[Mlir-commits] [mlir] [mlir][python] Add tests for gpu.launch(_func) ops (PR #163883)

[Maksim Levental]

================
@@ -157,3 +158,99 @@ def builder(func: gpu.GPUFuncOp) -> None:
     # CHECK:   %[[VAL_0:.*]] = gpu.global_id  x
     # CHECK:   gpu.return
     # CHECK: }
+
+# CHECK-LABEL: testGPULaunchFuncOp
+ at run
+def testGPULaunchFuncOp():
+    module = Module.create()
+
+    module.operation.attributes["gpu.container_module"] = UnitAttr.get()
+    with InsertionPoint(module.body):
+        gpu_module = gpu.GPUModuleOp("gpu_module")
+        block = gpu_module.bodyRegion.blocks.append()
+
+    with InsertionPoint(block):
+        gpu_func = gpu.GPUFuncOp(
+            FunctionType.get([], []),
+            "kernel",
+            body_builder=lambda func: gpu.return_([]),
+            kernel=True,
+        )
+
+    with InsertionPoint(module.body):
+        host = func.FuncOp(type=FunctionType.get([], []), name="host")
+
+    with InsertionPoint(host.add_entry_block()):
+        c1 = arith.constant(T.index(), 1)
+        grid_sizes = [c1] * 3
+        block_sizes = [c1] * 3
+        sym_ref = SymbolRefAttr.get([gpu_module.sym_name.value, gpu_func.name.value])
+        token_type = Type.parse("!gpu.async.token")
+        token = gpu.wait(async_token=token_type, async_dependencies=[])
+        token = gpu.launch_func(
----------------
makslevental wrote:

The python APIs aren't stable (ie we make no stability guarantees). So basically this same "breakage" occurs whenever we add one of these nicer builders. Also there's a simple "migration path": people can just import the generated original APIs directly from `_gpu_ops_gen`.

https://github.com/llvm/llvm-project/pull/163883