Parameterize test_link_params.

Author: Andrew Reusch

tests/python/unittest/test_link_params.py

@@ -39,15 +39,15 @@
 
 
 # The data types that are linkable.
-LINKABLE_DTYPES = (
-    [f"uint{b}" for b in (8, 16, 32, 64)]
-    + [f"int{b}" for b in (8, 16, 32, 64)]
-    + ["float32", "float64"]
+linkable_dtype = tvm.testing.parameter(
+    *([f"uint{b}" for b in (8, 16, 32, 64)]
+      + [f"int{b}" for b in (8, 16, 32, 64)]
+      + ["float32", "float64"])
 )
 
 
 def dtype_info(dtype):
-    """Lookup numpy type info for the given string dtype (of LINKABLE_DTYPES above)."""
+    """Lookup numpy type info for the given string dtype (of linkable_dtype params above)."""
     if "int" in dtype:
         return np.iinfo(getattr(np, dtype))
     else:
@@ -181,54 +181,53 @@ def _add_decl(name, dtype):
 
 
 @tvm.testing.requires_llvm
-def test_llvm_link_params():
-    for dtype in LINKABLE_DTYPES:
-        ir_mod, param_init = _make_mod_and_params(dtype)
-        rand_input = _make_random_tensor(dtype, INPUT_SHAPE)
-        main_func = ir_mod["main"]
-        target = "llvm --runtime=c --system-lib --link-params"
-        with tvm.transform.PassContext(opt_level=3):
-            lib = tvm.relay.build(ir_mod, target, params=param_init)
-
-            # NOTE: Need to export_library() and load_library() to link all the Module(llvm, ...)
-            # against one another.
-            temp_dir = tempfile.mkdtemp()
-            export_file = os.path.join(temp_dir, "lib.so")
-            lib.lib.export_library(export_file)
-            mod = tvm.runtime.load_module(export_file)
-            assert set(lib.params.keys()) == {"p0", "p1"}  # NOTE: op folded
-            assert mod.get_function("TVMSystemLibEntryPoint") != None
-
-            graph = json.loads(lib.graph_json)
-            for p in lib.params:
-                _verify_linked_param(dtype, lib, mod, graph, p) or found_one
-
-            # Wrap in function to explicitly deallocate the runtime.
-            def _run_linked(lib, mod):
-                graph_json, _, _ = lib
-                graph_rt = tvm.contrib.graph_executor.create(graph_json, mod, tvm.cpu(0))
-                graph_rt.set_input("rand_input", rand_input)  # NOTE: params not required.
-                graph_rt.run()
-                return graph_rt.get_output(0)
-
-            linked_output = _run_linked(lib, mod)
-
-        with tvm.transform.PassContext(opt_level=3):
-            lib = tvm.relay.build(ir_mod, "llvm --system-lib", params=param_init)
-
-            def _run_unlinked(lib):
-                graph_json, mod, lowered_params = lib
-                graph_rt = tvm.contrib.graph_executor.create(graph_json, mod, tvm.cpu(0))
-                graph_rt.set_input("rand_input", rand_input, **lowered_params)
-                graph_rt.run()
-                return graph_rt.get_output(0)
-
-            unlinked_output = _run_unlinked(lib)
-
-        if "int" in dtype:
-            np.testing.assert_equal(unlinked_output.numpy(), linked_output.numpy())
-        else:
-            np.testing.assert_allclose(unlinked_output.numpy(), linked_output.numpy())
+def test_llvm_link_params(linkable_dtype):
+    ir_mod, param_init = _make_mod_and_params(linkable_dtype)
+    rand_input = _make_random_tensor(linkable_dtype, INPUT_SHAPE)
+    main_func = ir_mod["main"]
+    target = "llvm --runtime=c --system-lib --link-params"
+    with tvm.transform.PassContext(opt_level=3):
+        lib = tvm.relay.build(ir_mod, target, params=param_init)
+
+        # NOTE: Need to export_library() and load_library() to link all the Module(llvm, ...)
+        # against one another.
+        temp_dir = tempfile.mkdtemp()
+        export_file = os.path.join(temp_dir, "lib.so")
+        lib.lib.export_library(export_file)
+        mod = tvm.runtime.load_module(export_file)
+        assert set(lib.params.keys()) == {"p0", "p1"}  # NOTE: op folded
+        assert mod.get_function("TVMSystemLibEntryPoint") != None
+
+        graph = json.loads(lib.graph_json)
+        for p in lib.params:
+            _verify_linked_param(linkable_dtype, lib, mod, graph, p) or found_one
+
+        # Wrap in function to explicitly deallocate the runtime.
+        def _run_linked(lib, mod):
+            graph_json, _, _ = lib
+            graph_rt = tvm.contrib.graph_executor.create(graph_json, mod, tvm.cpu(0))
+            graph_rt.set_input("rand_input", rand_input)  # NOTE: params not required.
+            graph_rt.run()
+            return graph_rt.get_output(0)
+
+        linked_output = _run_linked(lib, mod)
+
+    with tvm.transform.PassContext(opt_level=3):
+        lib = tvm.relay.build(ir_mod, "llvm --system-lib", params=param_init)
+
+        def _run_unlinked(lib):
+            graph_json, mod, lowered_params = lib
+            graph_rt = tvm.contrib.graph_executor.create(graph_json, mod, tvm.cpu(0))
+            graph_rt.set_input("rand_input", rand_input, **lowered_params)
+            graph_rt.run()
+            return graph_rt.get_output(0)
+
+        unlinked_output = _run_unlinked(lib)
+
+    if "int" in linkable_dtype:
+        np.testing.assert_equal(unlinked_output.numpy(), linked_output.numpy())
+    else:
+        np.testing.assert_allclose(unlinked_output.numpy(), linked_output.numpy())
 
 
 def _get_c_datatype(dtype):
@@ -263,137 +262,134 @@ def _format_c_value(dtype, width, x):
 HEX_NUM_RE = re.compile(r"[+\-]?(?:(?:0x[0-9A-Fa-f.p+-]+)|(?:INFINITY)|(?:NAN))")
 
 
-def test_c_link_params():
+def test_c_link_params(linkable_dtype):
     temp_dir = utils.tempdir()
-    for dtype in LINKABLE_DTYPES:
-        mod, param_init = _make_mod_and_params(dtype)
-        rand_input = _make_random_tensor(dtype, INPUT_SHAPE)
-        main_func = mod["main"]
-        target = "c --link-params"
-        with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
-            lib = tvm.relay.build(mod, target, params=param_init)
-            assert set(lib.params.keys()) == {"p0", "p1"}  # NOTE: op folded
-
-            src = lib.lib.get_source()
-            lib.lib.save(temp_dir.relpath("test.c"), "c")
-            c_dtype = _get_c_datatype(dtype)
-            src_lines = src.split("\n")
-            param = lib.params["p0"].numpy().reshape(np.prod(KERNEL_SHAPE))
-            param_def = f"static const {c_dtype} __tvm_param__p0[{np.prod(param.shape)}] = {{"
-            for i, line in enumerate(src_lines):
-                if line == param_def:
-                    i += 1
-                    break
-            else:
-                assert False, f'did not find parameter definition "{param_def}":\n{src}'
-
-            cursor = 0
-            width = dtype_info(dtype).bits // 4 + 2
-            if dtype.startswith("int"):
-                width += 1  # Account for sign
-
-            while "};" not in src_lines[i]:
-                for match in HEX_NUM_RE.finditer(src_lines[i]):
-                    assert match.group() == _format_c_value(dtype, width, param[cursor]), (
-                        f'p0 byte {cursor}: want "{_format_c_value(dtype, width, param[cursor])}" got '
-                        f'"{match.group(0)}"; full p0 follows:\n{src}'
-                    )
-                    cursor += 1
+    mod, param_init = _make_mod_and_params(linkable_dtype)
+    rand_input = _make_random_tensor(linkable_dtype, INPUT_SHAPE)
+    main_func = mod["main"]
+    target = "c --link-params"
+    with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
+        lib = tvm.relay.build(mod, target, params=param_init)
+        assert set(lib.params.keys()) == {"p0", "p1"}  # NOTE: op folded
+
+        src = lib.lib.get_source()
+        lib.lib.save(temp_dir.relpath("test.c"), "c")
+        c_dtype = _get_c_datatype(linkable_dtype)
+        src_lines = src.split("\n")
+        param = lib.params["p0"].numpy().reshape(np.prod(KERNEL_SHAPE))
+        param_def = f"static const {c_dtype} __tvm_param__p0[{np.prod(param.shape)}] = {{"
+        for i, line in enumerate(src_lines):
+            if line == param_def:
                 i += 1
-
-            assert cursor == np.prod(param.shape)
-
-            # Need a unique name per library to avoid dlopen caching the lib load.
-            lib_path = temp_dir.relpath(f"test-{dtype}-linked.so")
-            lib["remove_params"]().export_library(lib_path)
-            lib_mod = tvm.runtime.load_module(lib_path)
-
-            #            lib_mod = lib_factory['default']()
-            graph = json.loads(lib.graph_json)
-            for p in lib.params:
-                _verify_linked_param(dtype, lib, lib_mod, graph, p)
-
-            # Wrap in function to explicitly deallocate the runtime.
-            def _run_linked(lib_mod):
-                graph_rt = tvm.contrib.graph_executor.GraphModule(lib_mod["default"](tvm.cpu(0)))
-                graph_rt.set_input("rand_input", rand_input)  # NOTE: params not required.
-                graph_rt.run()
-
-                return graph_rt.get_output(0)
-
-            linked_output = _run_linked(lib_mod)
-
-        linked_params = lib.params
-        with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
-            lib = tvm.relay.build(mod, "c", params=param_init)
-            _, _, params = lib
-            # Need a unique name per library to avoid dlopen caching the lib load.
-            lib_path = temp_dir.relpath(f"test-{dtype}-unlinked.so")
-            lib.export_library(lib_path)
-            lib_mod = tvm.runtime.load_module(lib_path)
-
-            def _run_unlinked(lib_mod):
-                graph_rt = tvm.contrib.graph_executor.GraphModule(lib_mod["default"](tvm.cpu(0)))
-                graph_rt.set_input("rand_input", rand_input, **params)
-                graph_rt.run()
-                return graph_rt.get_output(0)
-
-            unlinked_output = _run_unlinked(lib_mod)
-
-        if "int" in dtype:
-            np.testing.assert_equal(unlinked_output.numpy(), linked_output.numpy())
+                break
         else:
-            np.testing.assert_allclose(unlinked_output.numpy(), linked_output.numpy())
+            assert False, f'did not find parameter definition "{param_def}":\n{src}'
+
+        cursor = 0
+        width = dtype_info(linkable_dtype).bits // 4 + 2
+        if linkable_dtype.startswith("int"):
+            width += 1  # Account for sign
+
+        while "};" not in src_lines[i]:
+            for match in HEX_NUM_RE.finditer(src_lines[i]):
+                assert match.group() == _format_c_value(linkable_dtype, width, param[cursor]), (
+                    f'p0 byte {cursor}: want "{_format_c_value(linkable_dtype, width, param[cursor])}" got '
+                    f'"{match.group(0)}"; full p0 follows:\n{src}'
+                )
+                cursor += 1
+            i += 1
+
+        assert cursor == np.prod(param.shape)
+
+        # Need a unique name per library to avoid dlopen caching the lib load.
+        lib_path = temp_dir.relpath(f"test-{linkable_dtype}-linked.so")
+        lib["remove_params"]().export_library(lib_path)
+        lib_mod = tvm.runtime.load_module(lib_path)
+
+        #            lib_mod = lib_factory['default']()
+        graph = json.loads(lib.graph_json)
+        for p in lib.params:
+            _verify_linked_param(linkable_dtype, lib, lib_mod, graph, p)
+
+        # Wrap in function to explicitly deallocate the runtime.
+        def _run_linked(lib_mod):
+            graph_rt = tvm.contrib.graph_executor.GraphModule(lib_mod["default"](tvm.cpu(0)))
+            graph_rt.set_input("rand_input", rand_input)  # NOTE: params not required.
+            graph_rt.run()
+
+            return graph_rt.get_output(0)
+
+        linked_output = _run_linked(lib_mod)
+
+    linked_params = lib.params
+    with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
+        lib = tvm.relay.build(mod, "c", params=param_init)
+        _, _, params = lib
+        # Need a unique name per library to avoid dlopen caching the lib load.
+        lib_path = temp_dir.relpath(f"test-{linkable_dtype}-unlinked.so")
+        lib.export_library(lib_path)
+        lib_mod = tvm.runtime.load_module(lib_path)
+
+        def _run_unlinked(lib_mod):
+            graph_rt = tvm.contrib.graph_executor.GraphModule(lib_mod["default"](tvm.cpu(0)))
+            graph_rt.set_input("rand_input", rand_input, **params)
+            graph_rt.run()
+            return graph_rt.get_output(0)
+
+        unlinked_output = _run_unlinked(lib_mod)
+
+    if "int" in linkable_dtype:
+        np.testing.assert_equal(unlinked_output.numpy(), linked_output.numpy())
+    else:
+        np.testing.assert_allclose(unlinked_output.numpy(), linked_output.numpy())
 
 
 @tvm.testing.requires_micro
-def test_crt_link_params():
+def test_crt_link_params(linkable_dtype):
     from tvm import micro
-
-    for dtype in LINKABLE_DTYPES:
-        mod, param_init = _make_mod_and_params(dtype)
-        rand_input = _make_random_tensor(dtype, INPUT_SHAPE)
-        main_func = mod["main"]
-        target = "c --system-lib --runtime=c --link-params"
-        with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
-            factory = tvm.relay.build(mod, target, params=param_init)
-            assert set(factory.get_params().keys()) == {"p0", "p1"}  # NOTE: op folded
-
-            temp_dir = tvm.contrib.utils.tempdir()
-            template_project_dir = os.path.join(
-                tvm.micro.get_standalone_crt_dir(), "template", "host"
-            )
-            project = tvm.micro.generate_project(
-                template_project_dir, factory, temp_dir / "project", {"verbose": 1}
+    mod, param_init = _make_mod_and_params(linkable_dtype)
+    rand_input = _make_random_tensor(linkable_dtype, INPUT_SHAPE)
+    main_func = mod["main"]
+    target = "c --system-lib --runtime=c --link-params"
+    with tvm.transform.PassContext(opt_level=3, config={"tir.disable_vectorize": True}):
+        factory = tvm.relay.build(mod, target, params=param_init)
+        assert set(factory.get_params().keys()) == {"p0", "p1"}  # NOTE: op folded
+
+        temp_dir = tvm.contrib.utils.tempdir()
+        template_project_dir = os.path.join(
+            tvm.micro.get_standalone_crt_dir(), "template", "host"
+        )
+        project = tvm.micro.generate_project(
+            template_project_dir, factory, temp_dir / "project", {"verbose": 1}
+        )
+        project.build()
+        project.flash()
+        with tvm.micro.Session(project.transport()) as sess:
+            graph_rt = tvm.micro.session.create_local_graph_executor(
+                factory.get_graph_json(), sess.get_system_lib(), sess.device
             )
-            project.build()
-            project.flash()
-            with tvm.micro.Session(project.transport()) as sess:
-                graph_rt = tvm.micro.session.create_local_graph_executor(
-                    factory.get_graph_json(), sess.get_system_lib(), sess.device
-                )
 
-                # NOTE: not setting params here.
-                graph_rt.set_input("rand_input", rand_input)
-                graph_rt.run()
-                linked_output = graph_rt.get_output(0).numpy()
+            # NOTE: not setting params here.
+            graph_rt.set_input("rand_input", rand_input)
+            graph_rt.run()
+            linked_output = graph_rt.get_output(0).numpy()
 
-        with tvm.transform.PassContext(opt_level=3):
-            lib = tvm.relay.build(mod, "llvm --system-lib", params=param_init)
+    with tvm.transform.PassContext(opt_level=3):
+        lib = tvm.relay.build(mod, "llvm --system-lib", params=param_init)
 
-            def _run_unlinked(lib):
-                graph_json, mod, lowered_params = lib
-                graph_rt = tvm.contrib.graph_executor.create(graph_json, mod, tvm.cpu(0))
-                graph_rt.set_input("rand_input", rand_input, **lowered_params)
-                graph_rt.run()
-                return graph_rt.get_output(0).numpy()
+        def _run_unlinked(lib):
+            graph_json, mod, lowered_params = lib
+            graph_rt = tvm.contrib.graph_executor.create(graph_json, mod, tvm.cpu(0))
+            graph_rt.set_input("rand_input", rand_input, **lowered_params)
+            graph_rt.run()
+            return graph_rt.get_output(0).numpy()
 
-            unlinked_output = _run_unlinked(lib)
+        unlinked_output = _run_unlinked(lib)
 
-        if "int" in dtype:
-            np.testing.assert_equal(unlinked_output, linked_output)
-        else:
-            np.testing.assert_allclose(unlinked_output, linked_output)
+    if "int" in linkable_dtype:
+        np.testing.assert_equal(unlinked_output, linked_output)
+    else:
+        np.testing.assert_allclose(unlinked_output, linked_output)
 
 
 if __name__ == "__main__":