| 34 | struct SM90_TMA_LOAD_3D |
| 35 | { |
| 36 | DEVICE static void |
| 37 | copy(void const* const desc_ptr, uint64_t& smem_mbar, |
| 38 | void const* const smem_ptr, |
| 39 | int32_t const& crd0, int32_t const& crd1, int32_t const& crd2) |
| 40 | { |
| 41 | uint64_t gmem_int_desc = reinterpret_cast<uint64_t>(desc_ptr); |
| 42 | uint32_t smem_int_mbar = cast_smem_ptr_to_uint(&smem_mbar); |
| 43 | uint32_t smem_int_ptr = cast_smem_ptr_to_uint(smem_ptr); |
| 44 | asm volatile ( |
| 45 | "cp.async.bulk.tensor.3d.shared::cluster.global.mbarrier::complete_tx::bytes" |
| 46 | " [%0], [%1, {%3, %4, %5}], [%2];" |
| 47 | : |
| 48 | : "r"(smem_int_ptr), "l"(gmem_int_desc), "r"(smem_int_mbar), |
| 49 | "r"(crd0), "r"(crd1), "r"(crd2) |
| 50 | : "memory"); |
| 51 | } |
| 52 | }; |
| 53 | |
| 54 | struct SM90_TMA_LOAD_MULTICAST_2D |
nothing calls this directly
no test coverage detected