diff --git a/paddle/fluid/inference/tensorrt/convert/roll_op.cc b/paddle/fluid/inference/tensorrt/convert/roll_op.cc
index df320c0abcdb65f5352de0187d2bc9842c8f4514..a94007999156a2af53c31a4a5fd2bc41a88452e9 100644
--- a/paddle/fluid/inference/tensorrt/convert/roll_op.cc
+++ b/paddle/fluid/inference/tensorrt/convert/roll_op.cc
@@ -28,7 +28,7 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 /*
- * Stack converter from fluid to tensorRT.
+ * Roll converter from fluid to tensorRT.
  */
 class RollOpConverter : public OpConverter {
  public:
@@ -53,7 +53,8 @@ class RollOpConverter : public OpConverter {
     }
     int axis_size = axis.size();
     for (int i = 0; i < axis_size; i++) {
-      start.d[axis[i]] = (-shifts[i]) % input_dims.d[axis[i]];
+      start.d[axis[i]] =
+          (input_dims.d[axis[i]] - shifts[i]) % input_dims.d[axis[i]];
     }
 
     nvinfer1::Dims stride;
@@ -70,11 +71,9 @@ class RollOpConverter : public OpConverter {
 
     auto output_name = op_desc.Output("Out")[0];
 
-    auto shape_layer = TRT_ENGINE_ADD_LAYER(engine_, Shape, *input);
-
     auto* layer =
         TRT_ENGINE_ADD_LAYER(engine_, Slice, *input, start, size, stride);
-    layer->setInput(2, *shape_layer->getOutput(0));
+    layer->setInput(2, *Shape(input));
 #if IS_TRT_VERSION_GE(7000)
     layer->setMode(nvinfer1::SliceMode::kWRAP);
 #endif
diff --git a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_roll.py b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_roll.py
index 8217b3e8d8506e602abcb0f77eb96e4962dcf5c5..a85cafbb5cce0f5b5d29841f8283762c77899dcf 100644
--- a/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_roll.py
+++ b/python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_roll.py
@@ -32,7 +32,7 @@ class TrtConvertRollTest(TrtLayerAutoScanTest):
 
     def sample_program_configs(self):
         def generate_input1(attrs: List[Dict[str, Any]]):
-            return np.ones([1, 56, 56, 192]).astype(np.float32)
+            return np.random.random([1, 56, 56, 192]).astype(np.float32)
 
         for axis in [[1, 2]]:
             for shifts in [[-1, -1], [-3, -3]]: