[0-size Tensor No.38] Add 0-size Tensor support for paddle.cross API (#72764)

DanielSun11 · web-flow · commit 16465cf689cf · 2025-05-21T11:19:37.000+08:00
* [0 size Tensor]Fix CrossGradKernel

* add  test case

* add cpu unittest case

* Update cross_grad_kernel.cu

* format
diff --git a/paddle/phi/kernels/cpu/cross_grad_kernel.cc b/paddle/phi/kernels/cpu/cross_grad_kernel.cc
@@ -110,6 +110,9 @@ void CrossGradKernel(const Context &dev_ctx,
 
   dev_ctx.template Alloc<T>(output_x_grad);
   dev_ctx.template Alloc<T>(output_y_grad);
+  if (numel == 0) {
+    return;
+  }
 
   for (int64_t i = 0; i < outer_loops; i++) {
     for (int64_t j = 0; j < 3; j++) {
diff --git a/paddle/phi/kernels/gpu/cross_grad_kernel.cu b/paddle/phi/kernels/gpu/cross_grad_kernel.cu
@@ -169,7 +169,9 @@ void CrossGradKernel(const Context& dev_ctx,
   const auto* input_out_grad_data = input_out_grad.data<T>();
   auto* output_x_grad_data = dev_ctx.template Alloc<T>(x_grad);
   auto* output_y_grad_data = dev_ctx.template Alloc<T>(y_grad);
-
+  if (numel == 0) {
+    return;
+  }
   backends::gpu::GpuLaunchConfig config =
       backends::gpu::GetGpuLaunchConfig1D(dev_ctx, numel / 3);
   auto index_calculator = phi::funcs::IndexCalculator<int64_t>(
diff --git a/test/legacy_test/test_cross_op.py b/test/legacy_test/test_cross_op.py
@@ -293,6 +293,92 @@ def test_dygraph_api(self):
         np.testing.assert_allclose(expect_out, np_z, rtol=1e-05)
 
 
+class TestCrossOpZeroSizeTest(TestCrossOp):
+    def initTestCase(self):
+        self.shape = (0, 3, 3)
+        self.dtype = np.float64
+        self.attr = {'dim': -1}
+
+    def init_output(self):
+        z_list = []
+        for i in range(0):
+            z_list.append(np.cross(self.inputs['X'][i], self.inputs['Y'][i]))
+        self.outputs = {'Out': np.array(z_list).reshape(self.shape)}
+
+
+class TestCrossOpZeroSizeTest1(TestCrossOp):
+    def initTestCase(self):
+        self.shape = (3, 0, 3)
+        self.dtype = np.float64
+        self.attr = {'dim': -1}
+
+    def init_output(self):
+        z_list = []
+        for i in range(3):
+            z_list.append(np.cross(self.inputs['X'][i], self.inputs['Y'][i]))
+        self.outputs = {'Out': np.array(z_list).reshape(self.shape)}
+
+
+class TestCrossOpZeroSizeTest2(TestCrossOp):
+    def initTestCase(self):
+        self.shape = (0, 0, 3)
+        self.dtype = np.float64
+        self.attr = {'dim': -1}
+
+    def init_output(self):
+        z_list = []
+        for i in range(0):
+            z_list.append(np.cross(self.inputs['X'][i], self.inputs['Y'][i]))
+        self.outputs = {'Out': np.array(z_list).reshape(self.shape)}
+
+
+class TestCrossOpZeroSizeCPUTest(TestCrossOp):
+    def initTestCase(self):
+        self.shape = (0, 0, 3)
+        self.dtype = np.float64
+        self.attr = {'dim': -1}
+
+    def init_output(self):
+        z_list = []
+        for i in range(0):
+            z_list.append(np.cross(self.inputs['X'][i], self.inputs['Y'][i]))
+        self.outputs = {'Out': np.array(z_list).reshape(self.shape)}
+
+    def test_check_output(self):
+        place = paddle.CPUPlace()
+        self.check_output_with_place(place, check_pir=True)
+
+    def test_check_grad_normal(self):
+        place = paddle.CPUPlace()
+        self.check_grad_with_place(place, ['X', 'Y'], 'Out', check_pir=True)
+
+
+class TestCrossOpZeroSizeCPUTest1(TestCrossOpZeroSizeCPUTest):
+    def initTestCase(self):
+        self.shape = (3, 0, 3)
+        self.dtype = np.float64
+        self.attr = {'dim': -1}
+
+    def init_output(self):
+        z_list = []
+        for i in range(3):
+            z_list.append(np.cross(self.inputs['X'][i], self.inputs['Y'][i]))
+        self.outputs = {'Out': np.array(z_list).reshape(self.shape)}
+
+
+class TestCrossOpZeroSizeCPUTest2(TestCrossOpZeroSizeCPUTest):
+    def initTestCase(self):
+        self.shape = (0, 0, 3)
+        self.dtype = np.float64
+        self.attr = {'dim': -1}
+
+    def init_output(self):
+        z_list = []
+        for i in range(0):
+            z_list.append(np.cross(self.inputs['X'][i], self.inputs['Y'][i]))
+        self.outputs = {'Out': np.array(z_list).reshape(self.shape)}
+
+
 if __name__ == '__main__':
     paddle.enable_static()
     unittest.main()
