Split benchmark suite for generic pointwise operators

benchmark/test_addcdiv.py

@@ -0,0 +1,24 @@
+import pytest
+import torch
+
+from . import attri_util as attrs
+from . import performance_utils as base
+
+
+def _input_fn(shape, cur_dtype, device):
+    inp1 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp2 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp3 = base.generate_tensor_input(shape, cur_dtype, device)
+
+    yield inp1, inp2, inp3, {"value": 0.5}
+
+
+@pytest.mark.addcdiv
+def test_addcdiv():
+    bench = base.GenericBenchmark(
+        op_name="addcdiv",
+        input_fn=_input_fn,
+        torch_op=torch.addcdiv,
+        dtypes=attrs.FLOAT_DTYPES,
+    )
+    bench.run()

benchmark/test_addcmul.py

@@ -0,0 +1,24 @@
+import pytest
+import torch
+
+from . import attri_util as attrs
+from . import performance_utils as base
+
+
+def _input_fn(shape, cur_dtype, device):
+    inp1 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp2 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp3 = base.generate_tensor_input(shape, cur_dtype, device)
+
+    yield inp1, inp2, inp3, {"value": 0.5}
+
+
+@pytest.mark.addcmul
+def test_addcmul():
+    bench = base.GenericBenchmark(
+        op_name="addcmul",
+        input_fn=_input_fn,
+        torch_op=torch.addcmul,
+        dtypes=attrs.FLOAT_DTYPES,
+    )
+    bench.run()

benchmark/test_clamp.py

@@ -0,0 +1,41 @@
+import pytest
+import torch
+
+from . import attri_util as attrs
+from . import performance_utils as base
+
+
+def _input_fn(shape, cur_dtype, device):
+    inp1 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp2 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp3 = base.generate_tensor_input(shape, cur_dtype, device)
+
+    yield inp1, inp2, inp3
+
+    if base.Config.bench_level == base.BenchLevel.COMPREHENSIVE:
+        # scalar or None situation
+        yield inp1, inp2, None
+        yield inp1, None, 3.14
+
+
+@pytest.mark.clamp
+def test_clamp():
+    bench = base.GenericBenchmark(
+        op_name="clamp",
+        input_fn=_input_fn,
+        torch_op=torch.clamp,
+        dtypes=attrs.FLOAT_DTYPES,
+    )
+    bench.run()
+
+
+@pytest.mark.clamp_
+def test_clamp_inplace():
+    bench = base.GenericBenchmark(
+        input_fn=_input_fn,
+        op_name="clamp_",
+        torch_op=torch.clamp_,
+        dtypes=attrs.FLOAT_DTYPES,
+        is_inplace=True,
+    )
+    bench.run()

benchmark/test_clamp_min.py

@@ -0,0 +1,39 @@
+import pytest
+import torch
+
+from . import attri_util as attrs
+from . import performance_utils as base
+
+
+def _input_fn(shape, cur_dtype, device):
+    inp1 = base.generate_tensor_input(shape, cur_dtype, device)
+    inp2 = base.generate_tensor_input(shape, cur_dtype, device)
+
+    yield inp1, inp2
+
+    if base.Config.bench_level == attrs.BenchLevel.COMPREHENSIVE:
+        # scalar situation
+        yield inp1, 3.14
+
+
+@pytest.mark.clamp_min
+def test_clamp_min():
+    bench = base.GenericBenchmark(
+        op_name="clamp_min",
+        input_fn=_input_fn,
+        torch_op=torch.clamp_min,
+        dtypes=attrs.FLOAT_DTYPES,
+    )
+    bench.run()
+
+
+@pytest.mark.clamp_min_
+def test_clamp_min_inplace():
+    bench = base.GenericBenchmark(
+        input_fn=_input_fn,
+        op_name="clamp_min_",
+        torch_op=torch.clamp_min_,
+        dtypes=attrs.FLOAT_DTYPES,
+        is_inplace=True,
+    )
+    bench.run()

benchmark/test_flip.py

@@ -0,0 +1,24 @@
+import pytest
+import torch
+
+from . import attri_util as attrs
+from . import performance_utils as base
+
+
+def _input_fn(shape, cur_dtype, device):
+    inp = base.generate_tensor_input(shape, cur_dtype, device)
+    if len(shape) > 1:
+        yield inp, {"dims": (0, 1)}
+    else:
+        yield inp, {"dims": (0,)}
+
+
+@pytest.mark.flip
+def test_flip():
+    bench = base.GenericBenchmark(
+        op_name="flip",
+        input_fn=_input_fn,
+        torch_op=torch.flip,
+        dtypes=attrs.FLOAT_DTYPES + attrs.INT_DTYPES,
+    )
+    bench.run()