From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id CDDE448960; Fri, 17 Oct 2025 12:34:25 +0200 (CEST) Received: from mails.dpdk.org (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 7B6DB42D27; Fri, 17 Oct 2025 12:34:21 +0200 (CEST) Received: from mgamail.intel.com (mgamail.intel.com [192.198.163.18]) by mails.dpdk.org (Postfix) with ESMTP id D32E140269 for ; Fri, 17 Oct 2025 12:34:19 +0200 (CEST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1760697260; x=1792233260; h=from:to:cc:subject:date:message-id:in-reply-to: references:mime-version:content-transfer-encoding; bh=17i6O1X6t57B4o49tZyf44nQ1JTBwjMMA7iPelgL9s0=; b=FfiqToJTSb1QFqpAy2l2OaIm8mLi5lnym769Lu8qullAISMotcg9lyEj /tgNSrLG8BgJAEiGptOJBEe9odaeZxKV8XIjh2ihOFH9pV94GaTsP0m51 tjGGgsqnu+Z4DIzYDtfJlXhyUdTtgY16kE9Zj5c439d7JuY4qJe+AGKah vsUm5zCLtQ13B+DMGf0rOJlqJapq24fHsaZFFPWSblTR01HBnDUr/4IY6 TCrtMC2F0S7LL04kif1Zuh7B6WAuQnX4PEeDpot7A49Tq42KzLfOizkUV YVVvX+K5RbiqrChQWyUa/GBLX8DR0Kq7LM4Z5f52wyc1S3mXuJw3Ng3vE w==; X-CSE-ConnectionGUID: 0JpY5xMeQiyVKY6pGIoLAA== X-CSE-MsgGUID: 2a/nr/wsT3KIrQ3rcR1h7A== X-IronPort-AV: E=McAfee;i="6800,10657,11584"; a="62115566" X-IronPort-AV: E=Sophos;i="6.19,236,1754982000"; d="scan'208";a="62115566" Received: from fmviesa001.fm.intel.com ([10.60.135.141]) by fmvoesa112.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 17 Oct 2025 03:34:19 -0700 X-CSE-ConnectionGUID: 0ShftJvAS/6wh3EvUU/ytA== X-CSE-MsgGUID: e5tnQxqCTOGcFLHGG2KINQ== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="6.19,236,1754982000"; d="scan'208";a="213671862" Received: from unknown (HELO srv24..) ([10.138.182.231]) by fmviesa001.fm.intel.com with ESMTP; 17 Oct 2025 03:34:18 -0700 From: Shaiq Wani To: dev@dpdk.org, bruce.richardson@intel.com Cc: aman.deep.singh@intel.com Subject: [PATCH v7 1/3] net/idpf: enable AVX2 for split queue Rx Date: Fri, 17 Oct 2025 16:04:13 +0530 Message-Id: <20251017103415.114400-2-shaiq.wani@intel.com> X-Mailer: git-send-email 2.34.1 In-Reply-To: <20251017103415.114400-1-shaiq.wani@intel.com> References: <20250917052658.582872-1-shaiq.wani@intel.com> <20251017103415.114400-1-shaiq.wani@intel.com> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org In case some CPUs don't support AVX512. Enable AVX2 for them to get better per-core performance. In the single queue model, the same descriptor queue is used by SW to post descriptors to the device and used by device to report completed descriptors to SW. While as the split queue model separates them into different queues for parallel processing and improved performance. Signed-off-by: Shaiq Wani --- drivers/net/intel/idpf/idpf_common_device.h | 1 + drivers/net/intel/idpf/idpf_common_rxtx.c | 59 +++++++++ drivers/net/intel/idpf/idpf_common_rxtx.h | 5 + .../net/intel/idpf/idpf_common_rxtx_avx2.c | 123 ++++++++++++++++++ .../net/intel/idpf/idpf_common_rxtx_avx512.c | 56 -------- 5 files changed, 188 insertions(+), 56 deletions(-) diff --git a/drivers/net/intel/idpf/idpf_common_device.h b/drivers/net/intel/idpf/idpf_common_device.h index b12bb8d9fc..fc911bf2e5 100644 --- a/drivers/net/intel/idpf/idpf_common_device.h +++ b/drivers/net/intel/idpf/idpf_common_device.h @@ -52,6 +52,7 @@ enum idpf_rx_func_type { IDPF_RX_SINGLEQ, IDPF_RX_SINGLEQ_SCATTERED, IDPF_RX_SINGLEQ_AVX2, + IDPF_RX_AVX2, IDPF_RX_AVX512, IDPF_RX_SINGLEQ_AVX512, IDPF_RX_MAX diff --git a/drivers/net/intel/idpf/idpf_common_rxtx.c b/drivers/net/intel/idpf/idpf_common_rxtx.c index a5d0795057..ab3d088899 100644 --- a/drivers/net/intel/idpf/idpf_common_rxtx.c +++ b/drivers/net/intel/idpf/idpf_common_rxtx.c @@ -250,6 +250,58 @@ idpf_qc_split_tx_complq_reset(struct ci_tx_queue *cq) cq->expected_gen_id = 1; } +RTE_EXPORT_INTERNAL_SYMBOL(idpf_splitq_rearm_common) +void +idpf_splitq_rearm_common(struct idpf_rx_queue *rx_bufq) +{ + struct rte_mbuf **rxp = &rx_bufq->sw_ring[rx_bufq->rxrearm_start]; + volatile union virtchnl2_rx_buf_desc *rxdp = rx_bufq->rx_ring; + uint16_t rx_id; + int i; + + rxdp += rx_bufq->rxrearm_start; + + /* Pull 'n' more MBUFs into the software ring */ + if (rte_mbuf_raw_alloc_bulk(rx_bufq->mp, + (void *)rxp, IDPF_RXQ_REARM_THRESH) < 0) { + if (rx_bufq->rxrearm_nb + IDPF_RXQ_REARM_THRESH >= + rx_bufq->nb_rx_desc) { + for (i = 0; i < IDPF_VPMD_DESCS_PER_LOOP; i++) { + rxp[i] = &rx_bufq->fake_mbuf; + rxdp[i] = (union virtchnl2_rx_buf_desc){0}; + } + } + rte_atomic_fetch_add_explicit(&rx_bufq->rx_stats.mbuf_alloc_failed, + IDPF_RXQ_REARM_THRESH, rte_memory_order_relaxed); + return; + } + + /* Initialize the mbufs in vector, process 8 mbufs in one loop */ + for (i = 0; i < IDPF_RXQ_REARM_THRESH; + i += 8, rxp += 8, rxdp += 8) { + rxdp[0].split_rd.pkt_addr = rxp[0]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[1].split_rd.pkt_addr = rxp[1]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[2].split_rd.pkt_addr = rxp[2]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[3].split_rd.pkt_addr = rxp[3]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[4].split_rd.pkt_addr = rxp[4]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[5].split_rd.pkt_addr = rxp[5]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[6].split_rd.pkt_addr = rxp[6]->buf_iova + RTE_PKTMBUF_HEADROOM; + rxdp[7].split_rd.pkt_addr = rxp[7]->buf_iova + RTE_PKTMBUF_HEADROOM; + } + + rx_bufq->rxrearm_start += IDPF_RXQ_REARM_THRESH; + if (rx_bufq->rxrearm_start >= rx_bufq->nb_rx_desc) + rx_bufq->rxrearm_start = 0; + + rx_bufq->rxrearm_nb -= IDPF_RXQ_REARM_THRESH; + + rx_id = (uint16_t)((rx_bufq->rxrearm_start == 0) ? + (rx_bufq->nb_rx_desc - 1) : (rx_bufq->rxrearm_start - 1)); + + /* Update the tail pointer on the NIC */ + IDPF_PCI_REG_WRITE(rx_bufq->qrx_tail, rx_id); +} + RTE_EXPORT_INTERNAL_SYMBOL(idpf_qc_single_tx_queue_reset) void idpf_qc_single_tx_queue_reset(struct ci_tx_queue *txq) @@ -1684,6 +1736,13 @@ const struct ci_rx_path_info idpf_rx_path_infos[] = { .rx_offloads = IDPF_RX_VECTOR_OFFLOADS, .simd_width = RTE_VECT_SIMD_256, .extra.single_queue = true}}, + [IDPF_RX_AVX2] = { + .pkt_burst = idpf_dp_splitq_recv_pkts_avx2, + .info = "Split AVX2 Vector", + .features = { + .rx_offloads = IDPF_RX_VECTOR_OFFLOADS, + .simd_width = RTE_VECT_SIMD_256, + }}, #ifdef CC_AVX512_SUPPORT [IDPF_RX_AVX512] = { .pkt_burst = idpf_dp_splitq_recv_pkts_avx512, diff --git a/drivers/net/intel/idpf/idpf_common_rxtx.h b/drivers/net/intel/idpf/idpf_common_rxtx.h index 3bc3323af4..87f6895c4c 100644 --- a/drivers/net/intel/idpf/idpf_common_rxtx.h +++ b/drivers/net/intel/idpf/idpf_common_rxtx.h @@ -203,6 +203,8 @@ void idpf_qc_split_tx_descq_reset(struct ci_tx_queue *txq); __rte_internal void idpf_qc_split_tx_complq_reset(struct ci_tx_queue *cq); __rte_internal +void idpf_splitq_rearm_common(struct idpf_rx_queue *rx_bufq); +__rte_internal void idpf_qc_single_tx_queue_reset(struct ci_tx_queue *txq); __rte_internal void idpf_qc_rx_queue_release(void *rxq); @@ -252,6 +254,9 @@ __rte_internal uint16_t idpf_dp_splitq_xmit_pkts_avx512(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts); __rte_internal +uint16_t idpf_dp_splitq_recv_pkts_avx2(void *rxq, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts); +__rte_internal uint16_t idpf_dp_singleq_recv_scatter_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts); __rte_internal diff --git a/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c b/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c index 21c8f79254..dbbb09afd7 100644 --- a/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c +++ b/drivers/net/intel/idpf/idpf_common_rxtx_avx2.c @@ -482,6 +482,129 @@ idpf_dp_singleq_recv_pkts_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, uint16 return _idpf_singleq_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts); } +uint16_t +idpf_dp_splitq_recv_pkts_avx2(void *rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) +{ + struct idpf_rx_queue *queue = (struct idpf_rx_queue *)rxq; + const uint32_t *ptype_tbl = queue->adapter->ptype_tbl; + struct rte_mbuf **sw_ring = &queue->bufq2->sw_ring[queue->rx_tail]; + volatile union virtchnl2_rx_desc *rxdp = + (volatile union virtchnl2_rx_desc *)queue->rx_ring + queue->rx_tail; + const __m256i mbuf_init = _mm256_set_epi64x(0, 0, 0, queue->mbuf_initializer); + + rte_prefetch0(rxdp); + nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, 4); /* 4 desc per AVX2 iteration */ + + if (queue->bufq2->rxrearm_nb > IDPF_RXQ_REARM_THRESH) + idpf_splitq_rearm_common(queue->bufq2); + + /* head gen check */ + uint64_t head_gen = rxdp->flex_adv_nic_3_wb.pktlen_gen_bufq_id; + if (((head_gen >> VIRTCHNL2_RX_FLEX_DESC_ADV_GEN_S) & + VIRTCHNL2_RX_FLEX_DESC_ADV_GEN_M) != queue->expected_gen_id) + return 0; + + uint16_t received = 0; + + /* Shuffle mask: picks fields from each 16-byte descriptor pair into the + * layout that will be merged into mbuf->rearm_data candidates. + */ + const __m256i shuf = _mm256_set_epi8( + /* high 128 bits (desc 3 then desc 2 lanes) */ + 0xFF, 0xFF, 0xFF, 0xFF, 11, 10, 5, 4, + 0xFF, 0xFF, 5, 4, 0xFF, 0xFF, 0xFF, 0xFF, + /* low 128 bits (desc 1 then desc 0 lanes) */ + 0xFF, 0xFF, 0xFF, 0xFF, 11, 10, 5, 4, + 0xFF, 0xFF, 5, 4, 0xFF, 0xFF, 0xFF, 0xFF + ); + + /* mask that clears bits 14 and 15 of the packet length word */ + const __m256i len_mask = _mm256_set_epi32( + 0xffffffff, 0xffffffff, 0xffff3fff, 0xffffffff, + 0xffffffff, 0xffffffff, 0xffff3fff, 0xffffffff + ); + + const __m256i ptype_mask = _mm256_set1_epi16(VIRTCHNL2_RX_FLEX_DESC_PTYPE_M); + + for (uint16_t i = 0; i < nb_pkts; + i += IDPF_VPMD_DESCS_PER_LOOP, + rxdp += IDPF_VPMD_DESCS_PER_LOOP) { + /* copy 4 mbuf pointers into rx_pkts[] */ + memcpy(&rx_pkts[i], &sw_ring[i], sizeof(rx_pkts[i]) * IDPF_VPMD_DESCS_PER_LOOP); + /* load four 128-bit descriptors */ + __m128i d0 = _mm_load_si128(RTE_CAST_PTR(const __m128i *, &rxdp[0])); + rte_compiler_barrier(); + __m128i d1 = _mm_load_si128(RTE_CAST_PTR(const __m128i *, &rxdp[1])); + rte_compiler_barrier(); + __m128i d2 = _mm_load_si128(RTE_CAST_PTR(const __m128i *, &rxdp[2])); + rte_compiler_barrier(); + __m128i d3 = _mm_load_si128(RTE_CAST_PTR(const __m128i *, &rxdp[3])); + + /* Build 256-bit descriptor-pairs */ + __m256i d01 = _mm256_set_m128i(d1, d0); /* low lane: d0, d1 */ + __m256i d23 = _mm256_set_m128i(d3, d2); /* high lane: d2, d3 */ + + /* mask off high pkt_len bits */ + __m256i desc01 = _mm256_and_si256(d01, len_mask); + __m256i desc23 = _mm256_and_si256(d23, len_mask); + + /* shuffle relevant bytes into mbuf rearm candidates */ + __m256i mb01 = _mm256_shuffle_epi8(desc01, shuf); + __m256i mb23 = _mm256_shuffle_epi8(desc23, shuf); + + /* extract ptypes from descriptors and translate via table */ + __m256i pt01 = _mm256_and_si256(d01, ptype_mask); + __m256i pt23 = _mm256_and_si256(d23, ptype_mask); + + uint16_t ptype0 = (uint16_t)_mm256_extract_epi16(pt01, 1); + uint16_t ptype1 = (uint16_t)_mm256_extract_epi16(pt01, 9); + uint16_t ptype2 = (uint16_t)_mm256_extract_epi16(pt23, 1); + uint16_t ptype3 = (uint16_t)_mm256_extract_epi16(pt23, 9); + + mb01 = _mm256_insert_epi32(mb01, (int)ptype_tbl[ptype1], 2); + mb01 = _mm256_insert_epi32(mb01, (int)ptype_tbl[ptype0], 0); + mb23 = _mm256_insert_epi32(mb23, (int)ptype_tbl[ptype3], 2); + mb23 = _mm256_insert_epi32(mb23, (int)ptype_tbl[ptype2], 0); + + /* build rearm vectors using mb01 and mb23 directly */ + __m256i rearm0 = _mm256_permute2f128_si256(mbuf_init, mb01, 0x20); + __m256i rearm1 = _mm256_blend_epi32(mbuf_init, mb01, 0xF0); + __m256i rearm2 = _mm256_permute2f128_si256(mbuf_init, mb23, 0x20); + __m256i rearm3 = _mm256_blend_epi32(mbuf_init, mb23, 0xF0); + + /* blind write rearm_data for all mbufs*/ + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data, rearm0); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data, rearm1); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data, rearm2); + _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data, rearm3); + + /* Check DD and GEN bits for all 4 descriptors */ + uint16_t burst = 0; + for (int j = 0; j < IDPF_VPMD_DESCS_PER_LOOP; ++j) { + uint64_t g = rxdp[j].flex_adv_nic_3_wb.pktlen_gen_bufq_id; + bool dd = (g & 1ULL) != 0ULL; + uint64_t gen = (g >> VIRTCHNL2_RX_FLEX_DESC_ADV_GEN_S) & + VIRTCHNL2_RX_FLEX_DESC_ADV_GEN_M; + if (dd && gen == queue->expected_gen_id) + burst++; + } + received += burst; + if (burst != IDPF_VPMD_DESCS_PER_LOOP) + break; + } + queue->rx_tail += received; + queue->expected_gen_id ^= ((queue->rx_tail & queue->nb_rx_desc) != 0); + queue->rx_tail &= (queue->nb_rx_desc - 1); + if ((queue->rx_tail & 1) == 1 && received > 1) { + queue->rx_tail--; + received--; + } + queue->bufq2->rxrearm_nb += received; + return received; +} + +RTE_EXPORT_INTERNAL_SYMBOL(idpf_dp_splitq_recv_pkts_avx2) + static inline void idpf_singleq_vtx1(volatile struct idpf_base_tx_desc *txdp, struct rte_mbuf *pkt, uint64_t flags) diff --git a/drivers/net/intel/idpf/idpf_common_rxtx_avx512.c b/drivers/net/intel/idpf/idpf_common_rxtx_avx512.c index bc2cadd738..d3a161c763 100644 --- a/drivers/net/intel/idpf/idpf_common_rxtx_avx512.c +++ b/drivers/net/intel/idpf/idpf_common_rxtx_avx512.c @@ -540,62 +540,6 @@ idpf_dp_singleq_recv_pkts_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, return _idpf_singleq_recv_raw_pkts_avx512(rx_queue, rx_pkts, nb_pkts); } -static __rte_always_inline void -idpf_splitq_rearm_common(struct idpf_rx_queue *rx_bufq) -{ - struct rte_mbuf **rxp = &rx_bufq->sw_ring[rx_bufq->rxrearm_start]; - volatile union virtchnl2_rx_buf_desc *rxdp = rx_bufq->rx_ring; - uint16_t rx_id; - int i; - - rxdp += rx_bufq->rxrearm_start; - - /* Pull 'n' more MBUFs into the software ring */ - if (rte_mbuf_raw_alloc_bulk(rx_bufq->mp, - (void *)rxp, - IDPF_RXQ_REARM_THRESH) < 0) { - if (rx_bufq->rxrearm_nb + IDPF_RXQ_REARM_THRESH >= - rx_bufq->nb_rx_desc) { - __m128i dma_addr0; - - dma_addr0 = _mm_setzero_si128(); - for (i = 0; i < IDPF_VPMD_DESCS_PER_LOOP; i++) { - rxp[i] = &rx_bufq->fake_mbuf; - _mm_store_si128(RTE_CAST_PTR(__m128i *, &rxdp[i]), - dma_addr0); - } - } - rte_atomic_fetch_add_explicit(&rx_bufq->rx_stats.mbuf_alloc_failed, - IDPF_RXQ_REARM_THRESH, rte_memory_order_relaxed); - return; - } - - /* Initialize the mbufs in vector, process 8 mbufs in one loop */ - for (i = 0; i < IDPF_RXQ_REARM_THRESH; - i += 8, rxp += 8, rxdp += 8) { - rxdp[0].split_rd.pkt_addr = rxp[0]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[1].split_rd.pkt_addr = rxp[1]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[2].split_rd.pkt_addr = rxp[2]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[3].split_rd.pkt_addr = rxp[3]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[4].split_rd.pkt_addr = rxp[4]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[5].split_rd.pkt_addr = rxp[5]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[6].split_rd.pkt_addr = rxp[6]->buf_iova + RTE_PKTMBUF_HEADROOM; - rxdp[7].split_rd.pkt_addr = rxp[7]->buf_iova + RTE_PKTMBUF_HEADROOM; - } - - rx_bufq->rxrearm_start += IDPF_RXQ_REARM_THRESH; - if (rx_bufq->rxrearm_start >= rx_bufq->nb_rx_desc) - rx_bufq->rxrearm_start = 0; - - rx_bufq->rxrearm_nb -= IDPF_RXQ_REARM_THRESH; - - rx_id = (uint16_t)((rx_bufq->rxrearm_start == 0) ? - (rx_bufq->nb_rx_desc - 1) : (rx_bufq->rxrearm_start - 1)); - - /* Update the tail pointer on the NIC */ - IDPF_PCI_REG_WRITE(rx_bufq->qrx_tail, rx_id); -} - static __rte_always_inline void idpf_splitq_rearm(struct idpf_rx_queue *rx_bufq) { -- 2.34.1