| 54 | |
| 55 | template <int32_t N> |
| 56 | DEVICE_FUNCTION void fromFloat(int32_t (&dst)[N], float const (&src)[2 * N]) |
| 57 | { |
| 58 | #pragma unroll |
| 59 | for (int32_t i = 0; i < N; ++i) |
| 60 | { |
| 61 | uint16_t lo, hi; |
| 62 | asm volatile("cvt.rn.f16.f32 %0, %1;" : "=h"(lo) : "f"(src[2 * i + 0])); |
| 63 | asm volatile("cvt.rn.f16.f32 %0, %1;" : "=h"(hi) : "f"(src[2 * i + 1])); |
| 64 | asm volatile("mov.b32 %0, {%1, %2};" : "=r"(dst[i]) : "h"(lo), "h"(hi)); |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | template <int32_t N> |
| 69 | DEVICE_FUNCTION void fromFloat(int32_t (&dst)[N], float const (&src)[4 * N], float scale) |