diff options
Diffstat (limited to 'volk/kernels/volk/volk_32u_popcnt.h')
-rw-r--r-- | volk/kernels/volk/volk_32u_popcnt.h | 36 |
1 files changed, 36 insertions, 0 deletions
diff --git a/volk/kernels/volk/volk_32u_popcnt.h b/volk/kernels/volk/volk_32u_popcnt.h new file mode 100644 index 000000000..978356972 --- /dev/null +++ b/volk/kernels/volk/volk_32u_popcnt.h @@ -0,0 +1,36 @@ +#ifndef INCLUDED_VOLK_32u_POPCNT_A16_H +#define INCLUDED_VOLK_32u_POPCNT_A16_H + +#include <stdio.h> +#include <inttypes.h> + + +#ifdef LV_HAVE_GENERIC + +static inline void volk_32u_popcnt_generic(uint32_t* ret, const uint32_t value) { + + // This is faster than a lookup table + uint32_t retVal = value; + + retVal = (retVal & 0x55555555) + (retVal >> 1 & 0x55555555); + retVal = (retVal & 0x33333333) + (retVal >> 2 & 0x33333333); + retVal = (retVal + (retVal >> 4)) & 0x0F0F0F0F; + retVal = (retVal + (retVal >> 8)); + retVal = (retVal + (retVal >> 16)) & 0x0000003F; + + *ret = retVal; +} + +#endif /*LV_HAVE_GENERIC*/ + +#ifdef LV_HAVE_SSE4_2 + +#include <nmmintrin.h> + +static inline void volk_32u_popcnt_a_sse4_2(uint32_t* ret, const uint32_t value) { + *ret = _mm_popcnt_u32(value); +} + +#endif /*LV_HAVE_SSE4_2*/ + +#endif /*INCLUDED_VOLK_32u_POPCNT_A16_H*/ |