[0-size Tensor No.69、70] Add 0-size Tensor support for fmax fmin API. (#72912)

DanielSun11 · web-flow · commit b48ac456cd83 · 2025-05-26T16:11:52.000+08:00
* fix 0 size Tensor for expand kernel in onednn

* format

* expand_grad support 0 size Tensor

* fix bug and add  unittest

* modify the unittest for onednn kernel

* modify expand unittest to force enable onednn op

* fix fmax fmin grad kernel
diff --git a/paddle/phi/kernels/funcs/elementwise_base.h b/paddle/phi/kernels/funcs/elementwise_base.h
@@ -373,6 +373,9 @@ void ElementwiseCompute(const CPUContext &dev_ctx,
                         DenseTensor *z,
                         int axis = -1) {
   dev_ctx.Alloc<OutType>(z);
+  if (z && z->numel() == 0) {
+    return;
+  }
   auto x_dims = x.dims();
   auto y_dims = y.dims();
   bool is_xsize_larger = true;
diff --git a/paddle/phi/kernels/impl/elementwise_grad_kernel_impl.h b/paddle/phi/kernels/impl/elementwise_grad_kernel_impl.h
@@ -746,6 +746,27 @@ void ElementwiseFMaxGradKernel(const Context& dev_ctx,
   auto x_dim = x.dims();
   auto y_dim = y.dims();
   int axis = -1;
+  if (out_grad.numel() == 0) {
+    if (x_grad) {
+      dev_ctx.template Alloc<T>(x_grad);
+      if (x_grad->numel() != 0) {
+        phi::Full<T, Context>(dev_ctx,
+                              phi::IntArray(common::vectorize(x_grad->dims())),
+                              0,
+                              x_grad);
+      }
+    }
+    if (y_grad) {
+      dev_ctx.template Alloc<T>(y_grad);
+      if (y_grad->numel() != 0) {
+        phi::Full<T, Context>(dev_ctx,
+                              phi::IntArray(common::vectorize(y_grad->dims())),
+                              0,
+                              y_grad);
+      }
+    }
+    return;
+  }
   if (x.dims() == y.dims()) {
     funcs::ElemwiseGradComputeNoBroadcast<Context,
                                           T,
@@ -791,6 +812,27 @@ void ElementwiseFMinGradKernel(const Context& dev_ctx,
                                DenseTensor* y_grad) {
   funcs::ElementwiseGradPreProcess(out_grad, x_grad);
   auto out = out_grad;  // Fake out, not used
+  if (out_grad.numel() == 0) {
+    if (x_grad) {
+      dev_ctx.template Alloc<T>(x_grad);
+      if (x_grad->numel() != 0) {
+        phi::Full<T, Context>(dev_ctx,
+                              phi::IntArray(common::vectorize(x_grad->dims())),
+                              0,
+                              x_grad);
+      }
+    }
+    if (y_grad) {
+      dev_ctx.template Alloc<T>(y_grad);
+      if (y_grad->numel() != 0) {
+        phi::Full<T, Context>(dev_ctx,
+                              phi::IntArray(common::vectorize(y_grad->dims())),
+                              0,
+                              y_grad);
+      }
+    }
+    return;
+  }
   auto x_dim = x.dims();
   auto y_dim = y.dims();
   int axis = -1;
diff --git a/test/legacy_test/test_fmax_op.py b/test/legacy_test/test_fmax_op.py
@@ -139,12 +139,16 @@ def setUp(self):
         # If x and y have the same value, the max() is not differentiable.
         # So we generate test data by the following method
         # to avoid them being too close to each other.
-        x = np.random.uniform(0.1, 1, [13, 17]).astype("float64")
-        sgn = np.random.choice([-1, 1], [13, 17]).astype("float64")
-        y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype("float64")
+        self.init_shape()
+        x = np.random.uniform(0.1, 1, self.shape).astype("float64")
+        sgn = np.random.choice([-1, 1], self.shape).astype("float64")
+        y = x + sgn * np.random.uniform(0.1, 1, self.shape).astype("float64")
         self.inputs = {'X': x, 'Y': y}
         self.outputs = {'Out': np.fmax(self.inputs['X'], self.inputs['Y'])}
 
+    def init_shape(self):
+        self.shape = [13, 17]
+
     def test_check_output(self):
         """test_check_output"""
         self.check_output(check_pir=True, check_symbol_infer=False)
@@ -286,5 +290,38 @@ def test_check_grad(self):
         )
 
 
+class TestElementwiseFmaxOpZeroSize(TestElementwiseFmaxOp):
+    def init_shape(self):
+        self.shape = [0, 15]
+
+
+class TestElementwiseFmaxOpZeroSize1(TestElementwiseFmaxOp):
+    def init_shape(self):
+        self.shape = [0, 15, 0]
+
+
+class ApiFMaxTestZeroSize(unittest.TestCase):
+    """ApiFMaxTest"""
+
+    def setUp(self):
+        """setUp"""
+        if core.is_compiled_with_cuda():
+            self.place = core.CUDAPlace(0)
+        else:
+            self.place = core.CPUPlace()
+
+        self.input_x = np.random.rand(0, 15).astype("float32")
+        self.input_y = np.random.rand(0, 15).astype("float32")
+        self.input_z = np.random.rand(1, 15).astype("float32")
+        self.input_a = np.random.rand(15, 0).astype('int64')
+        self.input_b = np.random.rand(15, 0, 1).astype('int64')
+        self.input_c = np.random.rand(15, 0, 2).astype('int64')
+
+        self.np_expected1 = np.fmax(self.input_x, self.input_y)
+        self.np_expected2 = np.fmax(self.input_x, self.input_z)
+        self.np_expected3 = np.fmax(self.input_a, self.input_c)
+        self.np_expected4 = np.fmax(self.input_b, self.input_c)
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/test/legacy_test/test_fmin_op.py b/test/legacy_test/test_fmin_op.py
@@ -141,12 +141,17 @@ def setUp(self):
         # If x and y have the same value, the min() is not differentiable.
         # So we generate test data by the following method
         # to avoid them being too close to each other.
-        x = np.random.uniform(0.1, 1, [13, 17]).astype("float64")
-        sgn = np.random.choice([-1, 1], [13, 17]).astype("float64")
-        y = x + sgn * np.random.uniform(0.1, 1, [13, 17]).astype("float64")
+        self.init_shape()
+        x = np.random.uniform(0.1, 1, self.shape).astype("float64")
+        sgn = np.random.choice([-1, 1], self.shape).astype("float64")
+        y = x + sgn * np.random.uniform(0.1, 1, self.shape).astype("float64")
         self.inputs = {'X': x, 'Y': y}
         self.outputs = {'Out': np.fmin(self.inputs['X'], self.inputs['Y'])}
 
+    def init_shape(self):
+        """init_shape"""
+        self.shape = [13, 17]
+
     def test_check_output(self):
         """test_check_output"""
         self.check_output(check_pir=True, check_symbol_infer=False)
@@ -288,6 +293,16 @@ def test_check_grad(self):
         )
 
 
+class TestElementwiseFminOpZeroSize(TestElementwiseFminOp):
+    def init_shape(self):
+        self.shape = [0, 9]
+
+
+class TestElementwiseFminOpZeroSize1(TestElementwiseFminOp):
+    def init_shape(self):
+        self.shape = [9, 0]
+
+
 if __name__ == "__main__":
     paddle.enable_static()
     unittest.main()
