The CpAsyncBulkTensor instruction submodule.
| 172 | |
| 173 | |
| 174 | class CpAsyncBulkTensorNamespace: |
| 175 | """The CpAsyncBulkTensor instruction submodule.""" |
| 176 | |
| 177 | def __init__(self): |
| 178 | self.g2c = _op_wrapper(_cuda_op.ptx_cp_async_bulk_tensor_global_to_cluster) |
| 179 | self.g2c_tile_gather4 = _op_wrapper( |
| 180 | _cuda_op.ptx_cp_async_bulk_tensor_tile_gather4_global_to_cluster |
| 181 | ) |
| 182 | self.s2g = _op_wrapper(_cuda_op.ptx_cp_async_bulk_tensor_shared_to_global) |
| 183 | self.s2g_reduce = _op_wrapper(_cuda_op.ptx_cp_async_bulk_tensor_shared_to_global_reduce) |
| 184 | self.g2c_prefetch = _op_wrapper( |
| 185 | _cuda_op.ptx_cp_async_bulk_tensor_global_to_cluster_prefetch |
| 186 | ) |
| 187 | |
| 188 | @staticmethod |
| 189 | def g2c_bar_addr( |
| 190 | dim, |
| 191 | dst_ptr, |
| 192 | bar_addr, |
| 193 | tensormap_addr, |
| 194 | cta_mask, |
| 195 | cta_group, |
| 196 | cache_hint, |
| 197 | *coords, |
| 198 | cache_policy=None, |
| 199 | ): |
| 200 | _cuda_op._choice("cta_group", cta_group, _cuda_op._TCGEN05_CTA_GROUP) |
| 201 | cache_policy, has_cache_policy = _cuda_op._resolve_cache_policy(cache_hint, cache_policy) |
| 202 | return _tir_op.call_intrin( |
| 203 | "", |
| 204 | "tirx.ptx_cp_async_bulk_tensor_global_to_cluster", |
| 205 | dim, |
| 206 | dst_ptr, |
| 207 | bar_addr, |
| 208 | tensormap_addr, |
| 209 | cta_mask, |
| 210 | cta_group, |
| 211 | cache_policy, |
| 212 | int(has_cache_policy), |
| 213 | 1, |
| 214 | *coords, |
| 215 | ) |
| 216 | |
| 217 | @staticmethod |
| 218 | def g2c_tile_gather4_bar_addr( |
| 219 | dim, |
| 220 | dst_ptr, |
| 221 | bar_addr, |
| 222 | tensormap_addr, |
| 223 | cta_mask, |
| 224 | cta_group, |
| 225 | cache_hint, |
| 226 | *coords, |
| 227 | cache_policy=None, |
| 228 | ): |
| 229 | _cuda_op._choice("cta_group", cta_group, _cuda_op._TCGEN05_CTA_GROUP) |
| 230 | cache_policy, has_cache_policy = _cuda_op._resolve_cache_policy(cache_hint, cache_policy) |
| 231 | return _tir_op.call_intrin( |
no outgoing calls
no test coverage detected
searching dependent graphs…