[mlir][xegpu] Refine layout assignment in XeGPU SIMT distribution.

mlir/include/mlir/Dialect/XeGPU/Transforms/Passes.td

@@ -27,10 +27,18 @@ def XeGPUSubgroupDistribute : Pass<"xegpu-subgroup-distribute"> {
   }];
   let dependentDialects = ["memref::MemRefDialect", "xegpu::XeGPUDialect",
                            "vector::VectorDialect"];
-  let options = [Option<
-      "printOnly", "print-analysis-only", "bool",
-      /*default=*/"false",
-      "Print the result of the subgroup map propagation analysis and exit.">];
+}
+
+def XeGPULayoutPropagate : Pass<"xegpu-layout-propagate"> {
+  let summary = "Propagate XeGPU layout information";
+  let description = [{
+    This pass propagates the XeGPU layout information accross ops. Starting
+    from a set of anchor operations (e.g. `dpas`, `store_nd`), this will
+    propagate the layouts required for operands and results to the producers or
+    consumers.
+  }];
+  let dependentDialects = ["memref::MemRefDialect", "xegpu::XeGPUDialect",
+                           "vector::VectorDialect"];
 }
 
 def XeGPUWgToSgDistribute : Pass<"xegpu-wg-to-sg-distribute"> {

mlir/include/mlir/Dialect/XeGPU/Utils/XeGPUUtils.h

@@ -24,6 +24,20 @@ class LayoutAttr;
 class TensorDescType;
 } // namespace xegpu
 
+namespace xegpu {
+/// HW dependent constants.
+/// TODO: These constants should be queried from the target information.
+namespace targetinfo {
+constexpr unsigned subgroupSize = 16; // How many lanes in a subgroup.
+/// If DPAS A or B operands have low precision element types they must be packed
+/// according to the following sizes.
+constexpr unsigned packedSizeInBitsForDefault =
+    16; // Minimum packing size per register for DPAS A.
+constexpr unsigned packedSizeInBitsForDpasB =
+    32; // Minimum packing size per register for DPAS B.
+} // namespace targetinfo
+} // namespace xegpu
+
 namespace xegpu {
 
 /// If tensor descriptor has a layout attribute it is used in SIMT mode.

mlir/lib/Dialect/XeGPU/Transforms/CMakeLists.txt

@@ -4,6 +4,7 @@ add_mlir_dialect_library(MLIRXeGPUTransforms
   XeGPUSubgroupDistribute.cpp
   XeGPUUnroll.cpp
   XeGPUWgToSgDistribute.cpp
+  XeGPULayoutPropagate.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/XeGPU