[Mlir-commits] [mlir] [MLIR][XeGPU] Add unroll patterns for XeGPU (1/N) (PR #137010)

[Charitha Saumya]

================
@@ -0,0 +1,98 @@
+// RUN: mlir-opt --test-xegpu-unrolling-patterns -split-input-file %s | FileCheck %s
+
+gpu.module @test {
+
+  // CHECK-LABEL: test_create_nd_tdesc
+  // CHECK-SAME: [[arg0:%.+]]: memref<24x32xf32>
+  // CHECK-COUNT-6: [[data:%.+]] = xegpu.create_nd_tdesc [[arg0]][{{.*}}] : memref<24x32xf32> -> !xegpu.tensor_desc<8x16xf32>
+  // CHECK: [[cast:%.+]] = builtin.unrealized_conversion_cast
+  // CHECK-SAME: !xegpu.tensor_desc<8x16xf32>, !xegpu.tensor_desc<8x16xf32>,
+  // CHECK-SAME: !xegpu.tensor_desc<8x16xf32>, !xegpu.tensor_desc<8x16xf32>,
+  // CHECK-SAME: !xegpu.tensor_desc<8x16xf32>, !xegpu.tensor_desc<8x16xf32>
+  // CHECK-SAME: to !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>> {__xetile_blocking_inner_block__ = array<i64: 8, 16>, __xetile_blocking_unpack__}
+  gpu.func @test_create_nd_tdesc(%src: memref<24x32xf32>) -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>> {
+    %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    gpu.return %tdesc : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+  }
+
+  //-----
+
+  // CHECK-LABEL: test_update_nd_tdesc
+  // CHECK-SAME: [[arg0:%.+]]: memref<24x32xf32>
+  // CHECK-COUNT-6: [[tdesc:%.+]] = xegpu.create_nd_tdesc [[arg0]][{{.*}}] : memref<24x32xf32> -> !xegpu.tensor_desc<8x16xf32>
+  // CHECK-COUNT-6: [[update:%.+]] = xegpu.update_nd_offset {{.*}} : !xegpu.tensor_desc<8x16xf32>
+  gpu.func @test_update_nd_tdesc(%src: memref<24x32xf32>) -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>> {
+    %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    %update = xegpu.update_nd_offset %tdesc, [0, 16] : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    gpu.return %update : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+  }
+
+  //-----
+
+  // CHECK-LABEL: test_prefetch_nd_tdesc
+  // CHECK-SAME: [[arg0:%.+]]: memref<24x32xf32>
+  // CHECK-COUNT-6: [[tdesc:%.+]] = xegpu.create_nd_tdesc [[arg0]][{{.*}}] : memref<24x32xf32> -> !xegpu.tensor_desc<8x16xf32>
+  // CHECK-COUNT-6: xegpu.prefetch_nd {{.*}} : !xegpu.tensor_desc<8x16xf32>
+  gpu.func @test_prefetch_nd_tdesc(%src: memref<24x32xf32>) {
+    %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    xegpu.prefetch_nd %tdesc : !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    gpu.return
+  }
+
+  //-----
+  // CHECK-LABEL: test_load_nd
+  // CHECK-SAME: [[arg0:%.+]]: memref<24x32xf32>
+  // CHECK-COUNT-6: [[tdesc:%.+]] = xegpu.create_nd_tdesc [[arg0]][{{.*}}] : memref<24x32xf32> -> !xegpu.tensor_desc<8x16xf32>
+  // CHECK-COUNT-6: [[ld:%.+]] = xegpu.load_nd {{.*}}  : !xegpu.tensor_desc<8x16xf32> -> vector<8x16xf32>
+  // CHECK-COUNT-6: [[insert:%.+]] = vector.insert_strided_slice {{.*}} : vector<8x16xf32> into vector<24x32xf32>
+  gpu.func @test_load_nd(%src: memref<24x32xf32>) -> vector<24x32xf32> {
+    %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    %ld = xegpu.load_nd %tdesc: !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>> -> vector<24x32xf32>
+    gpu.return %ld : vector<24x32xf32>
+  }
+
+  //-----
+
+  // CHECK-LABEL: test_store_nd
+  // CHECK-SAME: [[arg0:%.+]]: memref<24x32xf32>
+  // CHECK-COUNT-6: [[tdesc:%.+]] = xegpu.create_nd_tdesc [[arg0]][{{.*}}] : memref<24x32xf32> -> !xegpu.tensor_desc<8x16xf32>
+  // CHECK-COUNT-6: xegpu.store_nd {{.*}}  : vector<8x16xf32>, !xegpu.tensor_desc<8x16xf32>
+  gpu.func @test_store_nd(%src: memref<24x32xf32>) {
+    %tdesc = xegpu.create_nd_tdesc %src[0, 0] : memref<24x32xf32> -> !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    %data = arith.constant dense<9.0> : vector<24x32xf32>
+    xegpu.store_nd %data, %tdesc: vector<24x32xf32>, !xegpu.tensor_desc<24x32xf32, #xegpu.layout<inst_data = [8, 16]>>
+    gpu.return
+  }
+
+  //-----
+
+  // CHECK-LABEL: test_createNd_loadNd_storeNd
+  // CHECK-SAME: [[arg0:%.+]]: memref<24x32xf32>
+  //CHECK-COUNT-6: [[tdesc:%.+]] = xegpu.create_nd_tdesc [[arg0]][{{.*}}] : memref<24x32xf32> -> !xegpu.tensor_desc<8x16xf32>
+  //CHECK-COUNT-6: [[data:%.+]] = xegpu.load_nd {{.*}}  : !xegpu.tensor_desc<8x16xf32> -> vector<8x16xf32>
+  //CHECK-COUNT-6: [[insert:%.+]] = vector.insert_strided_slice {{.*}} : vector<8x16xf32> into vector<24x32xf32>
+  //CHECK: [[add:%.+]] = arith.addf {{.*}} : vector<24x32xf32>
----------------
charithaintc wrote:

in this case, what happens with the addi, doesn't it need unrolling as well?

https://github.com/llvm/llvm-project/pull/137010