[Rocm] fix test of random_crop_op & logsumexp

paddle/fluid/operators/random_crop_op.h

@@ -59,6 +59,7 @@ HOSTDEVICE inline void StridedMemcpy(const T* x, const size_t* x_dims, T* out,
   size_t offset_i = offsets[i];
 
   if (i == rank - 1) {
+#ifndef PADDLE_WITH_HIP
     PADDLE_ENFORCE(x_stride == 1,
                    "When i:%d == rank:%d - 1, x_stride of random_crop_op "
                    "expected to be 1, but got %ld. Please check input "
@@ -69,6 +70,22 @@ HOSTDEVICE inline void StridedMemcpy(const T* x, const size_t* x_dims, T* out,
                    "expected to be 1, but got %ld. Please check input "
                    "value.",
                    i, rank, out_stride);
+#else
+    if (x_stride != 1) {
+      printf(
+          "When i:%d == rank:%d - 1, x_stride of random_crop_op expected to be "
+          "1, but got %ld. Please check input value.\n",
+          i, rank, x_stride);
+      abort();
+    }
+    if (out_stride != 1) {
+      printf(
+          "When i:%d == rank:%d - 1, out_stride of random_crop_op expected to "
+          "be 1, but got %ld. Please check input value.\n",
+          i, rank, out_stride);
+      abort();
+    }
+#endif
     x += offset_i;
     for (size_t j = 0; j < out_dim_i; ++j) {
       *out++ = *x++;

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

@@ -50,15 +50,30 @@ def setUp(self):
             'keepdim': self.keepdim,
             'reduce_all': self.reduce_all
         }
+        self.user_defined_grads = None
+        self.user_defined_grad_outputs = None
+        self.set_attrs_addition()
 
     def set_attrs(self):
         pass
 
+    def set_attrs_addition(self):
+        pass
+
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X'], ['Out'])
+        self.check_grad(
+            ['X'], ['Out'],
+            user_defined_grads=self.user_defined_grads,
+            user_defined_grad_outputs=self.user_defined_grad_outputs)
+
+    def calc_grad(self):
+        dy = np.ones(1, dtype=self.dtype)
+        x = self.inputs['X']
+        y = self.outputs['Out']
+        return dy * np.exp(x - y)
 
 
 class TestLogsumexp_shape(TestLogsumexp):
@@ -75,6 +90,11 @@ class TestLogsumexp_axis_all(TestLogsumexp):
     def set_attrs(self):
         self.axis = [0, 1, 2, 3]
 
+    def set_attrs_addition(self):
+        if paddle.fluid.core.is_compiled_with_rocm():
+            self.user_defined_grads = [self.calc_grad()]
+            self.user_defined_grad_outputs = [np.ones(1, dtype=self.dtype)]
+
 
 class TestLogsumexp_keepdim(TestLogsumexp):
     def set_attrs(self):
@@ -85,6 +105,11 @@ class TestLogsumexp_reduce_all(TestLogsumexp):
     def set_attrs(self):
         self.reduce_all = True
 
+    def set_attrs_addition(self):
+        if paddle.fluid.core.is_compiled_with_rocm():
+            self.user_defined_grads = [self.calc_grad()]
+            self.user_defined_grad_outputs = [np.ones(1, dtype=self.dtype)]
+
 
 class TestLogsumexpError(unittest.TestCase):
     def test_errors(self):

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

@@ -44,8 +44,11 @@ def get_weight_names(self):
 
     def setUp(self):
         self.op_type = "rnn"
-        self.dtype = np.float64
-        self.sequence_length = np.array([12, 11, 10, 9, 8], dtype=np.int32)
+        self.dtype = np.float32 if fluid.core.is_compiled_with_rocm(
+        ) else np.float64
+        self.sequence_length = None if fluid.core.is_compiled_with_rocm(
+        ) else np.array(
+            [12, 11, 10, 9, 8], dtype=np.int32)
         self.num_layers = 1
         self.is_bidirec = False
         self.is_test = False
@@ -76,12 +79,19 @@ def setUp(self):
             time_major=True,
             direction=direction,
             dropout=self.dropout,
-            nonlinearity=self.mode)
+            nonlinearity=self.mode,
+            dtype=self.dtype)
 
         flat_w = get_params_for_net(rnn1)
 
         output, last_hidden = rnn1(input, sequence_length=self.sequence_length)
 
+        if fluid.core.is_compiled_with_rocm():
+            self._get_places = lambda: [fluid.core.CUDAPlace(0)]
+            if self.is_bidirec:
+                for i in range(0, len(flat_w), 4):
+                    flat_w[i + 1], flat_w[i + 2] = flat_w[i + 2], flat_w[i + 1]
+
         init_h = np.zeros((self.num_layers * self.direction_num, batch_size,
                            hidden_size)).astype(self.dtype)