llvm
4695c84c - [MLIR][XeGPU] Extend op definitions to support 3D+: load_nd, store_nd, prefetch_nd (#199811)

Commit
2 days ago
[MLIR][XeGPU] Extend op definitions to support 3D+: load_nd, store_nd, prefetch_nd (#199811) **Summary** Extend xegpu.load_nd, xegpu.store_nd, and xegpu.prefetch_nd operations to support 3D and higher-dimensional tensor descriptors with batch dimensions, enabling batched memory operations for workloads like [4, 8, 16] tensor loads/stores. **Changes** - Verifiers: Removed rank > 2 checks in LoadNdOp::verify() and StoreNdOp::verify() to allow 3D+ tensor descriptors - Documentation: Added comprehensive documentation explaining: Tensor descriptors can be 1D, 2D, 3D, or higher dimensional; Batch dimensions (leading dimensions) are unrolled to unit dimensions during lowering; Operations execute at 2D granularity at subgroup level to match 2D block IO hardware; Examples of 3D operations - Tests: Added unit tests for 3D operations (load_nd_3d, store_nd_3d, prefetch_nd_3d) --------- Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
Author
Parents
Loading