */ + const __m256i flags_mask = + _mm256_set1_epi32((1 << 2) | (1 << 11) | + (3 << 12) | (7 << 22)); + /** + * data to be shuffled by result of flag mask. If VLAN bit is set, + * (bit 2), then position 4 in this array will be used in the + * destination + */ + const __m256i vlan_flags_shuf = + _mm256_set_epi32(0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0, + 0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0); + /** + * data to be shuffled by result of flag mask, shifted down 11. + * If RSS/FDIR bits are set, shuffle moves appropriate flags in + * place. + */ + const __m256i rss_flags_shuf = + _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, + PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, + 0, 0, 0, 0, PKT_RX_FDIR, 0,/* end up 128-bits */ + 0, 0, 0, 0, 0, 0, 0, 0, + PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, + 0, 0, 0, 0, PKT_RX_FDIR, 0); + + /** + * data to be shuffled by the result of the flags mask shifted by 22 + * bits. This gives use the l3_l4 flags. + */ + const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, + /* shift right 1 bit to make sure it not exceed 255 */ + (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD) >> 1, + (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1, + (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1, + PKT_RX_IP_CKSUM_BAD >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1, + /* second 128-bits */ + 0, 0, 0, 0, 0, 0, 0, 0, + (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD) >> 1, + (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1, + (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1, + PKT_RX_IP_CKSUM_BAD >> 1, + (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1); + + const __m256i cksum_mask = + _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD | + PKT_RX_EIP_CKSUM_BAD); + + uint16_t i, received; + + for (i = 0, received = 0; i < nb_pkts; + i += IAVF_DESCS_PER_LOOP_AVX, + rxdp += IAVF_DESCS_PER_LOOP_AVX) { + /* step 1, copy over 8 mbuf pointers to rx_pkts array */ + _mm256_storeu_si256((void *)&rx_pkts[i], + _mm256_loadu_si256((void *)&sw_ring[i])); +#ifdef RTE_ARCH_X86_64 + _mm256_storeu_si256 + ((void *)&rx_pkts[i + 4], + _mm256_loadu_si256((void *)&sw_ring[i + 4])); +#endif + + __m512i raw_desc0_3, raw_desc4_7; + const __m128i raw_desc7 = + _mm_load_si128((void *)(rxdp + 7)); + rte_compiler_barrier(); + const __m128i raw_desc6 = + _mm_load_si128((void *)(rxdp + 6)); + rte_compiler_barrier(); + const __m128i raw_desc5 = + _mm_load_si128((void *)(rxdp + 5)); + rte_compiler_barrier(); + const __m128i raw_desc4 = + _mm_load_si128((void *)(rxdp + 4)); + rte_compiler_barrier(); + const __m128i raw_desc3 = + _mm_load_si128((void *)(rxdp + 3)); + rte_compiler_barrier(); + const __m128i raw_desc2 = + _mm_load_si128((void *)(rxdp + 2)); + rte_compiler_barrier(); + const __m128i raw_desc1 = + _mm_load_si128((void *)(rxdp + 1)); + rte_compiler_barrier(); + const __m128i raw_desc0 = + _mm_load_si128((void *)(rxdp + 0)); + + raw_desc4_7 = _mm512_broadcast_i32x4(raw_desc4); + raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc5, 1); + raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc6, 2); + raw_desc4_7 = _mm512_inserti32x4(raw_desc4_7, raw_desc7, 3); + raw_desc0_3 = _mm512_broadcast_i32x4(raw_desc0); + raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc1, 1); + raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc2, 2); + raw_desc0_3 = _mm512_inserti32x4(raw_desc0_3, raw_desc3, 3); + + if (split_packet) { + int j; + + for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++) + rte_mbuf_prefetch_part2(rx_pkts[i + j]); + } + + /** + * convert descriptors 4-7 into mbufs, adjusting length and + * re-arranging fields. Then write into the mbuf + */ + const __m512i len4_7 = _mm512_slli_epi32(raw_desc4_7, + PKTLEN_SHIFT); + const __m512i desc4_7 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK, + raw_desc4_7, + len4_7); + __m512i mb4_7 = _mm512_shuffle_epi8(desc4_7, shuf_msk); + + mb4_7 = _mm512_add_epi16(mb4_7, crc_adjust); + /** + * to get packet types, shift 64-bit values down 30 bits + * and so ptype is in lower 8-bits in each + */ + const __m512i ptypes4_7 = _mm512_srli_epi64(desc4_7, 30); + const __m256i ptypes6_7 = _mm512_extracti64x4_epi64(ptypes4_7, 1); + const __m256i ptypes4_5 = _mm512_extracti64x4_epi64(ptypes4_7, 0); + const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 24); + const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 8); + const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 24); + const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 8); + + const __m512i ptype4_7 = _mm512_set_epi32 + (0, 0, 0, type_table[ptype7], + 0, 0, 0, type_table[ptype6], + 0, 0, 0, type_table[ptype5], + 0, 0, 0, type_table[ptype4]); + mb4_7 = _mm512_mask_blend_epi32(0x1111, mb4_7, ptype4_7); + + /** + * convert descriptors 0-3 into mbufs, adjusting length and + * re-arranging fields. Then write into the mbuf + */ + const __m512i len0_3 = _mm512_slli_epi32(raw_desc0_3, + PKTLEN_SHIFT); + const __m512i desc0_3 = _mm512_mask_blend_epi16(IAVF_RX_LEN_MASK, + raw_desc0_3, + len0_3); + __m512i mb0_3 = _mm512_shuffle_epi8(desc0_3, shuf_msk); + + mb0_3 = _mm512_add_epi16(mb0_3, crc_adjust); + /* get the packet types */ + const __m512i ptypes0_3 = _mm512_srli_epi64(desc0_3, 30); + const __m256i ptypes2_3 = _mm512_extracti64x4_epi64(ptypes0_3, 1); + const __m256i ptypes0_1 = _mm512_extracti64x4_epi64(ptypes0_3, 0); + const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 24); + const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 8); + const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 24); + const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 8); + + const __m512i ptype0_3 = _mm512_set_epi32 + (0, 0, 0, type_table[ptype3], + 0, 0, 0, type_table[ptype2], + 0, 0, 0, type_table[ptype1], + 0, 0, 0, type_table[ptype0]); + mb0_3 = _mm512_mask_blend_epi32(0x1111, mb0_3, ptype0_3); + + /** + * use permute/extract to get status content + * After the operations, the packets status flags are in the + * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6] + */ + /* merge the status bits into one register */ + const __m512i status_permute_msk = _mm512_set_epi32 + (0, 0, 0, 0, + 0, 0, 0, 0, + 22, 30, 6, 14, + 18, 26, 2, 10); + const __m512i raw_status0_7 = _mm512_permutex2var_epi32 + (raw_desc4_7, status_permute_msk, raw_desc0_3); + __m256i status0_7 = _mm512_extracti64x4_epi64 + (raw_status0_7, 0); + + /* now do flag manipulation */ + + /* get only flag/error bits we want */ + const __m256i flag_bits = + _mm256_and_si256(status0_7, flags_mask); + /* set vlan and rss flags */ + const __m256i vlan_flags = + _mm256_shuffle_epi8(vlan_flags_shuf, flag_bits); + const __m256i rss_flags = + _mm256_shuffle_epi8(rss_flags_shuf, + _mm256_srli_epi32(flag_bits, 11)); + /** + * l3_l4_error flags, shuffle, then shift to correct adjustment + * of flags in flags_shuf, and finally mask out extra bits + */ + __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf, + _mm256_srli_epi32(flag_bits, 22)); + l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1); + l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask); + + /* merge flags */ + const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags, + _mm256_or_si256(rss_flags, vlan_flags)); + /** + * At this point, we have the 8 sets of flags in the low 16-bits + * of each 32-bit value in vlan0. + * We want to extract these, and merge them with the mbuf init + * data so we can do a single write to the mbuf to set the flags + * and all the other initialization fields. Extracting the + * appropriate flags means that we have to do a shift and blend + * for each mbuf before we do the write. However, we can also + * add in the previously computed rx_descriptor fields to + * make a single 256-bit write per mbuf + */ + /* check the structure matches expectations */ + RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) != + offsetof(struct rte_mbuf, rearm_data) + 8); + RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) != + RTE_ALIGN(offsetof(struct rte_mbuf, + rearm_data), + 16)); + /* build up data and do writes */ + __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5, + rearm6, rearm7; + const __m256i mb4_5 = _mm512_extracti64x4_epi64(mb4_7, 0); + const __m256i mb6_7 = _mm512_extracti64x4_epi64(mb4_7, 1); + const __m256i mb0_1 = _mm512_extracti64x4_epi64(mb0_3, 0); + const __m256i mb2_3 = _mm512_extracti64x4_epi64(mb0_3, 1); + + rearm6 = _mm256_blend_epi32(mbuf_init, + _mm256_slli_si256(mbuf_flags, 8), + 0x04); + rearm4 = _mm256_blend_epi32(mbuf_init, + _mm256_slli_si256(mbuf_flags, 4), + 0x04); + rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04); + rearm0 = _mm256_blend_epi32(mbuf_init, + _mm256_srli_si256(mbuf_flags, 4), + 0x04); + /* permute to add in the rx_descriptor e.g. rss fields */ + rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20); + rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20); + rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20); + rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20); + /* write to mbuf */ + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data, + rearm6); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data, + rearm4); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data, + rearm2); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data, + rearm0); + + /* repeat for the odd mbufs */ + const __m256i odd_flags = + _mm256_castsi128_si256 + (_mm256_extracti128_si256(mbuf_flags, 1)); + rearm7 = _mm256_blend_epi32(mbuf_init, + _mm256_slli_si256(odd_flags, 8), + 0x04); + rearm5 = _mm256_blend_epi32(mbuf_init, + _mm256_slli_si256(odd_flags, 4), + 0x04); + rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04); + rearm1 = _mm256_blend_epi32(mbuf_init, + _mm256_srli_si256(odd_flags, 4), + 0x04); + /* since odd mbufs are already in hi 128-bits use blend */ + rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0); + rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0); + rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0); + rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0); + /* again write to mbufs */ + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data, + rearm7); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data, + rearm5); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data, + rearm3); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data, + rearm1); + + /* extract and record EOP bit */ + if (split_packet) { + const __m128i eop_mask = + _mm_set1_epi16(1 << IAVF_RX_DESC_STATUS_EOF_SHIFT); + const __m256i eop_bits256 = _mm256_and_si256(status0_7, + eop_check); + /* pack status bits into a single 128-bit register */ + const __m128i eop_bits = + _mm_packus_epi32 + (_mm256_castsi256_si128(eop_bits256), + _mm256_extractf128_si256(eop_bits256, + 1)); + /** + * flip bits, and mask out the EOP bit, which is now + * a split-packet bit i.e. !EOP, rather than EOP one. + */ + __m128i split_bits = _mm_andnot_si128(eop_bits, + eop_mask); + /** + * eop bits are out of order, so we need to shuffle them + * back into order again. In doing so, only use low 8 + * bits, which acts like another pack instruction + * The original order is (hi->lo): 1,3,5,7,0,2,4,6 + * [Since we use epi8, the 16-bit positions are + * multiplied by 2 in the eop_shuffle value.] + */ + __m128i eop_shuffle = + _mm_set_epi8(/* zero hi 64b */ + 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, + /* move values to lo 64b */ + 8, 0, 10, 2, + 12, 4, 14, 6); + split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle); + *(uint64_t *)split_packet = + _mm_cvtsi128_si64(split_bits); + split_packet += IAVF_DESCS_PER_LOOP_AVX; + } + + /* perform dd_check */ + status0_7 = _mm256_and_si256(status0_7, dd_check); + status0_7 = _mm256_packs_epi32(status0_7, + _mm256_setzero_si256()); + + uint64_t burst = __builtin_popcountll + (_mm_cvtsi128_si64 + (_mm256_extracti128_si256 + (status0_7, 1))); + burst += __builtin_popcountll + (_mm_cvtsi128_si64 + (_mm256_castsi256_si128(status0_7))); + received += burst; + if (burst != IAVF_DESCS_PER_LOOP_AVX) + break; + } + + /* update tail pointers */ + rxq->rx_tail += received; + rxq->rx_tail &= (rxq->nb_rx_desc - 1); + if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep aligned */ + rxq->rx_tail--; + received--; + } + rxq->rxrearm_nb += received; + return received; +} + +/** + * Notice: + * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet + */ +uint16_t +iavf_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + return _iavf_recv_raw_pkts_vec_avx512(rx_queue, rx_pkts, nb_pkts, NULL); +} + +/** + * vPMD receive routine that reassembles single burst of 32 scattered packets + * Notice: + * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet + */ +static uint16_t +iavf_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + struct iavf_rx_queue *rxq = rx_queue; + uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0}; + + /* get some new buffers */ + uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx512(rxq, rx_pkts, nb_pkts, + split_flags); + if (nb_bufs == 0) + return 0; + + /* happy day case, full burst + no packets to be joined */ + const uint64_t *split_fl64 = (uint64_t *)split_flags; + + if (!rxq->pkt_first_seg && + split_fl64[0] == 0 && split_fl64[1] == 0 && + split_fl64[2] == 0 && split_fl64[3] == 0) + return nb_bufs; + + /* reassemble any packets that need reassembly*/ + unsigned int i = 0; + + if (!rxq->pkt_first_seg) { + /* find the first split flag, and only reassemble then*/ + while (i < nb_bufs && !split_flags[i]) + i++; + if (i == nb_bufs) + return nb_bufs; + rxq->pkt_first_seg = rx_pkts[i]; + } + return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i, + &split_flags[i]); +} + +/** + * vPMD receive routine that reassembles scattered packets. + * Main receive routine that can handle arbitrary burst sizes + * Notice: + * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet + */ +uint16_t +iavf_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + uint16_t retval = 0; + + while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) { + uint16_t burst = iavf_recv_scattered_burst_vec_avx512(rx_queue, + rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST); + retval += burst; + nb_pkts -= burst; + if (burst < IAVF_VPMD_RX_MAX_BURST) + return retval; + } + return retval + iavf_recv_scattered_burst_vec_avx512(rx_queue, + rx_pkts + retval, nb_pkts); +} diff --git a/drivers/net/iavf/meson.build b/drivers/net/iavf/meson.build index 33407c5..fd33b72 100644 --- a/drivers/net/iavf/meson.build +++ b/drivers/net/iavf/meson.build @@ -34,4 +34,21 @@ if arch_subdir == 'x86' c_args: [cflags, '-mavx2']) objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c') endif + + if not machine_args.contains('-mno-avx512f') + if dpdk_conf.has('RTE_MACHINE_CPUFLAG_AVX512F') + cflags += ['-DCC_AVX512_SUPPORT'] + cflags += ['-march=skylake-avx512'] + sources += files('iavf_rxtx_vec_avx512.c') + elif cc.has_argument('-mavx512f') + cflags += ['-DCC_AVX512_SUPPORT'] + iavf_avx512_lib = static_library('iavf_avx512_lib', + 'iavf_rxtx_vec_avx512.c', + dependencies: [static_rte_ethdev, + static_rte_kvargs, static_rte_hash], + include_directories: includes, + c_args: [cflags, '-mavx512f', '-march=skylake-avx512']) + objs += iavf_avx512_lib.extract_objects('iavf_rxtx_vec_avx512.c') + endif + endif endif -- 1.9.3