diff options
| author |   Lijian.Zhang <Lijian.Zhang@arm.com> | 2020-12-10 16:01:00 +0800 |
|---|---|---|
| committer |   Damjan Marion <damarion@cisco.com> | 2020-12-18 11:26:53 +0100 |
| commit | e3e16172070e16e48d88b3e61dd2f3c539655513 (patch) | |
| tree | cbf52c8e791a8d6ac45ffd2602a64d0ebb0f2643 /src/plugins | |
| parent | 942542f7c1c3aae62a88612b8702a45f0d3b0e35 (diff) | |
avf: optimized with NEON SIMD instruction
Optimize avf-input node processing function with NEON SIMD instruction.
Type: improvement
Change-Id: I3dd76ac659686209dda9b176fc426aeae639e99b
Signed-off-by: Lijian Zhang <Lijian.Zhang@arm.com>
Diffstat (limited to 'src/plugins')
| -rw-r--r-- | src/plugins/avf/input.c | 28 |
1 files changed, 27 insertions, 1 deletions
diff --git a/src/plugins/avf/input.c b/src/plugins/avf/input.c index 4e4f116304d..85f97ca3e49 100644 --- a/src/plugins/avf/input.c +++ b/src/plugins/avf/input.c @@ -253,6 +253,9 @@ avf_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node, #ifdef CLIB_HAVE_VEC256 u64x4 q1x4, or_q1x4 = { 0 }; u64x4 dd_eop_mask4 = u64x4_splat (AVF_RXD_STATUS_DD | AVF_RXD_STATUS_EOP); +#elif defined(CLIB_HAVE_VEC128) + u32x4 q1x4_lo, q1x4_hi, or_q1x4 = { 0 }; + u32x4 dd_eop_mask4 = u32x4_splat (AVF_RXD_STATUS_DD | AVF_RXD_STATUS_EOP); #endif /* is there anything on the ring */ @@ -300,6 +303,29 @@ avf_device_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node, or_q1x4 |= q1x4; u64x4_store_unaligned (q1x4, ptd->qw1s + n_rx_packets); +#elif defined(CLIB_HAVE_VEC128) + if (n_rx_packets >= AVF_RX_VECTOR_SZ - 4 || next >= size - 4) + goto one_by_one; + + q1x4_lo = + u32x4_gather ((void *) &d[0].qword[1], (void *) &d[1].qword[1], + (void *) &d[2].qword[1], (void *) &d[3].qword[1]); + + /* not all packets are ready or at least one of them is chained */ + if (!u32x4_is_equal (q1x4_lo & dd_eop_mask4, dd_eop_mask4)) + goto one_by_one; + + q1x4_hi = u32x4_gather ( + (void *) &d[0].qword[1] + 4, (void *) &d[1].qword[1] + 4, + (void *) &d[2].qword[1] + 4, (void *) &d[3].qword[1] + 4); + + or_q1x4 |= q1x4_lo; + ptd->qw1s[n_rx_packets + 0] = (u64) q1x4_hi[0] << 32 | (u64) q1x4_lo[0]; + ptd->qw1s[n_rx_packets + 1] = (u64) q1x4_hi[1] << 32 | (u64) q1x4_lo[1]; + ptd->qw1s[n_rx_packets + 2] = (u64) q1x4_hi[2] << 32 | (u64) q1x4_lo[2]; + ptd->qw1s[n_rx_packets + 3] = (u64) q1x4_hi[3] << 32 | (u64) q1x4_lo[3]; +#endif +#if defined(CLIB_HAVE_VEC256) || defined(CLIB_HAVE_VEC128) vlib_buffer_copy_indices (bi, rxq->bufs + next, 4); /* next */ @@ -359,7 +385,7 @@ no_more_desc: rxq->next = next; rxq->n_enqueued -= n_rx_packets + n_tail_desc; -#ifdef CLIB_HAVE_VEC256 +#if defined(CLIB_HAVE_VEC256) || defined(CLIB_HAVE_VEC128) or_qw1 |= or_q1x4[0] | or_q1x4[1] | or_q1x4[2] | or_q1x4[3]; #endif |