[Mlir-commits] [mlir] [MLIR][XeGPU] Add sg layout propagation (PR #170879)

[Tuomas Kärnä]

================
@@ -0,0 +1,131 @@
+// RUN: mlir-opt -xevm-attach-target='chip=pvc' -xegpu-propagate-layout="layout-kind=sg" -split-input-file %s | FileCheck %s
+
+gpu.module @test {
+  // CHECK-LABEL: store_nd
+  // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>
+  func.func @store_nd(%src: memref<256x128xf32>) {
+    // CHECK: %[[TDESC:.*]] = xegpu.create_nd_tdesc %[[ARG_0]] : memref<256x128xf32>
+    // CHECK-SAME: -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>>
+    // CHECK: %[[LOAD:.*]] = xegpu.load_nd %[[TDESC]] <{layout = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>}>
+    // CHECK-SAME: {layout_result_0 = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>}
+    // CHECK-SAME: : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>>
+    // CHECK-SAME: -> vector<256x128xf32>
+    // CHECK: xegpu.store_nd %[[LOAD]], %[[TDESC]] <{layout = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>}>
+    // CHECK-SAME: : vector<256x128xf32>, !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>>
+    %tdesc = xegpu.create_nd_tdesc %src : memref<256x128xf32> -> !xegpu.tensor_desc<256x128xf32>
+    %load = xegpu.load_nd %tdesc : !xegpu.tensor_desc<256x128xf32> -> vector<256x128xf32>
+    xegpu.store_nd %load, %tdesc {layout = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>}
----------------
tkarna wrote:

Shouldn't this be the anchor layout, i.e. `<{layout = ...}>` instead of attr dict `{layout = ...}`

https://github.com/llvm/llvm-project/pull/170879