Add torch.rsqrt check

tests/fixtures/misc/checker/rsqrt.py

@@ -0,0 +1,10 @@
+import torch
+
+
+a = torch.randn(5)
+b = 1 / torch.sqrt(a)
+b = 1.0 / torch.sqrt(a)
+b = a / torch.sqrt(a)
+# False negative
+b = 1 / a.sqrt()
+b = 1.0 / a.sqrt()

tests/fixtures/misc/checker/rsqrt.txt

@@ -0,0 +1,3 @@
+5:5 TOR109 Consider faster `a*torch.rsqrt(b)` instead of `a/torch.sqrt(b)`.
+6:5 TOR109 Consider faster `a*torch.rsqrt(b)` instead of `a/torch.sqrt(b)`.
+7:5 TOR109 Consider faster `a*torch.rsqrt(b)` instead of `a/torch.sqrt(b)`.

tests/test_torchfix.py

@@ -48,6 +48,7 @@ def pytest_generate_tests(metafunc):
                     "TOR106",
                     "TOR107",
                     "TOR108",
+                    "TOR109",
                 },
             ),
             (None, set(GET_ALL_ERROR_CODES()) - exclude_set),

torchfix/torchfix.py

@@ -18,6 +18,7 @@
     TorchScopedLibraryVisitor,
     TorchSynchronizedDataLoaderVisitor,
     TorchUnsafeLoadVisitor,
+    TorchRsqrtVisitor,
     TorchVisionDeprecatedPretrainedVisitor,
     TorchVisionDeprecatedToTensorVisitor,
     TorchVisionSingletonImportVisitor,
@@ -35,6 +36,7 @@
     TorchExpm1Visitor,
     TorchLog1pVisitor,
     TorchLogsumexpVisitor,
+    TorchRsqrtVisitor,
     TorchNonPublicAliasVisitor,
     TorchRequireGradVisitor,
     TorchReentrantCheckpointVisitor,

torchfix/visitors/__init__.py

@@ -6,6 +6,7 @@
     TorchLogsumexpVisitor,
     TorchReentrantCheckpointVisitor,
     TorchRequireGradVisitor,
+    TorchRsqrtVisitor,
 )
 from .nonpublic import TorchNonPublicAliasVisitor
 from .performance import (
@@ -30,6 +31,7 @@
     "TorchScopedLibraryVisitor",
     "TorchSynchronizedDataLoaderVisitor",
     "TorchUnsafeLoadVisitor",
+    "TorchRsqrtVisitor",
     "TorchVisionDeprecatedPretrainedVisitor",
     "TorchVisionDeprecatedToTensorVisitor",
     "TorchVisionSingletonImportVisitor",

torchfix/visitors/misc/__init__.py

@@ -184,7 +184,6 @@ def visit_Call(self, node):
                         )
                         == "torch.exp"
                     ):
-
                         # if `dim` is not provided or None for sum, skip:
                         # https://github.com/pytorch/pytorch/issues/144339
                         dim_arg = self.get_specific_arg(
@@ -201,3 +200,29 @@ def visit_Call(self, node):
                                     message=self.ERRORS[0].message(),
                                     replacement=None,
                                 )
+
+
+class TorchRsqrtVisitor(TorchVisitor):
+    """
+    Suggest using `a*torch.rsqrt(b)` instead of `a/torch.sqrt(b)`.
+    """
+
+    ERRORS = [
+        TorchError(
+            "TOR109",
+            ("Consider faster `a*torch.rsqrt(b)` instead of `a/torch.sqrt(b)`."),
+        )
+    ]
+
+    def visit_BinaryOperation(self, node):
+        if m.matches(
+            node,
+            m.BinaryOperation(operator=m.Divide(), right=m.Call()),
+        ):
+            if self.get_qualified_name_for_call(node.right) == "torch.sqrt":
+                self.add_violation(
+                    node,
+                    error_code=self.ERRORS[0].error_code,
+                    message=self.ERRORS[0].message(),
+                    replacement=None,
+                )