[Mlir-commits] [mlir] [MLIR] Add XeGPU dialect for Intel GPU (PR #78483)

[Mehdi Amini]

================
@@ -0,0 +1,170 @@
+//===- XeGPUTypes.td - XeGPU dialect types definition -------*- tablegen -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_DIALECT_XEGPU_IR_XEGPUTYPES_TD
+#define MLIR_DIALECT_XEGPU_IR_XEGPUTYPES_TD
+
+include "mlir/IR/BuiltinTypes.td"
+
+include "mlir/Dialect/XeGPU/IR/XeGPUAttrs.td"
+include "mlir/Dialect/XeGPU/IR/XeGPUDialect.td"
+
+// An Integer array attribute with fixed 2 elements.
+def XeGPU_IntType: AnyTypeOf<[I1, I8, I16, I32, I64, SI1, SI8, SI16, SI32, SI64, UI1, UI8, UI16, UI32, UI64]>;
+def XeGPU_FloatType: AnyTypeOf<[F16, F32, F64, BF16, TF32]>;
+def XeGPU_ScalarType: AnyTypeOf<[XeGPU_IntType, XeGPU_FloatType]>;
+def XeGPU_BaseAddrType: AnyTypeOf<[MemRefRankOf<[XeGPU_ScalarType], [1, 2]>, UI64, UI32, I64, I32]>;
+def XeGPU_DpasOpType: VectorOfRankAndType<[2, 3], [XeGPU_ScalarType]>;
+// def XeGPU_OffsetType: AnyTypeOf<[VectorOfRankAndType<[1], [Index]>, Index]>;
+def XeGPU_OffsetType: VectorOfRankAndType<[1], [Index]>;
+def XeGPU_MaskType: AnyTypeOf<[VectorOfRankAndType<[1,2], [I1]>, I1]>;
+def XeGPU_ValueType: AnyTypeOf<[VectorOfRankAndType<[1,2,3,4], [XeGPU_ScalarType]>, XeGPU_ScalarType]>;
+
+def XeGPU_Vector2DType: VectorOfRankAndType<[2], [XeGPU_ScalarType]>;
+
+// common base class for types in XeGPU dialect
+class XeGPUTypeDef<string name, string typeMnemonic, list<Trait> traits = [],
+                   string baseCppClass = "::mlir::Type">
+    : TypeDef<XeGPU_Dialect, name, traits, baseCppClass> {
+  let mnemonic = typeMnemonic;
+}
+
+// TensorDesc contains dim and element type info
+def XeGPU_TensorDesc: XeGPUTypeDef<"TensorDesc", "tensor_desc",
+        [ShapedTypeInterface], "::mlir::TensorType"> {
+  let summary = "TensorDesc describing all kinds of memory and tensors, including scatter tensor, 1d tensor, 2d tensor, … 5d tensor";
+  let description = [{
+    TensorDesc is a type designed to describe all kinds of memory, scatter tensor, 1d tensor, 2d tensor, … 5d tensor.
+    Different with the builtin tensor type in MLIR, it essentially only contains the meta data that describes a region
+    of the intereted data as well as some features that are unique to intel hardware features. It does not hold the data
+    directly by itself. It is designed to mainly support 2d block load/store and DPAS (matrix multiplication instruction)
+    on Intel GPU. It majorly encodes the following information:
+
+    * shape:  the sizes/shape of the intereted data block, e.g., 8x16 means 8 rows
+              and each row contains 16 continious data element. The rows could be
+              either continuous or not, depends on whether the encoding attribute
+              is set or not.
+    * element_type: the data type of the data element, e.g., f16, f32.
+
+    Similar to the builtin tensor, it also provides an optinal attribute to encoding the following information via the TensorDescAttr object:
+    * memory_scope (xegpu::MemoryScope): [optional] where the data is located, global memory or shared memory. It is default to Global.
+    * array_length (int): [optional] The number of continuous blocks with size as `shape`,
+               that will be loaded by block load at a time. It is default to 1.
+    * boundary_check (bool): [optional] indicates whether the operation detects the boundary and pads with zero for out-of-boundary access (default)
+    * scattered (xegpu::ScatteredAttr): [optional] It is a unit attribute. It can be only set as empty or ScatteredAttr, indicating
+                whether the TensorDesc is blocked (empty, default) or scattered (ScatteredAttr). If it is
+                blocked, rows are continuous in the correspoding dimention, otherwise, rows may be not continous.
+    * mapping (xegpu::SubGroupMapAttr): [optional] Used to guide compiler to distribute the workload into different threads. It is default to none.
+
+    For convinience, its attribute field can also take either "ScatteredAttr" or "SubGroupMapAttr" directly if and only
----------------
joker-eph wrote:

```suggestion
    For convenience, its attribute field can also take either "ScatteredAttr" or "SubGroupMapAttr" directly if and only
```

https://github.com/llvm/llvm-project/pull/78483