On Fri, Oct 09, 2026 at 12:36:30AM +0000, Sandeep Penigalapati wrote: > The ice PMD only supported Rx hardware timestamp offload > (RTE_ETH_RX_OFFLOAD_TIMESTAMP) on the scalar Rx path. Enabling the > offload forced a fallback from the AVX2/AVX512 vector Rx paths to the > scalar path, causing a significant performance drop on 800 series > adapters. > > Add Rx timestamp support to the AVX2 and AVX512 vector Rx paths, > mirroring the existing iavf implementation: the 32-bit timestamp is > read from the flex descriptor in the vectorized loop and converted to > 64 bits, with register rollover tracking, in a scalar pass over the > received packets after the loop. > > Advertise the timestamp offload only on the x86 vector paths that > implement it, via a new ICE_RX_VECTOR_OFFLOAD_TS_OFFLOADS mask, so on > Arm the request still falls back to the scalar path. > > Signed-off-by: Sandeep Penigalapati <[email protected]> Acked-by: Bruce Richardson <[email protected]>
Applied to dpdk-next-net-intel. Thanks, /Bruce > --- > v2: > - rebased to current main; fixed release notes and ice.ini conflicts > - kept the offload check on two lines >
