[Mlir-commits] [mlir] [MLIR][XeGPU] Updates XeGPU TensorDescAttr and Refine Gather/Scatter definition. (PR #109144)

[Chao Chen]

================
@@ -36,6 +36,13 @@ gpu.func @test_create_nd_tdesc_vc_4(%src: memref<2x24x32xf32>) {
   gpu.return
 }
 
+// CHECK: gpu.func @test_create_nd_tdesc_vc_5(%[[arg0:.*]]: memref<2x24x32xf32, 3>) {
+gpu.func @test_create_nd_tdesc_vc_5(%src: memref<2x24x32xf32, 3>) {
+  // CHECK: %[[REG:.*]] = xegpu.create_nd_tdesc %arg0[0, 0, 0] : memref<2x24x32xf32, 3> -> !xegpu.tensor_desc<16xf32, #xegpu.block_tdesc_attr<memory_space = slm>>
+  %1 = xegpu.create_nd_tdesc %src[0, 0, 0] : memref<2x24x32xf32, 3> -> !xegpu.tensor_desc<16xf32, #xegpu.block_tdesc_attr<memory_space = slm>>
----------------
chencha3 wrote:

I think this question is also related to your first question. We currently only support using IntegerAttr to express memory space for memref now. So #gpu.address_space doesn't work here. We need a sophisticated mechanism to do the transformations as in MemrefToLLVM and MemrefToSPIRV.  It is still under design and development. The changes in the PR just enable basic functionality support for SLM for current test cases. 

https://github.com/llvm/llvm-project/pull/109144