| 1 | #include "common.cuh" |
| 2 | |
| 3 | static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){ |
| 4 | const block_q4_0 * x = (const block_q4_0 *) vx; |
| 5 | |
| 6 | const float d = x[ib].d; |
| 7 | |
| 8 | const int vui = x[ib].qs[iqs]; |
| 9 | |
| 10 | v.x = vui & 0xF; |
| 11 | v.y = vui >> 4; |
| 12 | |
| 13 | v.x = (v.x - 8.0f) * d; |
| 14 | v.y = (v.y - 8.0f) * d; |
| 15 | } |
| 16 | |
| 17 | static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, float2 & v){ |
| 18 | const block_q4_1 * x = (const block_q4_1 *) vx; |
| 19 | |
| 20 | const float2 dm = __half22float2(x[ib].dm); |
| 21 | |
| 22 | const int vui = x[ib].qs[iqs]; |
| 23 | |
| 24 | v.x = vui & 0xF; |
| 25 | v.y = vui >> 4; |
| 26 | |
| 27 | v.x = (v.x * dm.x) + dm.y; |
| 28 | v.y = (v.y * dm.x) + dm.y; |
| 29 | } |
| 30 | |
| 31 | static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, float2 & v){ |
| 32 | const block_q5_0 * x = (const block_q5_0 *) vx; |
| 33 | |
| 34 | const float d = x[ib].d; |
| 35 | |
| 36 | uint32_t qh; |
| 37 | memcpy(&qh, x[ib].qh, sizeof(qh)); |
| 38 | |
| 39 | const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; |
| 40 | const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; |
| 41 | |
| 42 | v.x = ((x[ib].qs[iqs] & 0xf) | xh_0); |
| 43 | v.y = ((x[ib].qs[iqs] >> 4) | xh_1); |
| 44 | |
| 45 | v.x = (v.x - 16.0f) * d; |
| 46 | v.y = (v.y - 16.0f) * d; |
| 47 | } |
| 48 | |
| 49 | static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, float2 & v){ |
| 50 | const block_q5_1 * x = (const block_q5_1 *) vx; |
| 51 | |
| 52 | const float2 dm = __half22float2(x[ib].dm); |
| 53 | |
| 54 | uint32_t qh; |
| 55 | memcpy(&qh, x[ib].qh, sizeof(qh)); |
| 56 | |
| 57 | const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; |
| 58 | const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; |
| 59 | |
| 60 | v.x = ((x[ib].qs[iqs] & 0xf) | xh_0); |
| 61 | v.y = ((x[ib].qs[iqs] >> 4) | xh_1); |
| 62 | |
| 63 | v.x = (v.x * dm.x) + dm.y; |
| 64 | v.y = (v.y * dm.x) + dm.y; |
| 65 | } |
| 66 | |
| 67 | static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, float2 & v){ |
| 68 | const block_q8_0 * x = (const block_q8_0 *) vx; |
| 69 | |
| 70 | const float d = x[ib].d; |
| 71 | |
| 72 | v.x = x[ib].qs[iqs + 0]; |
| 73 | v.y = x[ib].qs[iqs + 1]; |
| 74 | |
| 75 | v.x *= d; |
| 76 | v.y *= d; |
| 77 | } |
| 78 | |