add support complex (#70635)

fangfangssj · web-flow · commit 961393d766fe · 2025-01-08T11:31:42.000+08:00
diff --git a/paddle/phi/kernels/cpu/activation_grad_kernel.cc b/paddle/phi/kernels/cpu/activation_grad_kernel.cc
@@ -307,7 +307,7 @@ PD_REGISTER_ACTIVATION_GRAD_KERNEL(mish_grad, MishGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(stanh_grad, STanhGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(reciprocal_grad,
                                                 ReciprocalGradKernel)
-PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_grad, SqrtGradKernel)
+PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(sqrt_grad, SqrtGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(rsqrt_grad, RsqrtGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(softplus_grad,
                                                 SoftplusGradKernel)
diff --git a/paddle/phi/kernels/cpu/activation_kernel.cc b/paddle/phi/kernels/cpu/activation_kernel.cc
@@ -212,7 +212,7 @@ PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(silu, SiluKernel)
 PD_REGISTER_ACTIVATION_KERNEL(mish, MishKernel)
 PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(stanh, STanhKernel)
 PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(reciprocal, ReciprocalKernel)
-PD_REGISTER_ACTIVATION_KERNEL(sqrt, SqrtKernel)
+PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(sqrt, SqrtKernel)
 PD_REGISTER_ACTIVATION_KERNEL(rsqrt, RsqrtKernel)
 PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(softplus, SoftplusKernel)
 
diff --git a/paddle/phi/kernels/funcs/activation_functor.h b/paddle/phi/kernels/funcs/activation_functor.h
@@ -759,6 +759,24 @@ struct SqrtGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct SqrtGradFunctor<ComplexType<T>>
+    : public BaseActivationFunctor<ComplexType<T>> {
+  template <typename Device,
+            typename X,
+            typename Out,
+            typename dOut,
+            typename dX>
+  void operator()(Device d, X x UNUSED, Out out, dOut dout, dX dx) const {
+    dx.device(d) =
+        dout * (static_cast<ComplexType<T>>(0.5) / out).unaryExpr(Conj<T>());
+  }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() {
+    return ActBwdOpFwdDeps::kDepOut;
+  }
+};
+
 // rsqrt(x) = x^(-1/2)
 template <typename T>
 struct RsqrtFunctor : public BaseActivationFunctor<T> {
@@ -4050,6 +4068,22 @@ struct CudaSqrtGradFunctor : public BaseActivationFunctor<T> {
   }
 };
 
+template <typename T>
+struct CudaSqrtGradFunctor<ComplexType<T>>
+    : public BaseActivationFunctor<ComplexType<T>> {
+  ComplexType<T> one_half = static_cast<ComplexType<T>>(0.5f);
+
+  // dx = dout * 0.5 / out
+  __device__ __forceinline__ ComplexType<T> operator()(
+      const ComplexType<T> dout, const ComplexType<T> out) const {
+    return dout * conj(one_half / out);
+  }
+
+  static constexpr ActBwdOpFwdDeps FwdDeps() {
+    return ActBwdOpFwdDeps::kDepOut;
+  }
+};
+
 template <typename T>
 struct CudaRsqrtFunctor : public BaseActivationFunctor<T> {
   using MPType = typename phi::dtype::MPTypeTrait<T>::Type;
diff --git a/paddle/phi/kernels/gpu/activation_grad_kernel.cu b/paddle/phi/kernels/gpu/activation_grad_kernel.cu
@@ -386,7 +386,7 @@ PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(softplus_grad,
                                                 SoftplusGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(softplus_double_grad,
                                                 SoftplusDoubleGradKernel)
-PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_grad, SqrtGradKernel)
+PD_REGISTER_ACTIVATION_GRAD_KERNEL_WITH_COMPLEX(sqrt_grad, SqrtGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(sqrt_double_grad, SqrtDoubleGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(rsqrt_grad, RsqrtGradKernel)
 PD_REGISTER_ACTIVATION_GRAD_KERNEL(rsqrt_double_grad, RsqrtDoubleGradKernel)
diff --git a/paddle/phi/kernels/gpu/activation_kernel.cu b/paddle/phi/kernels/gpu/activation_kernel.cu
@@ -261,7 +261,7 @@ PD_REGISTER_ACTIVATION_KERNEL(leaky_relu, LeakyReluKernel)
 PD_REGISTER_ACTIVATION_KERNEL(mish, MishKernel)
 PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(stanh, StanhKernel)
 PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(reciprocal, ReciprocalKernel)
-PD_REGISTER_ACTIVATION_KERNEL(sqrt, SqrtKernel)
+PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(sqrt, SqrtKernel)
 PD_REGISTER_ACTIVATION_KERNEL(rsqrt, RsqrtKernel)
 PD_REGISTER_ACTIVATION_KERNEL_WITH_COMPLEX(softplus, SoftplusKernel)
 
diff --git a/python/paddle/tensor/ops.py b/python/paddle/tensor/ops.py
@@ -1114,6 +1114,8 @@ def sqrt(x: Tensor, name: str | None = None) -> Tensor:
                 'int16',
                 'int32',
                 'int64',
+                'complex64',
+                'complex128',
             ],
             'sqrt',
         )
diff --git a/test/legacy_test/test_activation_op.py b/test/legacy_test/test_activation_op.py
@@ -1667,6 +1667,11 @@ def setUp(self):
 
         np.random.seed(1023)
         x = np.random.uniform(0.1, 1, self.shape).astype(self.dtype)
+        if self.dtype == np.complex64 or self.dtype == np.complex128:
+            x = (
+                np.random.uniform(-1, 1, self.shape)
+                + 1j * np.random.uniform(-1, 1, self.shape)
+            ).astype(self.dtype)
         out = np.sqrt(x)
 
         self.inputs = {'X': OpTest.np_dtype_to_base_dtype(x)}
@@ -1679,14 +1684,20 @@ def if_enable_cinn(self):
     def test_check_grad(self):
         if self.dtype == np.float16:
             return
-        self.check_grad(
-            ['X'],
-            'Out',
-            check_prim=True,
-            check_pir=True,
-            check_prim_pir=True,
-            check_pir_onednn=self.check_pir_onednn,
-        )
+        if self.dtype not in [np.complex64, np.complex128]:
+            self.check_grad(
+                ['X'],
+                'Out',
+                check_prim=True,
+                check_pir=True,
+                check_prim_pir=True,
+                check_pir_onednn=self.check_pir_onednn,
+            )
+        else:
+            self.check_grad(
+                ['X'],
+                'Out',
+            )
 
     def test_check_output(self):
         self.check_output(
@@ -1746,6 +1757,16 @@ def init_shape(self):
         self.shape = []
 
 
+class TestSqrt_Complex64(TestSqrt):
+    def init_dtype(self):
+        self.dtype = np.complex64
+
+
+class TestSqrt_Complex128(TestSqrt):
+    def init_dtype(self):
+        self.dtype = np.complex128
+
+
 @unittest.skipIf(
     not core.is_compiled_with_cuda() or core.is_compiled_with_rocm(),
     "core is not compiled with CUDA",

Original file line number	Diff line number	Diff line change
`@@ -1114,6 +1114,8 @@ def sqrt(x: Tensor, name: str \| None = None) -> Tensor:`
`1114`	`1114`	`'int16',`
`1115`	`1115`	`'int32',`
`1116`	`1116`	`'int64',`
	`1117`	`+ 'complex64',`
	`1118`	`+ 'complex128',`
`1117`	`1119`	`],`
`1118`	`1120`	`'sqrt',`
`1119`	`1121`	`)`