diff options
| author | Jonathan Wright <jonathan.wright@arm.com> | 2019-05-30 14:48:11 +0100 |
|---|---|---|
| committer | Jonathan Wright <jonathan.wright@arm.com> | 2019-05-30 16:36:19 +0100 |
| commit | 0aabb3f90fc8f4e477609a2c15d18804f14fb218 (patch) | |
| tree | e46245212bea0faca979fab18d23d0f0e1c84b6f /jmemmgr.c | |
| parent | fdcaeede1eae1dc9898212fdb963b0ae98ec43c8 (diff) | |
Add 32-byte memory alignment check for Arm NEON
ALIGN_SIZE is defined to be 32 bytes in order to accommodate AVX2's
alignment requirements. Arm NEON only requires 16-byte alignment but
software optimizations are possible if 32-byte alignment is assumed
for memory allocated by jmemmgr.c.
This patch introduces a compile-time check to make sure that
ALIGN_SIZE is a multiple of 32 bytes for Arm NEON systems.
Optimizations relying on this assumption for Arm NEON will follow
this CL.
Bug: 922430
Change-Id: I57f349eeffa704ec6c440f356a607964fc4fc8cc
Diffstat (limited to 'jmemmgr.c')
| -rw-r--r-- | jmemmgr.c | 4 |
1 files changed, 4 insertions, 0 deletions
@@ -101,6 +101,10 @@ round_up_pow2(size_t a, size_t b) #endif #if defined(__AVX2__) && (ALIGN_SIZE % 32) #error "AVX2 requires 32-byte alignment. ALIGN_SIZE is not a multiple of 32 bytes." +#elif defined(__ARM_NEON) && (ALIGN_SIZE % 32) + /* 32-byte alignment allows us to extract more performance from */ + /* fancy-upsampling algorithms when using NEON. */ + #error "NEON optimizations rely on 32-byte alignment. ALIGN_SIZE is not a multiple of 32 bytes." #endif #endif |