Re: [EXTERNAL] [PATCH v7 2/4] hash: optimize compare signature for NEON

[Yoan Picchi]

On 3/20/24 07:37, Pavan Nikhilesh Bhagavatula wrote:
Upon a successful comparison, NEON sets all the bits in the lane to 1
We can skip shifting by simply masking with specific masks.
Signed-off-by: Yoan Picchi <yoan.pic...@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.w...@arm.com>
Reviewed-by: Nathan Brown <nathan.br...@arm.com>
---
  lib/hash/arch/arm/compare_signatures.h | 24 +++++++++++-------------
  1 file changed, 11 insertions(+), 13 deletions(-)

diff --git a/lib/hash/arch/arm/compare_signatures.h
b/lib/hash/arch/arm/compare_signatures.h
index 1af6ba8190..b5a457f936 100644
--- a/lib/hash/arch/arm/compare_signatures.h
+++ b/lib/hash/arch/arm/compare_signatures.h
@@ -30,23 +30,21 @@ compare_signatures_dense(uint16_t
*hitmask_buffer,
        switch (sig_cmp_fn) {
  #if RTE_HASH_BUCKET_ENTRIES <= 8
        case RTE_HASH_COMPARE_NEON: {
-               uint16x8_t vmat, vsig, x;
-               int16x8_t shift = {0, 1, 2, 3, 4, 5, 6, 7};
-               uint16_t low, high;
+               uint16x8_t vmat, hit1, hit2;
+               const uint16x8_t mask = {0x1, 0x2, 0x4, 0x8, 0x10, 0x20,
0x40, 0x80};
+               const uint16x8_t vsig = vld1q_dup_u16((uint16_t const
*)&sig);

-               vsig = vld1q_dup_u16((uint16_t const *)&sig);
                /* Compare all signatures in the primary bucket */
-               vmat = vceqq_u16(vsig,
-                       vld1q_u16((uint16_t const *)prim_bucket_sigs));
-               x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x0001)),
shift);
-               low = (uint16_t)(vaddvq_u16(x));
+               vmat = vceqq_u16(vsig, vld1q_u16(prim_bucket_sigs));
+               hit1 = vandq_u16(vmat, mask);
+
                /* Compare all signatures in the secondary bucket */
-               vmat = vceqq_u16(vsig,
-                       vld1q_u16((uint16_t const *)sec_bucket_sigs));
-               x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x0001)),
shift);
-               high = (uint16_t)(vaddvq_u16(x));
-               *hitmask_buffer = low | high << RTE_HASH_BUCKET_ENTRIES;
+               vmat = vceqq_u16(vsig, vld1q_u16(sec_bucket_sigs));
+               hit2 = vandq_u16(vmat, mask);

+               hit2 = vshlq_n_u16(hit2, RTE_HASH_BUCKET_ENTRIES);
+               hit2 = vorrq_u16(hit1, hit2);
+               *hitmask_buffer = vaddvq_u16(hit2);

Since vaddv is expensive could you convert it to vshrn?

https://community.arm.com/arm-community-blogs/b/infrastructure-solutions-blog/posts/porting-x86-vector-bitmask-optimizations-to-arm-neon

https://github.com/DPDK/dpdk/blob/main/examples/l3fwd/l3fwd_neon.h#L226

Thank you for those links, it was a good read.
Unfortunatly I don't think it is a good use case here. A decent part of 
the speedup I get is by using a dense hitmask: ie every bit count with 
no padding. Using the vshrn would have 4 bits of padding, and stripping 
them would be more expensive than using a regular reduce.


                }
                break;
  #endif
--
2.25.1