| 14 | struct SM90_TMA_LOAD_2D |
| 15 | { |
| 16 | DEVICE static void |
| 17 | copy(void const* const desc_ptr, uint64_t& smem_mbar, |
| 18 | void const* const smem_ptr, |
| 19 | int32_t const& crd0, int32_t const& crd1) |
| 20 | { |
| 21 | uint64_t gmem_int_desc = reinterpret_cast<uint64_t>(desc_ptr); |
| 22 | uint32_t smem_int_mbar = cast_smem_ptr_to_uint(&smem_mbar); |
| 23 | uint32_t smem_int_ptr = cast_smem_ptr_to_uint(smem_ptr); |
| 24 | asm volatile ( |
| 25 | "cp.async.bulk.tensor.2d.shared::cluster.global.mbarrier::complete_tx::bytes" |
| 26 | " [%0], [%1, {%3, %4}], [%2];" |
| 27 | : |
| 28 | : "r"(smem_int_ptr), "l"(gmem_int_desc), "r"(smem_int_mbar), |
| 29 | "r"(crd0), "r"(crd1) |
| 30 | : "memory"); |
| 31 | } |
| 32 | }; |
| 33 | |
| 34 | struct SM90_TMA_LOAD_3D |
nothing calls this directly
no test coverage detected