[0-size Tensor No.98] Add 0-size Tensor support for kron

paddle/phi/kernels/impl/kron_grad_kernel_impl.h

@@ -14,9 +14,9 @@
 
 #pragma once
 
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/impl/kron_kernel_impl.h"
 #include "paddle/phi/kernels/reduce_sum_kernel.h"
-
 namespace phi {
 
 template <typename T>
@@ -266,6 +266,17 @@ void KronGradKernel(const Context &ctx,
                     const DenseTensor &out_grad,
                     DenseTensor *x_grad,
                     DenseTensor *y_grad) {
+  if (out_grad.numel() == 0) {
+    if (x_grad) {
+      phi::Full<T, Context>(
+          ctx, phi::IntArray(common::vectorize(x_grad->dims())), 0, x_grad);
+    }
+    if (y_grad) {
+      phi::Full<T, Context>(
+          ctx, phi::IntArray(common::vectorize(y_grad->dims())), 0, y_grad);
+    }
+    return;
+  }
   if (x_grad) {
     ctx.template Alloc<T>(x_grad);
   }

paddle/phi/kernels/impl/kron_kernel_impl.h

@@ -158,6 +158,9 @@ void KronKernel(const Context &ctx,
                 const DenseTensor &y,
                 DenseTensor *out) {
   ctx.template Alloc<T>(out);
+  if (out && out->numel() == 0) {
+    return;
+  }
 
   int ndims = out->dims().size();
   DenseTensor xx = UnsqueezeTo(x, ndims);

test/legacy_test/test_kron_op.py

@@ -144,6 +144,33 @@ def _init_dtype(self):
         return "float16"
 
 
+class TestKronOp_ZeroSize(OpTest):
+    def setUp(self):
+        self.op_type = "kron"
+        self.prim_op_type = "prim"
+        self.python_api = paddle.kron
+        self.public_python_api = paddle.kron
+        self.dtype = self._init_dtype()
+        x = np.random.uniform(size=(10, 0, 2)).astype(self.dtype)
+        y = np.random.uniform(size=(2, 3, 4, 3, 2)).astype(self.dtype)
+        out_ref = np.kron(x, y)
+        self.inputs = {'X': x, 'Y': y}
+        self.outputs = {'Out': out_ref}
+
+    def _init_dtype(self):
+        return "float64"
+
+    def test_check_output(self):
+        self.check_output(check_pir=True)
+
+    def test_check_grad(self):
+        self.check_grad(
+            ['X', 'Y'],
+            'Out',
+            check_pir=True,
+        )
+
+
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),