[v2] hash: optimize signature compare by using neon intrinsic
Checks
Commit Message
Implemented signature compare function based on neon intrinsic.
Hash bulk lookup had 3% - 6% performance gain after optimization.
Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Gavin Hu <gavin.hu@arm.com>
---
v2:
* Use 'rte_vect.h' instead of 'arm_neon.h'.
* Moved down variable declaration to avoid extra #ifdef.
lib/librte_hash/rte_cuckoo_hash.c | 29 ++++++++++++++++++++++++++++-
lib/librte_hash/rte_cuckoo_hash.h | 1 +
2 files changed, 29 insertions(+), 1 deletion(-)
Comments
Hi, Ruifeng,
I don't have an ARM Platform at hand to test but I gone through the code against a NEON instruction manual,
and the logic looks good to me. One small comment is the coding style for switch case, I think it is more common
to leave the braces on their own lines.
Otherwise:
Acked-by: Yipeng Wang<yipeng1.wang@intel.com>
>-----Original Message-----
>From: Ruifeng Wang [mailto:ruifeng.wang@arm.com]
>Sent: Monday, February 11, 2019 11:01 PM
>To: Wang, Yipeng1 <yipeng1.wang@intel.com>; jerinj@marvell.com
>Cc: dev@dpdk.org; honnappa.nagarahalli@arm.com; gavin.hu@arm.com; nd@arm.com; Ruifeng Wang <ruifeng.wang@arm.com>
>Subject: [PATCH v2] hash: optimize signature compare by using neon intrinsic
>
>Implemented signature compare function based on neon intrinsic.
>Hash bulk lookup had 3% - 6% performance gain after optimization.
>
>Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
>Reviewed-by: Gavin Hu <gavin.hu@arm.com>
Hi Yipeng,
Thanks for your review and acknowledgement.
I did think about braces on separate lines. One concern is, in this way, extra indent will be needed and
make it hard to keep code lines within 80 characters. So I made the style compromise.
Regards,
/Ruifeng
> -----Original Message-----
> From: Wang, Yipeng1 <yipeng1.wang@intel.com>
> Sent: Thursday, February 14, 2019 4:32
> To: Ruifeng Wang (Arm Technology China) <Ruifeng.Wang@arm.com>;
> jerinj@marvell.com
> Cc: dev@dpdk.org; Honnappa Nagarahalli
> <Honnappa.Nagarahalli@arm.com>; Gavin Hu (Arm Technology China)
> <Gavin.Hu@arm.com>; nd <nd@arm.com>
> Subject: RE: [PATCH v2] hash: optimize signature compare by using neon
> intrinsic
>
> Hi, Ruifeng,
>
> I don't have an ARM Platform at hand to test but I gone through the code
> against a NEON instruction manual, and the logic looks good to me. One small
> comment is the coding style for switch case, I think it is more common to
> leave the braces on their own lines.
>
> Otherwise:
> Acked-by: Yipeng Wang<yipeng1.wang@intel.com>
>
>
> >-----Original Message-----
> >From: Ruifeng Wang [mailto:ruifeng.wang@arm.com]
> >Sent: Monday, February 11, 2019 11:01 PM
> >To: Wang, Yipeng1 <yipeng1.wang@intel.com>; jerinj@marvell.com
> >Cc: dev@dpdk.org; honnappa.nagarahalli@arm.com; gavin.hu@arm.com;
> >nd@arm.com; Ruifeng Wang <ruifeng.wang@arm.com>
> >Subject: [PATCH v2] hash: optimize signature compare by using neon
> >intrinsic
> >
> >Implemented signature compare function based on neon intrinsic.
> >Hash bulk lookup had 3% - 6% performance gain after optimization.
> >
> >Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> >Reviewed-by: Gavin Hu <gavin.hu@arm.com>
On Tue, 2019-02-12 at 15:01 +0800, Ruifeng Wang wrote:
> -------------------------------------------------------------------
> ---
> Implemented signature compare function based on neon intrinsic.
> Hash bulk lookup had 3% - 6% performance gain after optimization.
>
> Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> Reviewed-by: Gavin Hu <gavin.hu@arm.com>
> ---
> v2:
> * Use 'rte_vect.h' instead of 'arm_neon.h'.
> * Moved down variable declaration to avoid extra #ifdef.
Acked-by: Jerin Jacob <jerinj@marvell.com>
25/02/2019 18:52, Jerin Jacob Kollanukkaran:
> On Tue, 2019-02-12 at 15:01 +0800, Ruifeng Wang wrote:
> > Implemented signature compare function based on neon intrinsic.
> > Hash bulk lookup had 3% - 6% performance gain after optimization.
> >
> > Signed-off-by: Ruifeng Wang <ruifeng.wang@arm.com>
> > Reviewed-by: Gavin Hu <gavin.hu@arm.com>
>
> Acked-by: Jerin Jacob <jerinj@marvell.com>
Applied, thanks
@@ -26,6 +26,7 @@
#include <rte_spinlock.h>
#include <rte_ring.h>
#include <rte_compat.h>
+#include <rte_vect.h>
#include "rte_hash.h"
#include "rte_cuckoo_hash.h"
@@ -407,6 +408,10 @@ rte_hash_create(const struct rte_hash_parameters *params)
if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE2))
h->sig_cmp_fn = RTE_HASH_COMPARE_SSE;
else
+#elif defined(RTE_ARCH_ARM64)
+ if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
+ h->sig_cmp_fn = RTE_HASH_COMPARE_NEON;
+ else
#endif
h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
@@ -1581,7 +1586,7 @@ compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
/* For match mask the first bit of every two bits indicates the match */
switch (sig_cmp_fn) {
-#ifdef RTE_MACHINE_CPUFLAG_SSE2
+#if defined(RTE_MACHINE_CPUFLAG_SSE2)
case RTE_HASH_COMPARE_SSE:
/* Compare all signatures in the bucket */
*prim_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
@@ -1594,6 +1599,28 @@ compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
(__m128i const *)sec_bkt->sig_current),
_mm_set1_epi16(sig)));
break;
+#elif defined(RTE_MACHINE_CPUFLAG_NEON)
+ case RTE_HASH_COMPARE_NEON: {
+ uint16x8_t vmat, vsig, x;
+ uint64x2_t x64;
+ int16x8_t shift = {-15, -13, -11, -9, -7, -5, -3, -1};
+
+ 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_bkt->sig_current));
+ x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
+ x64 = vpaddlq_u32(vpaddlq_u16(x));
+ *prim_hash_matches = (uint32_t)(vgetq_lane_u64(x64, 0) +
+ vgetq_lane_u64(x64, 1));
+ /* Compare all signatures in the secondary bucket */
+ vmat = vceqq_u16(vsig,
+ vld1q_u16((uint16_t const *)sec_bkt->sig_current));
+ x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
+ x64 = vpaddlq_u32(vpaddlq_u16(x));
+ *sec_hash_matches = (uint32_t)(vgetq_lane_u64(x64, 0) +
+ vgetq_lane_u64(x64, 1)); }
+ break;
#endif
default:
for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
@@ -141,6 +141,7 @@ struct rte_hash_key {
enum rte_hash_sig_compare_function {
RTE_HASH_COMPARE_SCALAR = 0,
RTE_HASH_COMPARE_SSE,
+ RTE_HASH_COMPARE_NEON,
RTE_HASH_COMPARE_NUM
};