From 3574865bf0952f27d70151ec0bcf2a65e3e8b537 Mon Sep 17 00:00:00 2001 From: Logan oos Even <46396513+Logan007@users.noreply.github.com> Date: Tue, 26 Jan 2021 21:23:08 +0545 Subject: [PATCH] disabled SPECK defaulting to NEON support on ARM (#602) --- doc/Building.md | 14 ++++++++++++-- include/speck.h | 2 +- src/speck.c | 2 +- 3 files changed, 14 insertions(+), 4 deletions(-) diff --git a/doc/Building.md b/doc/Building.md index 1047f1e..f538b7c 100644 --- a/doc/Building.md +++ b/doc/Building.md @@ -105,7 +105,7 @@ So far, the following portions of n2n's code benefit from hardware features: ``` AES: AES-NI ChaCha20: SSE2, SSSE3 -SPECK: SSE2, SSSE3, AVX2, NEON +SPECK: SSE2, SSSE3, AVX2, (NEON) Pearson Hashing: AES-NI Random Numbers: RDSEED, RDRND (not faster but more random seed) ``` @@ -146,4 +146,14 @@ If used with multiple supernodes, by default, an edge choses the least loaded su which of course can be combined with the compiler optimizations mentioned above… -Note that the activation of this strategy requires a sufficiently accurate local day-of-time clock. It probably will fail on smaller systems using `uclibc` (instead of `glibc`) whose day-of-time clock is said to not provide sub-second accuracy. \ No newline at end of file +Note that the activation of this strategy requires a sufficiently accurate local day-of-time clock. It probably will fail on smaller systems using `uclibc` (instead of `glibc`) whose day-of-time clock is said to not provide sub-second accuracy. + +## SPECK – ARM NEON Hardware Acceleration + +By default, SPECK does not take advantage of ARM NEON hardware acceleration even if compiled with `-march=native`. The reason is that the NEON implementation proved to be slower than the 64-bit scalar code on Raspberry Pi 3B+, see [here](https://github.com/ntop/n2n/issues/563). + +Your specific ARM mileage may vary, so it can be enabled by configuring the definition of the `SPECK_ARM_NEON` macro: + +`./configure CFLAGS="-DSPECK_ARM_NEON"` + +Just make sure that the correct architecture is set, too. `-march=native` usually works quite well. \ No newline at end of file diff --git a/include/speck.h b/include/speck.h index ff73200..ad1fbe3 100644 --- a/include/speck.h +++ b/include/speck.h @@ -72,7 +72,7 @@ typedef struct { } speck_context_t; -#elif defined (__ARM_NEON) // NEON support ------------------------------------------------------------------------ +#elif defined (__ARM_NEON) && defined (SPECK_ARM_NEON) // NEON support --------------------------------------- #include diff --git a/src/speck.c b/src/speck.c index 7ae80fa..f92b5e0 100644 --- a/src/speck.c +++ b/src/speck.c @@ -441,7 +441,7 @@ static int speck_expand_key (speck_context_t *ctx, const unsigned char *k, int k } -#elif defined (__ARM_NEON) // NEON support ------------------------------------------------------------------- +#elif defined (__ARM_NEON) && defined (SPECK_ARM_NEON) // NEON support --------------------------------------- #define LCS(x,r) (((x)<>(64-r)))