root / volk / include / volk / volk_64u_popcnt_aligned16.h @ 74f206ed
History | View | Annotate | Download (2 kB)
| 1 | #ifndef INCLUDED_VOLK_64u_POPCNT_ALIGNED16_H
|
|---|---|
| 2 | #define INCLUDED_VOLK_64u_POPCNT_ALIGNED16_H
|
| 3 | |
| 4 | #include <stdio.h> |
| 5 | #include <inttypes.h> |
| 6 | |
| 7 | |
| 8 | #if LV_HAVE_GENERIC
|
| 9 | |
| 10 | |
| 11 | static inline void volk_64u_popcnt_aligned16_generic(uint64_t* ret, const uint64_t value) { |
| 12 | |
| 13 | const uint32_t* valueVector = (const uint32_t*)&value; |
| 14 | |
| 15 | // This is faster than a lookup table
|
| 16 | uint32_t retVal = valueVector[0];
|
| 17 | |
| 18 | retVal = (retVal & 0x55555555) + (retVal >> 1 & 0x55555555); |
| 19 | retVal = (retVal & 0x33333333) + (retVal >> 2 & 0x33333333); |
| 20 | retVal = (retVal + (retVal >> 4)) & 0x0F0F0F0F; |
| 21 | retVal = (retVal + (retVal >> 8));
|
| 22 | retVal = (retVal + (retVal >> 16)) & 0x0000003F; |
| 23 | uint64_t retVal64 = retVal; |
| 24 | |
| 25 | retVal = valueVector[1];
|
| 26 | retVal = (retVal & 0x55555555) + (retVal >> 1 & 0x55555555); |
| 27 | retVal = (retVal & 0x33333333) + (retVal >> 2 & 0x33333333); |
| 28 | retVal = (retVal + (retVal >> 4)) & 0x0F0F0F0F; |
| 29 | retVal = (retVal + (retVal >> 8));
|
| 30 | retVal = (retVal + (retVal >> 16)) & 0x0000003F; |
| 31 | retVal64 += retVal; |
| 32 | |
| 33 | *ret = retVal64; |
| 34 | |
| 35 | } |
| 36 | |
| 37 | #endif /*LV_HAVE_GENERIC*/ |
| 38 | |
| 39 | #if LV_HAVE_SSE4_2
|
| 40 | |
| 41 | #include <nmmintrin.h> |
| 42 | |
| 43 | static inline void volk_64u_popcnt_aligned16_sse4_2(uint64_t* ret, const uint64_t value) { |
| 44 | #if LV_64
|
| 45 | *ret = _mm_popcnt_u64(value); |
| 46 | #else
|
| 47 | const uint32_t* valueVector = (const uint32_t*)&value; |
| 48 | |
| 49 | // This is faster than a lookup table
|
| 50 | uint32_t retVal = valueVector[0];
|
| 51 | |
| 52 | retVal = (retVal & 0x55555555) + (retVal >> 1 & 0x55555555); |
| 53 | retVal = (retVal & 0x33333333) + (retVal >> 2 & 0x33333333); |
| 54 | retVal = (retVal + (retVal >> 4)) & 0x0F0F0F0F; |
| 55 | retVal = (retVal + (retVal >> 8));
|
| 56 | retVal = (retVal + (retVal >> 16)) & 0x0000003F; |
| 57 | uint64_t retVal64 = retVal; |
| 58 | |
| 59 | retVal = valueVector[1];
|
| 60 | retVal = (retVal & 0x55555555) + (retVal >> 1 & 0x55555555); |
| 61 | retVal = (retVal & 0x33333333) + (retVal >> 2 & 0x33333333); |
| 62 | retVal = (retVal + (retVal >> 4)) & 0x0F0F0F0F; |
| 63 | retVal = (retVal + (retVal >> 8));
|
| 64 | retVal = (retVal + (retVal >> 16)) & 0x0000003F; |
| 65 | retVal64 += retVal; |
| 66 | |
| 67 | *ret = retVal64; |
| 68 | |
| 69 | #endif
|
| 70 | } |
| 71 | |
| 72 | #endif /*LV_HAVE_SSE4_2*/ |
| 73 | |
| 74 | #endif /*INCLUDED_VOLK_64u_POPCNT_ALIGNED16_H*/ |