[dpdk-dev] [PATCH] net/virtio: Add support for vectorized functions on Power systems

[David Christensen]

Added the file virtio_rxtx_simple_altivec.c which implements Altivec
code for the virtio vectorized RX functions. Updated the various build
files.

Cc: Maxime Coquelin <maxime.coquelin at redhat.com>
Cc: Tiwei Bie <tiwei.bie at intel.com>

Signed-off-by: David Christensen <drc at linux.vnet.ibm.com>
---
 drivers/net/virtio/Makefile                     |   2 +
 drivers/net/virtio/meson.build                  |   2 +
 drivers/net/virtio/virtio_rxtx_simple_altivec.c | 202 ++++++++++++++++++++++++
 3 files changed, 206 insertions(+)
 create mode 100644 drivers/net/virtio/virtio_rxtx_simple_altivec.c

diff --git a/drivers/net/virtio/Makefile b/drivers/net/virtio/Makefile
index 6c2c996..5144e7c 100644
--- a/drivers/net/virtio/Makefile
+++ b/drivers/net/virtio/Makefile
@@ -33,6 +33,8 @@ SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_simple.c
 
 ifeq ($(CONFIG_RTE_ARCH_X86),y)
 SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_simple_sse.c
+else ifeq ($(CONFIG_RTE_ARCH_PPC_64),y)
+SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_simple_altivec.c
 else ifneq ($(filter y,$(CONFIG_RTE_ARCH_ARM) $(CONFIG_RTE_ARCH_ARM64)),)
 SRCS-$(CONFIG_RTE_LIBRTE_VIRTIO_PMD) += virtio_rxtx_simple_neon.c
 endif
diff --git a/drivers/net/virtio/meson.build b/drivers/net/virtio/meson.build
index 7949054..04c7fdf 100644
--- a/drivers/net/virtio/meson.build
+++ b/drivers/net/virtio/meson.build
@@ -11,6 +11,8 @@ deps += ['kvargs', 'bus_pci']
 
 if arch_subdir == 'x86'
 	sources += files('virtio_rxtx_simple_sse.c')
+elif arch_subdir == 'ppc_64'
+	sources += files('virtio_rxtx_simple_altivec.c')
 elif arch_subdir == 'arm' and host_machine.cpu_family().startswith('aarch64')
 	sources += files('virtio_rxtx_simple_neon.c')
 endif
diff --git a/drivers/net/virtio/virtio_rxtx_simple_altivec.c b/drivers/net/virtio/virtio_rxtx_simple_altivec.c
new file mode 100644
index 0000000..f4eb4eb
--- /dev/null
+++ b/drivers/net/virtio/virtio_rxtx_simple_altivec.c
@@ -0,0 +1,202 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2010-2015 Intel Corporation
+ */
+
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+
+#include <altivec.h>
+
+#include <rte_byteorder.h>
+#include <rte_branch_prediction.h>
+#include <rte_cycles.h>
+#include <rte_ether.h>
+#include <rte_ethdev_driver.h>
+#include <rte_errno.h>
+#include <rte_memory.h>
+#include <rte_mempool.h>
+#include <rte_malloc.h>
+#include <rte_mbuf.h>
+#include <rte_prefetch.h>
+#include <rte_string_fns.h>
+
+#include "virtio_rxtx_simple.h"
+
+#define RTE_VIRTIO_DESC_PER_LOOP 8
+
+/* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
+ *
+ * This routine is for non-mergeable RX, one desc for each guest buffer.
+ * This routine is based on the RX ring layout optimization. Each entry in the
+ * avail ring points to the desc with the same index in the desc ring and this
+ * will never be changed in the driver.
+ *
+ * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
+ */
+uint16_t
+virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+	uint16_t nb_pkts)
+{
+	struct virtnet_rx *rxvq = rx_queue;
+	struct virtqueue *vq = rxvq->vq;
+	struct virtio_hw *hw = vq->hw;
+	uint16_t nb_used;
+	uint16_t desc_idx;
+	struct vring_used_elem *rused;
+	struct rte_mbuf **sw_ring;
+	struct rte_mbuf **sw_ring_end;
+	uint16_t nb_pkts_received = 0;
+	const vector unsigned char zero = {0};
+
+	const vector unsigned char shuf_msk1 = {
+		0xFF, 0xFF, 0xFF, 0xFF,	/* packet type */
+		4, 5, 0xFF, 0xFF, /* vlan tci */
+		4, 5,			/* dat len */
+		0xFF, 0xFF,		/* vlan tci */
+		0xFF, 0xFF, 0xFF, 0xFF
+	};
+
+	const vector unsigned char shuf_msk2 = {
+		0xFF, 0xFF, 0xFF, 0xFF,	/* packet type */
+		12, 13, 0xFF, 0xFF,	/* pkt len */
+		12, 13,			/* dat len */
+		0xFF, 0xFF,		/* vlan tci */
+		0xFF, 0xFF, 0xFF, 0xFF
+	};
+
+	/*
+	 * Subtract the header length.
+	 *  In which case do we need the header length in used->len ?
+	 */
+	const vector unsigned short len_adjust = {
+		0, 0,
+		(uint16_t)-vq->hw->vtnet_hdr_size, 0,
+		(uint16_t)-vq->hw->vtnet_hdr_size, 0,
+		0, 0
+	};
+
+	if (unlikely(hw->started == 0))
+		return nb_pkts_received;
+
+	if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
+		return 0;
+
+	nb_used = VIRTQUEUE_NUSED(vq);
+
+	rte_compiler_barrier();
+
+	if (unlikely(nb_used == 0))
+		return 0;
+
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
+	nb_used = RTE_MIN(nb_used, nb_pkts);
+
+	desc_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
+	rused = &vq->vq_split.ring.used->ring[desc_idx];
+	sw_ring  = &vq->sw_ring[desc_idx];
+	sw_ring_end = &vq->sw_ring[vq->vq_nentries];
+
+	rte_prefetch0(rused);
+
+	if (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
+		virtio_rxq_rearm_vec(rxvq);
+		if (unlikely(virtqueue_kick_prepare(vq)))
+			virtqueue_notify(vq);
+	}
+
+	for (nb_pkts_received = 0;
+		nb_pkts_received < nb_used;) {
+		vector unsigned char desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
+		vector unsigned char mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
+		vector unsigned char pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
+
+		mbp[0] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 0));
+		desc[0] = vec_vsx_ld(0, (unsigned char const *)(rused + 0));
+		*(vector unsigned char *)&rx_pkts[0] = mbp[0];
+
+		mbp[1] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 2));
+		desc[1] = vec_vsx_ld(0, (unsigned char const *)(rused + 2));
+		*(vector unsigned char *)&rx_pkts[2] = mbp[1];
+
+		mbp[2] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 4));
+		desc[2] = vec_vsx_ld(0, (unsigned char const *)(rused + 4));
+		*(vector unsigned char *)&rx_pkts[4] = mbp[2];
+
+		mbp[3] = vec_vsx_ld(0, (unsigned char const *)(sw_ring + 6));
+		desc[3] = vec_vsx_ld(0, (unsigned char const *)(rused + 6));
+		*(vector unsigned char *)&rx_pkts[6] = mbp[3];
+
+		pkt_mb[0] = vec_perm(desc[0], zero, shuf_msk1);
+		pkt_mb[1] = vec_perm(desc[0], zero, shuf_msk2);
+		pkt_mb[0] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[0] + len_adjust);
+		pkt_mb[1] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[1] + len_adjust);
+		*(vector unsigned char *)&rx_pkts[0]->rx_descriptor_fields1 =
+			pkt_mb[0];
+		*(vector unsigned char *)&rx_pkts[1]->rx_descriptor_fields1 =
+			pkt_mb[1];
+
+		pkt_mb[2] = vec_perm(desc[1], zero, shuf_msk1);
+		pkt_mb[3] = vec_perm(desc[1], zero, shuf_msk2);
+		pkt_mb[2] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[2] + len_adjust);
+		pkt_mb[3] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[3] + len_adjust);
+		*(vector unsigned char *)&rx_pkts[2]->rx_descriptor_fields1 =
+			pkt_mb[2];
+		*(vector unsigned char *)&rx_pkts[3]->rx_descriptor_fields1 =
+			pkt_mb[3];
+
+		pkt_mb[4] = vec_perm(desc[2], zero, shuf_msk1);
+		pkt_mb[5] = vec_perm(desc[2], zero, shuf_msk2);
+		pkt_mb[4] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[4] + len_adjust);
+		pkt_mb[5] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[5] + len_adjust);
+		*(vector unsigned char *)&rx_pkts[4]->rx_descriptor_fields1 =
+			pkt_mb[4];
+		*(vector unsigned char *)&rx_pkts[5]->rx_descriptor_fields1 =
+			pkt_mb[5];
+
+		pkt_mb[6] = vec_perm(desc[3], zero, shuf_msk1);
+		pkt_mb[7] = vec_perm(desc[3], zero, shuf_msk2);
+		pkt_mb[6] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[6] + len_adjust);
+		pkt_mb[7] = (vector unsigned char)
+			((vector unsigned short)pkt_mb[7] + len_adjust);
+		*(vector unsigned char *)&rx_pkts[6]->rx_descriptor_fields1 =
+			pkt_mb[6];
+		*(vector unsigned char *)&rx_pkts[7]->rx_descriptor_fields1 =
+			pkt_mb[7];
+
+		if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
+			if (sw_ring + nb_used <= sw_ring_end)
+				nb_pkts_received += nb_used;
+			else
+				nb_pkts_received += sw_ring_end - sw_ring;
+			break;
+		} else {
+			if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
+				sw_ring_end)) {
+				nb_pkts_received += sw_ring_end - sw_ring;
+				break;
+			} else {
+				nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
+
+				rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
+				sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
+				rused   += RTE_VIRTIO_DESC_PER_LOOP;
+				nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
+			}
+		}
+	}
+
+	vq->vq_used_cons_idx += nb_pkts_received;
+	vq->vq_free_cnt += nb_pkts_received;
+	rxvq->stats.packets += nb_pkts_received;
+	return nb_pkts_received;
+}
-- 
1.8.3.1