> > On Thu, Apr 29, 2021 at 08:31:03AM +0000, Ananyev, Konstantin wrote: > > Hi Stanislaw, > > > > > > > > On Wed, Apr 28, 2021 at 09:44:54PM +0000, Honnappa Nagarahalli wrote: > > > <snip> > > > > [Honnappa] Sorry, I do not understand this. I see that vector code is > > > > under compile time flag as below > > > > > > > > #if defined RTE_ARCH_X86 || defined __ARM_NEON > > > > l3fwd_em_send_packets(nb_rx, pkts_burst, > > > > portid, qconf); > > > > #else > > > > l3fwd_em_no_opt_send_packets(nb_rx, pkts_burst, > > > > portid, qconf); > > > > #endif > > > Take a look at the ifdef tree at the top of l3fwd_em.c, here: > > > http://git.dpdk.org/dpdk/tree/examples/l3fwd/l3fwd_em.c#n218 > > > > > > #if defined(__SSE2__) > > > ... > > > #else > > > #error No vector engine (SSE, NEON, ALTIVEC) available, check your > > > toolchain > > > #endif > > > > > > > I think it is just a flaw and needs to be fixed. > > Patch would help here 😊 > > Konstantin > > It looks as if implementing em_mask_key() is enough to get l3fwd > working. However to me this ifdef seems tricky. How should a scalar > implementation handle the xmm_t type? rte_xmm_t looks like an API > type/union, but both are not mentioned in documentation and are in > platform dependent rte_vect.h only. > So either I add another case for RISC-V or (what seems more proper) add > an else clause implementation. However then should I change this function > to take rte_xmm_t? If not is casting xmm_t to i.e. int32_t[] always > valid? Even if I change to rte_xmm_t, it's not a stable API type, is it? > So what guarantee do I have that it maps to int32_t bit-wise on every > platform? > > I think the semantic requirements of xmm_t typedef are a bit undefined as > well as the vector handling across the architectures (being something > rather arch specific). I don't have a clear idea on how to solve this > yet and I would not like to hijack this discussion with vector stuff. > > Though I may be missing some obvious solution here. Any idea is welcome. > :)
I think it should be possible to replace xmm_t with rte_xmm_t in ipv(4|6)_5tuple_host and make em_mask_key to take 'rte_xmm_t *' as a parameter/return value instead of xmm_t. With that in place scalar version seems straightforward. Of course perf regression test would be needed after such changes, but I think with '-O3' it should be no difference.
