[Mlir-commits] [mlir] [mlir][xegpu] Add support for `vector.extract_strided_slice` XeGPU SIMT distribution with partial offsets. (PR #171512)

[Jianhui Li]

================
@@ -880,6 +901,31 @@ gpu.func @vector_insert_strided_slice_1d(%laneid: index) {
   gpu.return
 }
 
+// CHECK-LABEL: gpu.func @vector_insert_strided_slice_different_ranks
+// CHECK-NEXT:      %[[W:.*]]:3 = gpu.warp_execute_on_lane_0(%{{.*}})[16] -> (vector<64x1xf32>, vector<1xf32>, vector<64x1xf32>) {
+// CHECK-NEXT:        %[[S:.*]] = "some_def"() : () -> vector<16xf32>
+// CHECK-NEXT:        %[[D:.*]] = "some_def"() : () -> vector<64x16xf32>
+// CHECK:             gpu.yield %{{.*}}, %[[S]], %[[D]] : vector<64x16xf32>, vector<16xf32>, vector<64x16xf32>
+// CHECK-NEXT:      }
+// CHECK-NEXT:      %[[T1:.*]] = vector.insert_strided_slice %[[W]]#1, %[[W]]#2
+// CHECK-SAME:        {offsets = [13, 0], strides = [1]} : vector<1xf32> into vector<64x1xf32>
+// CHECK-NEXT:      "some_use"(%[[T1]]) : (vector<64x1xf32>) -> ()
+gpu.func @vector_insert_strided_slice_different_ranks(%laneid: index) {
+  %r = gpu.warp_execute_on_lane_0(%laneid)[16] -> (vector<64x1xf32>) {
+    %0 = "some_def"() : () -> (vector<16xf32>)
+    %1 = "some_def"() : () -> (vector<64x16xf32>)
+    %2 = vector.insert_strided_slice %0, %1 { offsets = [13, 0],  strides = [1],
----------------
Jianhui-Li wrote:

nit: how you make this work by only changing the extract_strided_slice? 

https://github.com/llvm/llvm-project/pull/171512