Add reciprocal operator

Summary: Add reciprocal Differential Revision: D62000543
facebookincubator · Sep 4, 2024 · fc8c4fa · fc8c4fa
1 parent e2eff07
commit fc8c4fa
Show file tree

Hide file tree

Showing 7 changed files with 107 additions and 0 deletions.
diff --git a/fx2ait/fx2ait/converters/ait_converters.py b/fx2ait/fx2ait/converters/ait_converters.py
@@ -166,6 +166,17 @@ def acc_ops_floor(
     return elementwise(FuncEnum.FLOOR)(input_val)
 
 
+@ait_converter(acc_ops.reciprocal)
+def acc_ops_reciprocal(
+    target: Target,
+    args: Tuple[Argument, ...],
+    kwargs: Dict[str, Argument],
+    name: str,
+) -> ConverterOutput:
+    input_val = kwargs["input"]
+    return elementwise(FuncEnum.RECIPROCAL)(input_val)
+
+
 @ait_converter(acc_ops.add)
 def acc_ops_add(
     target: Target,

diff --git a/fx2ait/fx2ait/converters/utils.py b/fx2ait/fx2ait/converters/utils.py
@@ -166,6 +166,8 @@ def get_python_op_from_ait_constant_elementwise_op(
         return operator.floordiv
     elif op_type == FuncEnum.FLOOR:
         return math.floor
+    elif op_type == FuncEnum.RECIPROCAL:
+        return math.reciprocal
     else:
         raise RuntimeError(f"{op_type} is not supported yet!")
 

diff --git a/python/aitemplate/backend/backend_spec.py b/python/aitemplate/backend/backend_spec.py
@@ -326,6 +326,13 @@ class GPUBackendSpec(BackendSpec):
                 "bfloat16": "__floor",
                 "bfloat16_2": "__floor",
             },
+            FuncEnum.RECIPROCAL: {
+                "float": "__reciprocal",
+                "half": "__hreciprocal",
+                "half2": "__h2reciprocal",
+                "bfloat16": "__breciprocal",
+                "bfloat16_2": "__b2reciprocal",
+            },
             FuncEnum.CELU: {
                 "float": "fcelu",
                 "half": "hcelu",

diff --git a/python/aitemplate/backend/cuda/elementwise/custom_math.cuh b/python/aitemplate/backend/cuda/elementwise/custom_math.cuh
@@ -1043,6 +1043,34 @@ __device__ bfloat16_2 __floor(const bfloat16_2 a) {
 #endif
 }
 
+__device__ float __reciprocal(const float a) {
+  return 1.0f / a;
+}
+
+__device__ half __hreciprocal(const half a) {
+  return __hdiv(1.0f, a);
+}
+
+__device__ bfloat16 __breciprocal(const bfloat16 a) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+  return __hdiv(1.0f, a);
+#else
+  NOT_IMPLEMENTED();
+#endif
+}
+
+__device__ half2 __h2reciprocal(const half2 a) {
+  return half2(__hdiv(1.0f, a.x), __hdiv(1.0f, a.y));
+}
+
+__device__ bfloat16_2 __b2reciprocal(const bfloat16_2 a) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 800)
+  return bfloat16_2(__hdiv(1.0f, a.x), __hdiv(1.0f, a.y));
+#else
+  NOT_IMPLEMENTED();
+#endif
+}
+
 __device__ float fcelu(const float a, const float alpha) {
   return a > 0.f ? a : alpha * (expf(a / alpha) - 1.0f);
 }

diff --git a/python/aitemplate/compiler/ops/common/epilogue.py b/python/aitemplate/compiler/ops/common/epilogue.py
@@ -67,3 +67,4 @@ class FuncEnum(Enum):
     CELU = 29
     FLOOR = 30
     LOG1P = 31
+    RECIPROCAL = 32
diff --git a/python/aitemplate/compiler/ops/common/math.py b/python/aitemplate/compiler/ops/common/math.py
@@ -125,3 +125,7 @@ def celu(tensor: Any) -> Tensor:
 
 def floor(tensor: Any) -> Tensor:
     return OP_REGISTRY.get("FLOOR")(tensor)
+
+
+def reciprocal(tensor: Any) -> Tensor:
+    return OP_REGISTRY.get("RECIPROCAL")(tensor)
diff --git a/tests/unittest/ops/test_activation.py b/tests/unittest/ops/test_activation.py
@@ -45,6 +45,7 @@
     FuncEnum.RELU: torch.relu,
     FuncEnum.CELU: torch.celu,
     FuncEnum.FLOOR: torch.floor,
+    FuncEnum.RECIPROCAL: torch.reciprocal,
 }
 
 
@@ -437,6 +438,38 @@ def _test_fast_gelu(
         module.run_with_tensors([x1_pt], [x2])
         torch.testing.assert_close(x2, x2_pt, atol=1e-2, rtol=1e-2)
 
+    def _test_reciprocal(
+        self,
+        input_size,
+        test_name="reciprocal",
+        copy_op=False,
+        dtype="float16",
+    ):
+        assert len(input_size) == 2
+        X1 = Tensor(
+            shape=[IntImm(input_size[0]), IntImm(input_size[1])],
+            dtype=dtype,
+            name="input0",
+            is_input=True,
+        )
+        X2_op = ops.elementwise(FuncEnum.RECIPROCAL)
+
+        if copy_op:
+            X2_op = ops.elementwise(**X2_op._get_op_attributes())
+        X2 = X2_op(X1)
+        X2._attrs["is_output"] = True
+        X2._attrs["name"] = "output0"
+
+        target = detect_target()
+        module = compile_model(X2, target, "./tmp", f"{test_name}_{dtype}")
+
+        x1_pt = get_random_torch_tensor(input_size, dtype)
+        x2_pt = torch.floor(x1_pt)
+
+        x2 = torch.empty_like(x2_pt)
+        module.run_with_tensors([x1_pt], [x2])
+        torch.testing.assert_close(x2, x2_pt, atol=1e-2, rtol=1e-2)
+
     @parameterized.expand(
         **filter_test_cases_by_params(
             {
@@ -921,6 +954,27 @@ def test_fast_gelu(self, dtype):
             [256, 128], test_name="fast_gelu_4_copy_op", copy_op=True, dtype=dtype
         )
 
+    @parameterized.expand(
+        **filter_test_cases_by_params(
+            {
+                TestEnv.CUDA_LESS_THAN_SM80: [("float16"), ("float32")],
+                TestEnv.CUDA_SM80: [("bfloat16")],
+                TestEnv.ROCM: [("float16")],
+            }
+        )
+    )
+    def test_reciprocal(self, dtype):
+        self._test_simple_function(
+            [32, 128], FuncEnum.RECIPROCAL, test_name="reciprocal", dtype=dtype
+        )
+        self._test_simple_function(
+            [32, 128],
+            FuncEnum.RECIPROCAL,
+            test_name="reciprocal_copy_op",
+            copy_op=True,
+            dtype=dtype,
+        )
+
 
 if __name__ == "__main__":
     unittest.main()