diff options
Diffstat (limited to 'mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td')
| -rw-r--r-- | mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td | 15 |
1 files changed, 15 insertions, 0 deletions
diff --git a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td index 41fe0ea..a031a75 100644 --- a/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td +++ b/mlir/include/mlir/Dialect/XeGPU/IR/XeGPUOps.td @@ -406,13 +406,28 @@ def XeGPU_CreateDescOp: XeGPU_Op<"create_tdesc", [Pure, ViewLikeOpInterface]> { elements accessed for each offset, default is 1. Example 1. It assumes subgroup size is 4, and accesses a[0], a[16], a[32], a[64] + ``` %a = memref.alloc() : memref<1024xf32> %1 = xegpu.create_tdesc %a[0, 16, 32, 64]: memref<1024xf32> -> TensorDesc<4xf32> + ``` Example 2. It assumes subgroup size is 4, and each workitem access 8 elements. It will access totally 32 data elements: a[0:7], a[16:23], a[32:39], a[64:71] + ``` %0 = memref.alloc() : memref<1024xf32> %1 = xegpu.create_tdesc %0[0, 16, 32, 64] {chunk_size = 8}: memref<1024xf32> -> TensorDesc<4x8xf32> + ``` + + Example 3. It is similar to Example 2, but there is some overlaps among workitems. + It accesses: a[0:7], a[4:11], a[8:15], a[12:19] + ``` + %0 = memref.alloc() : memref<1024xf32> + %1 = xegpu.create_tdesc %0[0, 4, 8, 12] {chunk_size = 8}: memref<1024xf32> -> TensorDesc<4x8xf32> + ``` + + + + }]; let arguments = (ins XeGPU_BaseAddrType: $source, |