| 18 | namespace cuda { |
| 19 | |
| 20 | static inline unsigned bitCount(unsigned v) { |
| 21 | v = v - ((v >> 1U) & 0x55555555U); |
| 22 | v = (v & 0x33333333U) + ((v >> 2U) & 0x33333333U); |
| 23 | return (((v + (v >> 4U)) & 0xF0F0F0FU) * 0x1010101U) >> 24U; |
| 24 | } |
| 25 | |
| 26 | using af::dim4; |
| 27 |