[Mlir-commits] [mlir] [mlir][gpu]Add GPUToXeVM lowering pipeline pass. (PR #161216)

[Md Abdullah Shahneous Bari]

================
@@ -0,0 +1,149 @@
+// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=workgroup" \
+// RUN: | mlir-runner \
+// RUN:   --shared-libs=%mlir_levelzero_runtime \
+// RUN:   --shared-libs=%mlir_runner_utils \
+// RUN:   --entry-point-result=void \
+// RUN: | FileCheck %s
+
+#a = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], inst_data = [8, 16]>
+#b = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 64], inst_data = [16, 16]>
+#c = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 64], inst_data = [8, 16]>
+#a_prefetch = #xegpu.layout<sg_layout = [32, 1], sg_data = [8, 32], inst_data = [8, 16]>
+#b_prefetch = #xegpu.layout<sg_layout = [4, 8], sg_data = [8, 32], inst_data = [8, 16]>
+module @gemm attributes {gpu.container_module} {
+  func.func @test(%A: memref<256x256xf16>, %B: memref<256x256xf16>, %C: memref<256x256xf32>) -> memref<256x256xf32> attributes {llvm.emit_c_interface} {
+    %c1 = arith.constant 1 : index
+    %c4 = arith.constant 4 : index
+    %c8 = arith.constant 8 : index
+    %c16 = arith.constant 16 : index
+    %c32 = arith.constant 32 : index
+    %c64 = arith.constant 64 : index
+    %c128 = arith.constant 128 : index
+    %c512 = arith.constant 512 : index
+    %A_gpu = gpu.alloc () : memref<256x256xf16>
+    gpu.memcpy %A_gpu, %A : memref<256x256xf16>, memref<256x256xf16>
+    %B_gpu = gpu.alloc () : memref<256x256xf16>
+    gpu.memcpy %B_gpu, %B : memref<256x256xf16>, memref<256x256xf16>
+    %C_gpu = gpu.alloc () : memref<256x256xf32>
+    gpu.memcpy %C_gpu, %C : memref<256x256xf32>, memref<256x256xf32>
+    // NOTE: Here we can't use [8, 64] wi threads following the SG thread layout of [8, 4]. Because runtime will linearize the x dimension first (we need y dimension to be linearized first).
----------------
mshahneo wrote:

Done.

https://github.com/llvm/llvm-project/pull/161216