arm64 performance

[cba...@gmail.com]

Hi!

I was trying to speedup solve-field performance on a Cortex-A53 arm processor and came accross the function ctmf_helper using some architecture dependent simd/vector code.

They exist for SSE2,MMX and ALTIVEC but nothing for ARM.

I looked up arm64 docs and added a similar section, and it speeded up image2xy by 20% and my solve-field by 10%. I don't know enough to know if it is always safe, or if it needs

other changes for alignment in malloc etc, but it works for my case.

--------------------------------------

#include <altivec.h>
+#elif defined(USE_ARM64)
+#include <arm_neon.h>
 #endif


--------------


+#elif defined(USE_ARM64)
+static inline void histogram_add( const uint16_t x[16], uint16_t y[16] )
+{
+    *(uint16x4_t*) &y[0]  = vadd_u16( *(uint16x4_t*) &y[0],  *(uint16x4_t*) &x[0]  );
+    *(uint16x4_t*) &y[4]  = vadd_u16( *(uint16x4_t*) &y[4],  *(uint16x4_t*) &x[4]  );
+    *(uint16x4_t*) &y[8]  = vadd_u16( *(uint16x4_t*) &y[8],  *(uint16x4_t*) &x[8]  );
+    *(uint16x4_t*) &y[12] = vadd_u16( *(uint16x4_t*) &y[12], *(uint16x4_t*) &x[12] );
+}
 #else

---------------------


+#elif defined(USE_ARM64)
+static inline void histogram_sub( const uint16_t x[16], uint16_t y[16] )
+{
+    *(uint16x4_t*) &y[0]  = vsub_u16( *(uint16x4_t*) &y[0],  *(uint16x4_t*) &x[0]  );
+    *(uint16x4_t*) &y[4]  = vsub_u16( *(uint16x4_t*) &y[4],  *(uint16x4_t*) &x[4]  );
+    *(uint16x4_t*) &y[8]  = vsub_u16( *(uint16x4_t*) &y[8],  *(uint16x4_t*) &x[8]  );
+    *(uint16x4_t*) &y[12] = vsub_u16( *(uint16x4_t*) &y[12], *(uint16x4_t*) &x[12] );
+}
 #else

Thought I would pass it along, and check if any other perf changes/flags are known for arm platforms.

[Dustin Lang]

Hi,

Great -- would you be willing to send a pull request at the github repository,

https://github.com/dstndstn/astrometry.net

?

Thanks,

--dustin

[cba...@gmail.com]

Done !

[Dustin Lang]

Many thanks!!