[dpdk-dev] [PATCH] hash: optimize the softrss computation

[Vladimir Medvedkin]

Hi,

2017-08-22 15:02 GMT+03:00 Yangchao Zhou <zhouyates at gmail.com>:

> Use rte_bsf32 and fast bit unset operation to optimize the softrss
> computation.
> The following measurements shows improvement over the default
> softrss computation function.
>
> tuple lens old(cycles) new(cycles)
>     3        1225         337
>     9        3743         992
>
> Signed-off-by: Yangchao Zhou <zhouyates at gmail.com>
> ---
>  lib/librte_hash/rte_thash.h | 22 ++++++++++------------
>  1 file changed, 10 insertions(+), 12 deletions(-)
>
> diff --git a/lib/librte_hash/rte_thash.h b/lib/librte_hash/rte_thash.h
> index 2fffd61..4fa5e07 100644
> --- a/lib/librte_hash/rte_thash.h
> +++ b/lib/librte_hash/rte_thash.h
> @@ -207,15 +207,14 @@ static inline uint32_t
>  rte_softrss(uint32_t *input_tuple, uint32_t input_len,
>                 const uint8_t *rss_key)
>  {
> -       uint32_t i, j, ret = 0;
> +       uint32_t i, j, map, ret = 0;
>
>         for (j = 0; j < input_len; j++) {
> -               for (i = 0; i < 32; i++) {
> -                       if (input_tuple[j] & (1 << (31 - i))) {
> -                               ret ^= rte_cpu_to_be_32(((const uint32_t
> *)rss_key)[j]) << i |
> +               for (map = input_tuple[j]; map; map &= (map - 1)) {
> +                       i = rte_bsf32(map);
> +                       ret ^= rte_cpu_to_be_32(((const uint32_t
> *)rss_key)[j]) << (31 - i) |
>                                         (uint32_t)((uint64_t)(rte_cpu_to_be_32(((const
> uint32_t *)rss_key)[j + 1])) >>
> -                                       (32 - i));
> -                       }
> +                                       (i + 1));
>                 }
>         }
>         return ret;
> @@ -238,14 +237,13 @@ static inline uint32_t
>  rte_softrss_be(uint32_t *input_tuple, uint32_t input_len,
>                 const uint8_t *rss_key)
>  {
> -       uint32_t i, j, ret = 0;
> +       uint32_t i, j, map, ret = 0;
>
>         for (j = 0; j < input_len; j++) {
> -               for (i = 0; i < 32; i++) {
> -                       if (input_tuple[j] & (1 << (31 - i))) {
> -                               ret ^= ((const uint32_t *)rss_key)[j] << i
> |
> -                                       (uint32_t)((uint64_t)(((const
> uint32_t *)rss_key)[j + 1]) >> (32 - i));
> -                       }
> +               for (map = input_tuple[j]; map; map &= (map - 1)) {
> +                       i = rte_bsf32(map);
> +                       ret ^= ((const uint32_t *)rss_key)[j] << (31 - i) |
> +                               (uint32_t)((uint64_t)(((const uint32_t
> *)rss_key)[j + 1]) >> (i + 1));
>                 }
>         }
>         return ret;
> --
> 2.7.4
>
>
Looks good for me. Thanks!

Reviewed-by:  Medvedkin Vladimir <medvedkinv at gmail.com>

-- 
Regards,
Vladimir