chore: import upstream snapshot with attribution

tests/python/tirx-base/test_tir_ptx_cp_async.py

@@ -0,0 +1,81 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+import numpy as np
+import pytest
+
+import tvm
+import tvm.testing
+from tvm.script import tirx as T
+from tvm.testing import env
+
+
+@T.prim_func(s_tir=True)
+def ptx_cp_async(A: T.Buffer((32, 128), "float16"), B: T.Buffer((32, 128), "float16")) -> None:
+    T.func_attr({"global_symbol": "default_function", "tirx.noalias": True})
+    bx = T.env_thread("blockIdx.x")
+    tx = T.env_thread("threadIdx.x")
+    T.launch_thread(bx, 1)
+    T.launch_thread(tx, 32)
+    with T.sblock():
+        A_shared = T.sblock_alloc_buffer([32, 128], "float16", scope="shared")
+        T.reads(A[0:32, 0:128])
+        T.writes(B[0:32, 0:128])
+
+        for i in range(16):
+            T.evaluate(
+                T.ptx.cp_async.legacy(
+                    A_shared.data, tx * 128 + 8 * i, A.data, tx * 128 + 8 * i, 16, dtype="float16"
+                )
+            )
+
+        # TODO(masahi): Remove dtype requirement from TVMScript parser
+        T.evaluate(T.ptx.cp_async.commit_group(dtype=""))
+        T.evaluate(T.ptx.cp_async.wait_group(0, dtype=""))
+
+        for i in range(128):
+            B[tx, i] = A_shared[tx, i]
+
+
+@pytest.mark.gpu
+@pytest.mark.skipif(not env.has_cuda_compute(8), reason="need cuda compute >= 8.0")
+def test_ptx_cp_async():
+    f = ptx_cp_async
+
+    mod = tvm.compile(f, target="cuda")
+    A_np = np.random.rand(32, 128).astype("float16")
+    B_np = np.zeros((32, 128)).astype("float16")
+
+    def run_and_check():
+        dev = tvm.cuda(0)
+        A_nd = tvm.runtime.tensor(A_np, device=dev)
+        B_nd = tvm.runtime.tensor(B_np, device=dev)
+        mod(A_nd, B_nd)
+        tvm.testing.assert_allclose(B_nd.numpy(), A_np)
+
+    tvm.testing.run_with_gpu_lock(run_and_check)
+
+
+# Note: tests for the indexed barrier API (`create_barriers`,
+# `ptx_init_barrier_thread_count`, `ptx_arrive_barrier`, `ptx_wait_barrier`,
+# `ptx_cp_async_barrier`, `ptx_arrive_barrier_expect_tx`) were removed —
+# fork uses `ptx_mbarrier_*` instead and those intrinsics have no
+# users elsewhere in this codebase.
+
+
+if __name__ == "__main__":
+    test_ptx_cp_async()