use libdevice for tanh (#90889)

Per title
I see slight differences in perf with this implementation, where standalone tanh is slightly slower for a tensor of 4000000
 elements (20.4 us instead of 19.4us), other sizes are within noise.
 @bertmaher could you check if it affects your benchmarks?

Pull Request resolved: https://github.com/pytorch/pytorch/pull/90889
Approved by: https://github.com/bertmaher, https://github.com/anijain2305

torch/_inductor/codegen/cpp.py

@@ -416,6 +416,10 @@ class CppOverrides(OpOverrides):
     def expm1(x):
         return f"std::expm1({x})"
 
+    @staticmethod
+    def tanh(x):
+        return f"std::tanh({x})"
+
     @staticmethod
     def signbit(x):
         return f"std::signbit({x})"

torch/_inductor/codegen/triton.py

@@ -239,6 +239,10 @@ class TritonOverrides(OpOverrides):
     def expm1(x):
         return f"tl.libdevice.expm1({x})"
 
+    @staticmethod
+    def tanh(x):
+        return f"tl.libdevice.tanh({x})"
+
     @staticmethod
     def sigmoid(x):
         return f"tl.sigmoid({x})"

torch/_inductor/decomposition.py

@@ -126,11 +126,6 @@ def clamp(x, min=None, max=None):
     return x
 
 
-@register_decomposition([aten.tanh])
-def tanh(x):
-    return 2.0 / (1.0 + torch.exp(-2.0 * x)) - 1.0
-
-
 # TorchInductor-only decomposition. It should not be taken to core.
 # See https://github.com/pytorch/torchdynamo/pull/1120
 @register_decomposition([aten.floor_divide.default])

torch/_inductor/lowering.py

@@ -3621,6 +3621,11 @@ register_pointwise(
     type_promotion_kind=ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT,
 )
 
+register_pointwise(
+    aten.tanh,
+    type_promotion_kind=ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT,
+)
+
 register_pointwise(
     aten.log,
     type_promotion_kind=ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT,