hash: optimize Toeplitz RSS computation
authorYangchao Zhou <zhouyates@gmail.com>
Tue, 22 Aug 2017 12:02:35 +0000 (20:02 +0800)
committerThomas Monjalon <thomas@monjalon.net>
Sat, 7 Oct 2017 11:50:43 +0000 (13:50 +0200)
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@gmail.com>
Reviewed-by: Vladimir Medvedkin <medvedkinv@gmail.com>
lib/librte_hash/rte_thash.h

index 2fffd61..4fa5e07 100644 (file)
@@ -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;