[0-size Tensor No.157、167、176、177、179、201、215、222、223、230、266] Add 0-size Tensor support for some activation function API. (#73028)

DanielSun11 · web-flow · commit 34b14c15ee34 · 2025-06-12T19:07:05.000+08:00
* fix some activation kernel

* fix unittest

* fix code style

* fxi xpu build error
diff --git a/paddle/phi/kernels/cpu/activation_grad_kernel.cc b/paddle/phi/kernels/cpu/activation_grad_kernel.cc
@@ -211,6 +211,9 @@ void EluGradKernel(const Context& dev_ctx,
                    float alpha,
                    DenseTensor* dx) {
   dev_ctx.template Alloc<T>(dx);
+  if (dx->numel() == 0) {
+    return;
+  }
 
   auto x_flatten =
       EigenVector<T>::Flatten(GET_DATA_SAFELY(&x, "Input", "X", "elu_grad"));
diff --git a/paddle/phi/kernels/gpu/activation_grad_kernel.cu b/paddle/phi/kernels/gpu/activation_grad_kernel.cu
@@ -270,6 +270,9 @@ void EluGradKernel(const Context& dev_ctx,
                    float alpha,
                    DenseTensor* dx) {
   dev_ctx.template Alloc<T>(dx);
+  if (dx->numel() == 0) {
+    return;
+  }
   std::vector<const DenseTensor*> ins = {&dout, &out};
   std::vector<DenseTensor*> outs = {dx};
   if (alpha > 0) {
diff --git a/paddle/phi/kernels/xpu/activation_grad_kernel.cc b/paddle/phi/kernels/xpu/activation_grad_kernel.cc
@@ -709,6 +709,9 @@ void TanhGradKernel(const Context& dev_ctx,
                     DenseTensor* dx) {
   using XPUType = typename XPUTypeTrait<T>::Type;
   dev_ctx.template Alloc<T>(dx);
+  if (dx->numel() == 0) {
+    return;
+  }
   const XPUType* y_data = reinterpret_cast<const XPUType*>(out.data<T>());
   const XPUType* y_grad = reinterpret_cast<const XPUType*>(dout.data<T>());
   XPUType* x_grad = reinterpret_cast<XPUType*>(dx->data<T>());
diff --git a/paddle/phi/kernels/xpu/activation_kernel.cc b/paddle/phi/kernels/xpu/activation_kernel.cc
@@ -439,6 +439,9 @@ void EluKernel(const Context& dev_ctx,
                DenseTensor* out) {
   using XPUType = typename XPUTypeTrait<T>::Type;
   dev_ctx.template Alloc<T>(out);
+  if (out->numel() == 0) {
+    return;
+  }
   // template<typename T> int elu(Context* ctx, const T* x, T* y, int64_t len,
   // float alpha = 1.0f, const float* max_x = nullptr, float* max_y = nullptr)
   int r = xpu::elu(dev_ctx.x_context(),
diff --git a/test/legacy_test/test_activation_op.py b/test/legacy_test/test_activation_op.py
@@ -1496,13 +1496,17 @@ class TestHardtanhAPI(unittest.TestCase):
     # test paddle.nn.Hardtanh, paddle.nn.functional.hardtanh
     def setUp(self):
         np.random.seed(1024)
-        self.x_np = np.random.uniform(-3, 3, [10, 12]).astype('float32')
+        self.init_shape()
+        self.x_np = np.random.uniform(-3, 3, self.x_shape).astype('float32')
         self.place = (
             paddle.CUDAPlace(0)
             if paddle.is_compiled_with_cuda()
             else paddle.CPUPlace()
         )
 
+    def init_shape(self):
+        self.x_shape = [10, 12]
+
     def test_static_api(self):
         with static_guard():
             with paddle.static.program_guard(paddle.static.Program()):
diff --git a/test/legacy_test/test_activation_op_zero_size.py b/test/legacy_test/test_activation_op_zero_size.py
@@ -23,14 +23,19 @@
     TestAtan,
     TestAtanh,
     TestCeil,
+    TestCELU,
     TestCos,
     TestCosh,
+    TestELU,
     TestExpFp32_Prim,
     TestExpm1,
     TestFloor,
+    TestHardShrink,
+    TestHardSigmoid,
     TestHardSwish,
     TestLeakyRelu,
     TestLogSigmoid,
+    TestMish,
     TestReciprocal,
     TestRelu,
     TestRelu6,
@@ -40,13 +45,17 @@
     TestSilu,
     TestSin,
     TestSinh,
+    TestSoftplus,
+    TestSoftshrink,
     TestSoftsign,
     TestSqrt,
     TestSquare,
+    TestSTanh,
     TestSwish,
     TestTan,
     TestTanh,
     TestTanhshrink,
+    TestThresholdedRelu,
 )
 
 
@@ -103,6 +112,15 @@ def test_check_grad(self):
 create_test_zero_size_class(TestLogSigmoid)
 create_test_zero_size_class(TestFloor)
 create_test_zero_size_class(TestCeil)
+create_test_zero_size_class(TestELU)
+create_test_zero_size_class(TestCELU)
+create_test_zero_size_class(TestHardShrink)
+create_test_zero_size_class(TestHardSigmoid)
+create_test_zero_size_class(TestMish)
+create_test_zero_size_class(TestSoftplus)
+create_test_zero_size_class(TestSoftshrink)
+create_test_zero_size_class(TestSTanh)
+create_test_zero_size_class(TestThresholdedRelu)
 create_test_zero_size_class(TestExpFp32_Prim)
 create_test_zero_size_class(TestExpm1)
 create_test_zero_size_class(TestLeakyRelu)
diff --git a/test/legacy_test/test_selu_op.py b/test/legacy_test/test_selu_op.py
@@ -42,7 +42,6 @@ class SeluTest(OpTest):
     def setUp(self):
         self.op_type = "selu"
         self.python_api = paddle.nn.functional.selu
-        self.x_shape = [3, 5, 5, 10]
         self.init_x_shape()
         self.init_dtype()
 
@@ -73,7 +72,7 @@ def setUp(self):
         }
 
     def init_x_shape(self):
-        pass
+        self.x_shape = [3, 5, 5, 10]
 
     def init_dtype(self):
         self.dtype = np.float64
@@ -108,6 +107,21 @@ def test_check_grad(self):
         )
 
 
+class SeluTestZeroSize1(SeluTest):
+    def init_x_shape(self):
+        self.x_shape = [9, 0]
+
+
+class SeluTestZeroSize2(SeluTest):
+    def init_x_shape(self):
+        self.x_shape = [0, 0]
+
+
+class SeluTestZeroSize3(SeluTest):
+    def init_x_shape(self):
+        self.x_shape = [5, 0, 8]
+
+
 class TestSeluAPI(unittest.TestCase):
     # test paddle.nn.SELU, paddle.nn.functional.selu
     def setUp(self):
@@ -180,4 +194,5 @@ def test_errors(self):
 
 
 if __name__ == "__main__":
+    paddle.enable_static()
     unittest.main()
