[Mlir-commits] [mlir] [MLIR][XeGPU]Enhance Pack/Unpack for XeGPUUnroll (PR #163459)

[Jianhui Li]

================
@@ -145,8 +145,13 @@ XeGPUBlockingPass::getTileShape(const T &operandOrResult) const {
   xegpu::DistributeLayoutAttr layout =
       xegpu::getDistributeLayoutAttr(operandOrResult);
   if (layout && layout.isForSubgroup()) {
-    if (!layout.getEffectiveInstDataAsInt().empty())
-      return layout.getEffectiveInstDataAsInt();
+    if (!layout.getEffectiveInstDataAsInt().empty()) {
+      SmallVector<int64_t> instData = layout.getEffectiveInstDataAsInt();
+      // Remove all leading unit dimensions from inst_data
+      while (!instData.empty() && instData.front() == 1)
----------------
Jianhui-Li wrote:

It doesn't work for all xegpu operations. Say load_nd would need 2 dim inst_data since its output is 2d. Should this logic happen per operation after they getTileShape? 
 

https://github.com/llvm/llvm-project/pull/163459