[Zero-dim] Zero-dim Tensor for XPU prelu, softmax and log_softmax (#50433)

paddle/phi/kernels/xpu/log_softmax_grad_kernel.cc

@@ -33,6 +33,7 @@ void LogSoftmaxGradKernel(const Context& dev_ctx,
 
   // For 0D Tensor
   if (rank == 0) {
+    dev_ctx.template Alloc<T>(x_grad);
     phi::funcs::set_constant(dev_ctx, x_grad, 0.0);
     return;
   }

paddle/phi/kernels/xpu/log_softmax_kernel.cc

@@ -32,6 +32,7 @@ void LogSoftmaxKernel(const Context& dev_ctx,
 
   // For 0D Tensor
   if (rank == 0) {
+    dev_ctx.template Alloc<T>(out);
     phi::funcs::set_constant(dev_ctx, out, 0.0);
     return;
   }

paddle/phi/kernels/xpu/prelu_grad_kernel.cc

@@ -40,16 +40,24 @@ void PReluGradKernel(const Context& dev_ctx,
   auto x_rank = x_dim.size();
 
   std::vector<int> x_shape(x_rank);
-  for (int i = 0; i < x_rank; i++) {
-    x_shape[i] = x_dim[i];
+  if (x_rank == 0) {
+    x_shape = std::vector<int>({1});
+  } else {
+    for (int i = 0; i < x_rank; i++) {
+      x_shape[i] = x_dim[i];
+    }
   }
 
   auto alpha_dim = alpha.dims();
   auto alpha_rank = alpha_dim.size();
 
   std::vector<int> alpha_shape(alpha_rank);
-  for (int i = 0; i < x_rank; i++) {
-    alpha_shape[i] = alpha_dim[i];
+  if (alpha_rank == 0) {
+    alpha_shape = std::vector<int>({1});
+  } else {
+    for (int i = 0; i < x_rank; i++) {
+      alpha_shape[i] = alpha_dim[i];
+    }
   }
 
   // mode = 0: channel_nchw, slope_shape = {c}, default. meanwhile, xhsape = {n,

paddle/phi/kernels/xpu/prelu_kernel.cc

@@ -34,20 +34,26 @@ void PReluKernel(const Context& dev_ctx,
 
   auto x_dim = x.dims();
   auto x_rank = x_dim.size();
-
   std::vector<int> x_shape(x_rank);
 
-  for (int i = 0; i < x_rank; i++) {
-    x_shape[i] = x_dim[i];
+  if (x_rank == 0) {
+    x_shape = std::vector<int>({1});
+  } else {
+    for (int i = 0; i < x_rank; i++) {
+      x_shape[i] = x_dim[i];
+    }
   }
 
   auto alpha_dim = alpha.dims();
   auto alpha_rank = alpha_dim.size();
-
   std::vector<int> alpha_shape(x_rank, 1);  // same size with x_shape
 
-  for (int i = 0; i < alpha_rank; i++) {
-    alpha_shape[i] = alpha_dim[i];
+  if (x_rank == 0) {
+    alpha_shape = std::vector<int>({1});
+  } else {
+    for (int i = 0; i < alpha_rank; i++) {
+      alpha_shape[i] = alpha_dim[i];
+    }
   }
 
   int r = xpu::prelu(dev_ctx.x_context(),

paddle/phi/kernels/xpu/softmax_grad_kernel.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/phi/backends/xpu/enforce_xpu.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/funcs/axis_utils.h"
+#include "paddle/phi/kernels/funcs/math_function.h"
 
 namespace phi {
 
@@ -35,6 +36,12 @@ void SoftmaxGradKernel(const Context& dev_ctx,
     return;
   }
 
+  // For 0D Tensor
+  if (rank == 0) {
+    phi::funcs::set_constant(dev_ctx, x_grad, 0.0);
+    return;
+  }
+
   std::vector<int> x_dims;
   for (int i = 0; i < rank; i++) {
     x_dims.push_back(x_grad->dims()[i]);

python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py

@@ -1628,24 +1628,29 @@ class TestSundryAPI(unittest.TestCase):
         self.assertEqual(x.grad.shape, [])
 
     def test_prelu(self):
-        x = paddle.full([], 1.0, 'float32')
-        x.stop_gradient = False
-
-        w1 = paddle.to_tensor([0.25], dtype='float32')
-        out1 = paddle.nn.functional.prelu(x, w1)
+        x1 = paddle.full([], 1.0, 'float32')
+        x1.stop_gradient = False
+        w1 = paddle.full([], 0.25, dtype='float32')
+        out1 = paddle.nn.functional.prelu(x1, w1)
         out1.retain_grads()
         out1.backward()
         self.assertEqual(out1.shape, [])
+        self.assertEqual(out1.numpy(), 1.0)
         self.assertEqual(out1.grad.shape, [])
-        self.assertEqual(x.grad.shape, [])
+        self.assertEqual(x1.grad.shape, [])
+        self.assertEqual(x1.grad.numpy(), 1.0)
 
+        x2 = paddle.full([], -1.0, 'float32')
+        x2.stop_gradient = False
         w2 = paddle.full([], 0.25, dtype='float32')
-        out2 = paddle.nn.functional.prelu(x, w2)
+        out2 = paddle.nn.functional.prelu(x2, w2)
         out2.retain_grads()
         out2.backward()
         self.assertEqual(out2.shape, [])
+        self.assertEqual(out2.numpy(), -0.25)
         self.assertEqual(out2.grad.shape, [])
-        self.assertEqual(x.grad.shape, [])
+        self.assertEqual(x2.grad.shape, [])
+        self.assertEqual(x2.grad.numpy(), 0.25)
 
     def test_while_loop(self):
         def cond(i, x):

python/paddle/fluid/tests/unittests/xpu/test_zero_dim_tensor_xpu.py

@@ -84,6 +84,8 @@ unary_api_list = [
     paddle.lgamma,
     paddle.poisson,
     paddle.bernoulli,
+    paddle.nn.functional.softmax,
+    paddle.nn.functional.log_softmax,
 ]
 
 inplace_api_list = [
@@ -1033,6 +1035,33 @@ class TestSundryAPI(unittest.TestCase):
         out2.backward()
         self.assertEqual(out2.shape, [1])
 
+    def test_prelu(self):
+        x1 = paddle.full([], 1.0, 'float32')
+        x1.stop_gradient = False
+        w1 = paddle.full([], 0.25, dtype='float32')
+        w1.stop_gradient = False
+        out1 = paddle.nn.functional.prelu(x1, w1)
+        out1.retain_grads()
+        out1.backward()
+        self.assertEqual(out1.shape, [])
+        self.assertEqual(out1.numpy(), 1.0)
+        self.assertEqual(out1.grad.shape, [])
+        self.assertEqual(x1.grad.shape, [])
+        self.assertEqual(x1.grad.numpy(), 1.0)
+
+        x2 = paddle.full([], -1.0, 'float32')
+        x2.stop_gradient = False
+        w2 = paddle.full([], 0.25, dtype='float32')
+        w2.stop_gradient = False
+        out2 = paddle.nn.functional.prelu(x2, w2)
+        out2.retain_grads()
+        out2.backward()
+        self.assertEqual(out2.shape, [])
+        self.assertEqual(out2.numpy(), -0.25)
+        self.assertEqual(out2.grad.shape, [])
+        self.assertEqual(x2.grad.shape, [])
+        self.assertEqual(x2.grad.numpy(), 0.25)
+
 
 # Use to test API whose zero-dim input tensors don't have grad and not need to test backward in OpTest.
 class TestNoBackwardAPI(unittest.TestCase):